Commit 70c85b1bf1c07db4420e8431c3322bfeb87688ee | zig

author	Andrew Kelley <andrew@ziglang.org> 2024-04-25 14:59:29-07:00
committer	Andrew Kelley <andrew@ziglang.org> 2024-05-08 19:37:28-07:00
log	70c85b1bf1c07db4420e8431c3322bfeb87688ee
tree	646b310e667401cc2c44595e6eb69fc52637a8d9
parent	e4029b9943a834bf5fc23e620f0dfc0ea18648f4

update C language headers to LLVM 18

release/18.x branch, commit 78b99c73ee4b96fe9ce0e294d4632326afb2db42

103 files changed, 21212 insertions(+), 2508 deletions(-)

lib
- include

lib/include/__clang_cuda_device_functions.h+2-2

...	@@ -502,8 +502,8 @@ __DEVICE__ unsigned int __pm0(void) { return __nvvm_read_ptx_sreg_pm0(); }	...	@@ -502,8 +502,8 @@ __DEVICE__ unsigned int __pm0(void) { return __nvvm_read_ptx_sreg_pm0(); }
502	__DEVICE__ unsigned int __pm1(void) { return __nvvm_read_ptx_sreg_pm1(); }	502	__DEVICE__ unsigned int __pm1(void) { return __nvvm_read_ptx_sreg_pm1(); }
503	__DEVICE__ unsigned int __pm2(void) { return __nvvm_read_ptx_sreg_pm2(); }	503	__DEVICE__ unsigned int __pm2(void) { return __nvvm_read_ptx_sreg_pm2(); }
504	__DEVICE__ unsigned int __pm3(void) { return __nvvm_read_ptx_sreg_pm3(); }	504	__DEVICE__ unsigned int __pm3(void) { return __nvvm_read_ptx_sreg_pm3(); }
505	__DEVICE__ int __popc(int __a) { return __nv_popc(__a); }	505	__DEVICE__ int __popc(unsigned int __a) { return __nv_popc(__a); }
506	__DEVICE__ int __popcll(long long __a) { return __nv_popcll(__a); }	506	__DEVICE__ int __popcll(unsigned long long __a) { return __nv_popcll(__a); }
507	__DEVICE__ float __powf(float __a, float __b) {	507	__DEVICE__ float __powf(float __a, float __b) {
508	return __nv_fast_powf(__a, __b);	508	return __nv_fast_powf(__a, __b);
509	}	509	}

lib/include/__clang_cuda_libdevice_declares.h+2-2

...	@@ -285,8 +285,8 @@ __DEVICE__ double __nv_normcdfinv(double __a);	...	@@ -285,8 +285,8 @@ __DEVICE__ double __nv_normcdfinv(double __a);
285	__DEVICE__ float __nv_normcdfinvf(float __a);	285	__DEVICE__ float __nv_normcdfinvf(float __a);
286	__DEVICE__ float __nv_normf(int __a, const float *__b);	286	__DEVICE__ float __nv_normf(int __a, const float *__b);
287	__DEVICE__ double __nv_norm(int __a, const double *__b);	287	__DEVICE__ double __nv_norm(int __a, const double *__b);
288	__DEVICE__ int __nv_popc(int __a);	288	__DEVICE__ int __nv_popc(unsigned int __a);
289	__DEVICE__ int __nv_popcll(long long __a);	289	__DEVICE__ int __nv_popcll(unsigned long long __a);
290	__DEVICE__ double __nv_pow(double __a, double __b);	290	__DEVICE__ double __nv_pow(double __a, double __b);
291	__DEVICE__ float __nv_powf(float __a, float __b);	291	__DEVICE__ float __nv_powf(float __a, float __b);
292	__DEVICE__ double __nv_powi(double __a, int __b);	292	__DEVICE__ double __nv_powi(double __a, int __b);

lib/include/__clang_cuda_math.h+3-3

lib/include/__clang_cuda_runtime_wrapper.h+2-2

...	@@ -196,12 +196,12 @@ inline __host__ double __signbitd(double x) {	...	@@ -196,12 +196,12 @@ inline __host__ double __signbitd(double x) {
196		196
197	// math_function.hpp uses the __USE_FAST_MATH__ macro to determine whether we	197	// math_function.hpp uses the __USE_FAST_MATH__ macro to determine whether we
198	// get the slow-but-accurate or fast-but-inaccurate versions of functions like	198	// get the slow-but-accurate or fast-but-inaccurate versions of functions like
199	// sin and exp. This is controlled in clang by -fcuda-approx-transcendentals.	199	// sin and exp. This is controlled in clang by -fgpu-approx-transcendentals.
200	//	200	//
201	// device_functions.hpp uses __USE_FAST_MATH__ for a different purpose (fast vs.	201	// device_functions.hpp uses __USE_FAST_MATH__ for a different purpose (fast vs.
202	// slow divides), so we need to scope our define carefully here.	202	// slow divides), so we need to scope our define carefully here.
203	#pragma push_macro("__USE_FAST_MATH__")	203	#pragma push_macro("__USE_FAST_MATH__")
204	#if defined(__CLANG_CUDA_APPROX_TRANSCENDENTALS__)	204	#if defined(__CLANG_GPU_APPROX_TRANSCENDENTALS__)
205	#define __USE_FAST_MATH__ 1	205	#define __USE_FAST_MATH__ 1
206	#endif	206	#endif
207		207

lib/include/__clang_hip_math.h+192-171

 @@ -14,9 +14,6 @@
 #endif
 #if !defined(__HIPCC_RTC__)
-#if defined(__cplusplus)
-#include <algorithm>
-#endif
 #include <limits.h>
 #include <stdint.h>
 #ifdef __OPENMP_AMDGCN__
 @@ -32,6 +29,17 @@
 #define __DEVICE__ static __device__ inline __attribute__((always_inline))
 #endif
+// Device library provides fast low precision and slow full-recision
+// implementations for some functions. Which one gets selected depends on
+// __CLANG_GPU_APPROX_TRANSCENDENTALS__ which gets defined by clang if
+// -ffast-math or -fgpu-approx-transcendentals are in effect.
+#pragma push_macro("__FAST_OR_SLOW")
+#if defined(__CLANG_GPU_APPROX_TRANSCENDENTALS__)
+#define __FAST_OR_SLOW(fast, slow) fast
+#else
+#define __FAST_OR_SLOW(fast, slow) slow
+#endif
 // A few functions return bool type starting only in C++11.
 #pragma push_macro("__RETURN_TYPE")
 #ifdef __OPENMP_AMDGCN__
 @@ -139,21 +147,180 @@ uint64_t __make_mantissa(const char *__tagp __attribute__((nonnull))) {
 }
 // BEGIN FLOAT
+// BEGIN INTRINSICS
+__DEVICE__
+float __cosf(float __x) { return __ocml_native_cos_f32(__x); }
+__DEVICE__
+float __exp10f(float __x) {
+  const float __log2_10 = 0x1.a934f0p+1f;
+  return __builtin_amdgcn_exp2f(__log2_10 * __x);
+}
+__DEVICE__
+float __expf(float __x) {
+  const float __log2_e = 0x1.715476p+0;
+  return __builtin_amdgcn_exp2f(__log2_e * __x);
+}
+#if defined OCML_BASIC_ROUNDED_OPERATIONS
+__DEVICE__
+float __fadd_rd(float __x, float __y) { return __ocml_add_rtn_f32(__x, __y); }
+__DEVICE__
+float __fadd_rn(float __x, float __y) { return __ocml_add_rte_f32(__x, __y); }
+__DEVICE__
+float __fadd_ru(float __x, float __y) { return __ocml_add_rtp_f32(__x, __y); }
+__DEVICE__
+float __fadd_rz(float __x, float __y) { return __ocml_add_rtz_f32(__x, __y); }
+#else
+__DEVICE__
+float __fadd_rn(float __x, float __y) { return __x + __y; }
+#endif
+#if defined OCML_BASIC_ROUNDED_OPERATIONS
+__DEVICE__
+float __fdiv_rd(float __x, float __y) { return __ocml_div_rtn_f32(__x, __y); }
+__DEVICE__
+float __fdiv_rn(float __x, float __y) { return __ocml_div_rte_f32(__x, __y); }
+__DEVICE__
+float __fdiv_ru(float __x, float __y) { return __ocml_div_rtp_f32(__x, __y); }
+__DEVICE__
+float __fdiv_rz(float __x, float __y) { return __ocml_div_rtz_f32(__x, __y); }
+#else
+__DEVICE__
+float __fdiv_rn(float __x, float __y) { return __x / __y; }
+#endif
+__DEVICE__
+float __fdividef(float __x, float __y) { return __x / __y; }
+#if defined OCML_BASIC_ROUNDED_OPERATIONS
+__DEVICE__
+float __fmaf_rd(float __x, float __y, float __z) {
+  return __ocml_fma_rtn_f32(__x, __y, __z);
+}
+__DEVICE__
+float __fmaf_rn(float __x, float __y, float __z) {
+  return __ocml_fma_rte_f32(__x, __y, __z);
+}
+__DEVICE__
+float __fmaf_ru(float __x, float __y, float __z) {
+  return __ocml_fma_rtp_f32(__x, __y, __z);
+}
+__DEVICE__
+float __fmaf_rz(float __x, float __y, float __z) {
+  return __ocml_fma_rtz_f32(__x, __y, __z);
+}
+#else
+__DEVICE__
+float __fmaf_rn(float __x, float __y, float __z) {
+  return __builtin_fmaf(__x, __y, __z);
+}
+#endif
+#if defined OCML_BASIC_ROUNDED_OPERATIONS
+__DEVICE__
+float __fmul_rd(float __x, float __y) { return __ocml_mul_rtn_f32(__x, __y); }
+__DEVICE__
+float __fmul_rn(float __x, float __y) { return __ocml_mul_rte_f32(__x, __y); }
+__DEVICE__
+float __fmul_ru(float __x, float __y) { return __ocml_mul_rtp_f32(__x, __y); }
+__DEVICE__
+float __fmul_rz(float __x, float __y) { return __ocml_mul_rtz_f32(__x, __y); }
+#else
+__DEVICE__
+float __fmul_rn(float __x, float __y) { return __x * __y; }
+#endif
+#if defined OCML_BASIC_ROUNDED_OPERATIONS
+__DEVICE__
+float __frcp_rd(float __x) { return __ocml_div_rtn_f32(1.0f, __x); }
+__DEVICE__
+float __frcp_rn(float __x) { return __ocml_div_rte_f32(1.0f, __x); }
+__DEVICE__
+float __frcp_ru(float __x) { return __ocml_div_rtp_f32(1.0f, __x); }
+__DEVICE__
+float __frcp_rz(float __x) { return __ocml_div_rtz_f32(1.0f, __x); }
+#else
+__DEVICE__
+float __frcp_rn(float __x) { return 1.0f / __x; }
+#endif
+__DEVICE__
+float __frsqrt_rn(float __x) { return __builtin_amdgcn_rsqf(__x); }
+#if defined OCML_BASIC_ROUNDED_OPERATIONS
+__DEVICE__
+float __fsqrt_rd(float __x) { return __ocml_sqrt_rtn_f32(__x); }
+__DEVICE__
+float __fsqrt_rn(float __x) { return __ocml_sqrt_rte_f32(__x); }
+__DEVICE__
+float __fsqrt_ru(float __x) { return __ocml_sqrt_rtp_f32(__x); }
+__DEVICE__
+float __fsqrt_rz(float __x) { return __ocml_sqrt_rtz_f32(__x); }
+#else
+__DEVICE__
+float __fsqrt_rn(float __x) { return __ocml_native_sqrt_f32(__x); }
+#endif
+#if defined OCML_BASIC_ROUNDED_OPERATIONS
+__DEVICE__
+float __fsub_rd(float __x, float __y) { return __ocml_sub_rtn_f32(__x, __y); }
+__DEVICE__
+float __fsub_rn(float __x, float __y) { return __ocml_sub_rte_f32(__x, __y); }
+__DEVICE__
+float __fsub_ru(float __x, float __y) { return __ocml_sub_rtp_f32(__x, __y); }
+__DEVICE__
+float __fsub_rz(float __x, float __y) { return __ocml_sub_rtz_f32(__x, __y); }
+#else
+__DEVICE__
+float __fsub_rn(float __x, float __y) { return __x - __y; }
+#endif
+__DEVICE__
+float __log10f(float __x) { return __builtin_log10f(__x); }
+__DEVICE__
+float __log2f(float __x) { return __builtin_amdgcn_logf(__x); }
+__DEVICE__
+float __logf(float __x) { return __builtin_logf(__x); }
+__DEVICE__
+float __powf(float __x, float __y) { return __ocml_pow_f32(__x, __y); }
+__DEVICE__
+float __saturatef(float __x) { return (__x < 0) ? 0 : ((__x > 1) ? 1 : __x); }
+__DEVICE__
+void __sincosf(float __x, float *__sinptr, float *__cosptr) {
+  *__sinptr = __ocml_native_sin_f32(__x);
+  *__cosptr = __ocml_native_cos_f32(__x);
+}
+__DEVICE__
+float __sinf(float __x) { return __ocml_native_sin_f32(__x); }
+__DEVICE__
+float __tanf(float __x) {
+  return __sinf(__x) * __builtin_amdgcn_rcpf(__cosf(__x));
+}
+// END INTRINSICS
 #if defined(__cplusplus)
 __DEVICE__
 int abs(int __x) {
-  int __sgn = __x >> (sizeof(int) * CHAR_BIT - 1);
+  return __builtin_abs(__x);
-  return (__x ^ __sgn) - __sgn;
 }
 __DEVICE__
 long labs(long __x) {
-  long __sgn = __x >> (sizeof(long) * CHAR_BIT - 1);
+  return __builtin_labs(__x);
-  return (__x ^ __sgn) - __sgn;
 }
 __DEVICE__
 long long llabs(long long __x) {
-  long long __sgn = __x >> (sizeof(long long) * CHAR_BIT - 1);
+  return __builtin_llabs(__x);
-  return (__x ^ __sgn) - __sgn;
 }
 #endif
 @@ -188,7 +355,7 @@ __DEVICE__
 float copysignf(float __x, float __y) { return __builtin_copysignf(__x, __y); }
 __DEVICE__
-float cosf(float __x) { return __ocml_cos_f32(__x); }
+float cosf(float __x) { return __FAST_OR_SLOW(__cosf, __ocml_cos_f32)(__x); }
 __DEVICE__
 float coshf(float __x) { return __ocml_cosh_f32(__x); }
 @@ -321,13 +488,13 @@ __DEVICE__
 float log1pf(float __x) { return __ocml_log1p_f32(__x); }
 __DEVICE__
-float log2f(float __x) { return __builtin_log2f(__x); }
+float log2f(float __x) { return __FAST_OR_SLOW(__log2f, __ocml_log2_f32)(__x); }
 __DEVICE__
 float logbf(float __x) { return __ocml_logb_f32(__x); }
 __DEVICE__
-float logf(float __x) { return __builtin_logf(__x); }
+float logf(float __x) { return __FAST_OR_SLOW(__logf, __ocml_log_f32)(__x); }
 __DEVICE__
 long int lrintf(float __x) { return __builtin_rintf(__x); }
 @@ -401,7 +568,7 @@ float normf(int __dim,
     ++__a;
   }
-  return __ocml_sqrt_f32(__r);
+  return __builtin_sqrtf(__r);
 }
 __DEVICE__
 @@ -483,9 +650,13 @@ void sincosf(float __x, float *__sinptr, float *__cosptr) {
 #ifdef __OPENMP_AMDGCN__
 #pragma omp allocate(__tmp) allocator(omp_thread_mem_alloc)
 #endif
+#ifdef __CLANG_CUDA_APPROX_TRANSCENDENTALS__
+  __sincosf(__x, __sinptr, __cosptr);
+#else
   *__sinptr =
       __ocml_sincos_f32(__x, (__attribute__((address_space(5))) float *)&__tmp);
   *__cosptr = __tmp;
+#endif
 }
 __DEVICE__
 @@ -500,7 +671,7 @@ void sincospif(float __x, float *__sinptr, float *__cosptr) {
 }
 __DEVICE__
-float sinf(float __x) { return __ocml_sin_f32(__x); }
+float sinf(float __x) { return __FAST_OR_SLOW(__sinf, __ocml_sin_f32)(__x); }
 __DEVICE__
 float sinhf(float __x) { return __ocml_sinh_f32(__x); }
 @@ -509,7 +680,7 @@ __DEVICE__
 float sinpif(float __x) { return __ocml_sinpi_f32(__x); }
 __DEVICE__
-float sqrtf(float __x) { return __ocml_sqrt_f32(__x); }
+float sqrtf(float __x) { return __builtin_sqrtf(__x); }
 __DEVICE__
 float tanf(float __x) { return __ocml_tan_f32(__x); }
 @@ -551,158 +722,7 @@ float ynf(int __n, float __x) { // TODO: we could use Ahmes multiplication
   return __x1;
 }
-// BEGIN INTRINSICS
-__DEVICE__
-float __cosf(float __x) { return __ocml_native_cos_f32(__x); }
-__DEVICE__
-float __exp10f(float __x) { return __ocml_native_exp10_f32(__x); }
-__DEVICE__
-float __expf(float __x) { return __ocml_native_exp_f32(__x); }
-#if defined OCML_BASIC_ROUNDED_OPERATIONS
-__DEVICE__
-float __fadd_rd(float __x, float __y) { return __ocml_add_rtn_f32(__x, __y); }
-__DEVICE__
-float __fadd_rn(float __x, float __y) { return __ocml_add_rte_f32(__x, __y); }
-__DEVICE__
-float __fadd_ru(float __x, float __y) { return __ocml_add_rtp_f32(__x, __y); }
-__DEVICE__
-float __fadd_rz(float __x, float __y) { return __ocml_add_rtz_f32(__x, __y); }
-#else
-__DEVICE__
-float __fadd_rn(float __x, float __y) { return __x + __y; }
-#endif
-#if defined OCML_BASIC_ROUNDED_OPERATIONS
-__DEVICE__
-float __fdiv_rd(float __x, float __y) { return __ocml_div_rtn_f32(__x, __y); }
-__DEVICE__
-float __fdiv_rn(float __x, float __y) { return __ocml_div_rte_f32(__x, __y); }
-__DEVICE__
-float __fdiv_ru(float __x, float __y) { return __ocml_div_rtp_f32(__x, __y); }
-__DEVICE__
-float __fdiv_rz(float __x, float __y) { return __ocml_div_rtz_f32(__x, __y); }
-#else
-__DEVICE__
-float __fdiv_rn(float __x, float __y) { return __x / __y; }
-#endif
-__DEVICE__
-float __fdividef(float __x, float __y) { return __x / __y; }
-#if defined OCML_BASIC_ROUNDED_OPERATIONS
-__DEVICE__
-float __fmaf_rd(float __x, float __y, float __z) {
-  return __ocml_fma_rtn_f32(__x, __y, __z);
-}
-__DEVICE__
-float __fmaf_rn(float __x, float __y, float __z) {
-  return __ocml_fma_rte_f32(__x, __y, __z);
-}
-__DEVICE__
-float __fmaf_ru(float __x, float __y, float __z) {
-  return __ocml_fma_rtp_f32(__x, __y, __z);
-}
-__DEVICE__
-float __fmaf_rz(float __x, float __y, float __z) {
-  return __ocml_fma_rtz_f32(__x, __y, __z);
-}
-#else
-__DEVICE__
-float __fmaf_rn(float __x, float __y, float __z) {
-  return __builtin_fmaf(__x, __y, __z);
-}
-#endif
-#if defined OCML_BASIC_ROUNDED_OPERATIONS
-__DEVICE__
-float __fmul_rd(float __x, float __y) { return __ocml_mul_rtn_f32(__x, __y); }
-__DEVICE__
-float __fmul_rn(float __x, float __y) { return __ocml_mul_rte_f32(__x, __y); }
-__DEVICE__
-float __fmul_ru(float __x, float __y) { return __ocml_mul_rtp_f32(__x, __y); }
-__DEVICE__
-float __fmul_rz(float __x, float __y) { return __ocml_mul_rtz_f32(__x, __y); }
-#else
-__DEVICE__
-float __fmul_rn(float __x, float __y) { return __x * __y; }
-#endif
-#if defined OCML_BASIC_ROUNDED_OPERATIONS
-__DEVICE__
-float __frcp_rd(float __x) { return __ocml_div_rtn_f32(1.0f, __x); }
-__DEVICE__
-float __frcp_rn(float __x) { return __ocml_div_rte_f32(1.0f, __x); }
-__DEVICE__
-float __frcp_ru(float __x) { return __ocml_div_rtp_f32(1.0f, __x); }
-__DEVICE__
-float __frcp_rz(float __x) { return __ocml_div_rtz_f32(1.0f, __x); }
-#else
-__DEVICE__
-float __frcp_rn(float __x) { return 1.0f / __x; }
-#endif
-__DEVICE__
-float __frsqrt_rn(float __x) { return __builtin_amdgcn_rsqf(__x); }
-#if defined OCML_BASIC_ROUNDED_OPERATIONS
-__DEVICE__
-float __fsqrt_rd(float __x) { return __ocml_sqrt_rtn_f32(__x); }
-__DEVICE__
-float __fsqrt_rn(float __x) { return __ocml_sqrt_rte_f32(__x); }
-__DEVICE__
-float __fsqrt_ru(float __x) { return __ocml_sqrt_rtp_f32(__x); }
-__DEVICE__
-float __fsqrt_rz(float __x) { return __ocml_sqrt_rtz_f32(__x); }
-#else
-__DEVICE__
-float __fsqrt_rn(float __x) { return __ocml_native_sqrt_f32(__x); }
-#endif
-#if defined OCML_BASIC_ROUNDED_OPERATIONS
-__DEVICE__
-float __fsub_rd(float __x, float __y) { return __ocml_sub_rtn_f32(__x, __y); }
-__DEVICE__
-float __fsub_rn(float __x, float __y) { return __ocml_sub_rte_f32(__x, __y); }
-__DEVICE__
-float __fsub_ru(float __x, float __y) { return __ocml_sub_rtp_f32(__x, __y); }
-__DEVICE__
-float __fsub_rz(float __x, float __y) { return __ocml_sub_rtz_f32(__x, __y); }
-#else
-__DEVICE__
-float __fsub_rn(float __x, float __y) { return __x - __y; }
-#endif
-__DEVICE__
-float __log10f(float __x) { return __ocml_native_log10_f32(__x); }
-__DEVICE__
-float __log2f(float __x) { return __ocml_native_log2_f32(__x); }
-__DEVICE__
-float __logf(float __x) { return __ocml_native_log_f32(__x); }
-__DEVICE__
-float __powf(float __x, float __y) { return __ocml_pow_f32(__x, __y); }
-__DEVICE__
-float __saturatef(float __x) { return (__x < 0) ? 0 : ((__x > 1) ? 1 : __x); }
-__DEVICE__
-void __sincosf(float __x, float *__sinptr, float *__cosptr) {
-  *__sinptr = __ocml_native_sin_f32(__x);
-  *__cosptr = __ocml_native_cos_f32(__x);
-}
-__DEVICE__
-float __sinf(float __x) { return __ocml_native_sin_f32(__x); }
-__DEVICE__
-float __tanf(float __x) { return __ocml_tan_f32(__x); }
-// END INTRINSICS
 // END FLOAT
 // BEGIN DOUBLE
 @@ -941,7 +961,7 @@ double norm(int __dim,
     ++__a;
   }
-  return __ocml_sqrt_f64(__r);
+  return __builtin_sqrt(__r);
 }
 __DEVICE__
 @@ -1064,7 +1084,7 @@ __DEVICE__
 double sinpi(double __x) { return __ocml_sinpi_f64(__x); }
 __DEVICE__
-double sqrt(double __x) { return __ocml_sqrt_f64(__x); }
+double sqrt(double __x) { return __builtin_sqrt(__x); }
 __DEVICE__
 double tan(double __x) { return __ocml_tan_f64(__x); }
 @@ -1198,7 +1218,7 @@ __DEVICE__
 double __dsqrt_rz(double __x) { return __ocml_sqrt_rtz_f64(__x); }
 #else
 __DEVICE__
-double __dsqrt_rn(double __x) { return __ocml_sqrt_f64(__x); }
+double __dsqrt_rn(double __x) { return __builtin_sqrt(__x); }
 #endif
 #if defined OCML_BASIC_ROUNDED_OPERATIONS
 @@ -1288,16 +1308,17 @@ double min(double __x, double __y) { return __builtin_fmin(__x, __y); }
 #if !defined(__HIPCC_RTC__) && !defined(__OPENMP_AMDGCN__)
 __host__ inline static int min(int __arg1, int __arg2) {
-  return std::min(__arg1, __arg2);
+  return __arg1 < __arg2 ? __arg1 : __arg2;
 }
 __host__ inline static int max(int __arg1, int __arg2) {
-  return std::max(__arg1, __arg2);
+  return __arg1 > __arg2 ? __arg1 : __arg2;
 }
 #endif // !defined(__HIPCC_RTC__) && !defined(__OPENMP_AMDGCN__)
 #endif
 #pragma pop_macro("__DEVICE__")
 #pragma pop_macro("__RETURN_TYPE")
+#pragma pop_macro("__FAST_OR_SLOW")
 #endif // __CLANG_HIP_MATH_H__

lib/include/__clang_hip_runtime_wrapper.h+34-26

 @@ -47,28 +47,9 @@ extern "C" {
 #endif //__cplusplus
 #if !defined(__HIPCC_RTC__)
-#include <cmath>
-#include <cstdlib>
-#include <stdlib.h>
 #if __has_include("hip/hip_version.h")
 #include "hip/hip_version.h"
 #endif // __has_include("hip/hip_version.h")
-#else
-typedef __SIZE_TYPE__ size_t;
-// Define macros which are needed to declare HIP device API's without standard
-// C/C++ headers. This is for readability so that these API's can be written
-// the same way as non-hipRTC use case. These macros need to be popped so that
-// they do not pollute users' name space.
-#pragma push_macro("NULL")
-#pragma push_macro("uint32_t")
-#pragma push_macro("uint64_t")
-#pragma push_macro("CHAR_BIT")
-#pragma push_macro("INT_MAX")
-#define NULL (void *)0
-#define uint32_t __UINT32_TYPE__
-#define uint64_t __UINT64_TYPE__
-#define CHAR_BIT __CHAR_BIT__
-#define INT_MAX __INTMAX_MAX__
 #endif // __HIPCC_RTC__
 typedef __SIZE_TYPE__ __hip_size_t;
 @@ -78,11 +59,13 @@ extern "C" {
 #endif //__cplusplus
 #if HIP_VERSION_MAJOR * 100 + HIP_VERSION_MINOR >= 405
-extern "C" __device__ unsigned long long __ockl_dm_alloc(unsigned long long __size);
+__device__ unsigned long long __ockl_dm_alloc(unsigned long long __size);
-extern "C" __device__ void __ockl_dm_dealloc(unsigned long long __addr);
+__device__ void __ockl_dm_dealloc(unsigned long long __addr);
 #if __has_feature(address_sanitizer)
-extern "C" __device__ unsigned long long __asan_malloc_impl(unsigned long long __size, unsigned long long __pc);
+__device__ unsigned long long __asan_malloc_impl(unsigned long long __size,
-extern "C" __device__ void __asan_free_impl(unsigned long long __addr, unsigned long long __pc);
+                                                 unsigned long long __pc);
+__device__ void __asan_free_impl(unsigned long long __addr,
+                                 unsigned long long __pc);
 __attribute__((noinline, weak)) __device__ void *malloc(__hip_size_t __size) {
   unsigned long long __pc = (unsigned long long)__builtin_return_address(0);
   return (void *)__asan_malloc_impl(__size, __pc);
 @@ -91,7 +74,7 @@ __attribute__((noinline, weak)) __device__ void free(void *__ptr) {
   unsigned long long __pc = (unsigned long long)__builtin_return_address(0);
   __asan_free_impl((unsigned long long)__ptr, __pc);
 }
-#else
+#else // __has_feature(address_sanitizer)
 __attribute__((weak)) inline __device__ void *malloc(__hip_size_t __size) {
   return (void *) __ockl_dm_alloc(__size);
 }
 @@ -109,7 +92,7 @@ __attribute__((weak)) inline __device__ void *malloc(__hip_size_t __size) {
 __attribute__((weak)) inline __device__ void free(void *__ptr) {
   __hip_free(__ptr);
 }
-#else
+#else  // __HIP_ENABLE_DEVICE_MALLOC__
 __attribute__((weak)) inline __device__ void *malloc(__hip_size_t __size) {
   __builtin_trap();
   return (void *)0;
 @@ -117,13 +100,38 @@ __attribute__((weak)) inline __device__ void *malloc(__hip_size_t __size) {
 __attribute__((weak)) inline __device__ void free(void *__ptr) {
   __builtin_trap();
 }
-#endif
+#endif // __HIP_ENABLE_DEVICE_MALLOC__
 #endif // HIP version check
 #ifdef __cplusplus
 } // extern "C"
 #endif //__cplusplus
+#if !defined(__HIPCC_RTC__)
+#include <cmath>
+#include <cstdlib>
+#include <stdlib.h>
+#if __has_include("hip/hip_version.h")
+#include "hip/hip_version.h"
+#endif // __has_include("hip/hip_version.h")
+#else
+typedef __SIZE_TYPE__ size_t;
+// Define macros which are needed to declare HIP device API's without standard
+// C/C++ headers. This is for readability so that these API's can be written
+// the same way as non-hipRTC use case. These macros need to be popped so that
+// they do not pollute users' name space.
+#pragma push_macro("NULL")
+#pragma push_macro("uint32_t")
+#pragma push_macro("uint64_t")
+#pragma push_macro("CHAR_BIT")
+#pragma push_macro("INT_MAX")
+#define NULL (void *)0
+#define uint32_t __UINT32_TYPE__
+#define uint64_t __UINT64_TYPE__
+#define CHAR_BIT __CHAR_BIT__
+#define INT_MAX __INTMAX_MAX__
+#endif // __HIPCC_RTC__
 #include <__clang_hip_libdevice_declares.h>
 #include <__clang_hip_math.h>
 #include <__clang_hip_stdlib.h>

lib/include/__stdarg___gnuc_va_list.h created+13

...	@@ -0,0 +1,13 @@
1	/*===---- __stdarg___gnuc_va_list.h - Definition of __gnuc_va_list ---------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	#ifndef __GNUC_VA_LIST
11	#define __GNUC_VA_LIST
12	typedef __builtin_va_list __gnuc_va_list;
13	#endif

lib/include/__stdarg___va_copy.h created+12

...	@@ -0,0 +1,12 @@
1	/*===---- __stdarg___va_copy.h - Definition of __va_copy -------------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	#ifndef __va_copy
11	#define __va_copy(d, s) __builtin_va_copy(d, s)
12	#endif

lib/include/__stdarg_va_arg.h created+22

...	@@ -0,0 +1,22 @@
1	/*===---- __stdarg_va_arg.h - Definitions of va_start, va_arg, va_end-------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	#ifndef va_arg
11
12	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
13	/* C23 does not require the second parameter for va_start. */
14	#define va_start(ap, ...) __builtin_va_start(ap, 0)
15	#else
16	/* Versions before C23 do require the second parameter. */
17	#define va_start(ap, param) __builtin_va_start(ap, param)
18	#endif
19	#define va_end(ap) __builtin_va_end(ap)
20	#define va_arg(ap, type) __builtin_va_arg(ap, type)
21
22	#endif

lib/include/__stdarg_va_copy.h created+12

...	@@ -0,0 +1,12 @@
1	/*===---- __stdarg_va_copy.h - Definition of va_copy------------------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	#ifndef va_copy
11	#define va_copy(dest, src) __builtin_va_copy(dest, src)
12	#endif

lib/include/__stdarg_va_list.h created+13

...	@@ -0,0 +1,13 @@
1	/*===---- __stdarg_va_list.h - Definition of va_list -----------------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	#ifndef _VA_LIST
11	#define _VA_LIST
12	typedef __builtin_va_list va_list;
13	#endif

lib/include/__stddef_max_align_t.h+1-1

...	@@ -1,4 +1,4 @@	...	@@ -1,4 +1,4 @@
1	/*===---- __stddef_max_align_t.h - Definition of max_align_t for modules ---===	1	/*===---- __stddef_max_align_t.h - Definition of max_align_t ---------------===
2	*	2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.	3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.	4	* See https://llvm.org/LICENSE.txt for license information.

lib/include/__stddef_null.h created+29

...	@@ -0,0 +1,29 @@
1	/*===---- __stddef_null.h - Definition of NULL -----------------------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	#if !defined(NULL) \|\| !__building_module(_Builtin_stddef)
11
12	/* linux/stddef.h will define NULL to 0. glibc (and other) headers then define
13	* __need_NULL and rely on stddef.h to redefine NULL to the correct value again.
14	* Modules don't support redefining macros like that, but support that pattern
15	* in the non-modules case.
16	*/
17	#undef NULL
18
19	#ifdef __cplusplus
20	#if !defined(__MINGW32__) && !defined(_MSC_VER)
21	#define NULL __null
22	#else
23	#define NULL 0
24	#endif
25	#else
26	#define NULL ((void*)0)
27	#endif
28
29	#endif

lib/include/__stddef_nullptr_t.h created+29

...	@@ -0,0 +1,29 @@
1	/*===---- __stddef_nullptr_t.h - Definition of nullptr_t -------------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	/*
11	* When -fbuiltin-headers-in-system-modules is set this is a non-modular header
12	* and needs to behave as if it was textual.
13	*/
14	#if !defined(_NULLPTR_T) \|\| \
15	(__has_feature(modules) && !__building_module(_Builtin_stddef))
16	#define _NULLPTR_T
17
18	#ifdef __cplusplus
19	#if defined(_MSC_EXTENSIONS) && defined(_NATIVE_NULLPTR_SUPPORTED)
20	namespace std {
21	typedef decltype(nullptr) nullptr_t;
22	}
23	using ::std::nullptr_t;
24	#endif
25	#elif defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
26	typedef typeof(nullptr) nullptr_t;
27	#endif
28
29	#endif

lib/include/__stddef_offsetof.h created+17

...	@@ -0,0 +1,17 @@
1	/*===---- __stddef_offsetof.h - Definition of offsetof ---------------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	/*
11	* When -fbuiltin-headers-in-system-modules is set this is a non-modular header
12	* and needs to behave as if it was textual.
13	*/
14	#if !defined(offsetof) \|\| \
15	(__has_feature(modules) && !__building_module(_Builtin_stddef))
16	#define offsetof(t, d) __builtin_offsetof(t, d)
17	#endif

lib/include/__stddef_ptrdiff_t.h created+20

...	@@ -0,0 +1,20 @@
1	/*===---- __stddef_ptrdiff_t.h - Definition of ptrdiff_t -------------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	/*
11	* When -fbuiltin-headers-in-system-modules is set this is a non-modular header
12	* and needs to behave as if it was textual.
13	*/
14	#if !defined(_PTRDIFF_T) \|\| \
15	(__has_feature(modules) && !__building_module(_Builtin_stddef))
16	#define _PTRDIFF_T
17
18	typedef __PTRDIFF_TYPE__ ptrdiff_t;
19
20	#endif

lib/include/__stddef_rsize_t.h created+20

...	@@ -0,0 +1,20 @@
1	/*===---- __stddef_rsize_t.h - Definition of rsize_t -----------------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	/*
11	* When -fbuiltin-headers-in-system-modules is set this is a non-modular header
12	* and needs to behave as if it was textual.
13	*/
14	#if !defined(_RSIZE_T) \|\| \
15	(__has_feature(modules) && !__building_module(_Builtin_stddef))
16	#define _RSIZE_T
17
18	typedef __SIZE_TYPE__ rsize_t;
19
20	#endif

lib/include/__stddef_size_t.h created+20

...	@@ -0,0 +1,20 @@
1	/*===---- __stddef_size_t.h - Definition of size_t -------------------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	/*
11	* When -fbuiltin-headers-in-system-modules is set this is a non-modular header
12	* and needs to behave as if it was textual.
13	*/
14	#if !defined(_SIZE_T) \|\| \
15	(__has_feature(modules) && !__building_module(_Builtin_stddef))
16	#define _SIZE_T
17
18	typedef __SIZE_TYPE__ size_t;
19
20	#endif

lib/include/__stddef_unreachable.h created+21

...	@@ -0,0 +1,21 @@
1	/*===---- __stddef_unreachable.h - Definition of unreachable ---------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	#ifndef __cplusplus
11
12	/*
13	* When -fbuiltin-headers-in-system-modules is set this is a non-modular header
14	* and needs to behave as if it was textual.
15	*/
16	#if !defined(unreachable) \|\| \
17	(__has_feature(modules) && !__building_module(_Builtin_stddef))
18	#define unreachable() __builtin_unreachable()
19	#endif
20
21	#endif

lib/include/__stddef_wchar_t.h created+28

...	@@ -0,0 +1,28 @@
1	/*===---- __stddef_wchar.h - Definition of wchar_t -------------------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	#if !defined(__cplusplus) \|\| (defined(_MSC_VER) && !_NATIVE_WCHAR_T_DEFINED)
11
12	/*
13	* When -fbuiltin-headers-in-system-modules is set this is a non-modular header
14	* and needs to behave as if it was textual.
15	*/
16	#if !defined(_WCHAR_T) \|\| \
17	(__has_feature(modules) && !__building_module(_Builtin_stddef))
18	#define _WCHAR_T
19
20	#ifdef _MSC_EXTENSIONS
21	#define _WCHAR_T_DEFINED
22	#endif
23
24	typedef __WCHAR_TYPE__ wchar_t;
25
26	#endif
27
28	#endif

lib/include/__stddef_wint_t.h created+15

...	@@ -0,0 +1,15 @@
1	/*===---- __stddef_wint.h - Definition of wint_t ---------------------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	#ifndef _WINT_T
11	#define _WINT_T
12
13	typedef __WINT_TYPE__ wint_t;
14
15	#endif

lib/include/adcintrin.h created+160

...	@@ -0,0 +1,160 @@
1	/*===---- adcintrin.h - ADC intrinsics -------------------------------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	#ifndef __ADCINTRIN_H
11	#define __ADCINTRIN_H
12
13	#if !defined(__i386__) && !defined(__x86_64__)
14	#error "This header is only meant to be used on x86 and x64 architecture"
15	#endif
16
17	/* Define the default attributes for the functions in this file. */
18	#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__))
19
20	/* Use C++ inline semantics in C++, GNU inline for C mode. */
21	#if defined(__cplusplus)
22	#define __INLINE __inline
23	#else
24	#define __INLINE static __inline
25	#endif
26
27	#if defined(__cplusplus)
28	extern "C" {
29	#endif
30
31	/// Adds unsigned 32-bit integers \a __x and \a __y, plus 0 or 1 as indicated
32	/// by the carry flag \a __cf. Stores the unsigned 32-bit sum in the memory
33	/// at \a __p, and returns the 8-bit carry-out (carry flag).
34	///
35	/// \code{.operation}
36	/// temp := (__cf == 0) ? 0 : 1
37	/// Store32(__p, __x + __y + temp)
38	/// result := CF
39	/// \endcode
40	///
41	/// \headerfile <immintrin.h>
42	///
43	/// This intrinsic corresponds to the \c ADC instruction.
44	///
45	/// \param __cf
46	/// The 8-bit unsigned carry flag; any non-zero value indicates carry.
47	/// \param __x
48	/// A 32-bit unsigned addend.
49	/// \param __y
50	/// A 32-bit unsigned addend.
51	/// \param __p
52	/// Pointer to memory for storing the sum.
53	/// \returns The 8-bit unsigned carry-out value.
54	__INLINE unsigned char __DEFAULT_FN_ATTRS _addcarry_u32(unsigned char __cf,
55	unsigned int __x,
56	unsigned int __y,
57	unsigned int *__p) {
58	return __builtin_ia32_addcarryx_u32(__cf, __x, __y, __p);
59	}
60
61	/// Adds unsigned 32-bit integer \a __y to 0 or 1 as indicated by the carry
62	/// flag \a __cf, and subtracts the result from unsigned 32-bit integer
63	/// \a __x. Stores the unsigned 32-bit difference in the memory at \a __p,
64	/// and returns the 8-bit carry-out (carry or overflow flag).
65	///
66	/// \code{.operation}
67	/// temp := (__cf == 0) ? 0 : 1
68	/// Store32(__p, __x - (__y + temp))
69	/// result := CF
70	/// \endcode
71	///
72	/// \headerfile <immintrin.h>
73	///
74	/// This intrinsic corresponds to the \c SBB instruction.
75	///
76	/// \param __cf
77	/// The 8-bit unsigned carry flag; any non-zero value indicates carry.
78	/// \param __x
79	/// The 32-bit unsigned minuend.
80	/// \param __y
81	/// The 32-bit unsigned subtrahend.
82	/// \param __p
83	/// Pointer to memory for storing the difference.
84	/// \returns The 8-bit unsigned carry-out value.
85	__INLINE unsigned char __DEFAULT_FN_ATTRS _subborrow_u32(unsigned char __cf,
86	unsigned int __x,
87	unsigned int __y,
88	unsigned int *__p) {
89	return __builtin_ia32_subborrow_u32(__cf, __x, __y, __p);
90	}
91
92	#ifdef __x86_64__
93	/// Adds unsigned 64-bit integers \a __x and \a __y, plus 0 or 1 as indicated
94	/// by the carry flag \a __cf. Stores the unsigned 64-bit sum in the memory
95	/// at \a __p, and returns the 8-bit carry-out (carry flag).
96	///
97	/// \code{.operation}
98	/// temp := (__cf == 0) ? 0 : 1
99	/// Store64(__p, __x + __y + temp)
100	/// result := CF
101	/// \endcode
102	///
103	/// \headerfile <immintrin.h>
104	///
105	/// This intrinsic corresponds to the \c ADC instruction.
106	///
107	/// \param __cf
108	/// The 8-bit unsigned carry flag; any non-zero value indicates carry.
109	/// \param __x
110	/// A 64-bit unsigned addend.
111	/// \param __y
112	/// A 64-bit unsigned addend.
113	/// \param __p
114	/// Pointer to memory for storing the sum.
115	/// \returns The 8-bit unsigned carry-out value.
116	__INLINE unsigned char __DEFAULT_FN_ATTRS
117	_addcarry_u64(unsigned char __cf, unsigned long long __x,
118	unsigned long long __y, unsigned long long *__p) {
119	return __builtin_ia32_addcarryx_u64(__cf, __x, __y, __p);
120	}
121
122	/// Adds unsigned 64-bit integer \a __y to 0 or 1 as indicated by the carry
123	/// flag \a __cf, and subtracts the result from unsigned 64-bit integer
124	/// \a __x. Stores the unsigned 64-bit difference in the memory at \a __p,
125	/// and returns the 8-bit carry-out (carry or overflow flag).
126	///
127	/// \code{.operation}
128	/// temp := (__cf == 0) ? 0 : 1
129	/// Store64(__p, __x - (__y + temp))
130	/// result := CF
131	/// \endcode
132	///
133	/// \headerfile <immintrin.h>
134	///
135	/// This intrinsic corresponds to the \c ADC instruction.
136	///
137	/// \param __cf
138	/// The 8-bit unsigned carry flag; any non-zero value indicates carry.
139	/// \param __x
140	/// The 64-bit unsigned minuend.
141	/// \param __y
142	/// The 64-bit unsigned subtrahend.
143	/// \param __p
144	/// Pointer to memory for storing the difference.
145	/// \returns The 8-bit unsigned carry-out value.
146	__INLINE unsigned char __DEFAULT_FN_ATTRS
147	_subborrow_u64(unsigned char __cf, unsigned long long __x,
148	unsigned long long __y, unsigned long long *__p) {
149	return __builtin_ia32_subborrow_u64(__cf, __x, __y, __p);
150	}
151	#endif
152
153	#if defined(__cplusplus)
154	}
155	#endif
156
157	#undef __INLINE
158	#undef __DEFAULT_FN_ATTRS
159
160	#endif /* __ADCINTRIN_H */

lib/include/adxintrin.h+9-134

...	@@ -15,7 +15,8 @@	...	@@ -15,7 +15,8 @@
15	#define __ADXINTRIN_H	15	#define __ADXINTRIN_H
16		16
17	/* Define the default attributes for the functions in this file. */	17	/* Define the default attributes for the functions in this file. */
18	#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__))	18	#define __DEFAULT_FN_ATTRS \
		19	__attribute__((__always_inline__, __nodebug__, __target__("adx")))
19		20
20	/* Use C++ inline semantics in C++, GNU inline for C mode. */	21	/* Use C++ inline semantics in C++, GNU inline for C mode. */
21	#if defined(__cplusplus)	22	#if defined(__cplusplus)
...	@@ -53,10 +54,10 @@ extern "C" {	...	@@ -53,10 +54,10 @@ extern "C" {
53	/// \param __p	54	/// \param __p
54	/// Pointer to memory for storing the sum.	55	/// Pointer to memory for storing the sum.
55	/// \returns The 8-bit unsigned carry-out value.	56	/// \returns The 8-bit unsigned carry-out value.
56	__INLINE unsigned char	57	__INLINE unsigned char __DEFAULT_FN_ATTRS _addcarryx_u32(unsigned char __cf,
57	__attribute__((__always_inline__, __nodebug__, __target__("adx")))	58	unsigned int __x,
58	_addcarryx_u32(unsigned char __cf, unsigned int __x, unsigned int __y,	59	unsigned int __y,
59	unsigned int *__p) {	60	unsigned int *__p) {
60	return __builtin_ia32_addcarryx_u32(__cf, __x, __y, __p);	61	return __builtin_ia32_addcarryx_u32(__cf, __x, __y, __p);
61	}	62	}
62		63
...	@@ -84,137 +85,10 @@ __INLINE unsigned char	...	@@ -84,137 +85,10 @@ __INLINE unsigned char
84	/// \param __p	85	/// \param __p
85	/// Pointer to memory for storing the sum.	86	/// Pointer to memory for storing the sum.
86	/// \returns The 8-bit unsigned carry-out value.	87	/// \returns The 8-bit unsigned carry-out value.
87	__INLINE unsigned char
88	__attribute__((__always_inline__, __nodebug__, __target__("adx")))
89	_addcarryx_u64(unsigned char __cf, unsigned long long __x,
90	unsigned long long __y, unsigned long long *__p) {
91	return __builtin_ia32_addcarryx_u64(__cf, __x, __y, __p);
92	}
93	#endif
94
95	/* Intrinsics that are also available if __ADX__ is undefined. */
96
97	/// Adds unsigned 32-bit integers \a __x and \a __y, plus 0 or 1 as indicated
98	/// by the carry flag \a __cf. Stores the unsigned 32-bit sum in the memory
99	/// at \a __p, and returns the 8-bit carry-out (carry flag).
100	///
101	/// \code{.operation}
102	/// temp := (__cf == 0) ? 0 : 1
103	/// Store32(__p, __x + __y + temp)
104	/// result := CF
105	/// \endcode
106	///
107	/// \headerfile <immintrin.h>
108	///
109	/// This intrinsic corresponds to the \c ADC instruction.
110	///
111	/// \param __cf
112	/// The 8-bit unsigned carry flag; any non-zero value indicates carry.
113	/// \param __x
114	/// A 32-bit unsigned addend.
115	/// \param __y
116	/// A 32-bit unsigned addend.
117	/// \param __p
118	/// Pointer to memory for storing the sum.
119	/// \returns The 8-bit unsigned carry-out value.
120	__INLINE unsigned char __DEFAULT_FN_ATTRS _addcarry_u32(unsigned char __cf,
121	unsigned int __x,
122	unsigned int __y,
123	unsigned int *__p) {
124	return __builtin_ia32_addcarryx_u32(__cf, __x, __y, __p);
125	}
126
127	#ifdef __x86_64__
128	/// Adds unsigned 64-bit integers \a __x and \a __y, plus 0 or 1 as indicated
129	/// by the carry flag \a __cf. Stores the unsigned 64-bit sum in the memory
130	/// at \a __p, and returns the 8-bit carry-out (carry flag).
131	///
132	/// \code{.operation}
133	/// temp := (__cf == 0) ? 0 : 1
134	/// Store64(__p, __x + __y + temp)
135	/// result := CF
136	/// \endcode
137	///
138	/// \headerfile <immintrin.h>
139	///
140	/// This intrinsic corresponds to the \c ADC instruction.
141	///
142	/// \param __cf
143	/// The 8-bit unsigned carry flag; any non-zero value indicates carry.
144	/// \param __x
145	/// A 64-bit unsigned addend.
146	/// \param __y
147	/// A 64-bit unsigned addend.
148	/// \param __p
149	/// Pointer to memory for storing the sum.
150	/// \returns The 8-bit unsigned carry-out value.
151	__INLINE unsigned char __DEFAULT_FN_ATTRS	88	__INLINE unsigned char __DEFAULT_FN_ATTRS
152	_addcarry_u64(unsigned char __cf, unsigned long long __x,	89	_addcarryx_u64(unsigned char __cf, unsigned long long __x,
153	unsigned long long __y, unsigned long long *__p) {
154	return __builtin_ia32_addcarryx_u64(__cf, __x, __y, __p);
155	}
156	#endif
157
158	/// Adds unsigned 32-bit integer \a __y to 0 or 1 as indicated by the carry
159	/// flag \a __cf, and subtracts the result from unsigned 32-bit integer
160	/// \a __x. Stores the unsigned 32-bit difference in the memory at \a __p,
161	/// and returns the 8-bit carry-out (carry or overflow flag).
162	///
163	/// \code{.operation}
164	/// temp := (__cf == 0) ? 0 : 1
165	/// Store32(__p, __x - (__y + temp))
166	/// result := CF
167	/// \endcode
168	///
169	/// \headerfile <immintrin.h>
170	///
171	/// This intrinsic corresponds to the \c SBB instruction.
172	///
173	/// \param __cf
174	/// The 8-bit unsigned carry flag; any non-zero value indicates carry.
175	/// \param __x
176	/// The 32-bit unsigned minuend.
177	/// \param __y
178	/// The 32-bit unsigned subtrahend.
179	/// \param __p
180	/// Pointer to memory for storing the difference.
181	/// \returns The 8-bit unsigned carry-out value.
182	__INLINE unsigned char __DEFAULT_FN_ATTRS _subborrow_u32(unsigned char __cf,
183	unsigned int __x,
184	unsigned int __y,
185	unsigned int *__p) {
186	return __builtin_ia32_subborrow_u32(__cf, __x, __y, __p);
187	}
188
189	#ifdef __x86_64__
190	/// Adds unsigned 64-bit integer \a __y to 0 or 1 as indicated by the carry
191	/// flag \a __cf, and subtracts the result from unsigned 64-bit integer
192	/// \a __x. Stores the unsigned 64-bit difference in the memory at \a __p,
193	/// and returns the 8-bit carry-out (carry or overflow flag).
194	///
195	/// \code{.operation}
196	/// temp := (__cf == 0) ? 0 : 1
197	/// Store64(__p, __x - (__y + temp))
198	/// result := CF
199	/// \endcode
200	///
201	/// \headerfile <immintrin.h>
202	///
203	/// This intrinsic corresponds to the \c ADC instruction.
204	///
205	/// \param __cf
206	/// The 8-bit unsigned carry flag; any non-zero value indicates carry.
207	/// \param __x
208	/// The 64-bit unsigned minuend.
209	/// \param __y
210	/// The 64-bit unsigned subtrahend.
211	/// \param __p
212	/// Pointer to memory for storing the difference.
213	/// \returns The 8-bit unsigned carry-out value.
214	__INLINE unsigned char __DEFAULT_FN_ATTRS
215	_subborrow_u64(unsigned char __cf, unsigned long long __x,
216	unsigned long long __y, unsigned long long *__p) {	90	unsigned long long __y, unsigned long long *__p) {
217	return __builtin_ia32_subborrow_u64(__cf, __x, __y, __p);	91	return __builtin_ia32_addcarryx_u64(__cf, __x, __y, __p);
218	}	92	}
219	#endif	93	#endif
220		94
...	@@ -222,6 +96,7 @@ _subborrow_u64(unsigned char __cf, unsigned long long __x,	...	@@ -222,6 +96,7 @@ _subborrow_u64(unsigned char __cf, unsigned long long __x,
222	}	96	}
223	#endif	97	#endif
224		98
		99	#undef __INLINE
225	#undef __DEFAULT_FN_ATTRS	100	#undef __DEFAULT_FN_ATTRS
226		101
227	#endif /* __ADXINTRIN_H */	102	#endif /* __ADXINTRIN_H */

lib/include/altivec.h+31-12

 @@ -14647,67 +14647,86 @@ static __inline__ void __ATTRS_o_ai vec_stvrxl(vector float __a, int __b,
 static __inline__ vector signed char __ATTRS_o_ai vec_promote(signed char __a,
                                                               int __b) {
-  vector signed char __res = (vector signed char)(0);
+  const vector signed char __zero = (vector signed char)0;
-  __res[__b & 0x7] = __a;
+  vector signed char __res =
+      __builtin_shufflevector(__zero, __zero, -1, -1, -1, -1, -1, -1, -1, -1,
+                              -1, -1, -1, -1, -1, -1, -1, -1);
+  __res[__b & 0xf] = __a;
   return __res;
 }
 static __inline__ vector unsigned char __ATTRS_o_ai
 vec_promote(unsigned char __a, int __b) {
-  vector unsigned char __res = (vector unsigned char)(0);
+  const vector unsigned char __zero = (vector unsigned char)(0);
-  __res[__b & 0x7] = __a;
+  vector unsigned char __res =
+      __builtin_shufflevector(__zero, __zero, -1, -1, -1, -1, -1, -1, -1, -1,
+                              -1, -1, -1, -1, -1, -1, -1, -1);
+  __res[__b & 0xf] = __a;
   return __res;
 }
 static __inline__ vector short __ATTRS_o_ai vec_promote(short __a, int __b) {
-  vector short __res = (vector short)(0);
+  const vector short __zero = (vector short)(0);
+  vector short __res =
+      __builtin_shufflevector(__zero, __zero, -1, -1, -1, -1, -1, -1, -1, -1);
   __res[__b & 0x7] = __a;
   return __res;
 }
 static __inline__ vector unsigned short __ATTRS_o_ai
 vec_promote(unsigned short __a, int __b) {
-  vector unsigned short __res = (vector unsigned short)(0);
+  const vector unsigned short __zero = (vector unsigned short)(0);
+  vector unsigned short __res =
+      __builtin_shufflevector(__zero, __zero, -1, -1, -1, -1, -1, -1, -1, -1);
   __res[__b & 0x7] = __a;
   return __res;
 }
 static __inline__ vector int __ATTRS_o_ai vec_promote(int __a, int __b) {
-  vector int __res = (vector int)(0);
+  const vector int __zero = (vector int)(0);
+  vector int __res = __builtin_shufflevector(__zero, __zero, -1, -1, -1, -1);
   __res[__b & 0x3] = __a;
   return __res;
 }
 static __inline__ vector unsigned int __ATTRS_o_ai vec_promote(unsigned int __a,
                                                                int __b) {
-  vector unsigned int __res = (vector unsigned int)(0);
+  const vector unsigned int __zero = (vector unsigned int)(0);
+  vector unsigned int __res =
+      __builtin_shufflevector(__zero, __zero, -1, -1, -1, -1);
   __res[__b & 0x3] = __a;
   return __res;
 }
 static __inline__ vector float __ATTRS_o_ai vec_promote(float __a, int __b) {
-  vector float __res = (vector float)(0);
+  const vector float __zero = (vector float)(0);
+  vector float __res = __builtin_shufflevector(__zero, __zero, -1, -1, -1, -1);
   __res[__b & 0x3] = __a;
   return __res;
 }
 #ifdef __VSX__
 static __inline__ vector double __ATTRS_o_ai vec_promote(double __a, int __b) {
-  vector double __res = (vector double)(0);
+  const vector double __zero = (vector double)(0);
+  vector double __res = __builtin_shufflevector(__zero, __zero, -1, -1);
   __res[__b & 0x1] = __a;
   return __res;
 }
 static __inline__ vector signed long long __ATTRS_o_ai
 vec_promote(signed long long __a, int __b) {
-  vector signed long long __res = (vector signed long long)(0);
+  const vector signed long long __zero = (vector signed long long)(0);
+  vector signed long long __res =
+      __builtin_shufflevector(__zero, __zero, -1, -1);
   __res[__b & 0x1] = __a;
   return __res;
 }
 static __inline__ vector unsigned long long __ATTRS_o_ai
 vec_promote(unsigned long long __a, int __b) {
-  vector unsigned long long __res = (vector unsigned long long)(0);
+  const vector unsigned long long __zero = (vector unsigned long long)(0);
+  vector unsigned long long __res =
+      __builtin_shufflevector(__zero, __zero, -1, -1);
   __res[__b & 0x1] = __a;
   return __res;
 }

lib/include/ammintrin.h+4-4

...	@@ -155,9 +155,9 @@ _mm_insert_si64(__m128i __x, __m128i __y)	...	@@ -155,9 +155,9 @@ _mm_insert_si64(__m128i __x, __m128i __y)
155	/// \param __a	155	/// \param __a
156	/// The 64-bit double-precision floating-point register value to be stored.	156	/// The 64-bit double-precision floating-point register value to be stored.
157	static __inline__ void __DEFAULT_FN_ATTRS	157	static __inline__ void __DEFAULT_FN_ATTRS
158	_mm_stream_sd(double *__p, __m128d __a)	158	_mm_stream_sd(void *__p, __m128d __a)
159	{	159	{
160	__builtin_ia32_movntsd(__p, (__v2df)__a);	160	__builtin_ia32_movntsd((double *)__p, (__v2df)__a);
161	}	161	}
162		162
163	/// Stores a 32-bit single-precision floating-point value in a 32-bit	163	/// Stores a 32-bit single-precision floating-point value in a 32-bit
...	@@ -173,9 +173,9 @@ _mm_stream_sd(double *__p, __m128d __a)	...	@@ -173,9 +173,9 @@ _mm_stream_sd(double *__p, __m128d __a)
173	/// \param __a	173	/// \param __a
174	/// The 32-bit single-precision floating-point register value to be stored.	174	/// The 32-bit single-precision floating-point register value to be stored.
175	static __inline__ void __DEFAULT_FN_ATTRS	175	static __inline__ void __DEFAULT_FN_ATTRS
176	_mm_stream_ss(float *__p, __m128 __a)	176	_mm_stream_ss(void *__p, __m128 __a)
177	{	177	{
178	__builtin_ia32_movntss(__p, (__v4sf)__a);	178	__builtin_ia32_movntss((float *)__p, (__v4sf)__a);
179	}	179	}
180		180
181	#undef __DEFAULT_FN_ATTRS	181	#undef __DEFAULT_FN_ATTRS

lib/include/arm_acle.h+111-31

 @@ -4,6 +4,13 @@
  * See https://llvm.org/LICENSE.txt for license information.
  * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
  *
+ * The Arm C Language Extensions specifications can be found in the following
+ * link: https://github.com/ARM-software/acle/releases
+ *
+ * The ACLE section numbers are subject to change. When consulting the
+ * specifications, it is recommended to search using section titles if
+ * the section numbers look outdated.
+ *
  *===-----------------------------------------------------------------------===
  */
 @@ -20,8 +27,8 @@
 extern "C" {
 #endif
-/* 8 SYNCHRONIZATION, BARRIER AND HINT INTRINSICS */
+/* 7 SYNCHRONIZATION, BARRIER AND HINT INTRINSICS */
-/* 8.3 Memory barriers */
+/* 7.3 Memory barriers */
 #if !__has_builtin(__dmb)
 #define __dmb(i) __builtin_arm_dmb(i)
 #endif
 @@ -32,7 +39,7 @@ extern "C" {
 #define __isb(i) __builtin_arm_isb(i)
 #endif
-/* 8.4 Hints */
+/* 7.4 Hints */
 #if !__has_builtin(__wfi)
 static __inline__ void __attribute__((__always_inline__, __nodebug__)) __wfi(void) {
 @@ -68,7 +75,7 @@ static __inline__ void __attribute__((__always_inline__, __nodebug__)) __yield(v
 #define __dbg(t) __builtin_arm_dbg(t)
 #endif
-/* 8.5 Swap */
+/* 7.5 Swap */
 static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
 __swp(uint32_t __x, volatile uint32_t *__p) {
   uint32_t v;
 @@ -78,8 +85,8 @@ __swp(uint32_t __x, volatile uint32_t *__p) {
   return v;
 }
-/* 8.6 Memory prefetch intrinsics */
+/* 7.6 Memory prefetch intrinsics */
-/* 8.6.1 Data prefetch */
+/* 7.6.1 Data prefetch */
 #define __pld(addr) __pldx(0, 0, 0, addr)
 #if defined(__ARM_32BIT_STATE) && __ARM_32BIT_STATE
 @@ -90,7 +97,7 @@ __swp(uint32_t __x, volatile uint32_t *__p) {
   __builtin_arm_prefetch(addr, access_kind, cache_level, retention_policy, 1)
 #endif
-/* 8.6.2 Instruction prefetch */
+/* 7.6.2 Instruction prefetch */
 #define __pli(addr) __plix(0, 0, addr)
 #if defined(__ARM_32BIT_STATE) && __ARM_32BIT_STATE
 @@ -101,15 +108,15 @@ __swp(uint32_t __x, volatile uint32_t *__p) {
   __builtin_arm_prefetch(addr, 0, cache_level, retention_policy, 0)
 #endif
-/* 8.7 NOP */
+/* 7.7 NOP */
 #if !defined(_MSC_VER) || !defined(__aarch64__)
 static __inline__ void __attribute__((__always_inline__, __nodebug__)) __nop(void) {
   __builtin_arm_nop();
 }
 #endif
-/* 9 DATA-PROCESSING INTRINSICS */
+/* 8 DATA-PROCESSING INTRINSICS */
-/* 9.2 Miscellaneous data-processing intrinsics */
+/* 8.2 Miscellaneous data-processing intrinsics */
 /* ROR */
 static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
 __ror(uint32_t __x, uint32_t __y) {
 @@ -248,9 +255,7 @@ __rbitl(unsigned long __t) {
 #endif
 }
-/*
+/* 8.3 16-bit multiplications */
- * 9.3 16-bit multiplications
- */
 #if defined(__ARM_FEATURE_DSP) && __ARM_FEATURE_DSP
 static __inline__ int32_t __attribute__((__always_inline__,__nodebug__))
 __smulbb(int32_t __a, int32_t __b) {
 @@ -279,18 +284,18 @@ __smulwt(int32_t __a, int32_t __b) {
 #endif
 /*
- * 9.4 Saturating intrinsics
+ * 8.4 Saturating intrinsics
  *
  * FIXME: Change guard to their corresponding __ARM_FEATURE flag when Q flag
  * intrinsics are implemented and the flag is enabled.
  */
-/* 9.4.1 Width-specified saturation intrinsics */
+/* 8.4.1 Width-specified saturation intrinsics */
 #if defined(__ARM_FEATURE_SAT) && __ARM_FEATURE_SAT
 #define __ssat(x, y) __builtin_arm_ssat(x, y)
 #define __usat(x, y) __builtin_arm_usat(x, y)
 #endif
-/* 9.4.2 Saturating addition and subtraction intrinsics */
+/* 8.4.2 Saturating addition and subtraction intrinsics */
 #if defined(__ARM_FEATURE_DSP) && __ARM_FEATURE_DSP
 static __inline__ int32_t __attribute__((__always_inline__, __nodebug__))
 __qadd(int32_t __t, int32_t __v) {
 @@ -308,7 +313,7 @@ __qdbl(int32_t __t) {
 }
 #endif
-/* 9.4.3 Accumultating multiplications */
+/* 8.4.3 Accumultating multiplications */
 #if defined(__ARM_FEATURE_DSP) && __ARM_FEATURE_DSP
 static __inline__ int32_t __attribute__((__always_inline__, __nodebug__))
 __smlabb(int32_t __a, int32_t __b, int32_t __c) {
 @@ -337,13 +342,13 @@ __smlawt(int32_t __a, int32_t __b, int32_t __c) {
 #endif
-/* 9.5.4 Parallel 16-bit saturation */
+/* 8.5.4 Parallel 16-bit saturation */
 #if defined(__ARM_FEATURE_SIMD32) && __ARM_FEATURE_SIMD32
 #define __ssat16(x, y) __builtin_arm_ssat16(x, y)
 #define __usat16(x, y) __builtin_arm_usat16(x, y)
 #endif
-/* 9.5.5 Packing and unpacking */
+/* 8.5.5 Packing and unpacking */
 #if defined(__ARM_FEATURE_SIMD32) && __ARM_FEATURE_SIMD32
 typedef int32_t int8x4_t;
 typedef int32_t int16x2_t;
 @@ -368,7 +373,7 @@ __uxtb16(int8x4_t __a) {
 }
 #endif
-/* 9.5.6 Parallel selection */
+/* 8.5.6 Parallel selection */
 #if defined(__ARM_FEATURE_SIMD32) && __ARM_FEATURE_SIMD32
 static __inline__ uint8x4_t __attribute__((__always_inline__, __nodebug__))
 __sel(uint8x4_t __a, uint8x4_t __b) {
 @@ -376,7 +381,7 @@ __sel(uint8x4_t __a, uint8x4_t __b) {
 }
 #endif
-/* 9.5.7 Parallel 8-bit addition and subtraction */
+/* 8.5.7 Parallel 8-bit addition and subtraction */
 #if defined(__ARM_FEATURE_SIMD32) && __ARM_FEATURE_SIMD32
 static __inline__ int8x4_t __attribute__((__always_inline__, __nodebug__))
 __qadd8(int8x4_t __a, int8x4_t __b) {
 @@ -428,7 +433,7 @@ __usub8(uint8x4_t __a, uint8x4_t __b) {
 }
 #endif
-/* 9.5.8 Sum of 8-bit absolute differences */
+/* 8.5.8 Sum of 8-bit absolute differences */
 #if defined(__ARM_FEATURE_SIMD32) && __ARM_FEATURE_SIMD32
 static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
 __usad8(uint8x4_t __a, uint8x4_t __b) {
 @@ -440,7 +445,7 @@ __usada8(uint8x4_t __a, uint8x4_t __b, uint32_t __c) {
 }
 #endif
-/* 9.5.9 Parallel 16-bit addition and subtraction */
+/* 8.5.9 Parallel 16-bit addition and subtraction */
 #if defined(__ARM_FEATURE_SIMD32) && __ARM_FEATURE_SIMD32
 static __inline__ int16x2_t __attribute__((__always_inline__, __nodebug__))
 __qadd16(int16x2_t __a, int16x2_t __b) {
 @@ -540,7 +545,7 @@ __usub16(uint16x2_t __a, uint16x2_t __b) {
 }
 #endif
-/* 9.5.10 Parallel 16-bit multiplications */
+/* 8.5.10 Parallel 16-bit multiplications */
 #if defined(__ARM_FEATURE_SIMD32) && __ARM_FEATURE_SIMD32
 static __inline__ int32_t __attribute__((__always_inline__, __nodebug__))
 __smlad(int16x2_t __a, int16x2_t __b, int32_t __c) {
 @@ -592,7 +597,22 @@ __smusdx(int16x2_t __a, int16x2_t __b) {
 }
 #endif
-/* 9.7 CRC32 intrinsics */
+/* 8.6 Floating-point data-processing intrinsics */
+#if (defined(__ARM_FEATURE_DIRECTED_ROUNDING)    &&                         \
+  (__ARM_FEATURE_DIRECTED_ROUNDING))             &&                         \
+  (defined(__ARM_64BIT_STATE) && __ARM_64BIT_STATE)
+static __inline__ double __attribute__((__always_inline__, __nodebug__))
+__rintn(double __a) {
+  return __builtin_roundeven(__a);
+}
+static __inline__ float __attribute__((__always_inline__, __nodebug__))
+__rintnf(float __a) {
+  return __builtin_roundevenf(__a);
+}
+#endif
+/* 8.8 CRC32 intrinsics */
 #if (defined(__ARM_FEATURE_CRC32) && __ARM_FEATURE_CRC32) ||                   \
     (defined(__ARM_64BIT_STATE) && __ARM_64BIT_STATE)
 static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__, target("crc")))
 @@ -636,6 +656,7 @@ __crc32cd(uint32_t __a, uint64_t __b) {
 }
 #endif
+/* 8.6 Floating-point data-processing intrinsics */
 /* Armv8.3-A Javascript conversion intrinsic */
 #if defined(__ARM_64BIT_STATE) && __ARM_64BIT_STATE
 static __inline__ int32_t __attribute__((__always_inline__, __nodebug__, target("v8.3a")))
 @@ -687,7 +708,7 @@ __rint64x(double __a) {
 }
 #endif
-/* Armv8.7-A load/store 64-byte intrinsics */
+/* 8.9 Armv8.7-A load/store 64-byte intrinsics */
 #if defined(__ARM_64BIT_STATE) && __ARM_64BIT_STATE
 typedef struct {
     uint64_t val[8];
 @@ -713,7 +734,7 @@ __arm_st64bv0(void *__addr, data512_t __value) {
 }
 #endif
-/* 10.1 Special register intrinsics */
+/* 11.1 Special register intrinsics */
 #define __arm_rsr(sysreg) __builtin_arm_rsr(sysreg)
 #define __arm_rsr64(sysreg) __builtin_arm_rsr64(sysreg)
 #define __arm_rsr128(sysreg) __builtin_arm_rsr128(sysreg)
 @@ -727,7 +748,7 @@ __arm_st64bv0(void *__addr, data512_t __value) {
 #define __arm_wsrf(sysreg, v) __arm_wsr(sysreg, __builtin_bit_cast(uint32_t, v))
 #define __arm_wsrf64(sysreg, v) __arm_wsr64(sysreg, __builtin_bit_cast(uint64_t, v))
-/* Memory Tagging Extensions (MTE) Intrinsics */
+/* 10.3 Memory Tagging Extensions (MTE) Intrinsics */
 #if defined(__ARM_64BIT_STATE) && __ARM_64BIT_STATE
 #define __arm_mte_create_random_tag(__ptr, __mask)  __builtin_arm_irg(__ptr, __mask)
 #define __arm_mte_increment_tag(__ptr, __tag_offset)  __builtin_arm_addg(__ptr, __tag_offset)
 @@ -736,12 +757,71 @@ __arm_st64bv0(void *__addr, data512_t __value) {
 #define __arm_mte_set_tag(__ptr) __builtin_arm_stg(__ptr)
 #define __arm_mte_ptrdiff(__ptra, __ptrb) __builtin_arm_subp(__ptra, __ptrb)
-/* Memory Operations Intrinsics */
+/* 18 Memory Operations Intrinsics */
 #define __arm_mops_memset_tag(__tagged_address, __value, __size)    \
   __builtin_arm_mops_memset_tag(__tagged_address, __value, __size)
 #endif
-/* Transactional Memory Extension (TME) Intrinsics */
+/* 11.3 Coprocessor Intrinsics */
+#if defined(__ARM_FEATURE_COPROC)
+#if (__ARM_FEATURE_COPROC & 0x1)
+#if (__ARM_ARCH < 8)
+#define __arm_cdp(coproc, opc1, CRd, CRn, CRm, opc2)                           \
+  __builtin_arm_cdp(coproc, opc1, CRd, CRn, CRm, opc2)
+#endif /* __ARM_ARCH < 8 */
+#define __arm_ldc(coproc, CRd, p) __builtin_arm_ldc(coproc, CRd, p)
+#define __arm_stc(coproc, CRd, p) __builtin_arm_stc(coproc, CRd, p)
+#define __arm_mcr(coproc, opc1, value, CRn, CRm, opc2)                         \
+  __builtin_arm_mcr(coproc, opc1, value, CRn, CRm, opc2)
+#define __arm_mrc(coproc, opc1, CRn, CRm, opc2)                                \
+  __builtin_arm_mrc(coproc, opc1, CRn, CRm, opc2)
+#if (__ARM_ARCH != 4) && (__ARM_ARCH < 8)
+#define __arm_ldcl(coproc, CRd, p) __builtin_arm_ldcl(coproc, CRd, p)
+#define __arm_stcl(coproc, CRd, p) __builtin_arm_stcl(coproc, CRd, p)
+#endif /* (__ARM_ARCH != 4) && (__ARM_ARCH != 8) */
+#if (__ARM_ARCH_8M_MAIN__) || (__ARM_ARCH_8_1M_MAIN__)
+#define __arm_cdp(coproc, opc1, CRd, CRn, CRm, opc2)                           \
+  __builtin_arm_cdp(coproc, opc1, CRd, CRn, CRm, opc2)
+#define __arm_ldcl(coproc, CRd, p) __builtin_arm_ldcl(coproc, CRd, p)
+#define __arm_stcl(coproc, CRd, p) __builtin_arm_stcl(coproc, CRd, p)
+#endif /* ___ARM_ARCH_8M_MAIN__ */
+#endif /* __ARM_FEATURE_COPROC & 0x1 */
+#if (__ARM_FEATURE_COPROC & 0x2)
+#define __arm_cdp2(coproc, opc1, CRd, CRn, CRm, opc2)                          \
+  __builtin_arm_cdp2(coproc, opc1, CRd, CRn, CRm, opc2)
+#define __arm_ldc2(coproc, CRd, p) __builtin_arm_ldc2(coproc, CRd, p)
+#define __arm_stc2(coproc, CRd, p) __builtin_arm_stc2(coproc, CRd, p)
+#define __arm_ldc2l(coproc, CRd, p) __builtin_arm_ldc2l(coproc, CRd, p)
+#define __arm_stc2l(coproc, CRd, p) __builtin_arm_stc2l(coproc, CRd, p)
+#define __arm_mcr2(coproc, opc1, value, CRn, CRm, opc2)                        \
+  __builtin_arm_mcr2(coproc, opc1, value, CRn, CRm, opc2)
+#define __arm_mrc2(coproc, opc1, CRn, CRm, opc2)                               \
+  __builtin_arm_mrc2(coproc, opc1, CRn, CRm, opc2)
+#endif
+#if (__ARM_FEATURE_COPROC & 0x4)
+#define __arm_mcrr(coproc, opc1, value, CRm)                                   \
+  __builtin_arm_mcrr(coproc, opc1, value, CRm)
+#define __arm_mrrc(coproc, opc1, CRm) __builtin_arm_mrrc(coproc, opc1, CRm)
+#endif
+#if (__ARM_FEATURE_COPROC & 0x8)
+#define __arm_mcrr2(coproc, opc1, value, CRm)                                  \
+  __builtin_arm_mcrr2(coproc, opc1, value, CRm)
+#define __arm_mrrc2(coproc, opc1, CRm) __builtin_arm_mrrc2(coproc, opc1, CRm)
+#endif
+#endif // __ARM_FEATURE_COPROC
+/* 17 Transactional Memory Extension (TME) Intrinsics */
 #if defined(__ARM_FEATURE_TME) && __ARM_FEATURE_TME
 #define _TMFAILURE_REASON  0x00007fffu
 @@ -763,7 +843,7 @@ __arm_st64bv0(void *__addr, data512_t __value) {
 #endif /* __ARM_FEATURE_TME */
-/* Armv8.5-A Random number generation intrinsics */
+/* 8.7 Armv8.5-A Random number generation intrinsics */
 #if defined(__ARM_64BIT_STATE) && __ARM_64BIT_STATE
 static __inline__ int __attribute__((__always_inline__, __nodebug__, target("rand")))
 __rndr(uint64_t *__p) {

lib/include/arm_neon.h+85-327

...	@@ -35,12 +35,7 @@	...	@@ -35,12 +35,7 @@
35	#include <stdint.h>	35	#include <stdint.h>
36		36
37	#include <arm_bf16.h>	37	#include <arm_bf16.h>
38	typedef float float32_t;	38	#include <arm_vector_types.h>
39	typedef __fp16 float16_t;
40	#ifdef __aarch64__
41	typedef double float64_t;
42	#endif
43
44	#ifdef __aarch64__	39	#ifdef __aarch64__
45	typedef uint8_t poly8_t;	40	typedef uint8_t poly8_t;
46	typedef uint16_t poly16_t;	41	typedef uint16_t poly16_t;
...	@@ -51,30 +46,6 @@ typedef int8_t poly8_t;	...	@@ -51,30 +46,6 @@ typedef int8_t poly8_t;
51	typedef int16_t poly16_t;	46	typedef int16_t poly16_t;
52	typedef int64_t poly64_t;	47	typedef int64_t poly64_t;
53	#endif	48	#endif
54	typedef __attribute__((neon_vector_type(8))) int8_t int8x8_t;
55	typedef __attribute__((neon_vector_type(16))) int8_t int8x16_t;
56	typedef __attribute__((neon_vector_type(4))) int16_t int16x4_t;
57	typedef __attribute__((neon_vector_type(8))) int16_t int16x8_t;
58	typedef __attribute__((neon_vector_type(2))) int32_t int32x2_t;
59	typedef __attribute__((neon_vector_type(4))) int32_t int32x4_t;
60	typedef __attribute__((neon_vector_type(1))) int64_t int64x1_t;
61	typedef __attribute__((neon_vector_type(2))) int64_t int64x2_t;
62	typedef __attribute__((neon_vector_type(8))) uint8_t uint8x8_t;
63	typedef __attribute__((neon_vector_type(16))) uint8_t uint8x16_t;
64	typedef __attribute__((neon_vector_type(4))) uint16_t uint16x4_t;
65	typedef __attribute__((neon_vector_type(8))) uint16_t uint16x8_t;
66	typedef __attribute__((neon_vector_type(2))) uint32_t uint32x2_t;
67	typedef __attribute__((neon_vector_type(4))) uint32_t uint32x4_t;
68	typedef __attribute__((neon_vector_type(1))) uint64_t uint64x1_t;
69	typedef __attribute__((neon_vector_type(2))) uint64_t uint64x2_t;
70	typedef __attribute__((neon_vector_type(4))) float16_t float16x4_t;
71	typedef __attribute__((neon_vector_type(8))) float16_t float16x8_t;
72	typedef __attribute__((neon_vector_type(2))) float32_t float32x2_t;
73	typedef __attribute__((neon_vector_type(4))) float32_t float32x4_t;
74	#ifdef __aarch64__
75	typedef __attribute__((neon_vector_type(1))) float64_t float64x1_t;
76	typedef __attribute__((neon_vector_type(2))) float64_t float64x2_t;
77	#endif
78	typedef __attribute__((neon_polyvector_type(8))) poly8_t poly8x8_t;	49	typedef __attribute__((neon_polyvector_type(8))) poly8_t poly8x8_t;
79	typedef __attribute__((neon_polyvector_type(16))) poly8_t poly8x16_t;	50	typedef __attribute__((neon_polyvector_type(16))) poly8_t poly8x16_t;
80	typedef __attribute__((neon_polyvector_type(4))) poly16_t poly16x4_t;	51	typedef __attribute__((neon_polyvector_type(4))) poly16_t poly16x4_t;
...	@@ -82,96 +53,6 @@ typedef __attribute__((neon_polyvector_type(8))) poly16_t poly16x8_t;	...	@@ -82,96 +53,6 @@ typedef __attribute__((neon_polyvector_type(8))) poly16_t poly16x8_t;
82	typedef __attribute__((neon_polyvector_type(1))) poly64_t poly64x1_t;	53	typedef __attribute__((neon_polyvector_type(1))) poly64_t poly64x1_t;
83	typedef __attribute__((neon_polyvector_type(2))) poly64_t poly64x2_t;	54	typedef __attribute__((neon_polyvector_type(2))) poly64_t poly64x2_t;
84		55
85	typedef struct int8x8x2_t {
86	int8x8_t val[2];
87	} int8x8x2_t;
88
89	typedef struct int8x16x2_t {
90	int8x16_t val[2];
91	} int8x16x2_t;
92
93	typedef struct int16x4x2_t {
94	int16x4_t val[2];
95	} int16x4x2_t;
96
97	typedef struct int16x8x2_t {
98	int16x8_t val[2];
99	} int16x8x2_t;
100
101	typedef struct int32x2x2_t {
102	int32x2_t val[2];
103	} int32x2x2_t;
104
105	typedef struct int32x4x2_t {
106	int32x4_t val[2];
107	} int32x4x2_t;
108
109	typedef struct int64x1x2_t {
110	int64x1_t val[2];
111	} int64x1x2_t;
112
113	typedef struct int64x2x2_t {
114	int64x2_t val[2];
115	} int64x2x2_t;
116
117	typedef struct uint8x8x2_t {
118	uint8x8_t val[2];
119	} uint8x8x2_t;
120
121	typedef struct uint8x16x2_t {
122	uint8x16_t val[2];
123	} uint8x16x2_t;
124
125	typedef struct uint16x4x2_t {
126	uint16x4_t val[2];
127	} uint16x4x2_t;
128
129	typedef struct uint16x8x2_t {
130	uint16x8_t val[2];
131	} uint16x8x2_t;
132
133	typedef struct uint32x2x2_t {
134	uint32x2_t val[2];
135	} uint32x2x2_t;
136
137	typedef struct uint32x4x2_t {
138	uint32x4_t val[2];
139	} uint32x4x2_t;
140
141	typedef struct uint64x1x2_t {
142	uint64x1_t val[2];
143	} uint64x1x2_t;
144
145	typedef struct uint64x2x2_t {
146	uint64x2_t val[2];
147	} uint64x2x2_t;
148
149	typedef struct float16x4x2_t {
150	float16x4_t val[2];
151	} float16x4x2_t;
152
153	typedef struct float16x8x2_t {
154	float16x8_t val[2];
155	} float16x8x2_t;
156
157	typedef struct float32x2x2_t {
158	float32x2_t val[2];
159	} float32x2x2_t;
160
161	typedef struct float32x4x2_t {
162	float32x4_t val[2];
163	} float32x4x2_t;
164
165	#ifdef __aarch64__
166	typedef struct float64x1x2_t {
167	float64x1_t val[2];
168	} float64x1x2_t;
169
170	typedef struct float64x2x2_t {
171	float64x2_t val[2];
172	} float64x2x2_t;
173
174	#endif
175	typedef struct poly8x8x2_t {	56	typedef struct poly8x8x2_t {
176	poly8x8_t val[2];	57	poly8x8_t val[2];
177	} poly8x8x2_t;	58	} poly8x8x2_t;
...	@@ -196,96 +77,6 @@ typedef struct poly64x2x2_t {	...	@@ -196,96 +77,6 @@ typedef struct poly64x2x2_t {
196	poly64x2_t val[2];	77	poly64x2_t val[2];
197	} poly64x2x2_t;	78	} poly64x2x2_t;
198		79
199	typedef struct int8x8x3_t {
200	int8x8_t val[3];
201	} int8x8x3_t;
202
203	typedef struct int8x16x3_t {
204	int8x16_t val[3];
205	} int8x16x3_t;
206
207	typedef struct int16x4x3_t {
208	int16x4_t val[3];
209	} int16x4x3_t;
210
211	typedef struct int16x8x3_t {
212	int16x8_t val[3];
213	} int16x8x3_t;
214
215	typedef struct int32x2x3_t {
216	int32x2_t val[3];
217	} int32x2x3_t;
218
219	typedef struct int32x4x3_t {
220	int32x4_t val[3];
221	} int32x4x3_t;
222
223	typedef struct int64x1x3_t {
224	int64x1_t val[3];
225	} int64x1x3_t;
226
227	typedef struct int64x2x3_t {
228	int64x2_t val[3];
229	} int64x2x3_t;
230
231	typedef struct uint8x8x3_t {
232	uint8x8_t val[3];
233	} uint8x8x3_t;
234
235	typedef struct uint8x16x3_t {
236	uint8x16_t val[3];
237	} uint8x16x3_t;
238
239	typedef struct uint16x4x3_t {
240	uint16x4_t val[3];
241	} uint16x4x3_t;
242
243	typedef struct uint16x8x3_t {
244	uint16x8_t val[3];
245	} uint16x8x3_t;
246
247	typedef struct uint32x2x3_t {
248	uint32x2_t val[3];
249	} uint32x2x3_t;
250
251	typedef struct uint32x4x3_t {
252	uint32x4_t val[3];
253	} uint32x4x3_t;
254
255	typedef struct uint64x1x3_t {
256	uint64x1_t val[3];
257	} uint64x1x3_t;
258
259	typedef struct uint64x2x3_t {
260	uint64x2_t val[3];
261	} uint64x2x3_t;
262
263	typedef struct float16x4x3_t {
264	float16x4_t val[3];
265	} float16x4x3_t;
266
267	typedef struct float16x8x3_t {
268	float16x8_t val[3];
269	} float16x8x3_t;
270
271	typedef struct float32x2x3_t {
272	float32x2_t val[3];
273	} float32x2x3_t;
274
275	typedef struct float32x4x3_t {
276	float32x4_t val[3];
277	} float32x4x3_t;
278
279	#ifdef __aarch64__
280	typedef struct float64x1x3_t {
281	float64x1_t val[3];
282	} float64x1x3_t;
283
284	typedef struct float64x2x3_t {
285	float64x2_t val[3];
286	} float64x2x3_t;
287
288	#endif
289	typedef struct poly8x8x3_t {	80	typedef struct poly8x8x3_t {
290	poly8x8_t val[3];	81	poly8x8_t val[3];
291	} poly8x8x3_t;	82	} poly8x8x3_t;
...	@@ -310,96 +101,6 @@ typedef struct poly64x2x3_t {	...	@@ -310,96 +101,6 @@ typedef struct poly64x2x3_t {
310	poly64x2_t val[3];	101	poly64x2_t val[3];
311	} poly64x2x3_t;	102	} poly64x2x3_t;
312		103
313	typedef struct int8x8x4_t {
314	int8x8_t val[4];
315	} int8x8x4_t;
316
317	typedef struct int8x16x4_t {
318	int8x16_t val[4];
319	} int8x16x4_t;
320
321	typedef struct int16x4x4_t {
322	int16x4_t val[4];
323	} int16x4x4_t;
324
325	typedef struct int16x8x4_t {
326	int16x8_t val[4];
327	} int16x8x4_t;
328
329	typedef struct int32x2x4_t {
330	int32x2_t val[4];
331	} int32x2x4_t;
332
333	typedef struct int32x4x4_t {
334	int32x4_t val[4];
335	} int32x4x4_t;
336
337	typedef struct int64x1x4_t {
338	int64x1_t val[4];
339	} int64x1x4_t;
340
341	typedef struct int64x2x4_t {
342	int64x2_t val[4];
343	} int64x2x4_t;
344
345	typedef struct uint8x8x4_t {
346	uint8x8_t val[4];
347	} uint8x8x4_t;
348
349	typedef struct uint8x16x4_t {
350	uint8x16_t val[4];
351	} uint8x16x4_t;
352
353	typedef struct uint16x4x4_t {
354	uint16x4_t val[4];
355	} uint16x4x4_t;
356
357	typedef struct uint16x8x4_t {
358	uint16x8_t val[4];
359	} uint16x8x4_t;
360
361	typedef struct uint32x2x4_t {
362	uint32x2_t val[4];
363	} uint32x2x4_t;
364
365	typedef struct uint32x4x4_t {
366	uint32x4_t val[4];
367	} uint32x4x4_t;
368
369	typedef struct uint64x1x4_t {
370	uint64x1_t val[4];
371	} uint64x1x4_t;
372
373	typedef struct uint64x2x4_t {
374	uint64x2_t val[4];
375	} uint64x2x4_t;
376
377	typedef struct float16x4x4_t {
378	float16x4_t val[4];
379	} float16x4x4_t;
380
381	typedef struct float16x8x4_t {
382	float16x8_t val[4];
383	} float16x8x4_t;
384
385	typedef struct float32x2x4_t {
386	float32x2_t val[4];
387	} float32x2x4_t;
388
389	typedef struct float32x4x4_t {
390	float32x4_t val[4];
391	} float32x4x4_t;
392
393	#ifdef __aarch64__
394	typedef struct float64x1x4_t {
395	float64x1_t val[4];
396	} float64x1x4_t;
397
398	typedef struct float64x2x4_t {
399	float64x2_t val[4];
400	} float64x2x4_t;
401
402	#endif
403	typedef struct poly8x8x4_t {	104	typedef struct poly8x8x4_t {
404	poly8x8_t val[4];	105	poly8x8_t val[4];
405	} poly8x8x4_t;	106	} poly8x8x4_t;
...	@@ -424,33 +125,6 @@ typedef struct poly64x2x4_t {	...	@@ -424,33 +125,6 @@ typedef struct poly64x2x4_t {
424	poly64x2_t val[4];	125	poly64x2_t val[4];
425	} poly64x2x4_t;	126	} poly64x2x4_t;
426		127
427	typedef __attribute__((neon_vector_type(4))) bfloat16_t bfloat16x4_t;
428	typedef __attribute__((neon_vector_type(8))) bfloat16_t bfloat16x8_t;
429
430	typedef struct bfloat16x4x2_t {
431	bfloat16x4_t val[2];
432	} bfloat16x4x2_t;
433
434	typedef struct bfloat16x8x2_t {
435	bfloat16x8_t val[2];
436	} bfloat16x8x2_t;
437
438	typedef struct bfloat16x4x3_t {
439	bfloat16x4_t val[3];
440	} bfloat16x4x3_t;
441
442	typedef struct bfloat16x8x3_t {
443	bfloat16x8_t val[3];
444	} bfloat16x8x3_t;
445
446	typedef struct bfloat16x4x4_t {
447	bfloat16x4_t val[4];
448	} bfloat16x4x4_t;
449
450	typedef struct bfloat16x8x4_t {
451	bfloat16x8_t val[4];
452	} bfloat16x8x4_t;
453
454	#define __ai static __inline__ __attribute__((__always_inline__, __nodebug__))	128	#define __ai static __inline__ __attribute__((__always_inline__, __nodebug__))
455		129
456	#ifdef __LITTLE_ENDIAN__	130	#ifdef __LITTLE_ENDIAN__
...	@@ -66600,6 +66274,27 @@ __ai __attribute__((target("v8.5a"))) float32x2_t vrnd32x_f32(float32x2_t __p0)	...	@@ -66600,6 +66274,27 @@ __ai __attribute__((target("v8.5a"))) float32x2_t vrnd32x_f32(float32x2_t __p0)
66600	}	66274	}
66601	#endif	66275	#endif
66602		66276
		66277	#ifdef __LITTLE_ENDIAN__
		66278	__ai __attribute__((target("v8.5a"))) float64x2_t vrnd32xq_f64(float64x2_t __p0) {
		66279	float64x2_t __ret;
		66280	__ret = (float64x2_t) __builtin_neon_vrnd32xq_f64((int8x16_t)__p0, 42);
		66281	return __ret;
		66282	}
		66283	#else
		66284	__ai __attribute__((target("v8.5a"))) float64x2_t vrnd32xq_f64(float64x2_t __p0) {
		66285	float64x2_t __ret;
		66286	float64x2_t __rev0; __rev0 = __builtin_shufflevector(__p0, __p0, 1, 0);
		66287	__ret = (float64x2_t) __builtin_neon_vrnd32xq_f64((int8x16_t)__rev0, 42);
		66288	__ret = __builtin_shufflevector(__ret, __ret, 1, 0);
		66289	return __ret;
		66290	}
		66291	#endif
		66292
		66293	__ai __attribute__((target("v8.5a"))) float64x1_t vrnd32x_f64(float64x1_t __p0) {
		66294	float64x1_t __ret;
		66295	__ret = (float64x1_t) __builtin_neon_vrnd32x_f64((int8x8_t)__p0, 10);
		66296	return __ret;
		66297	}
66603	#ifdef __LITTLE_ENDIAN__	66298	#ifdef __LITTLE_ENDIAN__
66604	__ai __attribute__((target("v8.5a"))) float32x4_t vrnd32zq_f32(float32x4_t __p0) {	66299	__ai __attribute__((target("v8.5a"))) float32x4_t vrnd32zq_f32(float32x4_t __p0) {
66605	float32x4_t __ret;	66300	float32x4_t __ret;
...	@@ -66632,6 +66327,27 @@ __ai __attribute__((target("v8.5a"))) float32x2_t vrnd32z_f32(float32x2_t __p0)	...	@@ -66632,6 +66327,27 @@ __ai __attribute__((target("v8.5a"))) float32x2_t vrnd32z_f32(float32x2_t __p0)
66632	}	66327	}
66633	#endif	66328	#endif
66634		66329
		66330	#ifdef __LITTLE_ENDIAN__
		66331	__ai __attribute__((target("v8.5a"))) float64x2_t vrnd32zq_f64(float64x2_t __p0) {
		66332	float64x2_t __ret;
		66333	__ret = (float64x2_t) __builtin_neon_vrnd32zq_f64((int8x16_t)__p0, 42);
		66334	return __ret;
		66335	}
		66336	#else
		66337	__ai __attribute__((target("v8.5a"))) float64x2_t vrnd32zq_f64(float64x2_t __p0) {
		66338	float64x2_t __ret;
		66339	float64x2_t __rev0; __rev0 = __builtin_shufflevector(__p0, __p0, 1, 0);
		66340	__ret = (float64x2_t) __builtin_neon_vrnd32zq_f64((int8x16_t)__rev0, 42);
		66341	__ret = __builtin_shufflevector(__ret, __ret, 1, 0);
		66342	return __ret;
		66343	}
		66344	#endif
		66345
		66346	__ai __attribute__((target("v8.5a"))) float64x1_t vrnd32z_f64(float64x1_t __p0) {
		66347	float64x1_t __ret;
		66348	__ret = (float64x1_t) __builtin_neon_vrnd32z_f64((int8x8_t)__p0, 10);
		66349	return __ret;
		66350	}
66635	#ifdef __LITTLE_ENDIAN__	66351	#ifdef __LITTLE_ENDIAN__
66636	__ai __attribute__((target("v8.5a"))) float32x4_t vrnd64xq_f32(float32x4_t __p0) {	66352	__ai __attribute__((target("v8.5a"))) float32x4_t vrnd64xq_f32(float32x4_t __p0) {
66637	float32x4_t __ret;	66353	float32x4_t __ret;
...	@@ -66664,6 +66380,27 @@ __ai __attribute__((target("v8.5a"))) float32x2_t vrnd64x_f32(float32x2_t __p0)	...	@@ -66664,6 +66380,27 @@ __ai __attribute__((target("v8.5a"))) float32x2_t vrnd64x_f32(float32x2_t __p0)
66664	}	66380	}
66665	#endif	66381	#endif
66666		66382
		66383	#ifdef __LITTLE_ENDIAN__
		66384	__ai __attribute__((target("v8.5a"))) float64x2_t vrnd64xq_f64(float64x2_t __p0) {
		66385	float64x2_t __ret;
		66386	__ret = (float64x2_t) __builtin_neon_vrnd64xq_f64((int8x16_t)__p0, 42);
		66387	return __ret;
		66388	}
		66389	#else
		66390	__ai __attribute__((target("v8.5a"))) float64x2_t vrnd64xq_f64(float64x2_t __p0) {
		66391	float64x2_t __ret;
		66392	float64x2_t __rev0; __rev0 = __builtin_shufflevector(__p0, __p0, 1, 0);
		66393	__ret = (float64x2_t) __builtin_neon_vrnd64xq_f64((int8x16_t)__rev0, 42);
		66394	__ret = __builtin_shufflevector(__ret, __ret, 1, 0);
		66395	return __ret;
		66396	}
		66397	#endif
		66398
		66399	__ai __attribute__((target("v8.5a"))) float64x1_t vrnd64x_f64(float64x1_t __p0) {
		66400	float64x1_t __ret;
		66401	__ret = (float64x1_t) __builtin_neon_vrnd64x_f64((int8x8_t)__p0, 10);
		66402	return __ret;
		66403	}
66667	#ifdef __LITTLE_ENDIAN__	66404	#ifdef __LITTLE_ENDIAN__
66668	__ai __attribute__((target("v8.5a"))) float32x4_t vrnd64zq_f32(float32x4_t __p0) {	66405	__ai __attribute__((target("v8.5a"))) float32x4_t vrnd64zq_f32(float32x4_t __p0) {
66669	float32x4_t __ret;	66406	float32x4_t __ret;
...	@@ -66696,6 +66433,27 @@ __ai __attribute__((target("v8.5a"))) float32x2_t vrnd64z_f32(float32x2_t __p0)	...	@@ -66696,6 +66433,27 @@ __ai __attribute__((target("v8.5a"))) float32x2_t vrnd64z_f32(float32x2_t __p0)
66696	}	66433	}
66697	#endif	66434	#endif
66698		66435
		66436	#ifdef __LITTLE_ENDIAN__
		66437	__ai __attribute__((target("v8.5a"))) float64x2_t vrnd64zq_f64(float64x2_t __p0) {
		66438	float64x2_t __ret;
		66439	__ret = (float64x2_t) __builtin_neon_vrnd64zq_f64((int8x16_t)__p0, 42);
		66440	return __ret;
		66441	}
		66442	#else
		66443	__ai __attribute__((target("v8.5a"))) float64x2_t vrnd64zq_f64(float64x2_t __p0) {
		66444	float64x2_t __ret;
		66445	float64x2_t __rev0; __rev0 = __builtin_shufflevector(__p0, __p0, 1, 0);
		66446	__ret = (float64x2_t) __builtin_neon_vrnd64zq_f64((int8x16_t)__rev0, 42);
		66447	__ret = __builtin_shufflevector(__ret, __ret, 1, 0);
		66448	return __ret;
		66449	}
		66450	#endif
		66451
		66452	__ai __attribute__((target("v8.5a"))) float64x1_t vrnd64z_f64(float64x1_t __p0) {
		66453	float64x1_t __ret;
		66454	__ret = (float64x1_t) __builtin_neon_vrnd64z_f64((int8x8_t)__p0, 10);
		66455	return __ret;
		66456	}
66699	#endif	66457	#endif
66700	#if defined(__aarch64__) && defined(__ARM_FEATURE_DIRECTED_ROUNDING)	66458	#if defined(__aarch64__) && defined(__ARM_FEATURE_DIRECTED_ROUNDING)
66701	#ifdef __LITTLE_ENDIAN__	66459	#ifdef __LITTLE_ENDIAN__

lib/include/arm_sme.h created+2412

...	@@ -0,0 +1,2412 @@
1	/*===---- arm_sme.h - ARM SME intrinsics ------===
2	*
3	*
4	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
5	* See https://llvm.org/LICENSE.txt for license information.
6	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7	*
8	*===-----------------------------------------------------------------------===
9	*/
10
11	#ifndef __ARM_SME_H
12	#define __ARM_SME_H
13
14	#if !defined(__LITTLE_ENDIAN__)
15	#error "Big endian is currently not supported for arm_sme.h"
16	#endif
17	#include <arm_sve.h>
18
19	/* Function attributes */
20	#define __ai static __inline__ __attribute__((__always_inline__, __nodebug__))
21
22	#define __aio static __inline__ __attribute__((__always_inline__, __nodebug__, __overloadable__))
23
24	#ifdef __cplusplus
25	extern "C" {
26	#endif
27
28	void __arm_za_disable(void) __arm_streaming_compatible;
29
30	__ai bool __arm_has_sme(void) __arm_streaming_compatible {
31	uint64_t x0, x1;
32	__builtin_arm_get_sme_state(&x0, &x1);
33	return x0 & (1ULL << 63);
34	}
35
36	__ai bool __arm_in_streaming_mode(void) __arm_streaming_compatible {
37	uint64_t x0, x1;
38	__builtin_arm_get_sme_state(&x0, &x1);
39	return x0 & 1;
40	}
41
42	__ai __attribute__((target("sme"))) void svundef_za(void) __arm_streaming_compatible __arm_out("za") { }
43
44	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddha_za32_u32_m)))
45	void svaddha_za32_u32_m(uint64_t, svbool_t, svbool_t, svuint32_t);
46	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddha_za32_s32_m)))
47	void svaddha_za32_s32_m(uint64_t, svbool_t, svbool_t, svint32_t);
48	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddva_za32_u32_m)))
49	void svaddva_za32_u32_m(uint64_t, svbool_t, svbool_t, svuint32_t);
50	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddva_za32_s32_m)))
51	void svaddva_za32_s32_m(uint64_t, svbool_t, svbool_t, svint32_t);
52	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svcntsb)))
53	uint64_t svcntsb(void);
54	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svcntsd)))
55	uint64_t svcntsd(void);
56	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svcntsh)))
57	uint64_t svcntsh(void);
58	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svcntsw)))
59	uint64_t svcntsw(void);
60	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_vnum_za128)))
61	void svld1_hor_vnum_za128(uint64_t, uint32_t, svbool_t, void const *, int64_t);
62	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_vnum_za16)))
63	void svld1_hor_vnum_za16(uint64_t, uint32_t, svbool_t, void const *, int64_t);
64	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_vnum_za32)))
65	void svld1_hor_vnum_za32(uint64_t, uint32_t, svbool_t, void const *, int64_t);
66	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_vnum_za64)))
67	void svld1_hor_vnum_za64(uint64_t, uint32_t, svbool_t, void const *, int64_t);
68	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_vnum_za8)))
69	void svld1_hor_vnum_za8(uint64_t, uint32_t, svbool_t, void const *, int64_t);
70	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_za128)))
71	void svld1_hor_za128(uint64_t, uint32_t, svbool_t, void const *);
72	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_za16)))
73	void svld1_hor_za16(uint64_t, uint32_t, svbool_t, void const *);
74	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_za32)))
75	void svld1_hor_za32(uint64_t, uint32_t, svbool_t, void const *);
76	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_za64)))
77	void svld1_hor_za64(uint64_t, uint32_t, svbool_t, void const *);
78	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_za8)))
79	void svld1_hor_za8(uint64_t, uint32_t, svbool_t, void const *);
80	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_vnum_za128)))
81	void svld1_ver_vnum_za128(uint64_t, uint32_t, svbool_t, void const *, int64_t);
82	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_vnum_za16)))
83	void svld1_ver_vnum_za16(uint64_t, uint32_t, svbool_t, void const *, int64_t);
84	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_vnum_za32)))
85	void svld1_ver_vnum_za32(uint64_t, uint32_t, svbool_t, void const *, int64_t);
86	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_vnum_za64)))
87	void svld1_ver_vnum_za64(uint64_t, uint32_t, svbool_t, void const *, int64_t);
88	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_vnum_za8)))
89	void svld1_ver_vnum_za8(uint64_t, uint32_t, svbool_t, void const *, int64_t);
90	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_za128)))
91	void svld1_ver_za128(uint64_t, uint32_t, svbool_t, void const *);
92	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_za16)))
93	void svld1_ver_za16(uint64_t, uint32_t, svbool_t, void const *);
94	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_za32)))
95	void svld1_ver_za32(uint64_t, uint32_t, svbool_t, void const *);
96	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_za64)))
97	void svld1_ver_za64(uint64_t, uint32_t, svbool_t, void const *);
98	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_za8)))
99	void svld1_ver_za8(uint64_t, uint32_t, svbool_t, void const *);
100	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svldr_vnum_za)))
101	void svldr_vnum_za(uint32_t, void const *, int64_t);
102	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svldr_za)))
103	void svldr_za(uint32_t, void const *);
104	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_f16_m)))
105	void svmopa_za32_f16_m(uint64_t, svbool_t, svbool_t, svfloat16_t, svfloat16_t);
106	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_bf16_m)))
107	void svmopa_za32_bf16_m(uint64_t, svbool_t, svbool_t, svbfloat16_t, svbfloat16_t);
108	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_f32_m)))
109	void svmopa_za32_f32_m(uint64_t, svbool_t, svbool_t, svfloat32_t, svfloat32_t);
110	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_s8_m)))
111	void svmopa_za32_s8_m(uint64_t, svbool_t, svbool_t, svint8_t, svint8_t);
112	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_u8_m)))
113	void svmopa_za32_u8_m(uint64_t, svbool_t, svbool_t, svuint8_t, svuint8_t);
114	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_f16_m)))
115	void svmops_za32_f16_m(uint64_t, svbool_t, svbool_t, svfloat16_t, svfloat16_t);
116	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_bf16_m)))
117	void svmops_za32_bf16_m(uint64_t, svbool_t, svbool_t, svbfloat16_t, svbfloat16_t);
118	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_f32_m)))
119	void svmops_za32_f32_m(uint64_t, svbool_t, svbool_t, svfloat32_t, svfloat32_t);
120	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_s8_m)))
121	void svmops_za32_s8_m(uint64_t, svbool_t, svbool_t, svint8_t, svint8_t);
122	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_u8_m)))
123	void svmops_za32_u8_m(uint64_t, svbool_t, svbool_t, svuint8_t, svuint8_t);
124	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_u8_m)))
125	svuint8_t svread_hor_za128_u8_m(svuint8_t, svbool_t, uint64_t, uint32_t);
126	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_u32_m)))
127	svuint32_t svread_hor_za128_u32_m(svuint32_t, svbool_t, uint64_t, uint32_t);
128	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_u64_m)))
129	svuint64_t svread_hor_za128_u64_m(svuint64_t, svbool_t, uint64_t, uint32_t);
130	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_u16_m)))
131	svuint16_t svread_hor_za128_u16_m(svuint16_t, svbool_t, uint64_t, uint32_t);
132	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_bf16_m)))
133	svbfloat16_t svread_hor_za128_bf16_m(svbfloat16_t, svbool_t, uint64_t, uint32_t);
134	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_s8_m)))
135	svint8_t svread_hor_za128_s8_m(svint8_t, svbool_t, uint64_t, uint32_t);
136	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_f64_m)))
137	svfloat64_t svread_hor_za128_f64_m(svfloat64_t, svbool_t, uint64_t, uint32_t);
138	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_f32_m)))
139	svfloat32_t svread_hor_za128_f32_m(svfloat32_t, svbool_t, uint64_t, uint32_t);
140	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_f16_m)))
141	svfloat16_t svread_hor_za128_f16_m(svfloat16_t, svbool_t, uint64_t, uint32_t);
142	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_s32_m)))
143	svint32_t svread_hor_za128_s32_m(svint32_t, svbool_t, uint64_t, uint32_t);
144	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_s64_m)))
145	svint64_t svread_hor_za128_s64_m(svint64_t, svbool_t, uint64_t, uint32_t);
146	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_s16_m)))
147	svint16_t svread_hor_za128_s16_m(svint16_t, svbool_t, uint64_t, uint32_t);
148	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_u16_m)))
149	svuint16_t svread_hor_za16_u16_m(svuint16_t, svbool_t, uint64_t, uint32_t);
150	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_bf16_m)))
151	svbfloat16_t svread_hor_za16_bf16_m(svbfloat16_t, svbool_t, uint64_t, uint32_t);
152	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_f16_m)))
153	svfloat16_t svread_hor_za16_f16_m(svfloat16_t, svbool_t, uint64_t, uint32_t);
154	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_s16_m)))
155	svint16_t svread_hor_za16_s16_m(svint16_t, svbool_t, uint64_t, uint32_t);
156	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_u32_m)))
157	svuint32_t svread_hor_za32_u32_m(svuint32_t, svbool_t, uint64_t, uint32_t);
158	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_f32_m)))
159	svfloat32_t svread_hor_za32_f32_m(svfloat32_t, svbool_t, uint64_t, uint32_t);
160	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_s32_m)))
161	svint32_t svread_hor_za32_s32_m(svint32_t, svbool_t, uint64_t, uint32_t);
162	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_u64_m)))
163	svuint64_t svread_hor_za64_u64_m(svuint64_t, svbool_t, uint64_t, uint32_t);
164	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_f64_m)))
165	svfloat64_t svread_hor_za64_f64_m(svfloat64_t, svbool_t, uint64_t, uint32_t);
166	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_s64_m)))
167	svint64_t svread_hor_za64_s64_m(svint64_t, svbool_t, uint64_t, uint32_t);
168	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za8_u8_m)))
169	svuint8_t svread_hor_za8_u8_m(svuint8_t, svbool_t, uint64_t, uint32_t);
170	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za8_s8_m)))
171	svint8_t svread_hor_za8_s8_m(svint8_t, svbool_t, uint64_t, uint32_t);
172	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_u8_m)))
173	svuint8_t svread_ver_za128_u8_m(svuint8_t, svbool_t, uint64_t, uint32_t);
174	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_u32_m)))
175	svuint32_t svread_ver_za128_u32_m(svuint32_t, svbool_t, uint64_t, uint32_t);
176	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_u64_m)))
177	svuint64_t svread_ver_za128_u64_m(svuint64_t, svbool_t, uint64_t, uint32_t);
178	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_u16_m)))
179	svuint16_t svread_ver_za128_u16_m(svuint16_t, svbool_t, uint64_t, uint32_t);
180	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_bf16_m)))
181	svbfloat16_t svread_ver_za128_bf16_m(svbfloat16_t, svbool_t, uint64_t, uint32_t);
182	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_s8_m)))
183	svint8_t svread_ver_za128_s8_m(svint8_t, svbool_t, uint64_t, uint32_t);
184	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_f64_m)))
185	svfloat64_t svread_ver_za128_f64_m(svfloat64_t, svbool_t, uint64_t, uint32_t);
186	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_f32_m)))
187	svfloat32_t svread_ver_za128_f32_m(svfloat32_t, svbool_t, uint64_t, uint32_t);
188	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_f16_m)))
189	svfloat16_t svread_ver_za128_f16_m(svfloat16_t, svbool_t, uint64_t, uint32_t);
190	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_s32_m)))
191	svint32_t svread_ver_za128_s32_m(svint32_t, svbool_t, uint64_t, uint32_t);
192	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_s64_m)))
193	svint64_t svread_ver_za128_s64_m(svint64_t, svbool_t, uint64_t, uint32_t);
194	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_s16_m)))
195	svint16_t svread_ver_za128_s16_m(svint16_t, svbool_t, uint64_t, uint32_t);
196	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_u16_m)))
197	svuint16_t svread_ver_za16_u16_m(svuint16_t, svbool_t, uint64_t, uint32_t);
198	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_bf16_m)))
199	svbfloat16_t svread_ver_za16_bf16_m(svbfloat16_t, svbool_t, uint64_t, uint32_t);
200	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_f16_m)))
201	svfloat16_t svread_ver_za16_f16_m(svfloat16_t, svbool_t, uint64_t, uint32_t);
202	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_s16_m)))
203	svint16_t svread_ver_za16_s16_m(svint16_t, svbool_t, uint64_t, uint32_t);
204	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_u32_m)))
205	svuint32_t svread_ver_za32_u32_m(svuint32_t, svbool_t, uint64_t, uint32_t);
206	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_f32_m)))
207	svfloat32_t svread_ver_za32_f32_m(svfloat32_t, svbool_t, uint64_t, uint32_t);
208	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_s32_m)))
209	svint32_t svread_ver_za32_s32_m(svint32_t, svbool_t, uint64_t, uint32_t);
210	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_u64_m)))
211	svuint64_t svread_ver_za64_u64_m(svuint64_t, svbool_t, uint64_t, uint32_t);
212	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_f64_m)))
213	svfloat64_t svread_ver_za64_f64_m(svfloat64_t, svbool_t, uint64_t, uint32_t);
214	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_s64_m)))
215	svint64_t svread_ver_za64_s64_m(svint64_t, svbool_t, uint64_t, uint32_t);
216	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za8_u8_m)))
217	svuint8_t svread_ver_za8_u8_m(svuint8_t, svbool_t, uint64_t, uint32_t);
218	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za8_s8_m)))
219	svint8_t svread_ver_za8_s8_m(svint8_t, svbool_t, uint64_t, uint32_t);
220	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_vnum_za128)))
221	void svst1_hor_vnum_za128(uint64_t, uint32_t, svbool_t, void *, int64_t);
222	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_vnum_za16)))
223	void svst1_hor_vnum_za16(uint64_t, uint32_t, svbool_t, void *, int64_t);
224	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_vnum_za32)))
225	void svst1_hor_vnum_za32(uint64_t, uint32_t, svbool_t, void *, int64_t);
226	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_vnum_za64)))
227	void svst1_hor_vnum_za64(uint64_t, uint32_t, svbool_t, void *, int64_t);
228	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_vnum_za8)))
229	void svst1_hor_vnum_za8(uint64_t, uint32_t, svbool_t, void *, int64_t);
230	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_za128)))
231	void svst1_hor_za128(uint64_t, uint32_t, svbool_t, void *);
232	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_za16)))
233	void svst1_hor_za16(uint64_t, uint32_t, svbool_t, void *);
234	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_za32)))
235	void svst1_hor_za32(uint64_t, uint32_t, svbool_t, void *);
236	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_za64)))
237	void svst1_hor_za64(uint64_t, uint32_t, svbool_t, void *);
238	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_za8)))
239	void svst1_hor_za8(uint64_t, uint32_t, svbool_t, void *);
240	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_vnum_za128)))
241	void svst1_ver_vnum_za128(uint64_t, uint32_t, svbool_t, void *, int64_t);
242	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_vnum_za16)))
243	void svst1_ver_vnum_za16(uint64_t, uint32_t, svbool_t, void *, int64_t);
244	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_vnum_za32)))
245	void svst1_ver_vnum_za32(uint64_t, uint32_t, svbool_t, void *, int64_t);
246	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_vnum_za64)))
247	void svst1_ver_vnum_za64(uint64_t, uint32_t, svbool_t, void *, int64_t);
248	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_vnum_za8)))
249	void svst1_ver_vnum_za8(uint64_t, uint32_t, svbool_t, void *, int64_t);
250	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_za128)))
251	void svst1_ver_za128(uint64_t, uint32_t, svbool_t, void *);
252	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_za16)))
253	void svst1_ver_za16(uint64_t, uint32_t, svbool_t, void *);
254	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_za32)))
255	void svst1_ver_za32(uint64_t, uint32_t, svbool_t, void *);
256	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_za64)))
257	void svst1_ver_za64(uint64_t, uint32_t, svbool_t, void *);
258	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_za8)))
259	void svst1_ver_za8(uint64_t, uint32_t, svbool_t, void *);
260	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svstr_vnum_za)))
261	void svstr_vnum_za(uint32_t, void *, int64_t);
262	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svstr_za)))
263	void svstr_za(uint32_t, void *);
264	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumopa_za32_s8_m)))
265	void svsumopa_za32_s8_m(uint64_t, svbool_t, svbool_t, svint8_t, svuint8_t);
266	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumops_za32_s8_m)))
267	void svsumops_za32_s8_m(uint64_t, svbool_t, svbool_t, svint8_t, svuint8_t);
268	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmopa_za32_u8_m)))
269	void svusmopa_za32_u8_m(uint64_t, svbool_t, svbool_t, svuint8_t, svint8_t);
270	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmops_za32_u8_m)))
271	void svusmops_za32_u8_m(uint64_t, svbool_t, svbool_t, svuint8_t, svint8_t);
272	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_u8_m)))
273	void svwrite_hor_za128_u8_m(uint64_t, uint32_t, svbool_t, svuint8_t);
274	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_u32_m)))
275	void svwrite_hor_za128_u32_m(uint64_t, uint32_t, svbool_t, svuint32_t);
276	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_u64_m)))
277	void svwrite_hor_za128_u64_m(uint64_t, uint32_t, svbool_t, svuint64_t);
278	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_u16_m)))
279	void svwrite_hor_za128_u16_m(uint64_t, uint32_t, svbool_t, svuint16_t);
280	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_bf16_m)))
281	void svwrite_hor_za128_bf16_m(uint64_t, uint32_t, svbool_t, svbfloat16_t);
282	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_s8_m)))
283	void svwrite_hor_za128_s8_m(uint64_t, uint32_t, svbool_t, svint8_t);
284	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_f64_m)))
285	void svwrite_hor_za128_f64_m(uint64_t, uint32_t, svbool_t, svfloat64_t);
286	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_f32_m)))
287	void svwrite_hor_za128_f32_m(uint64_t, uint32_t, svbool_t, svfloat32_t);
288	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_f16_m)))
289	void svwrite_hor_za128_f16_m(uint64_t, uint32_t, svbool_t, svfloat16_t);
290	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_s32_m)))
291	void svwrite_hor_za128_s32_m(uint64_t, uint32_t, svbool_t, svint32_t);
292	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_s64_m)))
293	void svwrite_hor_za128_s64_m(uint64_t, uint32_t, svbool_t, svint64_t);
294	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_s16_m)))
295	void svwrite_hor_za128_s16_m(uint64_t, uint32_t, svbool_t, svint16_t);
296	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_u16_m)))
297	void svwrite_hor_za16_u16_m(uint64_t, uint32_t, svbool_t, svuint16_t);
298	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_bf16_m)))
299	void svwrite_hor_za16_bf16_m(uint64_t, uint32_t, svbool_t, svbfloat16_t);
300	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_f16_m)))
301	void svwrite_hor_za16_f16_m(uint64_t, uint32_t, svbool_t, svfloat16_t);
302	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_s16_m)))
303	void svwrite_hor_za16_s16_m(uint64_t, uint32_t, svbool_t, svint16_t);
304	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_u32_m)))
305	void svwrite_hor_za32_u32_m(uint64_t, uint32_t, svbool_t, svuint32_t);
306	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_f32_m)))
307	void svwrite_hor_za32_f32_m(uint64_t, uint32_t, svbool_t, svfloat32_t);
308	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_s32_m)))
309	void svwrite_hor_za32_s32_m(uint64_t, uint32_t, svbool_t, svint32_t);
310	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_u64_m)))
311	void svwrite_hor_za64_u64_m(uint64_t, uint32_t, svbool_t, svuint64_t);
312	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_f64_m)))
313	void svwrite_hor_za64_f64_m(uint64_t, uint32_t, svbool_t, svfloat64_t);
314	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_s64_m)))
315	void svwrite_hor_za64_s64_m(uint64_t, uint32_t, svbool_t, svint64_t);
316	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za8_u8_m)))
317	void svwrite_hor_za8_u8_m(uint64_t, uint32_t, svbool_t, svuint8_t);
318	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za8_s8_m)))
319	void svwrite_hor_za8_s8_m(uint64_t, uint32_t, svbool_t, svint8_t);
320	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_u8_m)))
321	void svwrite_ver_za128_u8_m(uint64_t, uint32_t, svbool_t, svuint8_t);
322	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_u32_m)))
323	void svwrite_ver_za128_u32_m(uint64_t, uint32_t, svbool_t, svuint32_t);
324	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_u64_m)))
325	void svwrite_ver_za128_u64_m(uint64_t, uint32_t, svbool_t, svuint64_t);
326	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_u16_m)))
327	void svwrite_ver_za128_u16_m(uint64_t, uint32_t, svbool_t, svuint16_t);
328	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_bf16_m)))
329	void svwrite_ver_za128_bf16_m(uint64_t, uint32_t, svbool_t, svbfloat16_t);
330	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_s8_m)))
331	void svwrite_ver_za128_s8_m(uint64_t, uint32_t, svbool_t, svint8_t);
332	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_f64_m)))
333	void svwrite_ver_za128_f64_m(uint64_t, uint32_t, svbool_t, svfloat64_t);
334	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_f32_m)))
335	void svwrite_ver_za128_f32_m(uint64_t, uint32_t, svbool_t, svfloat32_t);
336	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_f16_m)))
337	void svwrite_ver_za128_f16_m(uint64_t, uint32_t, svbool_t, svfloat16_t);
338	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_s32_m)))
339	void svwrite_ver_za128_s32_m(uint64_t, uint32_t, svbool_t, svint32_t);
340	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_s64_m)))
341	void svwrite_ver_za128_s64_m(uint64_t, uint32_t, svbool_t, svint64_t);
342	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_s16_m)))
343	void svwrite_ver_za128_s16_m(uint64_t, uint32_t, svbool_t, svint16_t);
344	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_u16_m)))
345	void svwrite_ver_za16_u16_m(uint64_t, uint32_t, svbool_t, svuint16_t);
346	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_bf16_m)))
347	void svwrite_ver_za16_bf16_m(uint64_t, uint32_t, svbool_t, svbfloat16_t);
348	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_f16_m)))
349	void svwrite_ver_za16_f16_m(uint64_t, uint32_t, svbool_t, svfloat16_t);
350	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_s16_m)))
351	void svwrite_ver_za16_s16_m(uint64_t, uint32_t, svbool_t, svint16_t);
352	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_u32_m)))
353	void svwrite_ver_za32_u32_m(uint64_t, uint32_t, svbool_t, svuint32_t);
354	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_f32_m)))
355	void svwrite_ver_za32_f32_m(uint64_t, uint32_t, svbool_t, svfloat32_t);
356	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_s32_m)))
357	void svwrite_ver_za32_s32_m(uint64_t, uint32_t, svbool_t, svint32_t);
358	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_u64_m)))
359	void svwrite_ver_za64_u64_m(uint64_t, uint32_t, svbool_t, svuint64_t);
360	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_f64_m)))
361	void svwrite_ver_za64_f64_m(uint64_t, uint32_t, svbool_t, svfloat64_t);
362	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_s64_m)))
363	void svwrite_ver_za64_s64_m(uint64_t, uint32_t, svbool_t, svint64_t);
364	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za8_u8_m)))
365	void svwrite_ver_za8_u8_m(uint64_t, uint32_t, svbool_t, svuint8_t);
366	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za8_s8_m)))
367	void svwrite_ver_za8_s8_m(uint64_t, uint32_t, svbool_t, svint8_t);
368	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svzero_mask_za)))
369	void svzero_mask_za(uint64_t);
370	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svzero_za)))
371	void svzero_za();
372	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddha_za32_u32_m)))
373	void svaddha_za32_m(uint64_t, svbool_t, svbool_t, svuint32_t);
374	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddha_za32_s32_m)))
375	void svaddha_za32_m(uint64_t, svbool_t, svbool_t, svint32_t);
376	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddva_za32_u32_m)))
377	void svaddva_za32_m(uint64_t, svbool_t, svbool_t, svuint32_t);
378	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddva_za32_s32_m)))
379	void svaddva_za32_m(uint64_t, svbool_t, svbool_t, svint32_t);
380	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_f16_m)))
381	void svmopa_za32_m(uint64_t, svbool_t, svbool_t, svfloat16_t, svfloat16_t);
382	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_bf16_m)))
383	void svmopa_za32_m(uint64_t, svbool_t, svbool_t, svbfloat16_t, svbfloat16_t);
384	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_f32_m)))
385	void svmopa_za32_m(uint64_t, svbool_t, svbool_t, svfloat32_t, svfloat32_t);
386	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_s8_m)))
387	void svmopa_za32_m(uint64_t, svbool_t, svbool_t, svint8_t, svint8_t);
388	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_u8_m)))
389	void svmopa_za32_m(uint64_t, svbool_t, svbool_t, svuint8_t, svuint8_t);
390	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_f16_m)))
391	void svmops_za32_m(uint64_t, svbool_t, svbool_t, svfloat16_t, svfloat16_t);
392	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_bf16_m)))
393	void svmops_za32_m(uint64_t, svbool_t, svbool_t, svbfloat16_t, svbfloat16_t);
394	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_f32_m)))
395	void svmops_za32_m(uint64_t, svbool_t, svbool_t, svfloat32_t, svfloat32_t);
396	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_s8_m)))
397	void svmops_za32_m(uint64_t, svbool_t, svbool_t, svint8_t, svint8_t);
398	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_u8_m)))
399	void svmops_za32_m(uint64_t, svbool_t, svbool_t, svuint8_t, svuint8_t);
400	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_u8_m)))
401	svuint8_t svread_hor_za128_m(svuint8_t, svbool_t, uint64_t, uint32_t);
402	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_u32_m)))
403	svuint32_t svread_hor_za128_m(svuint32_t, svbool_t, uint64_t, uint32_t);
404	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_u64_m)))
405	svuint64_t svread_hor_za128_m(svuint64_t, svbool_t, uint64_t, uint32_t);
406	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_u16_m)))
407	svuint16_t svread_hor_za128_m(svuint16_t, svbool_t, uint64_t, uint32_t);
408	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_bf16_m)))
409	svbfloat16_t svread_hor_za128_m(svbfloat16_t, svbool_t, uint64_t, uint32_t);
410	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_s8_m)))
411	svint8_t svread_hor_za128_m(svint8_t, svbool_t, uint64_t, uint32_t);
412	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_f64_m)))
413	svfloat64_t svread_hor_za128_m(svfloat64_t, svbool_t, uint64_t, uint32_t);
414	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_f32_m)))
415	svfloat32_t svread_hor_za128_m(svfloat32_t, svbool_t, uint64_t, uint32_t);
416	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_f16_m)))
417	svfloat16_t svread_hor_za128_m(svfloat16_t, svbool_t, uint64_t, uint32_t);
418	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_s32_m)))
419	svint32_t svread_hor_za128_m(svint32_t, svbool_t, uint64_t, uint32_t);
420	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_s64_m)))
421	svint64_t svread_hor_za128_m(svint64_t, svbool_t, uint64_t, uint32_t);
422	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_s16_m)))
423	svint16_t svread_hor_za128_m(svint16_t, svbool_t, uint64_t, uint32_t);
424	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_u16_m)))
425	svuint16_t svread_hor_za16_m(svuint16_t, svbool_t, uint64_t, uint32_t);
426	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_bf16_m)))
427	svbfloat16_t svread_hor_za16_m(svbfloat16_t, svbool_t, uint64_t, uint32_t);
428	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_f16_m)))
429	svfloat16_t svread_hor_za16_m(svfloat16_t, svbool_t, uint64_t, uint32_t);
430	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_s16_m)))
431	svint16_t svread_hor_za16_m(svint16_t, svbool_t, uint64_t, uint32_t);
432	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_u32_m)))
433	svuint32_t svread_hor_za32_m(svuint32_t, svbool_t, uint64_t, uint32_t);
434	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_f32_m)))
435	svfloat32_t svread_hor_za32_m(svfloat32_t, svbool_t, uint64_t, uint32_t);
436	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_s32_m)))
437	svint32_t svread_hor_za32_m(svint32_t, svbool_t, uint64_t, uint32_t);
438	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_u64_m)))
439	svuint64_t svread_hor_za64_m(svuint64_t, svbool_t, uint64_t, uint32_t);
440	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_f64_m)))
441	svfloat64_t svread_hor_za64_m(svfloat64_t, svbool_t, uint64_t, uint32_t);
442	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_s64_m)))
443	svint64_t svread_hor_za64_m(svint64_t, svbool_t, uint64_t, uint32_t);
444	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za8_u8_m)))
445	svuint8_t svread_hor_za8_m(svuint8_t, svbool_t, uint64_t, uint32_t);
446	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za8_s8_m)))
447	svint8_t svread_hor_za8_m(svint8_t, svbool_t, uint64_t, uint32_t);
448	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_u8_m)))
449	svuint8_t svread_ver_za128_m(svuint8_t, svbool_t, uint64_t, uint32_t);
450	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_u32_m)))
451	svuint32_t svread_ver_za128_m(svuint32_t, svbool_t, uint64_t, uint32_t);
452	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_u64_m)))
453	svuint64_t svread_ver_za128_m(svuint64_t, svbool_t, uint64_t, uint32_t);
454	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_u16_m)))
455	svuint16_t svread_ver_za128_m(svuint16_t, svbool_t, uint64_t, uint32_t);
456	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_bf16_m)))
457	svbfloat16_t svread_ver_za128_m(svbfloat16_t, svbool_t, uint64_t, uint32_t);
458	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_s8_m)))
459	svint8_t svread_ver_za128_m(svint8_t, svbool_t, uint64_t, uint32_t);
460	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_f64_m)))
461	svfloat64_t svread_ver_za128_m(svfloat64_t, svbool_t, uint64_t, uint32_t);
462	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_f32_m)))
463	svfloat32_t svread_ver_za128_m(svfloat32_t, svbool_t, uint64_t, uint32_t);
464	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_f16_m)))
465	svfloat16_t svread_ver_za128_m(svfloat16_t, svbool_t, uint64_t, uint32_t);
466	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_s32_m)))
467	svint32_t svread_ver_za128_m(svint32_t, svbool_t, uint64_t, uint32_t);
468	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_s64_m)))
469	svint64_t svread_ver_za128_m(svint64_t, svbool_t, uint64_t, uint32_t);
470	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_s16_m)))
471	svint16_t svread_ver_za128_m(svint16_t, svbool_t, uint64_t, uint32_t);
472	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_u16_m)))
473	svuint16_t svread_ver_za16_m(svuint16_t, svbool_t, uint64_t, uint32_t);
474	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_bf16_m)))
475	svbfloat16_t svread_ver_za16_m(svbfloat16_t, svbool_t, uint64_t, uint32_t);
476	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_f16_m)))
477	svfloat16_t svread_ver_za16_m(svfloat16_t, svbool_t, uint64_t, uint32_t);
478	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_s16_m)))
479	svint16_t svread_ver_za16_m(svint16_t, svbool_t, uint64_t, uint32_t);
480	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_u32_m)))
481	svuint32_t svread_ver_za32_m(svuint32_t, svbool_t, uint64_t, uint32_t);
482	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_f32_m)))
483	svfloat32_t svread_ver_za32_m(svfloat32_t, svbool_t, uint64_t, uint32_t);
484	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_s32_m)))
485	svint32_t svread_ver_za32_m(svint32_t, svbool_t, uint64_t, uint32_t);
486	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_u64_m)))
487	svuint64_t svread_ver_za64_m(svuint64_t, svbool_t, uint64_t, uint32_t);
488	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_f64_m)))
489	svfloat64_t svread_ver_za64_m(svfloat64_t, svbool_t, uint64_t, uint32_t);
490	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_s64_m)))
491	svint64_t svread_ver_za64_m(svint64_t, svbool_t, uint64_t, uint32_t);
492	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za8_u8_m)))
493	svuint8_t svread_ver_za8_m(svuint8_t, svbool_t, uint64_t, uint32_t);
494	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za8_s8_m)))
495	svint8_t svread_ver_za8_m(svint8_t, svbool_t, uint64_t, uint32_t);
496	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumopa_za32_s8_m)))
497	void svsumopa_za32_m(uint64_t, svbool_t, svbool_t, svint8_t, svuint8_t);
498	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumops_za32_s8_m)))
499	void svsumops_za32_m(uint64_t, svbool_t, svbool_t, svint8_t, svuint8_t);
500	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmopa_za32_u8_m)))
501	void svusmopa_za32_m(uint64_t, svbool_t, svbool_t, svuint8_t, svint8_t);
502	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmops_za32_u8_m)))
503	void svusmops_za32_m(uint64_t, svbool_t, svbool_t, svuint8_t, svint8_t);
504	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_u8_m)))
505	void svwrite_hor_za128_m(uint64_t, uint32_t, svbool_t, svuint8_t);
506	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_u32_m)))
507	void svwrite_hor_za128_m(uint64_t, uint32_t, svbool_t, svuint32_t);
508	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_u64_m)))
509	void svwrite_hor_za128_m(uint64_t, uint32_t, svbool_t, svuint64_t);
510	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_u16_m)))
511	void svwrite_hor_za128_m(uint64_t, uint32_t, svbool_t, svuint16_t);
512	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_bf16_m)))
513	void svwrite_hor_za128_m(uint64_t, uint32_t, svbool_t, svbfloat16_t);
514	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_s8_m)))
515	void svwrite_hor_za128_m(uint64_t, uint32_t, svbool_t, svint8_t);
516	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_f64_m)))
517	void svwrite_hor_za128_m(uint64_t, uint32_t, svbool_t, svfloat64_t);
518	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_f32_m)))
519	void svwrite_hor_za128_m(uint64_t, uint32_t, svbool_t, svfloat32_t);
520	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_f16_m)))
521	void svwrite_hor_za128_m(uint64_t, uint32_t, svbool_t, svfloat16_t);
522	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_s32_m)))
523	void svwrite_hor_za128_m(uint64_t, uint32_t, svbool_t, svint32_t);
524	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_s64_m)))
525	void svwrite_hor_za128_m(uint64_t, uint32_t, svbool_t, svint64_t);
526	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_s16_m)))
527	void svwrite_hor_za128_m(uint64_t, uint32_t, svbool_t, svint16_t);
528	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_u16_m)))
529	void svwrite_hor_za16_m(uint64_t, uint32_t, svbool_t, svuint16_t);
530	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_bf16_m)))
531	void svwrite_hor_za16_m(uint64_t, uint32_t, svbool_t, svbfloat16_t);
532	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_f16_m)))
533	void svwrite_hor_za16_m(uint64_t, uint32_t, svbool_t, svfloat16_t);
534	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_s16_m)))
535	void svwrite_hor_za16_m(uint64_t, uint32_t, svbool_t, svint16_t);
536	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_u32_m)))
537	void svwrite_hor_za32_m(uint64_t, uint32_t, svbool_t, svuint32_t);
538	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_f32_m)))
539	void svwrite_hor_za32_m(uint64_t, uint32_t, svbool_t, svfloat32_t);
540	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_s32_m)))
541	void svwrite_hor_za32_m(uint64_t, uint32_t, svbool_t, svint32_t);
542	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_u64_m)))
543	void svwrite_hor_za64_m(uint64_t, uint32_t, svbool_t, svuint64_t);
544	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_f64_m)))
545	void svwrite_hor_za64_m(uint64_t, uint32_t, svbool_t, svfloat64_t);
546	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_s64_m)))
547	void svwrite_hor_za64_m(uint64_t, uint32_t, svbool_t, svint64_t);
548	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za8_u8_m)))
549	void svwrite_hor_za8_m(uint64_t, uint32_t, svbool_t, svuint8_t);
550	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za8_s8_m)))
551	void svwrite_hor_za8_m(uint64_t, uint32_t, svbool_t, svint8_t);
552	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_u8_m)))
553	void svwrite_ver_za128_m(uint64_t, uint32_t, svbool_t, svuint8_t);
554	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_u32_m)))
555	void svwrite_ver_za128_m(uint64_t, uint32_t, svbool_t, svuint32_t);
556	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_u64_m)))
557	void svwrite_ver_za128_m(uint64_t, uint32_t, svbool_t, svuint64_t);
558	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_u16_m)))
559	void svwrite_ver_za128_m(uint64_t, uint32_t, svbool_t, svuint16_t);
560	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_bf16_m)))
561	void svwrite_ver_za128_m(uint64_t, uint32_t, svbool_t, svbfloat16_t);
562	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_s8_m)))
563	void svwrite_ver_za128_m(uint64_t, uint32_t, svbool_t, svint8_t);
564	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_f64_m)))
565	void svwrite_ver_za128_m(uint64_t, uint32_t, svbool_t, svfloat64_t);
566	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_f32_m)))
567	void svwrite_ver_za128_m(uint64_t, uint32_t, svbool_t, svfloat32_t);
568	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_f16_m)))
569	void svwrite_ver_za128_m(uint64_t, uint32_t, svbool_t, svfloat16_t);
570	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_s32_m)))
571	void svwrite_ver_za128_m(uint64_t, uint32_t, svbool_t, svint32_t);
572	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_s64_m)))
573	void svwrite_ver_za128_m(uint64_t, uint32_t, svbool_t, svint64_t);
574	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_s16_m)))
575	void svwrite_ver_za128_m(uint64_t, uint32_t, svbool_t, svint16_t);
576	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_u16_m)))
577	void svwrite_ver_za16_m(uint64_t, uint32_t, svbool_t, svuint16_t);
578	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_bf16_m)))
579	void svwrite_ver_za16_m(uint64_t, uint32_t, svbool_t, svbfloat16_t);
580	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_f16_m)))
581	void svwrite_ver_za16_m(uint64_t, uint32_t, svbool_t, svfloat16_t);
582	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_s16_m)))
583	void svwrite_ver_za16_m(uint64_t, uint32_t, svbool_t, svint16_t);
584	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_u32_m)))
585	void svwrite_ver_za32_m(uint64_t, uint32_t, svbool_t, svuint32_t);
586	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_f32_m)))
587	void svwrite_ver_za32_m(uint64_t, uint32_t, svbool_t, svfloat32_t);
588	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_s32_m)))
589	void svwrite_ver_za32_m(uint64_t, uint32_t, svbool_t, svint32_t);
590	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_u64_m)))
591	void svwrite_ver_za64_m(uint64_t, uint32_t, svbool_t, svuint64_t);
592	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_f64_m)))
593	void svwrite_ver_za64_m(uint64_t, uint32_t, svbool_t, svfloat64_t);
594	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_s64_m)))
595	void svwrite_ver_za64_m(uint64_t, uint32_t, svbool_t, svint64_t);
596	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za8_u8_m)))
597	void svwrite_ver_za8_m(uint64_t, uint32_t, svbool_t, svuint8_t);
598	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za8_s8_m)))
599	void svwrite_ver_za8_m(uint64_t, uint32_t, svbool_t, svint8_t);
600	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za64_f64_m)))
601	void svmopa_za64_f64_m(uint64_t, svbool_t, svbool_t, svfloat64_t, svfloat64_t);
602	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za64_f64_m)))
603	void svmops_za64_f64_m(uint64_t, svbool_t, svbool_t, svfloat64_t, svfloat64_t);
604	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za64_f64_m)))
605	void svmopa_za64_m(uint64_t, svbool_t, svbool_t, svfloat64_t, svfloat64_t);
606	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za64_f64_m)))
607	void svmops_za64_m(uint64_t, svbool_t, svbool_t, svfloat64_t, svfloat64_t);
608	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddha_za64_u64_m)))
609	void svaddha_za64_u64_m(uint64_t, svbool_t, svbool_t, svuint64_t);
610	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddha_za64_s64_m)))
611	void svaddha_za64_s64_m(uint64_t, svbool_t, svbool_t, svint64_t);
612	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddva_za64_u64_m)))
613	void svaddva_za64_u64_m(uint64_t, svbool_t, svbool_t, svuint64_t);
614	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddva_za64_s64_m)))
615	void svaddva_za64_s64_m(uint64_t, svbool_t, svbool_t, svint64_t);
616	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za64_s16_m)))
617	void svmopa_za64_s16_m(uint64_t, svbool_t, svbool_t, svint16_t, svint16_t);
618	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za64_u16_m)))
619	void svmopa_za64_u16_m(uint64_t, svbool_t, svbool_t, svuint16_t, svuint16_t);
620	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za64_s16_m)))
621	void svmops_za64_s16_m(uint64_t, svbool_t, svbool_t, svint16_t, svint16_t);
622	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za64_u16_m)))
623	void svmops_za64_u16_m(uint64_t, svbool_t, svbool_t, svuint16_t, svuint16_t);
624	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumopa_za64_s16_m)))
625	void svsumopa_za64_s16_m(uint64_t, svbool_t, svbool_t, svint16_t, svuint16_t);
626	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumops_za64_s16_m)))
627	void svsumops_za64_s16_m(uint64_t, svbool_t, svbool_t, svint16_t, svuint16_t);
628	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmopa_za64_u16_m)))
629	void svusmopa_za64_u16_m(uint64_t, svbool_t, svbool_t, svuint16_t, svint16_t);
630	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmops_za64_u16_m)))
631	void svusmops_za64_u16_m(uint64_t, svbool_t, svbool_t, svuint16_t, svint16_t);
632	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddha_za64_u64_m)))
633	void svaddha_za64_m(uint64_t, svbool_t, svbool_t, svuint64_t);
634	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddha_za64_s64_m)))
635	void svaddha_za64_m(uint64_t, svbool_t, svbool_t, svint64_t);
636	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddva_za64_u64_m)))
637	void svaddva_za64_m(uint64_t, svbool_t, svbool_t, svuint64_t);
638	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddva_za64_s64_m)))
639	void svaddva_za64_m(uint64_t, svbool_t, svbool_t, svint64_t);
640	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za64_s16_m)))
641	void svmopa_za64_m(uint64_t, svbool_t, svbool_t, svint16_t, svint16_t);
642	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za64_u16_m)))
643	void svmopa_za64_m(uint64_t, svbool_t, svbool_t, svuint16_t, svuint16_t);
644	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za64_s16_m)))
645	void svmops_za64_m(uint64_t, svbool_t, svbool_t, svint16_t, svint16_t);
646	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za64_u16_m)))
647	void svmops_za64_m(uint64_t, svbool_t, svbool_t, svuint16_t, svuint16_t);
648	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumopa_za64_s16_m)))
649	void svsumopa_za64_m(uint64_t, svbool_t, svbool_t, svint16_t, svuint16_t);
650	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumops_za64_s16_m)))
651	void svsumops_za64_m(uint64_t, svbool_t, svbool_t, svint16_t, svuint16_t);
652	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmopa_za64_u16_m)))
653	void svusmopa_za64_m(uint64_t, svbool_t, svbool_t, svuint16_t, svint16_t);
654	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmops_za64_u16_m)))
655	void svusmops_za64_m(uint64_t, svbool_t, svbool_t, svuint16_t, svint16_t);
656	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_single_za32_u32_vg1x2)))
657	void svadd_write_single_za32_u32_vg1x2(uint32_t, svuint32x2_t, svuint32_t);
658	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_single_za32_s32_vg1x2)))
659	void svadd_write_single_za32_s32_vg1x2(uint32_t, svint32x2_t, svint32_t);
660	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_single_za32_u32_vg1x4)))
661	void svadd_write_single_za32_u32_vg1x4(uint32_t, svuint32x4_t, svuint32_t);
662	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_single_za32_s32_vg1x4)))
663	void svadd_write_single_za32_s32_vg1x4(uint32_t, svint32x4_t, svint32_t);
664	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_za32_u32_vg1x2)))
665	void svadd_write_za32_u32_vg1x2(uint32_t, svuint32x2_t, svuint32x2_t);
666	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_za32_s32_vg1x2)))
667	void svadd_write_za32_s32_vg1x2(uint32_t, svint32x2_t, svint32x2_t);
668	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_za32_u32_vg1x4)))
669	void svadd_write_za32_u32_vg1x4(uint32_t, svuint32x4_t, svuint32x4_t);
670	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_za32_s32_vg1x4)))
671	void svadd_write_za32_s32_vg1x4(uint32_t, svint32x4_t, svint32x4_t);
672	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za32_u32_vg1x2)))
673	void svadd_za32_u32_vg1x2(uint32_t, svuint32x2_t);
674	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za32_f32_vg1x2)))
675	void svadd_za32_f32_vg1x2(uint32_t, svfloat32x2_t);
676	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za32_s32_vg1x2)))
677	void svadd_za32_s32_vg1x2(uint32_t, svint32x2_t);
678	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za32_u32_vg1x4)))
679	void svadd_za32_u32_vg1x4(uint32_t, svuint32x4_t);
680	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za32_f32_vg1x4)))
681	void svadd_za32_f32_vg1x4(uint32_t, svfloat32x4_t);
682	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za32_s32_vg1x4)))
683	void svadd_za32_s32_vg1x4(uint32_t, svint32x4_t);
684	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svbmopa_za32_u32_m)))
685	void svbmopa_za32_u32_m(uint64_t, svbool_t, svbool_t, svuint32_t, svuint32_t);
686	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svbmopa_za32_s32_m)))
687	void svbmopa_za32_s32_m(uint64_t, svbool_t, svbool_t, svint32_t, svint32_t);
688	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svbmops_za32_u32_m)))
689	void svbmops_za32_u32_m(uint64_t, svbool_t, svbool_t, svuint32_t, svuint32_t);
690	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svbmops_za32_s32_m)))
691	void svbmops_za32_s32_m(uint64_t, svbool_t, svbool_t, svint32_t, svint32_t);
692	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_bf16_vg1x2)))
693	void svdot_single_za32_bf16_vg1x2(uint32_t, svbfloat16x2_t, svbfloat16_t);
694	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_f16_vg1x2)))
695	void svdot_single_za32_f16_vg1x2(uint32_t, svfloat16x2_t, svfloat16_t);
696	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_s8_vg1x2)))
697	void svdot_single_za32_s8_vg1x2(uint32_t, svint8x2_t, svint8_t);
698	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_s16_vg1x2)))
699	void svdot_single_za32_s16_vg1x2(uint32_t, svint16x2_t, svint16_t);
700	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_u8_vg1x2)))
701	void svdot_single_za32_u8_vg1x2(uint32_t, svuint8x2_t, svuint8_t);
702	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_u16_vg1x2)))
703	void svdot_single_za32_u16_vg1x2(uint32_t, svuint16x2_t, svuint16_t);
704	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_bf16_vg1x4)))
705	void svdot_single_za32_bf16_vg1x4(uint32_t, svbfloat16x4_t, svbfloat16_t);
706	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_f16_vg1x4)))
707	void svdot_single_za32_f16_vg1x4(uint32_t, svfloat16x4_t, svfloat16_t);
708	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_s8_vg1x4)))
709	void svdot_single_za32_s8_vg1x4(uint32_t, svint8x4_t, svint8_t);
710	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_s16_vg1x4)))
711	void svdot_single_za32_s16_vg1x4(uint32_t, svint16x4_t, svint16_t);
712	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_u8_vg1x4)))
713	void svdot_single_za32_u8_vg1x4(uint32_t, svuint8x4_t, svuint8_t);
714	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_u16_vg1x4)))
715	void svdot_single_za32_u16_vg1x4(uint32_t, svuint16x4_t, svuint16_t);
716	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_bf16_vg1x2)))
717	void svdot_lane_za32_bf16_vg1x2(uint32_t, svbfloat16x2_t, svbfloat16_t, uint64_t);
718	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_f16_vg1x2)))
719	void svdot_lane_za32_f16_vg1x2(uint32_t, svfloat16x2_t, svfloat16_t, uint64_t);
720	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_s8_vg1x2)))
721	void svdot_lane_za32_s8_vg1x2(uint32_t, svint8x2_t, svint8_t, uint64_t);
722	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_s16_vg1x2)))
723	void svdot_lane_za32_s16_vg1x2(uint32_t, svint16x2_t, svint16_t, uint64_t);
724	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_u8_vg1x2)))
725	void svdot_lane_za32_u8_vg1x2(uint32_t, svuint8x2_t, svuint8_t, uint64_t);
726	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_u16_vg1x2)))
727	void svdot_lane_za32_u16_vg1x2(uint32_t, svuint16x2_t, svuint16_t, uint64_t);
728	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_bf16_vg1x4)))
729	void svdot_lane_za32_bf16_vg1x4(uint32_t, svbfloat16x4_t, svbfloat16_t, uint64_t);
730	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_f16_vg1x4)))
731	void svdot_lane_za32_f16_vg1x4(uint32_t, svfloat16x4_t, svfloat16_t, uint64_t);
732	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_s8_vg1x4)))
733	void svdot_lane_za32_s8_vg1x4(uint32_t, svint8x4_t, svint8_t, uint64_t);
734	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_s16_vg1x4)))
735	void svdot_lane_za32_s16_vg1x4(uint32_t, svint16x4_t, svint16_t, uint64_t);
736	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_u8_vg1x4)))
737	void svdot_lane_za32_u8_vg1x4(uint32_t, svuint8x4_t, svuint8_t, uint64_t);
738	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_u16_vg1x4)))
739	void svdot_lane_za32_u16_vg1x4(uint32_t, svuint16x4_t, svuint16_t, uint64_t);
740	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_bf16_vg1x2)))
741	void svdot_za32_bf16_vg1x2(uint32_t, svbfloat16x2_t, svbfloat16x2_t);
742	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_f16_vg1x2)))
743	void svdot_za32_f16_vg1x2(uint32_t, svfloat16x2_t, svfloat16x2_t);
744	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_s8_vg1x2)))
745	void svdot_za32_s8_vg1x2(uint32_t, svint8x2_t, svint8x2_t);
746	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_s16_vg1x2)))
747	void svdot_za32_s16_vg1x2(uint32_t, svint16x2_t, svint16x2_t);
748	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_u8_vg1x2)))
749	void svdot_za32_u8_vg1x2(uint32_t, svuint8x2_t, svuint8x2_t);
750	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_u16_vg1x2)))
751	void svdot_za32_u16_vg1x2(uint32_t, svuint16x2_t, svuint16x2_t);
752	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_bf16_vg1x4)))
753	void svdot_za32_bf16_vg1x4(uint32_t, svbfloat16x4_t, svbfloat16x4_t);
754	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_f16_vg1x4)))
755	void svdot_za32_f16_vg1x4(uint32_t, svfloat16x4_t, svfloat16x4_t);
756	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_s8_vg1x4)))
757	void svdot_za32_s8_vg1x4(uint32_t, svint8x4_t, svint8x4_t);
758	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_s16_vg1x4)))
759	void svdot_za32_s16_vg1x4(uint32_t, svint16x4_t, svint16x4_t);
760	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_u8_vg1x4)))
761	void svdot_za32_u8_vg1x4(uint32_t, svuint8x4_t, svuint8x4_t);
762	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_u16_vg1x4)))
763	void svdot_za32_u16_vg1x4(uint32_t, svuint16x4_t, svuint16x4_t);
764	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svldr_zt)))
765	void svldr_zt(uint64_t, void const *);
766	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_u8)))
767	svuint8_t svluti2_lane_zt_u8(uint64_t, svuint8_t, uint64_t);
768	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_u32)))
769	svuint32_t svluti2_lane_zt_u32(uint64_t, svuint8_t, uint64_t);
770	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_u16)))
771	svuint16_t svluti2_lane_zt_u16(uint64_t, svuint8_t, uint64_t);
772	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_bf16)))
773	svbfloat16_t svluti2_lane_zt_bf16(uint64_t, svuint8_t, uint64_t);
774	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_s8)))
775	svint8_t svluti2_lane_zt_s8(uint64_t, svuint8_t, uint64_t);
776	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_f32)))
777	svfloat32_t svluti2_lane_zt_f32(uint64_t, svuint8_t, uint64_t);
778	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_f16)))
779	svfloat16_t svluti2_lane_zt_f16(uint64_t, svuint8_t, uint64_t);
780	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_s32)))
781	svint32_t svluti2_lane_zt_s32(uint64_t, svuint8_t, uint64_t);
782	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_s16)))
783	svint16_t svluti2_lane_zt_s16(uint64_t, svuint8_t, uint64_t);
784	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_u8_x2)))
785	svuint8x2_t svluti2_lane_zt_u8_x2(uint64_t, svuint8_t, uint64_t);
786	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_u32_x2)))
787	svuint32x2_t svluti2_lane_zt_u32_x2(uint64_t, svuint8_t, uint64_t);
788	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_u16_x2)))
789	svuint16x2_t svluti2_lane_zt_u16_x2(uint64_t, svuint8_t, uint64_t);
790	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_bf16_x2)))
791	svbfloat16x2_t svluti2_lane_zt_bf16_x2(uint64_t, svuint8_t, uint64_t);
792	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_s8_x2)))
793	svint8x2_t svluti2_lane_zt_s8_x2(uint64_t, svuint8_t, uint64_t);
794	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_f32_x2)))
795	svfloat32x2_t svluti2_lane_zt_f32_x2(uint64_t, svuint8_t, uint64_t);
796	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_f16_x2)))
797	svfloat16x2_t svluti2_lane_zt_f16_x2(uint64_t, svuint8_t, uint64_t);
798	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_s32_x2)))
799	svint32x2_t svluti2_lane_zt_s32_x2(uint64_t, svuint8_t, uint64_t);
800	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_s16_x2)))
801	svint16x2_t svluti2_lane_zt_s16_x2(uint64_t, svuint8_t, uint64_t);
802	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_u8_x4)))
803	svuint8x4_t svluti2_lane_zt_u8_x4(uint64_t, svuint8_t, uint64_t);
804	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_u32_x4)))
805	svuint32x4_t svluti2_lane_zt_u32_x4(uint64_t, svuint8_t, uint64_t);
806	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_u16_x4)))
807	svuint16x4_t svluti2_lane_zt_u16_x4(uint64_t, svuint8_t, uint64_t);
808	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_bf16_x4)))
809	svbfloat16x4_t svluti2_lane_zt_bf16_x4(uint64_t, svuint8_t, uint64_t);
810	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_s8_x4)))
811	svint8x4_t svluti2_lane_zt_s8_x4(uint64_t, svuint8_t, uint64_t);
812	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_f32_x4)))
813	svfloat32x4_t svluti2_lane_zt_f32_x4(uint64_t, svuint8_t, uint64_t);
814	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_f16_x4)))
815	svfloat16x4_t svluti2_lane_zt_f16_x4(uint64_t, svuint8_t, uint64_t);
816	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_s32_x4)))
817	svint32x4_t svluti2_lane_zt_s32_x4(uint64_t, svuint8_t, uint64_t);
818	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti2_lane_zt_s16_x4)))
819	svint16x4_t svluti2_lane_zt_s16_x4(uint64_t, svuint8_t, uint64_t);
820	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_u8)))
821	svuint8_t svluti4_lane_zt_u8(uint64_t, svuint8_t, uint64_t);
822	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_u32)))
823	svuint32_t svluti4_lane_zt_u32(uint64_t, svuint8_t, uint64_t);
824	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_u16)))
825	svuint16_t svluti4_lane_zt_u16(uint64_t, svuint8_t, uint64_t);
826	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_bf16)))
827	svbfloat16_t svluti4_lane_zt_bf16(uint64_t, svuint8_t, uint64_t);
828	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_s8)))
829	svint8_t svluti4_lane_zt_s8(uint64_t, svuint8_t, uint64_t);
830	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_f32)))
831	svfloat32_t svluti4_lane_zt_f32(uint64_t, svuint8_t, uint64_t);
832	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_f16)))
833	svfloat16_t svluti4_lane_zt_f16(uint64_t, svuint8_t, uint64_t);
834	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_s32)))
835	svint32_t svluti4_lane_zt_s32(uint64_t, svuint8_t, uint64_t);
836	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_s16)))
837	svint16_t svluti4_lane_zt_s16(uint64_t, svuint8_t, uint64_t);
838	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_u8_x2)))
839	svuint8x2_t svluti4_lane_zt_u8_x2(uint64_t, svuint8_t, uint64_t);
840	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_u32_x2)))
841	svuint32x2_t svluti4_lane_zt_u32_x2(uint64_t, svuint8_t, uint64_t);
842	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_u16_x2)))
843	svuint16x2_t svluti4_lane_zt_u16_x2(uint64_t, svuint8_t, uint64_t);
844	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_bf16_x2)))
845	svbfloat16x2_t svluti4_lane_zt_bf16_x2(uint64_t, svuint8_t, uint64_t);
846	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_s8_x2)))
847	svint8x2_t svluti4_lane_zt_s8_x2(uint64_t, svuint8_t, uint64_t);
848	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_f32_x2)))
849	svfloat32x2_t svluti4_lane_zt_f32_x2(uint64_t, svuint8_t, uint64_t);
850	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_f16_x2)))
851	svfloat16x2_t svluti4_lane_zt_f16_x2(uint64_t, svuint8_t, uint64_t);
852	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_s32_x2)))
853	svint32x2_t svluti4_lane_zt_s32_x2(uint64_t, svuint8_t, uint64_t);
854	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_s16_x2)))
855	svint16x2_t svluti4_lane_zt_s16_x2(uint64_t, svuint8_t, uint64_t);
856	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_u32_x4)))
857	svuint32x4_t svluti4_lane_zt_u32_x4(uint64_t, svuint8_t, uint64_t);
858	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_u16_x4)))
859	svuint16x4_t svluti4_lane_zt_u16_x4(uint64_t, svuint8_t, uint64_t);
860	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_bf16_x4)))
861	svbfloat16x4_t svluti4_lane_zt_bf16_x4(uint64_t, svuint8_t, uint64_t);
862	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_f32_x4)))
863	svfloat32x4_t svluti4_lane_zt_f32_x4(uint64_t, svuint8_t, uint64_t);
864	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_f16_x4)))
865	svfloat16x4_t svluti4_lane_zt_f16_x4(uint64_t, svuint8_t, uint64_t);
866	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_s32_x4)))
867	svint32x4_t svluti4_lane_zt_s32_x4(uint64_t, svuint8_t, uint64_t);
868	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svluti4_lane_zt_s16_x4)))
869	svint16x4_t svluti4_lane_zt_s16_x4(uint64_t, svuint8_t, uint64_t);
870	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_f32_vg1x2)))
871	void svmla_single_za32_f32_vg1x2(uint32_t, svfloat32x2_t, svfloat32_t);
872	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_f32_vg1x4)))
873	void svmla_single_za32_f32_vg1x4(uint32_t, svfloat32x4_t, svfloat32_t);
874	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_bf16_vg2x2)))
875	void svmla_single_za32_bf16_vg2x2(uint32_t, svbfloat16x2_t, svbfloat16_t);
876	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_f16_vg2x2)))
877	void svmla_single_za32_f16_vg2x2(uint32_t, svfloat16x2_t, svfloat16_t);
878	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_s16_vg2x2)))
879	void svmla_single_za32_s16_vg2x2(uint32_t, svint16x2_t, svint16_t);
880	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_u16_vg2x2)))
881	void svmla_single_za32_u16_vg2x2(uint32_t, svuint16x2_t, svuint16_t);
882	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_bf16_vg2x4)))
883	void svmla_single_za32_bf16_vg2x4(uint32_t, svbfloat16x4_t, svbfloat16_t);
884	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_f16_vg2x4)))
885	void svmla_single_za32_f16_vg2x4(uint32_t, svfloat16x4_t, svfloat16_t);
886	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_s16_vg2x4)))
887	void svmla_single_za32_s16_vg2x4(uint32_t, svint16x4_t, svint16_t);
888	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_u16_vg2x4)))
889	void svmla_single_za32_u16_vg2x4(uint32_t, svuint16x4_t, svuint16_t);
890	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_s8_vg4x2)))
891	void svmla_single_za32_s8_vg4x2(uint32_t, svint8x2_t, svint8_t);
892	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_u8_vg4x2)))
893	void svmla_single_za32_u8_vg4x2(uint32_t, svuint8x2_t, svuint8_t);
894	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_s8_vg4x4)))
895	void svmla_single_za32_s8_vg4x4(uint32_t, svint8x4_t, svint8_t);
896	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_u8_vg4x4)))
897	void svmla_single_za32_u8_vg4x4(uint32_t, svuint8x4_t, svuint8_t);
898	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_f32_vg1x2)))
899	void svmla_lane_za32_f32_vg1x2(uint32_t, svfloat32x2_t, svfloat32_t, uint64_t);
900	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_f32_vg1x4)))
901	void svmla_lane_za32_f32_vg1x4(uint32_t, svfloat32x4_t, svfloat32_t, uint64_t);
902	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_bf16_vg2x1)))
903	void svmla_lane_za32_bf16_vg2x1(uint32_t, svbfloat16_t, svbfloat16_t, uint64_t);
904	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_f16_vg2x1)))
905	void svmla_lane_za32_f16_vg2x1(uint32_t, svfloat16_t, svfloat16_t, uint64_t);
906	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_s16_vg2x1)))
907	void svmla_lane_za32_s16_vg2x1(uint32_t, svint16_t, svint16_t, uint64_t);
908	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_u16_vg2x1)))
909	void svmla_lane_za32_u16_vg2x1(uint32_t, svuint16_t, svuint16_t, uint64_t);
910	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_bf16_vg2x2)))
911	void svmla_lane_za32_bf16_vg2x2(uint32_t, svbfloat16x2_t, svbfloat16_t, uint64_t);
912	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_f16_vg2x2)))
913	void svmla_lane_za32_f16_vg2x2(uint32_t, svfloat16x2_t, svfloat16_t, uint64_t);
914	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_s16_vg2x2)))
915	void svmla_lane_za32_s16_vg2x2(uint32_t, svint16x2_t, svint16_t, uint64_t);
916	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_u16_vg2x2)))
917	void svmla_lane_za32_u16_vg2x2(uint32_t, svuint16x2_t, svuint16_t, uint64_t);
918	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_bf16_vg2x4)))
919	void svmla_lane_za32_bf16_vg2x4(uint32_t, svbfloat16x4_t, svbfloat16_t, uint64_t);
920	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_f16_vg2x4)))
921	void svmla_lane_za32_f16_vg2x4(uint32_t, svfloat16x4_t, svfloat16_t, uint64_t);
922	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_s16_vg2x4)))
923	void svmla_lane_za32_s16_vg2x4(uint32_t, svint16x4_t, svint16_t, uint64_t);
924	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_u16_vg2x4)))
925	void svmla_lane_za32_u16_vg2x4(uint32_t, svuint16x4_t, svuint16_t, uint64_t);
926	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_s8_vg4x1)))
927	void svmla_lane_za32_s8_vg4x1(uint32_t, svint8_t, svint8_t, uint64_t);
928	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_u8_vg4x1)))
929	void svmla_lane_za32_u8_vg4x1(uint32_t, svuint8_t, svuint8_t, uint64_t);
930	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_s8_vg4x2)))
931	void svmla_lane_za32_s8_vg4x2(uint32_t, svint8x2_t, svint8_t, uint64_t);
932	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_u8_vg4x2)))
933	void svmla_lane_za32_u8_vg4x2(uint32_t, svuint8x2_t, svuint8_t, uint64_t);
934	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_s8_vg4x4)))
935	void svmla_lane_za32_s8_vg4x4(uint32_t, svint8x4_t, svint8_t, uint64_t);
936	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_u8_vg4x4)))
937	void svmla_lane_za32_u8_vg4x4(uint32_t, svuint8x4_t, svuint8_t, uint64_t);
938	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_f32_vg1x2)))
939	void svmla_za32_f32_vg1x2(uint32_t, svfloat32x2_t, svfloat32x2_t);
940	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_f32_vg1x4)))
941	void svmla_za32_f32_vg1x4(uint32_t, svfloat32x4_t, svfloat32x4_t);
942	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_bf16_vg2x1)))
943	void svmla_za32_bf16_vg2x1(uint32_t, svbfloat16_t, svbfloat16_t);
944	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_f16_vg2x1)))
945	void svmla_za32_f16_vg2x1(uint32_t, svfloat16_t, svfloat16_t);
946	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_s16_vg2x1)))
947	void svmla_za32_s16_vg2x1(uint32_t, svint16_t, svint16_t);
948	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_u16_vg2x1)))
949	void svmla_za32_u16_vg2x1(uint32_t, svuint16_t, svuint16_t);
950	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_bf16_vg2x2)))
951	void svmla_za32_bf16_vg2x2(uint32_t, svbfloat16x2_t, svbfloat16x2_t);
952	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_f16_vg2x2)))
953	void svmla_za32_f16_vg2x2(uint32_t, svfloat16x2_t, svfloat16x2_t);
954	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_s16_vg2x2)))
955	void svmla_za32_s16_vg2x2(uint32_t, svint16x2_t, svint16x2_t);
956	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_u16_vg2x2)))
957	void svmla_za32_u16_vg2x2(uint32_t, svuint16x2_t, svuint16x2_t);
958	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_bf16_vg2x4)))
959	void svmla_za32_bf16_vg2x4(uint32_t, svbfloat16x4_t, svbfloat16x4_t);
960	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_f16_vg2x4)))
961	void svmla_za32_f16_vg2x4(uint32_t, svfloat16x4_t, svfloat16x4_t);
962	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_s16_vg2x4)))
963	void svmla_za32_s16_vg2x4(uint32_t, svint16x4_t, svint16x4_t);
964	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_u16_vg2x4)))
965	void svmla_za32_u16_vg2x4(uint32_t, svuint16x4_t, svuint16x4_t);
966	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_s8_vg4x1)))
967	void svmla_za32_s8_vg4x1(uint32_t, svint8_t, svint8_t);
968	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_u8_vg4x1)))
969	void svmla_za32_u8_vg4x1(uint32_t, svuint8_t, svuint8_t);
970	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_s8_vg4x2)))
971	void svmla_za32_s8_vg4x2(uint32_t, svint8x2_t, svint8x2_t);
972	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_u8_vg4x2)))
973	void svmla_za32_u8_vg4x2(uint32_t, svuint8x2_t, svuint8x2_t);
974	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_s8_vg4x4)))
975	void svmla_za32_s8_vg4x4(uint32_t, svint8x4_t, svint8x4_t);
976	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_u8_vg4x4)))
977	void svmla_za32_u8_vg4x4(uint32_t, svuint8x4_t, svuint8x4_t);
978	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_f32_vg1x2)))
979	void svmls_single_za32_f32_vg1x2(uint32_t, svfloat32x2_t, svfloat32_t);
980	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_f32_vg1x4)))
981	void svmls_single_za32_f32_vg1x4(uint32_t, svfloat32x4_t, svfloat32_t);
982	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_bf16_vg2x2)))
983	void svmls_single_za32_bf16_vg2x2(uint32_t, svbfloat16x2_t, svbfloat16_t);
984	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_f16_vg2x2)))
985	void svmls_single_za32_f16_vg2x2(uint32_t, svfloat16x2_t, svfloat16_t);
986	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_s16_vg2x2)))
987	void svmls_single_za32_s16_vg2x2(uint32_t, svint16x2_t, svint16_t);
988	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_u16_vg2x2)))
989	void svmls_single_za32_u16_vg2x2(uint32_t, svuint16x2_t, svuint16_t);
990	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_bf16_vg2x4)))
991	void svmls_single_za32_bf16_vg2x4(uint32_t, svbfloat16x4_t, svbfloat16_t);
992	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_f16_vg2x4)))
993	void svmls_single_za32_f16_vg2x4(uint32_t, svfloat16x4_t, svfloat16_t);
994	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_s16_vg2x4)))
995	void svmls_single_za32_s16_vg2x4(uint32_t, svint16x4_t, svint16_t);
996	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_u16_vg2x4)))
997	void svmls_single_za32_u16_vg2x4(uint32_t, svuint16x4_t, svuint16_t);
998	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_s8_vg4x2)))
999	void svmls_single_za32_s8_vg4x2(uint32_t, svint8x2_t, svint8_t);
1000	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_u8_vg4x2)))
1001	void svmls_single_za32_u8_vg4x2(uint32_t, svuint8x2_t, svuint8_t);
1002	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_s8_vg4x4)))
1003	void svmls_single_za32_s8_vg4x4(uint32_t, svint8x4_t, svint8_t);
1004	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_u8_vg4x4)))
1005	void svmls_single_za32_u8_vg4x4(uint32_t, svuint8x4_t, svuint8_t);
1006	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_f32_vg1x2)))
1007	void svmls_lane_za32_f32_vg1x2(uint32_t, svfloat32x2_t, svfloat32_t, uint64_t);
1008	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_f32_vg1x4)))
1009	void svmls_lane_za32_f32_vg1x4(uint32_t, svfloat32x4_t, svfloat32_t, uint64_t);
1010	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_bf16_vg2x1)))
1011	void svmls_lane_za32_bf16_vg2x1(uint32_t, svbfloat16_t, svbfloat16_t, uint64_t);
1012	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_f16_vg2x1)))
1013	void svmls_lane_za32_f16_vg2x1(uint32_t, svfloat16_t, svfloat16_t, uint64_t);
1014	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_s16_vg2x1)))
1015	void svmls_lane_za32_s16_vg2x1(uint32_t, svint16_t, svint16_t, uint64_t);
1016	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_u16_vg2x1)))
1017	void svmls_lane_za32_u16_vg2x1(uint32_t, svuint16_t, svuint16_t, uint64_t);
1018	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_bf16_vg2x2)))
1019	void svmls_lane_za32_bf16_vg2x2(uint32_t, svbfloat16x2_t, svbfloat16_t, uint64_t);
1020	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_f16_vg2x2)))
1021	void svmls_lane_za32_f16_vg2x2(uint32_t, svfloat16x2_t, svfloat16_t, uint64_t);
1022	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_s16_vg2x2)))
1023	void svmls_lane_za32_s16_vg2x2(uint32_t, svint16x2_t, svint16_t, uint64_t);
1024	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_u16_vg2x2)))
1025	void svmls_lane_za32_u16_vg2x2(uint32_t, svuint16x2_t, svuint16_t, uint64_t);
1026	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_bf16_vg2x4)))
1027	void svmls_lane_za32_bf16_vg2x4(uint32_t, svbfloat16x4_t, svbfloat16_t, uint64_t);
1028	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_f16_vg2x4)))
1029	void svmls_lane_za32_f16_vg2x4(uint32_t, svfloat16x4_t, svfloat16_t, uint64_t);
1030	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_s16_vg2x4)))
1031	void svmls_lane_za32_s16_vg2x4(uint32_t, svint16x4_t, svint16_t, uint64_t);
1032	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_u16_vg2x4)))
1033	void svmls_lane_za32_u16_vg2x4(uint32_t, svuint16x4_t, svuint16_t, uint64_t);
1034	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_s8_vg4x1)))
1035	void svmls_lane_za32_s8_vg4x1(uint32_t, svint8_t, svint8_t, uint64_t);
1036	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_u8_vg4x1)))
1037	void svmls_lane_za32_u8_vg4x1(uint32_t, svuint8_t, svuint8_t, uint64_t);
1038	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_s8_vg4x2)))
1039	void svmls_lane_za32_s8_vg4x2(uint32_t, svint8x2_t, svint8_t, uint64_t);
1040	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_u8_vg4x2)))
1041	void svmls_lane_za32_u8_vg4x2(uint32_t, svuint8x2_t, svuint8_t, uint64_t);
1042	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_s8_vg4x4)))
1043	void svmls_lane_za32_s8_vg4x4(uint32_t, svint8x4_t, svint8_t, uint64_t);
1044	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_u8_vg4x4)))
1045	void svmls_lane_za32_u8_vg4x4(uint32_t, svuint8x4_t, svuint8_t, uint64_t);
1046	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_f32_vg1x2)))
1047	void svmls_za32_f32_vg1x2(uint32_t, svfloat32x2_t, svfloat32x2_t);
1048	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_f32_vg1x4)))
1049	void svmls_za32_f32_vg1x4(uint32_t, svfloat32x4_t, svfloat32x4_t);
1050	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_bf16_vg2x1)))
1051	void svmls_za32_bf16_vg2x1(uint32_t, svbfloat16_t, svbfloat16_t);
1052	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_f16_vg2x1)))
1053	void svmls_za32_f16_vg2x1(uint32_t, svfloat16_t, svfloat16_t);
1054	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_s16_vg2x1)))
1055	void svmls_za32_s16_vg2x1(uint32_t, svint16_t, svint16_t);
1056	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_u16_vg2x1)))
1057	void svmls_za32_u16_vg2x1(uint32_t, svuint16_t, svuint16_t);
1058	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_bf16_vg2x2)))
1059	void svmls_za32_bf16_vg2x2(uint32_t, svbfloat16x2_t, svbfloat16x2_t);
1060	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_f16_vg2x2)))
1061	void svmls_za32_f16_vg2x2(uint32_t, svfloat16x2_t, svfloat16x2_t);
1062	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_s16_vg2x2)))
1063	void svmls_za32_s16_vg2x2(uint32_t, svint16x2_t, svint16x2_t);
1064	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_u16_vg2x2)))
1065	void svmls_za32_u16_vg2x2(uint32_t, svuint16x2_t, svuint16x2_t);
1066	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_bf16_vg2x4)))
1067	void svmls_za32_bf16_vg2x4(uint32_t, svbfloat16x4_t, svbfloat16x4_t);
1068	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_f16_vg2x4)))
1069	void svmls_za32_f16_vg2x4(uint32_t, svfloat16x4_t, svfloat16x4_t);
1070	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_s16_vg2x4)))
1071	void svmls_za32_s16_vg2x4(uint32_t, svint16x4_t, svint16x4_t);
1072	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_u16_vg2x4)))
1073	void svmls_za32_u16_vg2x4(uint32_t, svuint16x4_t, svuint16x4_t);
1074	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_s8_vg4x1)))
1075	void svmls_za32_s8_vg4x1(uint32_t, svint8_t, svint8_t);
1076	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_u8_vg4x1)))
1077	void svmls_za32_u8_vg4x1(uint32_t, svuint8_t, svuint8_t);
1078	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_s8_vg4x2)))
1079	void svmls_za32_s8_vg4x2(uint32_t, svint8x2_t, svint8x2_t);
1080	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_u8_vg4x2)))
1081	void svmls_za32_u8_vg4x2(uint32_t, svuint8x2_t, svuint8x2_t);
1082	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_s8_vg4x4)))
1083	void svmls_za32_s8_vg4x4(uint32_t, svint8x4_t, svint8x4_t);
1084	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_u8_vg4x4)))
1085	void svmls_za32_u8_vg4x4(uint32_t, svuint8x4_t, svuint8x4_t);
1086	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_s16_m)))
1087	void svmopa_za32_s16_m(uint64_t, svbool_t, svbool_t, svint16_t, svint16_t);
1088	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_u16_m)))
1089	void svmopa_za32_u16_m(uint64_t, svbool_t, svbool_t, svuint16_t, svuint16_t);
1090	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_s16_m)))
1091	void svmops_za32_s16_m(uint64_t, svbool_t, svbool_t, svint16_t, svint16_t);
1092	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_u16_m)))
1093	void svmops_za32_u16_m(uint64_t, svbool_t, svbool_t, svuint16_t, svuint16_t);
1094	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_u16_vg2)))
1095	svuint16x2_t svread_hor_za16_u16_vg2(uint64_t, uint32_t);
1096	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_bf16_vg2)))
1097	svbfloat16x2_t svread_hor_za16_bf16_vg2(uint64_t, uint32_t);
1098	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_f16_vg2)))
1099	svfloat16x2_t svread_hor_za16_f16_vg2(uint64_t, uint32_t);
1100	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_s16_vg2)))
1101	svint16x2_t svread_hor_za16_s16_vg2(uint64_t, uint32_t);
1102	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_u16_vg4)))
1103	svuint16x4_t svread_hor_za16_u16_vg4(uint64_t, uint32_t);
1104	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_bf16_vg4)))
1105	svbfloat16x4_t svread_hor_za16_bf16_vg4(uint64_t, uint32_t);
1106	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_f16_vg4)))
1107	svfloat16x4_t svread_hor_za16_f16_vg4(uint64_t, uint32_t);
1108	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_s16_vg4)))
1109	svint16x4_t svread_hor_za16_s16_vg4(uint64_t, uint32_t);
1110	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_u32_vg2)))
1111	svuint32x2_t svread_hor_za32_u32_vg2(uint64_t, uint32_t);
1112	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_f32_vg2)))
1113	svfloat32x2_t svread_hor_za32_f32_vg2(uint64_t, uint32_t);
1114	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_s32_vg2)))
1115	svint32x2_t svread_hor_za32_s32_vg2(uint64_t, uint32_t);
1116	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_u32_vg4)))
1117	svuint32x4_t svread_hor_za32_u32_vg4(uint64_t, uint32_t);
1118	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_f32_vg4)))
1119	svfloat32x4_t svread_hor_za32_f32_vg4(uint64_t, uint32_t);
1120	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_s32_vg4)))
1121	svint32x4_t svread_hor_za32_s32_vg4(uint64_t, uint32_t);
1122	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_u64_vg2)))
1123	svuint64x2_t svread_hor_za64_u64_vg2(uint64_t, uint32_t);
1124	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_f64_vg2)))
1125	svfloat64x2_t svread_hor_za64_f64_vg2(uint64_t, uint32_t);
1126	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_s64_vg2)))
1127	svint64x2_t svread_hor_za64_s64_vg2(uint64_t, uint32_t);
1128	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_u64_vg4)))
1129	svuint64x4_t svread_hor_za64_u64_vg4(uint64_t, uint32_t);
1130	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_f64_vg4)))
1131	svfloat64x4_t svread_hor_za64_f64_vg4(uint64_t, uint32_t);
1132	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_s64_vg4)))
1133	svint64x4_t svread_hor_za64_s64_vg4(uint64_t, uint32_t);
1134	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za8_u8_vg2)))
1135	svuint8x2_t svread_hor_za8_u8_vg2(uint64_t, uint32_t);
1136	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za8_s8_vg2)))
1137	svint8x2_t svread_hor_za8_s8_vg2(uint64_t, uint32_t);
1138	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za8_u8_vg4)))
1139	svuint8x4_t svread_hor_za8_u8_vg4(uint64_t, uint32_t);
1140	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za8_s8_vg4)))
1141	svint8x4_t svread_hor_za8_s8_vg4(uint64_t, uint32_t);
1142	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_u16_vg2)))
1143	svuint16x2_t svread_ver_za16_u16_vg2(uint64_t, uint32_t);
1144	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_bf16_vg2)))
1145	svbfloat16x2_t svread_ver_za16_bf16_vg2(uint64_t, uint32_t);
1146	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_f16_vg2)))
1147	svfloat16x2_t svread_ver_za16_f16_vg2(uint64_t, uint32_t);
1148	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_s16_vg2)))
1149	svint16x2_t svread_ver_za16_s16_vg2(uint64_t, uint32_t);
1150	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_u16_vg4)))
1151	svuint16x4_t svread_ver_za16_u16_vg4(uint64_t, uint32_t);
1152	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_bf16_vg4)))
1153	svbfloat16x4_t svread_ver_za16_bf16_vg4(uint64_t, uint32_t);
1154	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_f16_vg4)))
1155	svfloat16x4_t svread_ver_za16_f16_vg4(uint64_t, uint32_t);
1156	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_s16_vg4)))
1157	svint16x4_t svread_ver_za16_s16_vg4(uint64_t, uint32_t);
1158	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_u32_vg2)))
1159	svuint32x2_t svread_ver_za32_u32_vg2(uint64_t, uint32_t);
1160	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_f32_vg2)))
1161	svfloat32x2_t svread_ver_za32_f32_vg2(uint64_t, uint32_t);
1162	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_s32_vg2)))
1163	svint32x2_t svread_ver_za32_s32_vg2(uint64_t, uint32_t);
1164	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_u32_vg4)))
1165	svuint32x4_t svread_ver_za32_u32_vg4(uint64_t, uint32_t);
1166	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_f32_vg4)))
1167	svfloat32x4_t svread_ver_za32_f32_vg4(uint64_t, uint32_t);
1168	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_s32_vg4)))
1169	svint32x4_t svread_ver_za32_s32_vg4(uint64_t, uint32_t);
1170	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_u64_vg2)))
1171	svuint64x2_t svread_ver_za64_u64_vg2(uint64_t, uint32_t);
1172	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_f64_vg2)))
1173	svfloat64x2_t svread_ver_za64_f64_vg2(uint64_t, uint32_t);
1174	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_s64_vg2)))
1175	svint64x2_t svread_ver_za64_s64_vg2(uint64_t, uint32_t);
1176	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_u64_vg4)))
1177	svuint64x4_t svread_ver_za64_u64_vg4(uint64_t, uint32_t);
1178	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_f64_vg4)))
1179	svfloat64x4_t svread_ver_za64_f64_vg4(uint64_t, uint32_t);
1180	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_s64_vg4)))
1181	svint64x4_t svread_ver_za64_s64_vg4(uint64_t, uint32_t);
1182	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za8_u8_vg2)))
1183	svuint8x2_t svread_ver_za8_u8_vg2(uint64_t, uint32_t);
1184	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za8_s8_vg2)))
1185	svint8x2_t svread_ver_za8_s8_vg2(uint64_t, uint32_t);
1186	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za8_u8_vg4)))
1187	svuint8x4_t svread_ver_za8_u8_vg4(uint64_t, uint32_t);
1188	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za8_s8_vg4)))
1189	svint8x4_t svread_ver_za8_s8_vg4(uint64_t, uint32_t);
1190	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za16_u16_vg1x2)))
1191	svuint16x2_t svread_za16_u16_vg1x2(uint32_t);
1192	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za16_bf16_vg1x2)))
1193	svbfloat16x2_t svread_za16_bf16_vg1x2(uint32_t);
1194	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za16_f16_vg1x2)))
1195	svfloat16x2_t svread_za16_f16_vg1x2(uint32_t);
1196	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za16_s16_vg1x2)))
1197	svint16x2_t svread_za16_s16_vg1x2(uint32_t);
1198	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za16_u16_vg1x4)))
1199	svuint16x4_t svread_za16_u16_vg1x4(uint32_t);
1200	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za16_bf16_vg1x4)))
1201	svbfloat16x4_t svread_za16_bf16_vg1x4(uint32_t);
1202	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za16_f16_vg1x4)))
1203	svfloat16x4_t svread_za16_f16_vg1x4(uint32_t);
1204	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za16_s16_vg1x4)))
1205	svint16x4_t svread_za16_s16_vg1x4(uint32_t);
1206	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za32_u32_vg1x2)))
1207	svuint32x2_t svread_za32_u32_vg1x2(uint32_t);
1208	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za32_f32_vg1x2)))
1209	svfloat32x2_t svread_za32_f32_vg1x2(uint32_t);
1210	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za32_s32_vg1x2)))
1211	svint32x2_t svread_za32_s32_vg1x2(uint32_t);
1212	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za32_u32_vg1x4)))
1213	svuint32x4_t svread_za32_u32_vg1x4(uint32_t);
1214	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za32_f32_vg1x4)))
1215	svfloat32x4_t svread_za32_f32_vg1x4(uint32_t);
1216	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za32_s32_vg1x4)))
1217	svint32x4_t svread_za32_s32_vg1x4(uint32_t);
1218	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za64_u64_vg1x2)))
1219	svuint64x2_t svread_za64_u64_vg1x2(uint32_t);
1220	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za64_f64_vg1x2)))
1221	svfloat64x2_t svread_za64_f64_vg1x2(uint32_t);
1222	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za64_s64_vg1x2)))
1223	svint64x2_t svread_za64_s64_vg1x2(uint32_t);
1224	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za64_u64_vg1x4)))
1225	svuint64x4_t svread_za64_u64_vg1x4(uint32_t);
1226	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za64_f64_vg1x4)))
1227	svfloat64x4_t svread_za64_f64_vg1x4(uint32_t);
1228	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za64_s64_vg1x4)))
1229	svint64x4_t svread_za64_s64_vg1x4(uint32_t);
1230	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za8_u8_vg1x2)))
1231	svuint8x2_t svread_za8_u8_vg1x2(uint32_t);
1232	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za8_s8_vg1x2)))
1233	svint8x2_t svread_za8_s8_vg1x2(uint32_t);
1234	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za8_u8_vg1x4)))
1235	svuint8x4_t svread_za8_u8_vg1x4(uint32_t);
1236	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_za8_s8_vg1x4)))
1237	svint8x4_t svread_za8_s8_vg1x4(uint32_t);
1238	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svstr_zt)))
1239	void svstr_zt(uint64_t, void *);
1240	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_single_za32_u32_vg1x2)))
1241	void svsub_write_single_za32_u32_vg1x2(uint32_t, svuint32x2_t, svuint32_t);
1242	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_single_za32_s32_vg1x2)))
1243	void svsub_write_single_za32_s32_vg1x2(uint32_t, svint32x2_t, svint32_t);
1244	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_single_za32_u32_vg1x4)))
1245	void svsub_write_single_za32_u32_vg1x4(uint32_t, svuint32x4_t, svuint32_t);
1246	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_single_za32_s32_vg1x4)))
1247	void svsub_write_single_za32_s32_vg1x4(uint32_t, svint32x4_t, svint32_t);
1248	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_za32_u32_vg1x2)))
1249	void svsub_write_za32_u32_vg1x2(uint32_t, svuint32x2_t, svuint32x2_t);
1250	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_za32_s32_vg1x2)))
1251	void svsub_write_za32_s32_vg1x2(uint32_t, svint32x2_t, svint32x2_t);
1252	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_za32_u32_vg1x4)))
1253	void svsub_write_za32_u32_vg1x4(uint32_t, svuint32x4_t, svuint32x4_t);
1254	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_za32_s32_vg1x4)))
1255	void svsub_write_za32_s32_vg1x4(uint32_t, svint32x4_t, svint32x4_t);
1256	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za32_u32_vg1x2)))
1257	void svsub_za32_u32_vg1x2(uint32_t, svuint32x2_t);
1258	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za32_f32_vg1x2)))
1259	void svsub_za32_f32_vg1x2(uint32_t, svfloat32x2_t);
1260	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za32_s32_vg1x2)))
1261	void svsub_za32_s32_vg1x2(uint32_t, svint32x2_t);
1262	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za32_u32_vg1x4)))
1263	void svsub_za32_u32_vg1x4(uint32_t, svuint32x4_t);
1264	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za32_f32_vg1x4)))
1265	void svsub_za32_f32_vg1x4(uint32_t, svfloat32x4_t);
1266	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za32_s32_vg1x4)))
1267	void svsub_za32_s32_vg1x4(uint32_t, svint32x4_t);
1268	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsudot_single_za32_s8_vg1x2)))
1269	void svsudot_single_za32_s8_vg1x2(uint32_t, svint8x2_t, svuint8_t);
1270	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsudot_single_za32_s8_vg1x4)))
1271	void svsudot_single_za32_s8_vg1x4(uint32_t, svint8x4_t, svuint8_t);
1272	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsudot_lane_za32_s8_vg1x2)))
1273	void svsudot_lane_za32_s8_vg1x2(uint32_t, svint8x2_t, svuint8_t, uint64_t);
1274	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsudot_lane_za32_s8_vg1x4)))
1275	void svsudot_lane_za32_s8_vg1x4(uint32_t, svint8x4_t, svuint8_t, uint64_t);
1276	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsudot_za32_s8_vg1x2)))
1277	void svsudot_za32_s8_vg1x2(uint32_t, svint8x2_t, svuint8x2_t);
1278	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsudot_za32_s8_vg1x4)))
1279	void svsudot_za32_s8_vg1x4(uint32_t, svint8x4_t, svuint8x4_t);
1280	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumla_single_za32_s8_vg4x2)))
1281	void svsumla_single_za32_s8_vg4x2(uint32_t, svint8x2_t, svuint8_t);
1282	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumla_single_za32_s8_vg4x4)))
1283	void svsumla_single_za32_s8_vg4x4(uint32_t, svint8x4_t, svuint8_t);
1284	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumla_lane_za32_s8_vg4x1)))
1285	void svsumla_lane_za32_s8_vg4x1(uint32_t, svint8_t, svuint8_t, uint64_t);
1286	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumla_lane_za32_s8_vg4x2)))
1287	void svsumla_lane_za32_s8_vg4x2(uint32_t, svint8x2_t, svuint8_t, uint64_t);
1288	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumla_lane_za32_s8_vg4x4)))
1289	void svsumla_lane_za32_s8_vg4x4(uint32_t, svint8x4_t, svuint8_t, uint64_t);
1290	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumla_za32_s8_vg4x1)))
1291	void svsumla_za32_s8_vg4x1(uint32_t, svint8_t, svuint8_t);
1292	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumla_za32_s8_vg4x2)))
1293	void svsumla_za32_s8_vg4x2(uint32_t, svint8x2_t, svuint8x2_t);
1294	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumla_za32_s8_vg4x4)))
1295	void svsumla_za32_s8_vg4x4(uint32_t, svint8x4_t, svuint8x4_t);
1296	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsuvdot_lane_za32_s8_vg1x4)))
1297	void svsuvdot_lane_za32_s8_vg1x4(uint32_t, svint8x4_t, svint8_t, uint64_t);
1298	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusdot_single_za32_u8_vg1x2)))
1299	void svusdot_single_za32_u8_vg1x2(uint32_t, svuint8x2_t, svint8_t);
1300	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusdot_single_za32_u8_vg1x4)))
1301	void svusdot_single_za32_u8_vg1x4(uint32_t, svuint8x4_t, svint8_t);
1302	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusdot_lane_za32_u8_vg1x2)))
1303	void svusdot_lane_za32_u8_vg1x2(uint32_t, svuint8x2_t, svint8_t, uint64_t);
1304	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusdot_lane_za32_u8_vg1x4)))
1305	void svusdot_lane_za32_u8_vg1x4(uint32_t, svuint8x4_t, svint8_t, uint64_t);
1306	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusdot_za32_u8_vg1x2)))
1307	void svusdot_za32_u8_vg1x2(uint32_t, svuint8x2_t, svint8x2_t);
1308	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusdot_za32_u8_vg1x4)))
1309	void svusdot_za32_u8_vg1x4(uint32_t, svuint8x4_t, svint8x4_t);
1310	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmla_single_za32_u8_vg4x2)))
1311	void svusmla_single_za32_u8_vg4x2(uint32_t, svuint8x2_t, svint8_t);
1312	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmla_single_za32_u8_vg4x4)))
1313	void svusmla_single_za32_u8_vg4x4(uint32_t, svuint8x4_t, svint8_t);
1314	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmla_lane_za32_u8_vg4x1)))
1315	void svusmla_lane_za32_u8_vg4x1(uint32_t, svuint8_t, svint8_t, uint64_t);
1316	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmla_lane_za32_u8_vg4x2)))
1317	void svusmla_lane_za32_u8_vg4x2(uint32_t, svuint8x2_t, svint8_t, uint64_t);
1318	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmla_lane_za32_u8_vg4x4)))
1319	void svusmla_lane_za32_u8_vg4x4(uint32_t, svuint8x4_t, svint8_t, uint64_t);
1320	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmla_za32_u8_vg4x1)))
1321	void svusmla_za32_u8_vg4x1(uint32_t, svuint8_t, svint8_t);
1322	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmla_za32_u8_vg4x2)))
1323	void svusmla_za32_u8_vg4x2(uint32_t, svuint8x2_t, svint8x2_t);
1324	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmla_za32_u8_vg4x4)))
1325	void svusmla_za32_u8_vg4x4(uint32_t, svuint8x4_t, svint8x4_t);
1326	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusvdot_lane_za32_u8_vg1x4)))
1327	void svusvdot_lane_za32_u8_vg1x4(uint32_t, svuint8x4_t, svuint8_t, uint64_t);
1328	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svvdot_lane_za32_bf16_vg1x2)))
1329	void svvdot_lane_za32_bf16_vg1x2(uint32_t, svbfloat16x2_t, svbfloat16_t, uint64_t);
1330	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svvdot_lane_za32_f16_vg1x2)))
1331	void svvdot_lane_za32_f16_vg1x2(uint32_t, svfloat16x2_t, svfloat16_t, uint64_t);
1332	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svvdot_lane_za32_s16_vg1x2)))
1333	void svvdot_lane_za32_s16_vg1x2(uint32_t, svint16x2_t, svint16_t, uint64_t);
1334	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svvdot_lane_za32_u16_vg1x2)))
1335	void svvdot_lane_za32_u16_vg1x2(uint32_t, svuint16x2_t, svuint16_t, uint64_t);
1336	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svvdot_lane_za32_s8_vg1x4)))
1337	void svvdot_lane_za32_s8_vg1x4(uint32_t, svint8x4_t, svint8_t, uint64_t);
1338	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svvdot_lane_za32_u8_vg1x4)))
1339	void svvdot_lane_za32_u8_vg1x4(uint32_t, svuint8x4_t, svuint8_t, uint64_t);
1340	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_u16_vg2)))
1341	void svwrite_hor_za16_u16_vg2(uint64_t, uint32_t, svuint16x2_t);
1342	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_bf16_vg2)))
1343	void svwrite_hor_za16_bf16_vg2(uint64_t, uint32_t, svbfloat16x2_t);
1344	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_f16_vg2)))
1345	void svwrite_hor_za16_f16_vg2(uint64_t, uint32_t, svfloat16x2_t);
1346	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_s16_vg2)))
1347	void svwrite_hor_za16_s16_vg2(uint64_t, uint32_t, svint16x2_t);
1348	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_u16_vg4)))
1349	void svwrite_hor_za16_u16_vg4(uint64_t, uint32_t, svuint16x4_t);
1350	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_bf16_vg4)))
1351	void svwrite_hor_za16_bf16_vg4(uint64_t, uint32_t, svbfloat16x4_t);
1352	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_f16_vg4)))
1353	void svwrite_hor_za16_f16_vg4(uint64_t, uint32_t, svfloat16x4_t);
1354	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_s16_vg4)))
1355	void svwrite_hor_za16_s16_vg4(uint64_t, uint32_t, svint16x4_t);
1356	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_u32_vg2)))
1357	void svwrite_hor_za32_u32_vg2(uint64_t, uint32_t, svuint32x2_t);
1358	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_f32_vg2)))
1359	void svwrite_hor_za32_f32_vg2(uint64_t, uint32_t, svfloat32x2_t);
1360	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_s32_vg2)))
1361	void svwrite_hor_za32_s32_vg2(uint64_t, uint32_t, svint32x2_t);
1362	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_u32_vg4)))
1363	void svwrite_hor_za32_u32_vg4(uint64_t, uint32_t, svuint32x4_t);
1364	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_f32_vg4)))
1365	void svwrite_hor_za32_f32_vg4(uint64_t, uint32_t, svfloat32x4_t);
1366	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_s32_vg4)))
1367	void svwrite_hor_za32_s32_vg4(uint64_t, uint32_t, svint32x4_t);
1368	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_u64_vg2)))
1369	void svwrite_hor_za64_u64_vg2(uint64_t, uint32_t, svuint64x2_t);
1370	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_f64_vg2)))
1371	void svwrite_hor_za64_f64_vg2(uint64_t, uint32_t, svfloat64x2_t);
1372	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_s64_vg2)))
1373	void svwrite_hor_za64_s64_vg2(uint64_t, uint32_t, svint64x2_t);
1374	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_u64_vg4)))
1375	void svwrite_hor_za64_u64_vg4(uint64_t, uint32_t, svuint64x4_t);
1376	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_f64_vg4)))
1377	void svwrite_hor_za64_f64_vg4(uint64_t, uint32_t, svfloat64x4_t);
1378	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_s64_vg4)))
1379	void svwrite_hor_za64_s64_vg4(uint64_t, uint32_t, svint64x4_t);
1380	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za8_u8_vg2)))
1381	void svwrite_hor_za8_u8_vg2(uint64_t, uint32_t, svuint8x2_t);
1382	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za8_s8_vg2)))
1383	void svwrite_hor_za8_s8_vg2(uint64_t, uint32_t, svint8x2_t);
1384	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za8_u8_vg4)))
1385	void svwrite_hor_za8_u8_vg4(uint64_t, uint32_t, svuint8x4_t);
1386	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za8_s8_vg4)))
1387	void svwrite_hor_za8_s8_vg4(uint64_t, uint32_t, svint8x4_t);
1388	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_u16_vg2)))
1389	void svwrite_ver_za16_u16_vg2(uint64_t, uint32_t, svuint16x2_t);
1390	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_bf16_vg2)))
1391	void svwrite_ver_za16_bf16_vg2(uint64_t, uint32_t, svbfloat16x2_t);
1392	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_f16_vg2)))
1393	void svwrite_ver_za16_f16_vg2(uint64_t, uint32_t, svfloat16x2_t);
1394	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_s16_vg2)))
1395	void svwrite_ver_za16_s16_vg2(uint64_t, uint32_t, svint16x2_t);
1396	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_u16_vg4)))
1397	void svwrite_ver_za16_u16_vg4(uint64_t, uint32_t, svuint16x4_t);
1398	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_bf16_vg4)))
1399	void svwrite_ver_za16_bf16_vg4(uint64_t, uint32_t, svbfloat16x4_t);
1400	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_f16_vg4)))
1401	void svwrite_ver_za16_f16_vg4(uint64_t, uint32_t, svfloat16x4_t);
1402	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_s16_vg4)))
1403	void svwrite_ver_za16_s16_vg4(uint64_t, uint32_t, svint16x4_t);
1404	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_u32_vg2)))
1405	void svwrite_ver_za32_u32_vg2(uint64_t, uint32_t, svuint32x2_t);
1406	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_f32_vg2)))
1407	void svwrite_ver_za32_f32_vg2(uint64_t, uint32_t, svfloat32x2_t);
1408	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_s32_vg2)))
1409	void svwrite_ver_za32_s32_vg2(uint64_t, uint32_t, svint32x2_t);
1410	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_u32_vg4)))
1411	void svwrite_ver_za32_u32_vg4(uint64_t, uint32_t, svuint32x4_t);
1412	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_f32_vg4)))
1413	void svwrite_ver_za32_f32_vg4(uint64_t, uint32_t, svfloat32x4_t);
1414	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_s32_vg4)))
1415	void svwrite_ver_za32_s32_vg4(uint64_t, uint32_t, svint32x4_t);
1416	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_u64_vg2)))
1417	void svwrite_ver_za64_u64_vg2(uint64_t, uint32_t, svuint64x2_t);
1418	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_f64_vg2)))
1419	void svwrite_ver_za64_f64_vg2(uint64_t, uint32_t, svfloat64x2_t);
1420	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_s64_vg2)))
1421	void svwrite_ver_za64_s64_vg2(uint64_t, uint32_t, svint64x2_t);
1422	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_u64_vg4)))
1423	void svwrite_ver_za64_u64_vg4(uint64_t, uint32_t, svuint64x4_t);
1424	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_f64_vg4)))
1425	void svwrite_ver_za64_f64_vg4(uint64_t, uint32_t, svfloat64x4_t);
1426	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_s64_vg4)))
1427	void svwrite_ver_za64_s64_vg4(uint64_t, uint32_t, svint64x4_t);
1428	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za8_u8_vg2)))
1429	void svwrite_ver_za8_u8_vg2(uint64_t, uint32_t, svuint8x2_t);
1430	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za8_s8_vg2)))
1431	void svwrite_ver_za8_s8_vg2(uint64_t, uint32_t, svint8x2_t);
1432	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za8_u8_vg4)))
1433	void svwrite_ver_za8_u8_vg4(uint64_t, uint32_t, svuint8x4_t);
1434	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za8_s8_vg4)))
1435	void svwrite_ver_za8_s8_vg4(uint64_t, uint32_t, svint8x4_t);
1436	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za16_u16_vg1x2)))
1437	void svwrite_za16_u16_vg1x2(uint32_t, svuint16x2_t);
1438	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za16_bf16_vg1x2)))
1439	void svwrite_za16_bf16_vg1x2(uint32_t, svbfloat16x2_t);
1440	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za16_f16_vg1x2)))
1441	void svwrite_za16_f16_vg1x2(uint32_t, svfloat16x2_t);
1442	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za16_s16_vg1x2)))
1443	void svwrite_za16_s16_vg1x2(uint32_t, svint16x2_t);
1444	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za16_u16_vg1x4)))
1445	void svwrite_za16_u16_vg1x4(uint32_t, svuint16x4_t);
1446	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za16_bf16_vg1x4)))
1447	void svwrite_za16_bf16_vg1x4(uint32_t, svbfloat16x4_t);
1448	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za16_f16_vg1x4)))
1449	void svwrite_za16_f16_vg1x4(uint32_t, svfloat16x4_t);
1450	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za16_s16_vg1x4)))
1451	void svwrite_za16_s16_vg1x4(uint32_t, svint16x4_t);
1452	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za32_u32_vg1x2)))
1453	void svwrite_za32_u32_vg1x2(uint32_t, svuint32x2_t);
1454	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za32_f32_vg1x2)))
1455	void svwrite_za32_f32_vg1x2(uint32_t, svfloat32x2_t);
1456	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za32_s32_vg1x2)))
1457	void svwrite_za32_s32_vg1x2(uint32_t, svint32x2_t);
1458	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za32_u32_vg1x4)))
1459	void svwrite_za32_u32_vg1x4(uint32_t, svuint32x4_t);
1460	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za32_f32_vg1x4)))
1461	void svwrite_za32_f32_vg1x4(uint32_t, svfloat32x4_t);
1462	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za32_s32_vg1x4)))
1463	void svwrite_za32_s32_vg1x4(uint32_t, svint32x4_t);
1464	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za64_u64_vg1x2)))
1465	void svwrite_za64_u64_vg1x2(uint32_t, svuint64x2_t);
1466	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za64_f64_vg1x2)))
1467	void svwrite_za64_f64_vg1x2(uint32_t, svfloat64x2_t);
1468	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za64_s64_vg1x2)))
1469	void svwrite_za64_s64_vg1x2(uint32_t, svint64x2_t);
1470	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za64_u64_vg1x4)))
1471	void svwrite_za64_u64_vg1x4(uint32_t, svuint64x4_t);
1472	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za64_f64_vg1x4)))
1473	void svwrite_za64_f64_vg1x4(uint32_t, svfloat64x4_t);
1474	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za64_s64_vg1x4)))
1475	void svwrite_za64_s64_vg1x4(uint32_t, svint64x4_t);
1476	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za8_u8_vg1x2)))
1477	void svwrite_za8_u8_vg1x2(uint32_t, svuint8x2_t);
1478	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za8_s8_vg1x2)))
1479	void svwrite_za8_s8_vg1x2(uint32_t, svint8x2_t);
1480	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za8_u8_vg1x4)))
1481	void svwrite_za8_u8_vg1x4(uint32_t, svuint8x4_t);
1482	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za8_s8_vg1x4)))
1483	void svwrite_za8_s8_vg1x4(uint32_t, svint8x4_t);
1484	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svzero_zt)))
1485	void svzero_zt(uint64_t);
1486	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_single_za32_u32_vg1x2)))
1487	void svadd_write_za32_vg1x2(uint32_t, svuint32x2_t, svuint32_t);
1488	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_single_za32_s32_vg1x2)))
1489	void svadd_write_za32_vg1x2(uint32_t, svint32x2_t, svint32_t);
1490	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_single_za32_u32_vg1x4)))
1491	void svadd_write_za32_vg1x4(uint32_t, svuint32x4_t, svuint32_t);
1492	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_single_za32_s32_vg1x4)))
1493	void svadd_write_za32_vg1x4(uint32_t, svint32x4_t, svint32_t);
1494	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_za32_u32_vg1x2)))
1495	void svadd_write_za32_vg1x2(uint32_t, svuint32x2_t, svuint32x2_t);
1496	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_za32_s32_vg1x2)))
1497	void svadd_write_za32_vg1x2(uint32_t, svint32x2_t, svint32x2_t);
1498	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_za32_u32_vg1x4)))
1499	void svadd_write_za32_vg1x4(uint32_t, svuint32x4_t, svuint32x4_t);
1500	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_za32_s32_vg1x4)))
1501	void svadd_write_za32_vg1x4(uint32_t, svint32x4_t, svint32x4_t);
1502	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za32_u32_vg1x2)))
1503	void svadd_za32_vg1x2(uint32_t, svuint32x2_t);
1504	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za32_f32_vg1x2)))
1505	void svadd_za32_vg1x2(uint32_t, svfloat32x2_t);
1506	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za32_s32_vg1x2)))
1507	void svadd_za32_vg1x2(uint32_t, svint32x2_t);
1508	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za32_u32_vg1x4)))
1509	void svadd_za32_vg1x4(uint32_t, svuint32x4_t);
1510	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za32_f32_vg1x4)))
1511	void svadd_za32_vg1x4(uint32_t, svfloat32x4_t);
1512	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za32_s32_vg1x4)))
1513	void svadd_za32_vg1x4(uint32_t, svint32x4_t);
1514	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svbmopa_za32_u32_m)))
1515	void svbmopa_za32_m(uint64_t, svbool_t, svbool_t, svuint32_t, svuint32_t);
1516	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svbmopa_za32_s32_m)))
1517	void svbmopa_za32_m(uint64_t, svbool_t, svbool_t, svint32_t, svint32_t);
1518	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svbmops_za32_u32_m)))
1519	void svbmops_za32_m(uint64_t, svbool_t, svbool_t, svuint32_t, svuint32_t);
1520	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svbmops_za32_s32_m)))
1521	void svbmops_za32_m(uint64_t, svbool_t, svbool_t, svint32_t, svint32_t);
1522	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_bf16_vg1x2)))
1523	void svdot_za32_vg1x2(uint32_t, svbfloat16x2_t, svbfloat16_t);
1524	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_f16_vg1x2)))
1525	void svdot_za32_vg1x2(uint32_t, svfloat16x2_t, svfloat16_t);
1526	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_s8_vg1x2)))
1527	void svdot_za32_vg1x2(uint32_t, svint8x2_t, svint8_t);
1528	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_s16_vg1x2)))
1529	void svdot_za32_vg1x2(uint32_t, svint16x2_t, svint16_t);
1530	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_u8_vg1x2)))
1531	void svdot_za32_vg1x2(uint32_t, svuint8x2_t, svuint8_t);
1532	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_u16_vg1x2)))
1533	void svdot_za32_vg1x2(uint32_t, svuint16x2_t, svuint16_t);
1534	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_bf16_vg1x4)))
1535	void svdot_za32_vg1x4(uint32_t, svbfloat16x4_t, svbfloat16_t);
1536	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_f16_vg1x4)))
1537	void svdot_za32_vg1x4(uint32_t, svfloat16x4_t, svfloat16_t);
1538	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_s8_vg1x4)))
1539	void svdot_za32_vg1x4(uint32_t, svint8x4_t, svint8_t);
1540	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_s16_vg1x4)))
1541	void svdot_za32_vg1x4(uint32_t, svint16x4_t, svint16_t);
1542	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_u8_vg1x4)))
1543	void svdot_za32_vg1x4(uint32_t, svuint8x4_t, svuint8_t);
1544	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za32_u16_vg1x4)))
1545	void svdot_za32_vg1x4(uint32_t, svuint16x4_t, svuint16_t);
1546	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_bf16_vg1x2)))
1547	void svdot_lane_za32_vg1x2(uint32_t, svbfloat16x2_t, svbfloat16_t, uint64_t);
1548	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_f16_vg1x2)))
1549	void svdot_lane_za32_vg1x2(uint32_t, svfloat16x2_t, svfloat16_t, uint64_t);
1550	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_s8_vg1x2)))
1551	void svdot_lane_za32_vg1x2(uint32_t, svint8x2_t, svint8_t, uint64_t);
1552	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_s16_vg1x2)))
1553	void svdot_lane_za32_vg1x2(uint32_t, svint16x2_t, svint16_t, uint64_t);
1554	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_u8_vg1x2)))
1555	void svdot_lane_za32_vg1x2(uint32_t, svuint8x2_t, svuint8_t, uint64_t);
1556	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_u16_vg1x2)))
1557	void svdot_lane_za32_vg1x2(uint32_t, svuint16x2_t, svuint16_t, uint64_t);
1558	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_bf16_vg1x4)))
1559	void svdot_lane_za32_vg1x4(uint32_t, svbfloat16x4_t, svbfloat16_t, uint64_t);
1560	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_f16_vg1x4)))
1561	void svdot_lane_za32_vg1x4(uint32_t, svfloat16x4_t, svfloat16_t, uint64_t);
1562	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_s8_vg1x4)))
1563	void svdot_lane_za32_vg1x4(uint32_t, svint8x4_t, svint8_t, uint64_t);
1564	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_s16_vg1x4)))
1565	void svdot_lane_za32_vg1x4(uint32_t, svint16x4_t, svint16_t, uint64_t);
1566	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_u8_vg1x4)))
1567	void svdot_lane_za32_vg1x4(uint32_t, svuint8x4_t, svuint8_t, uint64_t);
1568	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za32_u16_vg1x4)))
1569	void svdot_lane_za32_vg1x4(uint32_t, svuint16x4_t, svuint16_t, uint64_t);
1570	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_bf16_vg1x2)))
1571	void svdot_za32_vg1x2(uint32_t, svbfloat16x2_t, svbfloat16x2_t);
1572	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_f16_vg1x2)))
1573	void svdot_za32_vg1x2(uint32_t, svfloat16x2_t, svfloat16x2_t);
1574	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_s8_vg1x2)))
1575	void svdot_za32_vg1x2(uint32_t, svint8x2_t, svint8x2_t);
1576	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_s16_vg1x2)))
1577	void svdot_za32_vg1x2(uint32_t, svint16x2_t, svint16x2_t);
1578	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_u8_vg1x2)))
1579	void svdot_za32_vg1x2(uint32_t, svuint8x2_t, svuint8x2_t);
1580	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_u16_vg1x2)))
1581	void svdot_za32_vg1x2(uint32_t, svuint16x2_t, svuint16x2_t);
1582	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_bf16_vg1x4)))
1583	void svdot_za32_vg1x4(uint32_t, svbfloat16x4_t, svbfloat16x4_t);
1584	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_f16_vg1x4)))
1585	void svdot_za32_vg1x4(uint32_t, svfloat16x4_t, svfloat16x4_t);
1586	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_s8_vg1x4)))
1587	void svdot_za32_vg1x4(uint32_t, svint8x4_t, svint8x4_t);
1588	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_s16_vg1x4)))
1589	void svdot_za32_vg1x4(uint32_t, svint16x4_t, svint16x4_t);
1590	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_u8_vg1x4)))
1591	void svdot_za32_vg1x4(uint32_t, svuint8x4_t, svuint8x4_t);
1592	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za32_u16_vg1x4)))
1593	void svdot_za32_vg1x4(uint32_t, svuint16x4_t, svuint16x4_t);
1594	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_f32_vg1x2)))
1595	void svmla_za32_vg1x2(uint32_t, svfloat32x2_t, svfloat32_t);
1596	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_f32_vg1x4)))
1597	void svmla_za32_vg1x4(uint32_t, svfloat32x4_t, svfloat32_t);
1598	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_bf16_vg2x2)))
1599	void svmla_za32_vg2x2(uint32_t, svbfloat16x2_t, svbfloat16_t);
1600	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_f16_vg2x2)))
1601	void svmla_za32_vg2x2(uint32_t, svfloat16x2_t, svfloat16_t);
1602	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_s16_vg2x2)))
1603	void svmla_za32_vg2x2(uint32_t, svint16x2_t, svint16_t);
1604	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_u16_vg2x2)))
1605	void svmla_za32_vg2x2(uint32_t, svuint16x2_t, svuint16_t);
1606	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_bf16_vg2x4)))
1607	void svmla_za32_vg2x4(uint32_t, svbfloat16x4_t, svbfloat16_t);
1608	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_f16_vg2x4)))
1609	void svmla_za32_vg2x4(uint32_t, svfloat16x4_t, svfloat16_t);
1610	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_s16_vg2x4)))
1611	void svmla_za32_vg2x4(uint32_t, svint16x4_t, svint16_t);
1612	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_u16_vg2x4)))
1613	void svmla_za32_vg2x4(uint32_t, svuint16x4_t, svuint16_t);
1614	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_s8_vg4x2)))
1615	void svmla_za32_vg4x2(uint32_t, svint8x2_t, svint8_t);
1616	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_u8_vg4x2)))
1617	void svmla_za32_vg4x2(uint32_t, svuint8x2_t, svuint8_t);
1618	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_s8_vg4x4)))
1619	void svmla_za32_vg4x4(uint32_t, svint8x4_t, svint8_t);
1620	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za32_u8_vg4x4)))
1621	void svmla_za32_vg4x4(uint32_t, svuint8x4_t, svuint8_t);
1622	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_f32_vg1x2)))
1623	void svmla_lane_za32_vg1x2(uint32_t, svfloat32x2_t, svfloat32_t, uint64_t);
1624	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_f32_vg1x4)))
1625	void svmla_lane_za32_vg1x4(uint32_t, svfloat32x4_t, svfloat32_t, uint64_t);
1626	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_bf16_vg2x1)))
1627	void svmla_lane_za32_vg2x1(uint32_t, svbfloat16_t, svbfloat16_t, uint64_t);
1628	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_f16_vg2x1)))
1629	void svmla_lane_za32_vg2x1(uint32_t, svfloat16_t, svfloat16_t, uint64_t);
1630	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_s16_vg2x1)))
1631	void svmla_lane_za32_vg2x1(uint32_t, svint16_t, svint16_t, uint64_t);
1632	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_u16_vg2x1)))
1633	void svmla_lane_za32_vg2x1(uint32_t, svuint16_t, svuint16_t, uint64_t);
1634	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_bf16_vg2x2)))
1635	void svmla_lane_za32_vg2x2(uint32_t, svbfloat16x2_t, svbfloat16_t, uint64_t);
1636	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_f16_vg2x2)))
1637	void svmla_lane_za32_vg2x2(uint32_t, svfloat16x2_t, svfloat16_t, uint64_t);
1638	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_s16_vg2x2)))
1639	void svmla_lane_za32_vg2x2(uint32_t, svint16x2_t, svint16_t, uint64_t);
1640	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_u16_vg2x2)))
1641	void svmla_lane_za32_vg2x2(uint32_t, svuint16x2_t, svuint16_t, uint64_t);
1642	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_bf16_vg2x4)))
1643	void svmla_lane_za32_vg2x4(uint32_t, svbfloat16x4_t, svbfloat16_t, uint64_t);
1644	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_f16_vg2x4)))
1645	void svmla_lane_za32_vg2x4(uint32_t, svfloat16x4_t, svfloat16_t, uint64_t);
1646	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_s16_vg2x4)))
1647	void svmla_lane_za32_vg2x4(uint32_t, svint16x4_t, svint16_t, uint64_t);
1648	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_u16_vg2x4)))
1649	void svmla_lane_za32_vg2x4(uint32_t, svuint16x4_t, svuint16_t, uint64_t);
1650	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_s8_vg4x1)))
1651	void svmla_lane_za32_vg4x1(uint32_t, svint8_t, svint8_t, uint64_t);
1652	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_u8_vg4x1)))
1653	void svmla_lane_za32_vg4x1(uint32_t, svuint8_t, svuint8_t, uint64_t);
1654	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_s8_vg4x2)))
1655	void svmla_lane_za32_vg4x2(uint32_t, svint8x2_t, svint8_t, uint64_t);
1656	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_u8_vg4x2)))
1657	void svmla_lane_za32_vg4x2(uint32_t, svuint8x2_t, svuint8_t, uint64_t);
1658	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_s8_vg4x4)))
1659	void svmla_lane_za32_vg4x4(uint32_t, svint8x4_t, svint8_t, uint64_t);
1660	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za32_u8_vg4x4)))
1661	void svmla_lane_za32_vg4x4(uint32_t, svuint8x4_t, svuint8_t, uint64_t);
1662	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_f32_vg1x2)))
1663	void svmla_za32_vg1x2(uint32_t, svfloat32x2_t, svfloat32x2_t);
1664	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_f32_vg1x4)))
1665	void svmla_za32_vg1x4(uint32_t, svfloat32x4_t, svfloat32x4_t);
1666	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_bf16_vg2x1)))
1667	void svmla_za32_vg2x1(uint32_t, svbfloat16_t, svbfloat16_t);
1668	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_f16_vg2x1)))
1669	void svmla_za32_vg2x1(uint32_t, svfloat16_t, svfloat16_t);
1670	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_s16_vg2x1)))
1671	void svmla_za32_vg2x1(uint32_t, svint16_t, svint16_t);
1672	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_u16_vg2x1)))
1673	void svmla_za32_vg2x1(uint32_t, svuint16_t, svuint16_t);
1674	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_bf16_vg2x2)))
1675	void svmla_za32_vg2x2(uint32_t, svbfloat16x2_t, svbfloat16x2_t);
1676	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_f16_vg2x2)))
1677	void svmla_za32_vg2x2(uint32_t, svfloat16x2_t, svfloat16x2_t);
1678	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_s16_vg2x2)))
1679	void svmla_za32_vg2x2(uint32_t, svint16x2_t, svint16x2_t);
1680	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_u16_vg2x2)))
1681	void svmla_za32_vg2x2(uint32_t, svuint16x2_t, svuint16x2_t);
1682	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_bf16_vg2x4)))
1683	void svmla_za32_vg2x4(uint32_t, svbfloat16x4_t, svbfloat16x4_t);
1684	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_f16_vg2x4)))
1685	void svmla_za32_vg2x4(uint32_t, svfloat16x4_t, svfloat16x4_t);
1686	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_s16_vg2x4)))
1687	void svmla_za32_vg2x4(uint32_t, svint16x4_t, svint16x4_t);
1688	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_u16_vg2x4)))
1689	void svmla_za32_vg2x4(uint32_t, svuint16x4_t, svuint16x4_t);
1690	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_s8_vg4x1)))
1691	void svmla_za32_vg4x1(uint32_t, svint8_t, svint8_t);
1692	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_u8_vg4x1)))
1693	void svmla_za32_vg4x1(uint32_t, svuint8_t, svuint8_t);
1694	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_s8_vg4x2)))
1695	void svmla_za32_vg4x2(uint32_t, svint8x2_t, svint8x2_t);
1696	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_u8_vg4x2)))
1697	void svmla_za32_vg4x2(uint32_t, svuint8x2_t, svuint8x2_t);
1698	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_s8_vg4x4)))
1699	void svmla_za32_vg4x4(uint32_t, svint8x4_t, svint8x4_t);
1700	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za32_u8_vg4x4)))
1701	void svmla_za32_vg4x4(uint32_t, svuint8x4_t, svuint8x4_t);
1702	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_f32_vg1x2)))
1703	void svmls_za32_vg1x2(uint32_t, svfloat32x2_t, svfloat32_t);
1704	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_f32_vg1x4)))
1705	void svmls_za32_vg1x4(uint32_t, svfloat32x4_t, svfloat32_t);
1706	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_bf16_vg2x2)))
1707	void svmls_za32_vg2x2(uint32_t, svbfloat16x2_t, svbfloat16_t);
1708	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_f16_vg2x2)))
1709	void svmls_za32_vg2x2(uint32_t, svfloat16x2_t, svfloat16_t);
1710	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_s16_vg2x2)))
1711	void svmls_za32_vg2x2(uint32_t, svint16x2_t, svint16_t);
1712	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_u16_vg2x2)))
1713	void svmls_za32_vg2x2(uint32_t, svuint16x2_t, svuint16_t);
1714	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_bf16_vg2x4)))
1715	void svmls_za32_vg2x4(uint32_t, svbfloat16x4_t, svbfloat16_t);
1716	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_f16_vg2x4)))
1717	void svmls_za32_vg2x4(uint32_t, svfloat16x4_t, svfloat16_t);
1718	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_s16_vg2x4)))
1719	void svmls_za32_vg2x4(uint32_t, svint16x4_t, svint16_t);
1720	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_u16_vg2x4)))
1721	void svmls_za32_vg2x4(uint32_t, svuint16x4_t, svuint16_t);
1722	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_s8_vg4x2)))
1723	void svmls_za32_vg4x2(uint32_t, svint8x2_t, svint8_t);
1724	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_u8_vg4x2)))
1725	void svmls_za32_vg4x2(uint32_t, svuint8x2_t, svuint8_t);
1726	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_s8_vg4x4)))
1727	void svmls_za32_vg4x4(uint32_t, svint8x4_t, svint8_t);
1728	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za32_u8_vg4x4)))
1729	void svmls_za32_vg4x4(uint32_t, svuint8x4_t, svuint8_t);
1730	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_f32_vg1x2)))
1731	void svmls_lane_za32_vg1x2(uint32_t, svfloat32x2_t, svfloat32_t, uint64_t);
1732	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_f32_vg1x4)))
1733	void svmls_lane_za32_vg1x4(uint32_t, svfloat32x4_t, svfloat32_t, uint64_t);
1734	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_bf16_vg2x1)))
1735	void svmls_lane_za32_vg2x1(uint32_t, svbfloat16_t, svbfloat16_t, uint64_t);
1736	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_f16_vg2x1)))
1737	void svmls_lane_za32_vg2x1(uint32_t, svfloat16_t, svfloat16_t, uint64_t);
1738	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_s16_vg2x1)))
1739	void svmls_lane_za32_vg2x1(uint32_t, svint16_t, svint16_t, uint64_t);
1740	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_u16_vg2x1)))
1741	void svmls_lane_za32_vg2x1(uint32_t, svuint16_t, svuint16_t, uint64_t);
1742	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_bf16_vg2x2)))
1743	void svmls_lane_za32_vg2x2(uint32_t, svbfloat16x2_t, svbfloat16_t, uint64_t);
1744	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_f16_vg2x2)))
1745	void svmls_lane_za32_vg2x2(uint32_t, svfloat16x2_t, svfloat16_t, uint64_t);
1746	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_s16_vg2x2)))
1747	void svmls_lane_za32_vg2x2(uint32_t, svint16x2_t, svint16_t, uint64_t);
1748	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_u16_vg2x2)))
1749	void svmls_lane_za32_vg2x2(uint32_t, svuint16x2_t, svuint16_t, uint64_t);
1750	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_bf16_vg2x4)))
1751	void svmls_lane_za32_vg2x4(uint32_t, svbfloat16x4_t, svbfloat16_t, uint64_t);
1752	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_f16_vg2x4)))
1753	void svmls_lane_za32_vg2x4(uint32_t, svfloat16x4_t, svfloat16_t, uint64_t);
1754	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_s16_vg2x4)))
1755	void svmls_lane_za32_vg2x4(uint32_t, svint16x4_t, svint16_t, uint64_t);
1756	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_u16_vg2x4)))
1757	void svmls_lane_za32_vg2x4(uint32_t, svuint16x4_t, svuint16_t, uint64_t);
1758	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_s8_vg4x1)))
1759	void svmls_lane_za32_vg4x1(uint32_t, svint8_t, svint8_t, uint64_t);
1760	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_u8_vg4x1)))
1761	void svmls_lane_za32_vg4x1(uint32_t, svuint8_t, svuint8_t, uint64_t);
1762	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_s8_vg4x2)))
1763	void svmls_lane_za32_vg4x2(uint32_t, svint8x2_t, svint8_t, uint64_t);
1764	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_u8_vg4x2)))
1765	void svmls_lane_za32_vg4x2(uint32_t, svuint8x2_t, svuint8_t, uint64_t);
1766	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_s8_vg4x4)))
1767	void svmls_lane_za32_vg4x4(uint32_t, svint8x4_t, svint8_t, uint64_t);
1768	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za32_u8_vg4x4)))
1769	void svmls_lane_za32_vg4x4(uint32_t, svuint8x4_t, svuint8_t, uint64_t);
1770	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_f32_vg1x2)))
1771	void svmls_za32_vg1x2(uint32_t, svfloat32x2_t, svfloat32x2_t);
1772	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_f32_vg1x4)))
1773	void svmls_za32_vg1x4(uint32_t, svfloat32x4_t, svfloat32x4_t);
1774	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_bf16_vg2x1)))
1775	void svmls_za32_vg2x1(uint32_t, svbfloat16_t, svbfloat16_t);
1776	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_f16_vg2x1)))
1777	void svmls_za32_vg2x1(uint32_t, svfloat16_t, svfloat16_t);
1778	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_s16_vg2x1)))
1779	void svmls_za32_vg2x1(uint32_t, svint16_t, svint16_t);
1780	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_u16_vg2x1)))
1781	void svmls_za32_vg2x1(uint32_t, svuint16_t, svuint16_t);
1782	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_bf16_vg2x2)))
1783	void svmls_za32_vg2x2(uint32_t, svbfloat16x2_t, svbfloat16x2_t);
1784	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_f16_vg2x2)))
1785	void svmls_za32_vg2x2(uint32_t, svfloat16x2_t, svfloat16x2_t);
1786	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_s16_vg2x2)))
1787	void svmls_za32_vg2x2(uint32_t, svint16x2_t, svint16x2_t);
1788	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_u16_vg2x2)))
1789	void svmls_za32_vg2x2(uint32_t, svuint16x2_t, svuint16x2_t);
1790	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_bf16_vg2x4)))
1791	void svmls_za32_vg2x4(uint32_t, svbfloat16x4_t, svbfloat16x4_t);
1792	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_f16_vg2x4)))
1793	void svmls_za32_vg2x4(uint32_t, svfloat16x4_t, svfloat16x4_t);
1794	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_s16_vg2x4)))
1795	void svmls_za32_vg2x4(uint32_t, svint16x4_t, svint16x4_t);
1796	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_u16_vg2x4)))
1797	void svmls_za32_vg2x4(uint32_t, svuint16x4_t, svuint16x4_t);
1798	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_s8_vg4x1)))
1799	void svmls_za32_vg4x1(uint32_t, svint8_t, svint8_t);
1800	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_u8_vg4x1)))
1801	void svmls_za32_vg4x1(uint32_t, svuint8_t, svuint8_t);
1802	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_s8_vg4x2)))
1803	void svmls_za32_vg4x2(uint32_t, svint8x2_t, svint8x2_t);
1804	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_u8_vg4x2)))
1805	void svmls_za32_vg4x2(uint32_t, svuint8x2_t, svuint8x2_t);
1806	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_s8_vg4x4)))
1807	void svmls_za32_vg4x4(uint32_t, svint8x4_t, svint8x4_t);
1808	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za32_u8_vg4x4)))
1809	void svmls_za32_vg4x4(uint32_t, svuint8x4_t, svuint8x4_t);
1810	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_s16_m)))
1811	void svmopa_za32_m(uint64_t, svbool_t, svbool_t, svint16_t, svint16_t);
1812	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_u16_m)))
1813	void svmopa_za32_m(uint64_t, svbool_t, svbool_t, svuint16_t, svuint16_t);
1814	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_s16_m)))
1815	void svmops_za32_m(uint64_t, svbool_t, svbool_t, svint16_t, svint16_t);
1816	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_u16_m)))
1817	void svmops_za32_m(uint64_t, svbool_t, svbool_t, svuint16_t, svuint16_t);
1818	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_single_za32_u32_vg1x2)))
1819	void svsub_write_za32_vg1x2(uint32_t, svuint32x2_t, svuint32_t);
1820	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_single_za32_s32_vg1x2)))
1821	void svsub_write_za32_vg1x2(uint32_t, svint32x2_t, svint32_t);
1822	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_single_za32_u32_vg1x4)))
1823	void svsub_write_za32_vg1x4(uint32_t, svuint32x4_t, svuint32_t);
1824	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_single_za32_s32_vg1x4)))
1825	void svsub_write_za32_vg1x4(uint32_t, svint32x4_t, svint32_t);
1826	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_za32_u32_vg1x2)))
1827	void svsub_write_za32_vg1x2(uint32_t, svuint32x2_t, svuint32x2_t);
1828	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_za32_s32_vg1x2)))
1829	void svsub_write_za32_vg1x2(uint32_t, svint32x2_t, svint32x2_t);
1830	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_za32_u32_vg1x4)))
1831	void svsub_write_za32_vg1x4(uint32_t, svuint32x4_t, svuint32x4_t);
1832	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_za32_s32_vg1x4)))
1833	void svsub_write_za32_vg1x4(uint32_t, svint32x4_t, svint32x4_t);
1834	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za32_u32_vg1x2)))
1835	void svsub_za32_vg1x2(uint32_t, svuint32x2_t);
1836	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za32_f32_vg1x2)))
1837	void svsub_za32_vg1x2(uint32_t, svfloat32x2_t);
1838	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za32_s32_vg1x2)))
1839	void svsub_za32_vg1x2(uint32_t, svint32x2_t);
1840	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za32_u32_vg1x4)))
1841	void svsub_za32_vg1x4(uint32_t, svuint32x4_t);
1842	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za32_f32_vg1x4)))
1843	void svsub_za32_vg1x4(uint32_t, svfloat32x4_t);
1844	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za32_s32_vg1x4)))
1845	void svsub_za32_vg1x4(uint32_t, svint32x4_t);
1846	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsudot_single_za32_s8_vg1x2)))
1847	void svsudot_za32_vg1x2(uint32_t, svint8x2_t, svuint8_t);
1848	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsudot_single_za32_s8_vg1x4)))
1849	void svsudot_za32_vg1x4(uint32_t, svint8x4_t, svuint8_t);
1850	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsudot_lane_za32_s8_vg1x2)))
1851	void svsudot_lane_za32_vg1x2(uint32_t, svint8x2_t, svuint8_t, uint64_t);
1852	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsudot_lane_za32_s8_vg1x4)))
1853	void svsudot_lane_za32_vg1x4(uint32_t, svint8x4_t, svuint8_t, uint64_t);
1854	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsudot_za32_s8_vg1x2)))
1855	void svsudot_za32_vg1x2(uint32_t, svint8x2_t, svuint8x2_t);
1856	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsudot_za32_s8_vg1x4)))
1857	void svsudot_za32_vg1x4(uint32_t, svint8x4_t, svuint8x4_t);
1858	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumla_single_za32_s8_vg4x2)))
1859	void svsumla_za32_vg4x2(uint32_t, svint8x2_t, svuint8_t);
1860	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumla_single_za32_s8_vg4x4)))
1861	void svsumla_za32_vg4x4(uint32_t, svint8x4_t, svuint8_t);
1862	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumla_lane_za32_s8_vg4x1)))
1863	void svsumla_lane_za32_vg4x1(uint32_t, svint8_t, svuint8_t, uint64_t);
1864	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumla_lane_za32_s8_vg4x2)))
1865	void svsumla_lane_za32_vg4x2(uint32_t, svint8x2_t, svuint8_t, uint64_t);
1866	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumla_lane_za32_s8_vg4x4)))
1867	void svsumla_lane_za32_vg4x4(uint32_t, svint8x4_t, svuint8_t, uint64_t);
1868	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumla_za32_s8_vg4x1)))
1869	void svsumla_za32_vg4x1(uint32_t, svint8_t, svuint8_t);
1870	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumla_za32_s8_vg4x2)))
1871	void svsumla_za32_vg4x2(uint32_t, svint8x2_t, svuint8x2_t);
1872	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumla_za32_s8_vg4x4)))
1873	void svsumla_za32_vg4x4(uint32_t, svint8x4_t, svuint8x4_t);
1874	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsuvdot_lane_za32_s8_vg1x4)))
1875	void svsuvdot_lane_za32_vg1x4(uint32_t, svint8x4_t, svint8_t, uint64_t);
1876	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusdot_single_za32_u8_vg1x2)))
1877	void svusdot_za32_vg1x2(uint32_t, svuint8x2_t, svint8_t);
1878	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusdot_single_za32_u8_vg1x4)))
1879	void svusdot_za32_vg1x4(uint32_t, svuint8x4_t, svint8_t);
1880	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusdot_lane_za32_u8_vg1x2)))
1881	void svusdot_lane_za32_vg1x2(uint32_t, svuint8x2_t, svint8_t, uint64_t);
1882	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusdot_lane_za32_u8_vg1x4)))
1883	void svusdot_lane_za32_vg1x4(uint32_t, svuint8x4_t, svint8_t, uint64_t);
1884	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusdot_za32_u8_vg1x2)))
1885	void svusdot_za32_vg1x2(uint32_t, svuint8x2_t, svint8x2_t);
1886	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusdot_za32_u8_vg1x4)))
1887	void svusdot_za32_vg1x4(uint32_t, svuint8x4_t, svint8x4_t);
1888	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmla_single_za32_u8_vg4x2)))
1889	void svusmla_za32_vg4x2(uint32_t, svuint8x2_t, svint8_t);
1890	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmla_single_za32_u8_vg4x4)))
1891	void svusmla_za32_vg4x4(uint32_t, svuint8x4_t, svint8_t);
1892	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmla_lane_za32_u8_vg4x1)))
1893	void svusmla_lane_za32_vg4x1(uint32_t, svuint8_t, svint8_t, uint64_t);
1894	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmla_lane_za32_u8_vg4x2)))
1895	void svusmla_lane_za32_vg4x2(uint32_t, svuint8x2_t, svint8_t, uint64_t);
1896	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmla_lane_za32_u8_vg4x4)))
1897	void svusmla_lane_za32_vg4x4(uint32_t, svuint8x4_t, svint8_t, uint64_t);
1898	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmla_za32_u8_vg4x1)))
1899	void svusmla_za32_vg4x1(uint32_t, svuint8_t, svint8_t);
1900	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmla_za32_u8_vg4x2)))
1901	void svusmla_za32_vg4x2(uint32_t, svuint8x2_t, svint8x2_t);
1902	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmla_za32_u8_vg4x4)))
1903	void svusmla_za32_vg4x4(uint32_t, svuint8x4_t, svint8x4_t);
1904	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusvdot_lane_za32_u8_vg1x4)))
1905	void svusvdot_lane_za32_vg1x4(uint32_t, svuint8x4_t, svuint8_t, uint64_t);
1906	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svvdot_lane_za32_bf16_vg1x2)))
1907	void svvdot_lane_za32_vg1x2(uint32_t, svbfloat16x2_t, svbfloat16_t, uint64_t);
1908	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svvdot_lane_za32_f16_vg1x2)))
1909	void svvdot_lane_za32_vg1x2(uint32_t, svfloat16x2_t, svfloat16_t, uint64_t);
1910	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svvdot_lane_za32_s16_vg1x2)))
1911	void svvdot_lane_za32_vg1x2(uint32_t, svint16x2_t, svint16_t, uint64_t);
1912	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svvdot_lane_za32_u16_vg1x2)))
1913	void svvdot_lane_za32_vg1x2(uint32_t, svuint16x2_t, svuint16_t, uint64_t);
1914	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svvdot_lane_za32_s8_vg1x4)))
1915	void svvdot_lane_za32_vg1x4(uint32_t, svint8x4_t, svint8_t, uint64_t);
1916	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svvdot_lane_za32_u8_vg1x4)))
1917	void svvdot_lane_za32_vg1x4(uint32_t, svuint8x4_t, svuint8_t, uint64_t);
1918	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_u16_vg2)))
1919	void svwrite_hor_za16_vg2(uint64_t, uint32_t, svuint16x2_t);
1920	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_bf16_vg2)))
1921	void svwrite_hor_za16_vg2(uint64_t, uint32_t, svbfloat16x2_t);
1922	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_f16_vg2)))
1923	void svwrite_hor_za16_vg2(uint64_t, uint32_t, svfloat16x2_t);
1924	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_s16_vg2)))
1925	void svwrite_hor_za16_vg2(uint64_t, uint32_t, svint16x2_t);
1926	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_u16_vg4)))
1927	void svwrite_hor_za16_vg4(uint64_t, uint32_t, svuint16x4_t);
1928	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_bf16_vg4)))
1929	void svwrite_hor_za16_vg4(uint64_t, uint32_t, svbfloat16x4_t);
1930	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_f16_vg4)))
1931	void svwrite_hor_za16_vg4(uint64_t, uint32_t, svfloat16x4_t);
1932	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_s16_vg4)))
1933	void svwrite_hor_za16_vg4(uint64_t, uint32_t, svint16x4_t);
1934	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_u32_vg2)))
1935	void svwrite_hor_za32_vg2(uint64_t, uint32_t, svuint32x2_t);
1936	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_f32_vg2)))
1937	void svwrite_hor_za32_vg2(uint64_t, uint32_t, svfloat32x2_t);
1938	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_s32_vg2)))
1939	void svwrite_hor_za32_vg2(uint64_t, uint32_t, svint32x2_t);
1940	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_u32_vg4)))
1941	void svwrite_hor_za32_vg4(uint64_t, uint32_t, svuint32x4_t);
1942	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_f32_vg4)))
1943	void svwrite_hor_za32_vg4(uint64_t, uint32_t, svfloat32x4_t);
1944	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_s32_vg4)))
1945	void svwrite_hor_za32_vg4(uint64_t, uint32_t, svint32x4_t);
1946	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_u64_vg2)))
1947	void svwrite_hor_za64_vg2(uint64_t, uint32_t, svuint64x2_t);
1948	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_f64_vg2)))
1949	void svwrite_hor_za64_vg2(uint64_t, uint32_t, svfloat64x2_t);
1950	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_s64_vg2)))
1951	void svwrite_hor_za64_vg2(uint64_t, uint32_t, svint64x2_t);
1952	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_u64_vg4)))
1953	void svwrite_hor_za64_vg4(uint64_t, uint32_t, svuint64x4_t);
1954	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_f64_vg4)))
1955	void svwrite_hor_za64_vg4(uint64_t, uint32_t, svfloat64x4_t);
1956	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_s64_vg4)))
1957	void svwrite_hor_za64_vg4(uint64_t, uint32_t, svint64x4_t);
1958	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za8_u8_vg2)))
1959	void svwrite_hor_za8_vg2(uint64_t, uint32_t, svuint8x2_t);
1960	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za8_s8_vg2)))
1961	void svwrite_hor_za8_vg2(uint64_t, uint32_t, svint8x2_t);
1962	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za8_u8_vg4)))
1963	void svwrite_hor_za8_vg4(uint64_t, uint32_t, svuint8x4_t);
1964	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za8_s8_vg4)))
1965	void svwrite_hor_za8_vg4(uint64_t, uint32_t, svint8x4_t);
1966	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_u16_vg2)))
1967	void svwrite_ver_za16_vg2(uint64_t, uint32_t, svuint16x2_t);
1968	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_bf16_vg2)))
1969	void svwrite_ver_za16_vg2(uint64_t, uint32_t, svbfloat16x2_t);
1970	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_f16_vg2)))
1971	void svwrite_ver_za16_vg2(uint64_t, uint32_t, svfloat16x2_t);
1972	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_s16_vg2)))
1973	void svwrite_ver_za16_vg2(uint64_t, uint32_t, svint16x2_t);
1974	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_u16_vg4)))
1975	void svwrite_ver_za16_vg4(uint64_t, uint32_t, svuint16x4_t);
1976	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_bf16_vg4)))
1977	void svwrite_ver_za16_vg4(uint64_t, uint32_t, svbfloat16x4_t);
1978	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_f16_vg4)))
1979	void svwrite_ver_za16_vg4(uint64_t, uint32_t, svfloat16x4_t);
1980	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_s16_vg4)))
1981	void svwrite_ver_za16_vg4(uint64_t, uint32_t, svint16x4_t);
1982	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_u32_vg2)))
1983	void svwrite_ver_za32_vg2(uint64_t, uint32_t, svuint32x2_t);
1984	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_f32_vg2)))
1985	void svwrite_ver_za32_vg2(uint64_t, uint32_t, svfloat32x2_t);
1986	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_s32_vg2)))
1987	void svwrite_ver_za32_vg2(uint64_t, uint32_t, svint32x2_t);
1988	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_u32_vg4)))
1989	void svwrite_ver_za32_vg4(uint64_t, uint32_t, svuint32x4_t);
1990	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_f32_vg4)))
1991	void svwrite_ver_za32_vg4(uint64_t, uint32_t, svfloat32x4_t);
1992	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_s32_vg4)))
1993	void svwrite_ver_za32_vg4(uint64_t, uint32_t, svint32x4_t);
1994	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_u64_vg2)))
1995	void svwrite_ver_za64_vg2(uint64_t, uint32_t, svuint64x2_t);
1996	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_f64_vg2)))
1997	void svwrite_ver_za64_vg2(uint64_t, uint32_t, svfloat64x2_t);
1998	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_s64_vg2)))
1999	void svwrite_ver_za64_vg2(uint64_t, uint32_t, svint64x2_t);
2000	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_u64_vg4)))
2001	void svwrite_ver_za64_vg4(uint64_t, uint32_t, svuint64x4_t);
2002	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_f64_vg4)))
2003	void svwrite_ver_za64_vg4(uint64_t, uint32_t, svfloat64x4_t);
2004	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_s64_vg4)))
2005	void svwrite_ver_za64_vg4(uint64_t, uint32_t, svint64x4_t);
2006	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za8_u8_vg2)))
2007	void svwrite_ver_za8_vg2(uint64_t, uint32_t, svuint8x2_t);
2008	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za8_s8_vg2)))
2009	void svwrite_ver_za8_vg2(uint64_t, uint32_t, svint8x2_t);
2010	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za8_u8_vg4)))
2011	void svwrite_ver_za8_vg4(uint64_t, uint32_t, svuint8x4_t);
2012	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za8_s8_vg4)))
2013	void svwrite_ver_za8_vg4(uint64_t, uint32_t, svint8x4_t);
2014	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za16_u16_vg1x2)))
2015	void svwrite_za16_vg1x2(uint32_t, svuint16x2_t);
2016	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za16_bf16_vg1x2)))
2017	void svwrite_za16_vg1x2(uint32_t, svbfloat16x2_t);
2018	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za16_f16_vg1x2)))
2019	void svwrite_za16_vg1x2(uint32_t, svfloat16x2_t);
2020	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za16_s16_vg1x2)))
2021	void svwrite_za16_vg1x2(uint32_t, svint16x2_t);
2022	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za16_u16_vg1x4)))
2023	void svwrite_za16_vg1x4(uint32_t, svuint16x4_t);
2024	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za16_bf16_vg1x4)))
2025	void svwrite_za16_vg1x4(uint32_t, svbfloat16x4_t);
2026	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za16_f16_vg1x4)))
2027	void svwrite_za16_vg1x4(uint32_t, svfloat16x4_t);
2028	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za16_s16_vg1x4)))
2029	void svwrite_za16_vg1x4(uint32_t, svint16x4_t);
2030	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za32_u32_vg1x2)))
2031	void svwrite_za32_vg1x2(uint32_t, svuint32x2_t);
2032	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za32_f32_vg1x2)))
2033	void svwrite_za32_vg1x2(uint32_t, svfloat32x2_t);
2034	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za32_s32_vg1x2)))
2035	void svwrite_za32_vg1x2(uint32_t, svint32x2_t);
2036	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za32_u32_vg1x4)))
2037	void svwrite_za32_vg1x4(uint32_t, svuint32x4_t);
2038	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za32_f32_vg1x4)))
2039	void svwrite_za32_vg1x4(uint32_t, svfloat32x4_t);
2040	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za32_s32_vg1x4)))
2041	void svwrite_za32_vg1x4(uint32_t, svint32x4_t);
2042	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za64_u64_vg1x2)))
2043	void svwrite_za64_vg1x2(uint32_t, svuint64x2_t);
2044	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za64_f64_vg1x2)))
2045	void svwrite_za64_vg1x2(uint32_t, svfloat64x2_t);
2046	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za64_s64_vg1x2)))
2047	void svwrite_za64_vg1x2(uint32_t, svint64x2_t);
2048	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za64_u64_vg1x4)))
2049	void svwrite_za64_vg1x4(uint32_t, svuint64x4_t);
2050	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za64_f64_vg1x4)))
2051	void svwrite_za64_vg1x4(uint32_t, svfloat64x4_t);
2052	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za64_s64_vg1x4)))
2053	void svwrite_za64_vg1x4(uint32_t, svint64x4_t);
2054	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za8_u8_vg1x2)))
2055	void svwrite_za8_vg1x2(uint32_t, svuint8x2_t);
2056	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za8_s8_vg1x2)))
2057	void svwrite_za8_vg1x2(uint32_t, svint8x2_t);
2058	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za8_u8_vg1x4)))
2059	void svwrite_za8_vg1x4(uint32_t, svuint8x4_t);
2060	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_za8_s8_vg1x4)))
2061	void svwrite_za8_vg1x4(uint32_t, svint8x4_t);
2062	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za64_f64_vg1x2)))
2063	void svadd_za64_f64_vg1x2(uint32_t, svfloat64x2_t);
2064	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za64_f64_vg1x4)))
2065	void svadd_za64_f64_vg1x4(uint32_t, svfloat64x4_t);
2066	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za64_f64_vg1x2)))
2067	void svmla_single_za64_f64_vg1x2(uint32_t, svfloat64x2_t, svfloat64_t);
2068	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za64_f64_vg1x4)))
2069	void svmla_single_za64_f64_vg1x4(uint32_t, svfloat64x4_t, svfloat64_t);
2070	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za64_f64_vg1x2)))
2071	void svmla_lane_za64_f64_vg1x2(uint32_t, svfloat64x2_t, svfloat64_t, uint64_t);
2072	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za64_f64_vg1x4)))
2073	void svmla_lane_za64_f64_vg1x4(uint32_t, svfloat64x4_t, svfloat64_t, uint64_t);
2074	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za64_f64_vg1x2)))
2075	void svmla_za64_f64_vg1x2(uint32_t, svfloat64x2_t, svfloat64x2_t);
2076	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za64_f64_vg1x4)))
2077	void svmla_za64_f64_vg1x4(uint32_t, svfloat64x4_t, svfloat64x4_t);
2078	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za64_f64_vg1x2)))
2079	void svmls_single_za64_f64_vg1x2(uint32_t, svfloat64x2_t, svfloat64_t);
2080	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za64_f64_vg1x4)))
2081	void svmls_single_za64_f64_vg1x4(uint32_t, svfloat64x4_t, svfloat64_t);
2082	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za64_f64_vg1x2)))
2083	void svmls_lane_za64_f64_vg1x2(uint32_t, svfloat64x2_t, svfloat64_t, uint64_t);
2084	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za64_f64_vg1x4)))
2085	void svmls_lane_za64_f64_vg1x4(uint32_t, svfloat64x4_t, svfloat64_t, uint64_t);
2086	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za64_f64_vg1x2)))
2087	void svmls_za64_f64_vg1x2(uint32_t, svfloat64x2_t, svfloat64x2_t);
2088	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za64_f64_vg1x4)))
2089	void svmls_za64_f64_vg1x4(uint32_t, svfloat64x4_t, svfloat64x4_t);
2090	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za64_f64_vg1x2)))
2091	void svsub_za64_f64_vg1x2(uint32_t, svfloat64x2_t);
2092	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za64_f64_vg1x4)))
2093	void svsub_za64_f64_vg1x4(uint32_t, svfloat64x4_t);
2094	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za64_f64_vg1x2)))
2095	void svadd_za64_vg1x2(uint32_t, svfloat64x2_t);
2096	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za64_f64_vg1x4)))
2097	void svadd_za64_vg1x4(uint32_t, svfloat64x4_t);
2098	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za64_f64_vg1x2)))
2099	void svmla_za64_vg1x2(uint32_t, svfloat64x2_t, svfloat64_t);
2100	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za64_f64_vg1x4)))
2101	void svmla_za64_vg1x4(uint32_t, svfloat64x4_t, svfloat64_t);
2102	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za64_f64_vg1x2)))
2103	void svmla_lane_za64_vg1x2(uint32_t, svfloat64x2_t, svfloat64_t, uint64_t);
2104	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za64_f64_vg1x4)))
2105	void svmla_lane_za64_vg1x4(uint32_t, svfloat64x4_t, svfloat64_t, uint64_t);
2106	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za64_f64_vg1x2)))
2107	void svmla_za64_vg1x2(uint32_t, svfloat64x2_t, svfloat64x2_t);
2108	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za64_f64_vg1x4)))
2109	void svmla_za64_vg1x4(uint32_t, svfloat64x4_t, svfloat64x4_t);
2110	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za64_f64_vg1x2)))
2111	void svmls_za64_vg1x2(uint32_t, svfloat64x2_t, svfloat64_t);
2112	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za64_f64_vg1x4)))
2113	void svmls_za64_vg1x4(uint32_t, svfloat64x4_t, svfloat64_t);
2114	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za64_f64_vg1x2)))
2115	void svmls_lane_za64_vg1x2(uint32_t, svfloat64x2_t, svfloat64_t, uint64_t);
2116	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za64_f64_vg1x4)))
2117	void svmls_lane_za64_vg1x4(uint32_t, svfloat64x4_t, svfloat64_t, uint64_t);
2118	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za64_f64_vg1x2)))
2119	void svmls_za64_vg1x2(uint32_t, svfloat64x2_t, svfloat64x2_t);
2120	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za64_f64_vg1x4)))
2121	void svmls_za64_vg1x4(uint32_t, svfloat64x4_t, svfloat64x4_t);
2122	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za64_f64_vg1x2)))
2123	void svsub_za64_vg1x2(uint32_t, svfloat64x2_t);
2124	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za64_f64_vg1x4)))
2125	void svsub_za64_vg1x4(uint32_t, svfloat64x4_t);
2126	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_single_za64_u64_vg1x2)))
2127	void svadd_write_single_za64_u64_vg1x2(uint32_t, svuint64x2_t, svuint64_t);
2128	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_single_za64_s64_vg1x2)))
2129	void svadd_write_single_za64_s64_vg1x2(uint32_t, svint64x2_t, svint64_t);
2130	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_single_za64_u64_vg1x4)))
2131	void svadd_write_single_za64_u64_vg1x4(uint32_t, svuint64x4_t, svuint64_t);
2132	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_single_za64_s64_vg1x4)))
2133	void svadd_write_single_za64_s64_vg1x4(uint32_t, svint64x4_t, svint64_t);
2134	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_za64_u64_vg1x2)))
2135	void svadd_write_za64_u64_vg1x2(uint32_t, svuint64x2_t, svuint64x2_t);
2136	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_za64_s64_vg1x2)))
2137	void svadd_write_za64_s64_vg1x2(uint32_t, svint64x2_t, svint64x2_t);
2138	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_za64_u64_vg1x4)))
2139	void svadd_write_za64_u64_vg1x4(uint32_t, svuint64x4_t, svuint64x4_t);
2140	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_za64_s64_vg1x4)))
2141	void svadd_write_za64_s64_vg1x4(uint32_t, svint64x4_t, svint64x4_t);
2142	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za64_u64_vg1x2)))
2143	void svadd_za64_u64_vg1x2(uint32_t, svuint64x2_t);
2144	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za64_s64_vg1x2)))
2145	void svadd_za64_s64_vg1x2(uint32_t, svint64x2_t);
2146	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za64_u64_vg1x4)))
2147	void svadd_za64_u64_vg1x4(uint32_t, svuint64x4_t);
2148	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za64_s64_vg1x4)))
2149	void svadd_za64_s64_vg1x4(uint32_t, svint64x4_t);
2150	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za64_s16_vg1x2)))
2151	void svdot_single_za64_s16_vg1x2(uint32_t, svint16x2_t, svint16_t);
2152	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za64_u16_vg1x2)))
2153	void svdot_single_za64_u16_vg1x2(uint32_t, svuint16x2_t, svuint16_t);
2154	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za64_s16_vg1x4)))
2155	void svdot_single_za64_s16_vg1x4(uint32_t, svint16x4_t, svint16_t);
2156	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za64_u16_vg1x4)))
2157	void svdot_single_za64_u16_vg1x4(uint32_t, svuint16x4_t, svuint16_t);
2158	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za64_s16_vg1x2)))
2159	void svdot_lane_za64_s16_vg1x2(uint32_t, svint16x2_t, svint16_t, uint64_t);
2160	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za64_u16_vg1x2)))
2161	void svdot_lane_za64_u16_vg1x2(uint32_t, svuint16x2_t, svuint16_t, uint64_t);
2162	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za64_s16_vg1x4)))
2163	void svdot_lane_za64_s16_vg1x4(uint32_t, svint16x4_t, svint16_t, uint64_t);
2164	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za64_u16_vg1x4)))
2165	void svdot_lane_za64_u16_vg1x4(uint32_t, svuint16x4_t, svuint16_t, uint64_t);
2166	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za64_s16_vg1x2)))
2167	void svdot_za64_s16_vg1x2(uint32_t, svint16x2_t, svint16x2_t);
2168	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za64_u16_vg1x2)))
2169	void svdot_za64_u16_vg1x2(uint32_t, svuint16x2_t, svuint16x2_t);
2170	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za64_s16_vg1x4)))
2171	void svdot_za64_s16_vg1x4(uint32_t, svint16x4_t, svint16x4_t);
2172	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za64_u16_vg1x4)))
2173	void svdot_za64_u16_vg1x4(uint32_t, svuint16x4_t, svuint16x4_t);
2174	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za64_s16_vg4x2)))
2175	void svmla_single_za64_s16_vg4x2(uint32_t, svint16x2_t, svint16_t);
2176	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za64_u16_vg4x2)))
2177	void svmla_single_za64_u16_vg4x2(uint32_t, svuint16x2_t, svuint16_t);
2178	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za64_s16_vg4x4)))
2179	void svmla_single_za64_s16_vg4x4(uint32_t, svint16x4_t, svint16_t);
2180	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za64_u16_vg4x4)))
2181	void svmla_single_za64_u16_vg4x4(uint32_t, svuint16x4_t, svuint16_t);
2182	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za64_s16_vg4x1)))
2183	void svmla_lane_za64_s16_vg4x1(uint32_t, svint16_t, svint16_t, uint64_t);
2184	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za64_u16_vg4x1)))
2185	void svmla_lane_za64_u16_vg4x1(uint32_t, svuint16_t, svuint16_t, uint64_t);
2186	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za64_s16_vg4x2)))
2187	void svmla_lane_za64_s16_vg4x2(uint32_t, svint16x2_t, svint16_t, uint64_t);
2188	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za64_u16_vg4x2)))
2189	void svmla_lane_za64_u16_vg4x2(uint32_t, svuint16x2_t, svuint16_t, uint64_t);
2190	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za64_s16_vg4x4)))
2191	void svmla_lane_za64_s16_vg4x4(uint32_t, svint16x4_t, svint16_t, uint64_t);
2192	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za64_u16_vg4x4)))
2193	void svmla_lane_za64_u16_vg4x4(uint32_t, svuint16x4_t, svuint16_t, uint64_t);
2194	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za64_s16_vg4x1)))
2195	void svmla_za64_s16_vg4x1(uint32_t, svint16_t, svint16_t);
2196	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za64_u16_vg4x1)))
2197	void svmla_za64_u16_vg4x1(uint32_t, svuint16_t, svuint16_t);
2198	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za64_s16_vg4x2)))
2199	void svmla_za64_s16_vg4x2(uint32_t, svint16x2_t, svint16x2_t);
2200	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za64_u16_vg4x2)))
2201	void svmla_za64_u16_vg4x2(uint32_t, svuint16x2_t, svuint16x2_t);
2202	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za64_s16_vg4x4)))
2203	void svmla_za64_s16_vg4x4(uint32_t, svint16x4_t, svint16x4_t);
2204	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za64_u16_vg4x4)))
2205	void svmla_za64_u16_vg4x4(uint32_t, svuint16x4_t, svuint16x4_t);
2206	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za64_s16_vg4x2)))
2207	void svmls_single_za64_s16_vg4x2(uint32_t, svint16x2_t, svint16_t);
2208	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za64_u16_vg4x2)))
2209	void svmls_single_za64_u16_vg4x2(uint32_t, svuint16x2_t, svuint16_t);
2210	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za64_s16_vg4x4)))
2211	void svmls_single_za64_s16_vg4x4(uint32_t, svint16x4_t, svint16_t);
2212	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za64_u16_vg4x4)))
2213	void svmls_single_za64_u16_vg4x4(uint32_t, svuint16x4_t, svuint16_t);
2214	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za64_s16_vg4x1)))
2215	void svmls_lane_za64_s16_vg4x1(uint32_t, svint16_t, svint16_t, uint64_t);
2216	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za64_u16_vg4x1)))
2217	void svmls_lane_za64_u16_vg4x1(uint32_t, svuint16_t, svuint16_t, uint64_t);
2218	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za64_s16_vg4x2)))
2219	void svmls_lane_za64_s16_vg4x2(uint32_t, svint16x2_t, svint16_t, uint64_t);
2220	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za64_u16_vg4x2)))
2221	void svmls_lane_za64_u16_vg4x2(uint32_t, svuint16x2_t, svuint16_t, uint64_t);
2222	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za64_s16_vg4x4)))
2223	void svmls_lane_za64_s16_vg4x4(uint32_t, svint16x4_t, svint16_t, uint64_t);
2224	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za64_u16_vg4x4)))
2225	void svmls_lane_za64_u16_vg4x4(uint32_t, svuint16x4_t, svuint16_t, uint64_t);
2226	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za64_s16_vg4x1)))
2227	void svmls_za64_s16_vg4x1(uint32_t, svint16_t, svint16_t);
2228	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za64_u16_vg4x1)))
2229	void svmls_za64_u16_vg4x1(uint32_t, svuint16_t, svuint16_t);
2230	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za64_s16_vg4x2)))
2231	void svmls_za64_s16_vg4x2(uint32_t, svint16x2_t, svint16x2_t);
2232	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za64_u16_vg4x2)))
2233	void svmls_za64_u16_vg4x2(uint32_t, svuint16x2_t, svuint16x2_t);
2234	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za64_s16_vg4x4)))
2235	void svmls_za64_s16_vg4x4(uint32_t, svint16x4_t, svint16x4_t);
2236	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za64_u16_vg4x4)))
2237	void svmls_za64_u16_vg4x4(uint32_t, svuint16x4_t, svuint16x4_t);
2238	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_single_za64_u64_vg1x2)))
2239	void svsub_write_single_za64_u64_vg1x2(uint32_t, svuint64x2_t, svuint64_t);
2240	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_single_za64_s64_vg1x2)))
2241	void svsub_write_single_za64_s64_vg1x2(uint32_t, svint64x2_t, svint64_t);
2242	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_single_za64_u64_vg1x4)))
2243	void svsub_write_single_za64_u64_vg1x4(uint32_t, svuint64x4_t, svuint64_t);
2244	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_single_za64_s64_vg1x4)))
2245	void svsub_write_single_za64_s64_vg1x4(uint32_t, svint64x4_t, svint64_t);
2246	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_za64_u64_vg1x2)))
2247	void svsub_write_za64_u64_vg1x2(uint32_t, svuint64x2_t, svuint64x2_t);
2248	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_za64_s64_vg1x2)))
2249	void svsub_write_za64_s64_vg1x2(uint32_t, svint64x2_t, svint64x2_t);
2250	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_za64_u64_vg1x4)))
2251	void svsub_write_za64_u64_vg1x4(uint32_t, svuint64x4_t, svuint64x4_t);
2252	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_za64_s64_vg1x4)))
2253	void svsub_write_za64_s64_vg1x4(uint32_t, svint64x4_t, svint64x4_t);
2254	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za64_u64_vg1x2)))
2255	void svsub_za64_u64_vg1x2(uint32_t, svuint64x2_t);
2256	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za64_s64_vg1x2)))
2257	void svsub_za64_s64_vg1x2(uint32_t, svint64x2_t);
2258	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za64_u64_vg1x4)))
2259	void svsub_za64_u64_vg1x4(uint32_t, svuint64x4_t);
2260	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za64_s64_vg1x4)))
2261	void svsub_za64_s64_vg1x4(uint32_t, svint64x4_t);
2262	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svvdot_lane_za64_s16_vg1x4)))
2263	void svvdot_lane_za64_s16_vg1x4(uint32_t, svint16x4_t, svint16_t, uint64_t);
2264	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svvdot_lane_za64_u16_vg1x4)))
2265	void svvdot_lane_za64_u16_vg1x4(uint32_t, svuint16x4_t, svuint16_t, uint64_t);
2266	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_single_za64_u64_vg1x2)))
2267	void svadd_write_za64_vg1x2(uint32_t, svuint64x2_t, svuint64_t);
2268	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_single_za64_s64_vg1x2)))
2269	void svadd_write_za64_vg1x2(uint32_t, svint64x2_t, svint64_t);
2270	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_single_za64_u64_vg1x4)))
2271	void svadd_write_za64_vg1x4(uint32_t, svuint64x4_t, svuint64_t);
2272	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_single_za64_s64_vg1x4)))
2273	void svadd_write_za64_vg1x4(uint32_t, svint64x4_t, svint64_t);
2274	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_za64_u64_vg1x2)))
2275	void svadd_write_za64_vg1x2(uint32_t, svuint64x2_t, svuint64x2_t);
2276	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_za64_s64_vg1x2)))
2277	void svadd_write_za64_vg1x2(uint32_t, svint64x2_t, svint64x2_t);
2278	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_za64_u64_vg1x4)))
2279	void svadd_write_za64_vg1x4(uint32_t, svuint64x4_t, svuint64x4_t);
2280	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_write_za64_s64_vg1x4)))
2281	void svadd_write_za64_vg1x4(uint32_t, svint64x4_t, svint64x4_t);
2282	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za64_u64_vg1x2)))
2283	void svadd_za64_vg1x2(uint32_t, svuint64x2_t);
2284	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za64_s64_vg1x2)))
2285	void svadd_za64_vg1x2(uint32_t, svint64x2_t);
2286	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za64_u64_vg1x4)))
2287	void svadd_za64_vg1x4(uint32_t, svuint64x4_t);
2288	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svadd_za64_s64_vg1x4)))
2289	void svadd_za64_vg1x4(uint32_t, svint64x4_t);
2290	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za64_s16_vg1x2)))
2291	void svdot_za64_vg1x2(uint32_t, svint16x2_t, svint16_t);
2292	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za64_u16_vg1x2)))
2293	void svdot_za64_vg1x2(uint32_t, svuint16x2_t, svuint16_t);
2294	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za64_s16_vg1x4)))
2295	void svdot_za64_vg1x4(uint32_t, svint16x4_t, svint16_t);
2296	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_single_za64_u16_vg1x4)))
2297	void svdot_za64_vg1x4(uint32_t, svuint16x4_t, svuint16_t);
2298	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za64_s16_vg1x2)))
2299	void svdot_lane_za64_vg1x2(uint32_t, svint16x2_t, svint16_t, uint64_t);
2300	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za64_u16_vg1x2)))
2301	void svdot_lane_za64_vg1x2(uint32_t, svuint16x2_t, svuint16_t, uint64_t);
2302	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za64_s16_vg1x4)))
2303	void svdot_lane_za64_vg1x4(uint32_t, svint16x4_t, svint16_t, uint64_t);
2304	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_lane_za64_u16_vg1x4)))
2305	void svdot_lane_za64_vg1x4(uint32_t, svuint16x4_t, svuint16_t, uint64_t);
2306	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za64_s16_vg1x2)))
2307	void svdot_za64_vg1x2(uint32_t, svint16x2_t, svint16x2_t);
2308	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za64_u16_vg1x2)))
2309	void svdot_za64_vg1x2(uint32_t, svuint16x2_t, svuint16x2_t);
2310	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za64_s16_vg1x4)))
2311	void svdot_za64_vg1x4(uint32_t, svint16x4_t, svint16x4_t);
2312	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svdot_za64_u16_vg1x4)))
2313	void svdot_za64_vg1x4(uint32_t, svuint16x4_t, svuint16x4_t);
2314	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za64_s16_vg4x2)))
2315	void svmla_za64_vg4x2(uint32_t, svint16x2_t, svint16_t);
2316	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za64_u16_vg4x2)))
2317	void svmla_za64_vg4x2(uint32_t, svuint16x2_t, svuint16_t);
2318	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za64_s16_vg4x4)))
2319	void svmla_za64_vg4x4(uint32_t, svint16x4_t, svint16_t);
2320	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_single_za64_u16_vg4x4)))
2321	void svmla_za64_vg4x4(uint32_t, svuint16x4_t, svuint16_t);
2322	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za64_s16_vg4x1)))
2323	void svmla_lane_za64_vg4x1(uint32_t, svint16_t, svint16_t, uint64_t);
2324	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za64_u16_vg4x1)))
2325	void svmla_lane_za64_vg4x1(uint32_t, svuint16_t, svuint16_t, uint64_t);
2326	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za64_s16_vg4x2)))
2327	void svmla_lane_za64_vg4x2(uint32_t, svint16x2_t, svint16_t, uint64_t);
2328	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za64_u16_vg4x2)))
2329	void svmla_lane_za64_vg4x2(uint32_t, svuint16x2_t, svuint16_t, uint64_t);
2330	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za64_s16_vg4x4)))
2331	void svmla_lane_za64_vg4x4(uint32_t, svint16x4_t, svint16_t, uint64_t);
2332	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_lane_za64_u16_vg4x4)))
2333	void svmla_lane_za64_vg4x4(uint32_t, svuint16x4_t, svuint16_t, uint64_t);
2334	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za64_s16_vg4x1)))
2335	void svmla_za64_vg4x1(uint32_t, svint16_t, svint16_t);
2336	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za64_u16_vg4x1)))
2337	void svmla_za64_vg4x1(uint32_t, svuint16_t, svuint16_t);
2338	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za64_s16_vg4x2)))
2339	void svmla_za64_vg4x2(uint32_t, svint16x2_t, svint16x2_t);
2340	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za64_u16_vg4x2)))
2341	void svmla_za64_vg4x2(uint32_t, svuint16x2_t, svuint16x2_t);
2342	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za64_s16_vg4x4)))
2343	void svmla_za64_vg4x4(uint32_t, svint16x4_t, svint16x4_t);
2344	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmla_za64_u16_vg4x4)))
2345	void svmla_za64_vg4x4(uint32_t, svuint16x4_t, svuint16x4_t);
2346	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za64_s16_vg4x2)))
2347	void svmls_za64_vg4x2(uint32_t, svint16x2_t, svint16_t);
2348	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za64_u16_vg4x2)))
2349	void svmls_za64_vg4x2(uint32_t, svuint16x2_t, svuint16_t);
2350	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za64_s16_vg4x4)))
2351	void svmls_za64_vg4x4(uint32_t, svint16x4_t, svint16_t);
2352	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_single_za64_u16_vg4x4)))
2353	void svmls_za64_vg4x4(uint32_t, svuint16x4_t, svuint16_t);
2354	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za64_s16_vg4x1)))
2355	void svmls_lane_za64_vg4x1(uint32_t, svint16_t, svint16_t, uint64_t);
2356	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za64_u16_vg4x1)))
2357	void svmls_lane_za64_vg4x1(uint32_t, svuint16_t, svuint16_t, uint64_t);
2358	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za64_s16_vg4x2)))
2359	void svmls_lane_za64_vg4x2(uint32_t, svint16x2_t, svint16_t, uint64_t);
2360	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za64_u16_vg4x2)))
2361	void svmls_lane_za64_vg4x2(uint32_t, svuint16x2_t, svuint16_t, uint64_t);
2362	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za64_s16_vg4x4)))
2363	void svmls_lane_za64_vg4x4(uint32_t, svint16x4_t, svint16_t, uint64_t);
2364	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_lane_za64_u16_vg4x4)))
2365	void svmls_lane_za64_vg4x4(uint32_t, svuint16x4_t, svuint16_t, uint64_t);
2366	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za64_s16_vg4x1)))
2367	void svmls_za64_vg4x1(uint32_t, svint16_t, svint16_t);
2368	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za64_u16_vg4x1)))
2369	void svmls_za64_vg4x1(uint32_t, svuint16_t, svuint16_t);
2370	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za64_s16_vg4x2)))
2371	void svmls_za64_vg4x2(uint32_t, svint16x2_t, svint16x2_t);
2372	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za64_u16_vg4x2)))
2373	void svmls_za64_vg4x2(uint32_t, svuint16x2_t, svuint16x2_t);
2374	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za64_s16_vg4x4)))
2375	void svmls_za64_vg4x4(uint32_t, svint16x4_t, svint16x4_t);
2376	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmls_za64_u16_vg4x4)))
2377	void svmls_za64_vg4x4(uint32_t, svuint16x4_t, svuint16x4_t);
2378	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_single_za64_u64_vg1x2)))
2379	void svsub_write_za64_vg1x2(uint32_t, svuint64x2_t, svuint64_t);
2380	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_single_za64_s64_vg1x2)))
2381	void svsub_write_za64_vg1x2(uint32_t, svint64x2_t, svint64_t);
2382	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_single_za64_u64_vg1x4)))
2383	void svsub_write_za64_vg1x4(uint32_t, svuint64x4_t, svuint64_t);
2384	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_single_za64_s64_vg1x4)))
2385	void svsub_write_za64_vg1x4(uint32_t, svint64x4_t, svint64_t);
2386	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_za64_u64_vg1x2)))
2387	void svsub_write_za64_vg1x2(uint32_t, svuint64x2_t, svuint64x2_t);
2388	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_za64_s64_vg1x2)))
2389	void svsub_write_za64_vg1x2(uint32_t, svint64x2_t, svint64x2_t);
2390	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_za64_u64_vg1x4)))
2391	void svsub_write_za64_vg1x4(uint32_t, svuint64x4_t, svuint64x4_t);
2392	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_write_za64_s64_vg1x4)))
2393	void svsub_write_za64_vg1x4(uint32_t, svint64x4_t, svint64x4_t);
2394	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za64_u64_vg1x2)))
2395	void svsub_za64_vg1x2(uint32_t, svuint64x2_t);
2396	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za64_s64_vg1x2)))
2397	void svsub_za64_vg1x2(uint32_t, svint64x2_t);
2398	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za64_u64_vg1x4)))
2399	void svsub_za64_vg1x4(uint32_t, svuint64x4_t);
2400	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsub_za64_s64_vg1x4)))
2401	void svsub_za64_vg1x4(uint32_t, svint64x4_t);
2402	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svvdot_lane_za64_s16_vg1x4)))
2403	void svvdot_lane_za64_vg1x4(uint32_t, svint16x4_t, svint16_t, uint64_t);
2404	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svvdot_lane_za64_u16_vg1x4)))
2405	void svvdot_lane_za64_vg1x4(uint32_t, svuint16x4_t, svuint16_t, uint64_t);
2406	#ifdef __cplusplus
2407	} // extern "C"
2408	#endif
2409
2410	#undef __ai
2411
2412	#endif /* __ARM_SME_H */

lib/include/arm_sme_draft_spec_subject_to_change.h deleted-642

...	@@ -1,642 +0,0 @@
1	/*===---- arm_sme_draft_spec_subject_to_change.h - ARM SME intrinsics ------===
2	*
3	*
4	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
5	* See https://llvm.org/LICENSE.txt for license information.
6	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7	*
8	*===-----------------------------------------------------------------------===
9	*/
10
11	#ifndef __ARM_SME_H
12	#define __ARM_SME_H
13
14	#if !defined(__LITTLE_ENDIAN__)
15	#error "Big endian is currently not supported for arm_sme_draft_spec_subject_to_change.h"
16	#endif
17	#include <arm_sve.h>
18
19	/* Function attributes */
20	#define __ai static __inline__ __attribute__((__always_inline__, __nodebug__))
21
22	#define __aio static __inline__ __attribute__((__always_inline__, __nodebug__, __overloadable__))
23
24	#ifdef __cplusplus
25	extern "C" {
26	#endif
27
28	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddha_za32_u32_m), arm_streaming, arm_shared_za))
29	void svaddha_za32_u32_m(uint64_t, svbool_t, svbool_t, svuint32_t);
30	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddha_za32_s32_m), arm_streaming, arm_shared_za))
31	void svaddha_za32_s32_m(uint64_t, svbool_t, svbool_t, svint32_t);
32	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddva_za32_u32_m), arm_streaming, arm_shared_za))
33	void svaddva_za32_u32_m(uint64_t, svbool_t, svbool_t, svuint32_t);
34	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddva_za32_s32_m), arm_streaming, arm_shared_za))
35	void svaddva_za32_s32_m(uint64_t, svbool_t, svbool_t, svint32_t);
36	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svcntsb), arm_streaming_compatible, arm_preserves_za))
37	uint64_t svcntsb(void);
38	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svcntsd), arm_streaming_compatible, arm_preserves_za))
39	uint64_t svcntsd(void);
40	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svcntsh), arm_streaming_compatible, arm_preserves_za))
41	uint64_t svcntsh(void);
42	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svcntsw), arm_streaming_compatible, arm_preserves_za))
43	uint64_t svcntsw(void);
44	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_vnum_za128), arm_streaming, arm_shared_za))
45	void svld1_hor_vnum_za128(uint64_t, uint32_t, uint64_t, svbool_t, void const *, int64_t);
46	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_vnum_za16), arm_streaming, arm_shared_za))
47	void svld1_hor_vnum_za16(uint64_t, uint32_t, uint64_t, svbool_t, void const *, int64_t);
48	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_vnum_za32), arm_streaming, arm_shared_za))
49	void svld1_hor_vnum_za32(uint64_t, uint32_t, uint64_t, svbool_t, void const *, int64_t);
50	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_vnum_za64), arm_streaming, arm_shared_za))
51	void svld1_hor_vnum_za64(uint64_t, uint32_t, uint64_t, svbool_t, void const *, int64_t);
52	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_vnum_za8), arm_streaming, arm_shared_za))
53	void svld1_hor_vnum_za8(uint64_t, uint32_t, uint64_t, svbool_t, void const *, int64_t);
54	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_za128), arm_streaming, arm_shared_za))
55	void svld1_hor_za128(uint64_t, uint32_t, uint64_t, svbool_t, void const *);
56	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_za16), arm_streaming, arm_shared_za))
57	void svld1_hor_za16(uint64_t, uint32_t, uint64_t, svbool_t, void const *);
58	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_za32), arm_streaming, arm_shared_za))
59	void svld1_hor_za32(uint64_t, uint32_t, uint64_t, svbool_t, void const *);
60	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_za64), arm_streaming, arm_shared_za))
61	void svld1_hor_za64(uint64_t, uint32_t, uint64_t, svbool_t, void const *);
62	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_hor_za8), arm_streaming, arm_shared_za))
63	void svld1_hor_za8(uint64_t, uint32_t, uint64_t, svbool_t, void const *);
64	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_vnum_za128), arm_streaming, arm_shared_za))
65	void svld1_ver_vnum_za128(uint64_t, uint32_t, uint64_t, svbool_t, void const *, int64_t);
66	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_vnum_za16), arm_streaming, arm_shared_za))
67	void svld1_ver_vnum_za16(uint64_t, uint32_t, uint64_t, svbool_t, void const *, int64_t);
68	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_vnum_za32), arm_streaming, arm_shared_za))
69	void svld1_ver_vnum_za32(uint64_t, uint32_t, uint64_t, svbool_t, void const *, int64_t);
70	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_vnum_za64), arm_streaming, arm_shared_za))
71	void svld1_ver_vnum_za64(uint64_t, uint32_t, uint64_t, svbool_t, void const *, int64_t);
72	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_vnum_za8), arm_streaming, arm_shared_za))
73	void svld1_ver_vnum_za8(uint64_t, uint32_t, uint64_t, svbool_t, void const *, int64_t);
74	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_za128), arm_streaming, arm_shared_za))
75	void svld1_ver_za128(uint64_t, uint32_t, uint64_t, svbool_t, void const *);
76	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_za16), arm_streaming, arm_shared_za))
77	void svld1_ver_za16(uint64_t, uint32_t, uint64_t, svbool_t, void const *);
78	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_za32), arm_streaming, arm_shared_za))
79	void svld1_ver_za32(uint64_t, uint32_t, uint64_t, svbool_t, void const *);
80	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_za64), arm_streaming, arm_shared_za))
81	void svld1_ver_za64(uint64_t, uint32_t, uint64_t, svbool_t, void const *);
82	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svld1_ver_za8), arm_streaming, arm_shared_za))
83	void svld1_ver_za8(uint64_t, uint32_t, uint64_t, svbool_t, void const *);
84	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_f16_m), arm_streaming, arm_shared_za))
85	void svmopa_za32_f16_m(uint64_t, svbool_t, svbool_t, svfloat16_t, svfloat16_t);
86	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_bf16_m), arm_streaming, arm_shared_za))
87	void svmopa_za32_bf16_m(uint64_t, svbool_t, svbool_t, svbfloat16_t, svbfloat16_t);
88	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_f32_m), arm_streaming, arm_shared_za))
89	void svmopa_za32_f32_m(uint64_t, svbool_t, svbool_t, svfloat32_t, svfloat32_t);
90	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_s8_m), arm_streaming, arm_shared_za))
91	void svmopa_za32_s8_m(uint64_t, svbool_t, svbool_t, svint8_t, svint8_t);
92	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_u8_m), arm_streaming, arm_shared_za))
93	void svmopa_za32_u8_m(uint64_t, svbool_t, svbool_t, svuint8_t, svuint8_t);
94	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_f16_m), arm_streaming, arm_shared_za))
95	void svmops_za32_f16_m(uint64_t, svbool_t, svbool_t, svfloat16_t, svfloat16_t);
96	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_bf16_m), arm_streaming, arm_shared_za))
97	void svmops_za32_bf16_m(uint64_t, svbool_t, svbool_t, svbfloat16_t, svbfloat16_t);
98	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_f32_m), arm_streaming, arm_shared_za))
99	void svmops_za32_f32_m(uint64_t, svbool_t, svbool_t, svfloat32_t, svfloat32_t);
100	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_s8_m), arm_streaming, arm_shared_za))
101	void svmops_za32_s8_m(uint64_t, svbool_t, svbool_t, svint8_t, svint8_t);
102	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_u8_m), arm_streaming, arm_shared_za))
103	void svmops_za32_u8_m(uint64_t, svbool_t, svbool_t, svuint8_t, svuint8_t);
104	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_u8_m), arm_streaming, arm_shared_za, arm_preserves_za))
105	svuint8_t svread_hor_za128_u8_m(svuint8_t, svbool_t, uint64_t, uint32_t, uint64_t);
106	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_u32_m), arm_streaming, arm_shared_za, arm_preserves_za))
107	svuint32_t svread_hor_za128_u32_m(svuint32_t, svbool_t, uint64_t, uint32_t, uint64_t);
108	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_u64_m), arm_streaming, arm_shared_za, arm_preserves_za))
109	svuint64_t svread_hor_za128_u64_m(svuint64_t, svbool_t, uint64_t, uint32_t, uint64_t);
110	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_u16_m), arm_streaming, arm_shared_za, arm_preserves_za))
111	svuint16_t svread_hor_za128_u16_m(svuint16_t, svbool_t, uint64_t, uint32_t, uint64_t);
112	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_bf16_m), arm_streaming, arm_shared_za, arm_preserves_za))
113	svbfloat16_t svread_hor_za128_bf16_m(svbfloat16_t, svbool_t, uint64_t, uint32_t, uint64_t);
114	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_s8_m), arm_streaming, arm_shared_za, arm_preserves_za))
115	svint8_t svread_hor_za128_s8_m(svint8_t, svbool_t, uint64_t, uint32_t, uint64_t);
116	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_f64_m), arm_streaming, arm_shared_za, arm_preserves_za))
117	svfloat64_t svread_hor_za128_f64_m(svfloat64_t, svbool_t, uint64_t, uint32_t, uint64_t);
118	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_f32_m), arm_streaming, arm_shared_za, arm_preserves_za))
119	svfloat32_t svread_hor_za128_f32_m(svfloat32_t, svbool_t, uint64_t, uint32_t, uint64_t);
120	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_f16_m), arm_streaming, arm_shared_za, arm_preserves_za))
121	svfloat16_t svread_hor_za128_f16_m(svfloat16_t, svbool_t, uint64_t, uint32_t, uint64_t);
122	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_s32_m), arm_streaming, arm_shared_za, arm_preserves_za))
123	svint32_t svread_hor_za128_s32_m(svint32_t, svbool_t, uint64_t, uint32_t, uint64_t);
124	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_s64_m), arm_streaming, arm_shared_za, arm_preserves_za))
125	svint64_t svread_hor_za128_s64_m(svint64_t, svbool_t, uint64_t, uint32_t, uint64_t);
126	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_s16_m), arm_streaming, arm_shared_za, arm_preserves_za))
127	svint16_t svread_hor_za128_s16_m(svint16_t, svbool_t, uint64_t, uint32_t, uint64_t);
128	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_u16_m), arm_streaming, arm_shared_za, arm_preserves_za))
129	svuint16_t svread_hor_za16_u16_m(svuint16_t, svbool_t, uint64_t, uint32_t, uint64_t);
130	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_bf16_m), arm_streaming, arm_shared_za, arm_preserves_za))
131	svbfloat16_t svread_hor_za16_bf16_m(svbfloat16_t, svbool_t, uint64_t, uint32_t, uint64_t);
132	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_f16_m), arm_streaming, arm_shared_za, arm_preserves_za))
133	svfloat16_t svread_hor_za16_f16_m(svfloat16_t, svbool_t, uint64_t, uint32_t, uint64_t);
134	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_s16_m), arm_streaming, arm_shared_za, arm_preserves_za))
135	svint16_t svread_hor_za16_s16_m(svint16_t, svbool_t, uint64_t, uint32_t, uint64_t);
136	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_u32_m), arm_streaming, arm_shared_za, arm_preserves_za))
137	svuint32_t svread_hor_za32_u32_m(svuint32_t, svbool_t, uint64_t, uint32_t, uint64_t);
138	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_f32_m), arm_streaming, arm_shared_za, arm_preserves_za))
139	svfloat32_t svread_hor_za32_f32_m(svfloat32_t, svbool_t, uint64_t, uint32_t, uint64_t);
140	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_s32_m), arm_streaming, arm_shared_za, arm_preserves_za))
141	svint32_t svread_hor_za32_s32_m(svint32_t, svbool_t, uint64_t, uint32_t, uint64_t);
142	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_u64_m), arm_streaming, arm_shared_za, arm_preserves_za))
143	svuint64_t svread_hor_za64_u64_m(svuint64_t, svbool_t, uint64_t, uint32_t, uint64_t);
144	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_f64_m), arm_streaming, arm_shared_za, arm_preserves_za))
145	svfloat64_t svread_hor_za64_f64_m(svfloat64_t, svbool_t, uint64_t, uint32_t, uint64_t);
146	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_s64_m), arm_streaming, arm_shared_za, arm_preserves_za))
147	svint64_t svread_hor_za64_s64_m(svint64_t, svbool_t, uint64_t, uint32_t, uint64_t);
148	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za8_u8_m), arm_streaming, arm_shared_za, arm_preserves_za))
149	svuint8_t svread_hor_za8_u8_m(svuint8_t, svbool_t, uint64_t, uint32_t, uint64_t);
150	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za8_s8_m), arm_streaming, arm_shared_za, arm_preserves_za))
151	svint8_t svread_hor_za8_s8_m(svint8_t, svbool_t, uint64_t, uint32_t, uint64_t);
152	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_u8_m), arm_streaming, arm_shared_za, arm_preserves_za))
153	svuint8_t svread_ver_za128_u8_m(svuint8_t, svbool_t, uint64_t, uint32_t, uint64_t);
154	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_u32_m), arm_streaming, arm_shared_za, arm_preserves_za))
155	svuint32_t svread_ver_za128_u32_m(svuint32_t, svbool_t, uint64_t, uint32_t, uint64_t);
156	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_u64_m), arm_streaming, arm_shared_za, arm_preserves_za))
157	svuint64_t svread_ver_za128_u64_m(svuint64_t, svbool_t, uint64_t, uint32_t, uint64_t);
158	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_u16_m), arm_streaming, arm_shared_za, arm_preserves_za))
159	svuint16_t svread_ver_za128_u16_m(svuint16_t, svbool_t, uint64_t, uint32_t, uint64_t);
160	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_bf16_m), arm_streaming, arm_shared_za, arm_preserves_za))
161	svbfloat16_t svread_ver_za128_bf16_m(svbfloat16_t, svbool_t, uint64_t, uint32_t, uint64_t);
162	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_s8_m), arm_streaming, arm_shared_za, arm_preserves_za))
163	svint8_t svread_ver_za128_s8_m(svint8_t, svbool_t, uint64_t, uint32_t, uint64_t);
164	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_f64_m), arm_streaming, arm_shared_za, arm_preserves_za))
165	svfloat64_t svread_ver_za128_f64_m(svfloat64_t, svbool_t, uint64_t, uint32_t, uint64_t);
166	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_f32_m), arm_streaming, arm_shared_za, arm_preserves_za))
167	svfloat32_t svread_ver_za128_f32_m(svfloat32_t, svbool_t, uint64_t, uint32_t, uint64_t);
168	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_f16_m), arm_streaming, arm_shared_za, arm_preserves_za))
169	svfloat16_t svread_ver_za128_f16_m(svfloat16_t, svbool_t, uint64_t, uint32_t, uint64_t);
170	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_s32_m), arm_streaming, arm_shared_za, arm_preserves_za))
171	svint32_t svread_ver_za128_s32_m(svint32_t, svbool_t, uint64_t, uint32_t, uint64_t);
172	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_s64_m), arm_streaming, arm_shared_za, arm_preserves_za))
173	svint64_t svread_ver_za128_s64_m(svint64_t, svbool_t, uint64_t, uint32_t, uint64_t);
174	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_s16_m), arm_streaming, arm_shared_za, arm_preserves_za))
175	svint16_t svread_ver_za128_s16_m(svint16_t, svbool_t, uint64_t, uint32_t, uint64_t);
176	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_u16_m), arm_streaming, arm_shared_za, arm_preserves_za))
177	svuint16_t svread_ver_za16_u16_m(svuint16_t, svbool_t, uint64_t, uint32_t, uint64_t);
178	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_bf16_m), arm_streaming, arm_shared_za, arm_preserves_za))
179	svbfloat16_t svread_ver_za16_bf16_m(svbfloat16_t, svbool_t, uint64_t, uint32_t, uint64_t);
180	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_f16_m), arm_streaming, arm_shared_za, arm_preserves_za))
181	svfloat16_t svread_ver_za16_f16_m(svfloat16_t, svbool_t, uint64_t, uint32_t, uint64_t);
182	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_s16_m), arm_streaming, arm_shared_za, arm_preserves_za))
183	svint16_t svread_ver_za16_s16_m(svint16_t, svbool_t, uint64_t, uint32_t, uint64_t);
184	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_u32_m), arm_streaming, arm_shared_za, arm_preserves_za))
185	svuint32_t svread_ver_za32_u32_m(svuint32_t, svbool_t, uint64_t, uint32_t, uint64_t);
186	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_f32_m), arm_streaming, arm_shared_za, arm_preserves_za))
187	svfloat32_t svread_ver_za32_f32_m(svfloat32_t, svbool_t, uint64_t, uint32_t, uint64_t);
188	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_s32_m), arm_streaming, arm_shared_za, arm_preserves_za))
189	svint32_t svread_ver_za32_s32_m(svint32_t, svbool_t, uint64_t, uint32_t, uint64_t);
190	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_u64_m), arm_streaming, arm_shared_za, arm_preserves_za))
191	svuint64_t svread_ver_za64_u64_m(svuint64_t, svbool_t, uint64_t, uint32_t, uint64_t);
192	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_f64_m), arm_streaming, arm_shared_za, arm_preserves_za))
193	svfloat64_t svread_ver_za64_f64_m(svfloat64_t, svbool_t, uint64_t, uint32_t, uint64_t);
194	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_s64_m), arm_streaming, arm_shared_za, arm_preserves_za))
195	svint64_t svread_ver_za64_s64_m(svint64_t, svbool_t, uint64_t, uint32_t, uint64_t);
196	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za8_u8_m), arm_streaming, arm_shared_za, arm_preserves_za))
197	svuint8_t svread_ver_za8_u8_m(svuint8_t, svbool_t, uint64_t, uint32_t, uint64_t);
198	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za8_s8_m), arm_streaming, arm_shared_za, arm_preserves_za))
199	svint8_t svread_ver_za8_s8_m(svint8_t, svbool_t, uint64_t, uint32_t, uint64_t);
200	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_vnum_za128), arm_streaming, arm_shared_za, arm_preserves_za))
201	void svst1_hor_vnum_za128(uint64_t, uint32_t, uint64_t, svbool_t, void *, int64_t);
202	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_vnum_za16), arm_streaming, arm_shared_za, arm_preserves_za))
203	void svst1_hor_vnum_za16(uint64_t, uint32_t, uint64_t, svbool_t, void *, int64_t);
204	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_vnum_za32), arm_streaming, arm_shared_za, arm_preserves_za))
205	void svst1_hor_vnum_za32(uint64_t, uint32_t, uint64_t, svbool_t, void *, int64_t);
206	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_vnum_za64), arm_streaming, arm_shared_za, arm_preserves_za))
207	void svst1_hor_vnum_za64(uint64_t, uint32_t, uint64_t, svbool_t, void *, int64_t);
208	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_vnum_za8), arm_streaming, arm_shared_za, arm_preserves_za))
209	void svst1_hor_vnum_za8(uint64_t, uint32_t, uint64_t, svbool_t, void *, int64_t);
210	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_za128), arm_streaming, arm_shared_za, arm_preserves_za))
211	void svst1_hor_za128(uint64_t, uint32_t, uint64_t, svbool_t, void *);
212	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_za16), arm_streaming, arm_shared_za, arm_preserves_za))
213	void svst1_hor_za16(uint64_t, uint32_t, uint64_t, svbool_t, void *);
214	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_za32), arm_streaming, arm_shared_za, arm_preserves_za))
215	void svst1_hor_za32(uint64_t, uint32_t, uint64_t, svbool_t, void *);
216	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_za64), arm_streaming, arm_shared_za, arm_preserves_za))
217	void svst1_hor_za64(uint64_t, uint32_t, uint64_t, svbool_t, void *);
218	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_hor_za8), arm_streaming, arm_shared_za, arm_preserves_za))
219	void svst1_hor_za8(uint64_t, uint32_t, uint64_t, svbool_t, void *);
220	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_vnum_za128), arm_streaming, arm_shared_za, arm_preserves_za))
221	void svst1_ver_vnum_za128(uint64_t, uint32_t, uint64_t, svbool_t, void *, int64_t);
222	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_vnum_za16), arm_streaming, arm_shared_za, arm_preserves_za))
223	void svst1_ver_vnum_za16(uint64_t, uint32_t, uint64_t, svbool_t, void *, int64_t);
224	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_vnum_za32), arm_streaming, arm_shared_za, arm_preserves_za))
225	void svst1_ver_vnum_za32(uint64_t, uint32_t, uint64_t, svbool_t, void *, int64_t);
226	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_vnum_za64), arm_streaming, arm_shared_za, arm_preserves_za))
227	void svst1_ver_vnum_za64(uint64_t, uint32_t, uint64_t, svbool_t, void *, int64_t);
228	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_vnum_za8), arm_streaming, arm_shared_za, arm_preserves_za))
229	void svst1_ver_vnum_za8(uint64_t, uint32_t, uint64_t, svbool_t, void *, int64_t);
230	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_za128), arm_streaming, arm_shared_za, arm_preserves_za))
231	void svst1_ver_za128(uint64_t, uint32_t, uint64_t, svbool_t, void *);
232	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_za16), arm_streaming, arm_shared_za, arm_preserves_za))
233	void svst1_ver_za16(uint64_t, uint32_t, uint64_t, svbool_t, void *);
234	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_za32), arm_streaming, arm_shared_za, arm_preserves_za))
235	void svst1_ver_za32(uint64_t, uint32_t, uint64_t, svbool_t, void *);
236	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_za64), arm_streaming, arm_shared_za, arm_preserves_za))
237	void svst1_ver_za64(uint64_t, uint32_t, uint64_t, svbool_t, void *);
238	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svst1_ver_za8), arm_streaming, arm_shared_za, arm_preserves_za))
239	void svst1_ver_za8(uint64_t, uint32_t, uint64_t, svbool_t, void *);
240	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumopa_za32_s8_m), arm_streaming, arm_shared_za))
241	void svsumopa_za32_s8_m(uint64_t, svbool_t, svbool_t, svint8_t, svuint8_t);
242	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumops_za32_s8_m), arm_streaming, arm_shared_za))
243	void svsumops_za32_s8_m(uint64_t, svbool_t, svbool_t, svint8_t, svuint8_t);
244	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmopa_za32_u8_m), arm_streaming, arm_shared_za))
245	void svusmopa_za32_u8_m(uint64_t, svbool_t, svbool_t, svuint8_t, svint8_t);
246	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmops_za32_u8_m), arm_streaming, arm_shared_za))
247	void svusmops_za32_u8_m(uint64_t, svbool_t, svbool_t, svuint8_t, svint8_t);
248	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_u8_m), arm_streaming, arm_shared_za))
249	void svwrite_hor_za128_u8_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint8_t);
250	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_u32_m), arm_streaming, arm_shared_za))
251	void svwrite_hor_za128_u32_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint32_t);
252	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_u64_m), arm_streaming, arm_shared_za))
253	void svwrite_hor_za128_u64_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint64_t);
254	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_u16_m), arm_streaming, arm_shared_za))
255	void svwrite_hor_za128_u16_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint16_t);
256	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_bf16_m), arm_streaming, arm_shared_za))
257	void svwrite_hor_za128_bf16_m(uint64_t, uint32_t, uint64_t, svbool_t, svbfloat16_t);
258	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_s8_m), arm_streaming, arm_shared_za))
259	void svwrite_hor_za128_s8_m(uint64_t, uint32_t, uint64_t, svbool_t, svint8_t);
260	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_f64_m), arm_streaming, arm_shared_za))
261	void svwrite_hor_za128_f64_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat64_t);
262	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_f32_m), arm_streaming, arm_shared_za))
263	void svwrite_hor_za128_f32_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat32_t);
264	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_f16_m), arm_streaming, arm_shared_za))
265	void svwrite_hor_za128_f16_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat16_t);
266	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_s32_m), arm_streaming, arm_shared_za))
267	void svwrite_hor_za128_s32_m(uint64_t, uint32_t, uint64_t, svbool_t, svint32_t);
268	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_s64_m), arm_streaming, arm_shared_za))
269	void svwrite_hor_za128_s64_m(uint64_t, uint32_t, uint64_t, svbool_t, svint64_t);
270	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_s16_m), arm_streaming, arm_shared_za))
271	void svwrite_hor_za128_s16_m(uint64_t, uint32_t, uint64_t, svbool_t, svint16_t);
272	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_u16_m), arm_streaming, arm_shared_za))
273	void svwrite_hor_za16_u16_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint16_t);
274	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_bf16_m), arm_streaming, arm_shared_za))
275	void svwrite_hor_za16_bf16_m(uint64_t, uint32_t, uint64_t, svbool_t, svbfloat16_t);
276	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_f16_m), arm_streaming, arm_shared_za))
277	void svwrite_hor_za16_f16_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat16_t);
278	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_s16_m), arm_streaming, arm_shared_za))
279	void svwrite_hor_za16_s16_m(uint64_t, uint32_t, uint64_t, svbool_t, svint16_t);
280	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_u32_m), arm_streaming, arm_shared_za))
281	void svwrite_hor_za32_u32_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint32_t);
282	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_f32_m), arm_streaming, arm_shared_za))
283	void svwrite_hor_za32_f32_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat32_t);
284	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_s32_m), arm_streaming, arm_shared_za))
285	void svwrite_hor_za32_s32_m(uint64_t, uint32_t, uint64_t, svbool_t, svint32_t);
286	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_u64_m), arm_streaming, arm_shared_za))
287	void svwrite_hor_za64_u64_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint64_t);
288	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_f64_m), arm_streaming, arm_shared_za))
289	void svwrite_hor_za64_f64_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat64_t);
290	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_s64_m), arm_streaming, arm_shared_za))
291	void svwrite_hor_za64_s64_m(uint64_t, uint32_t, uint64_t, svbool_t, svint64_t);
292	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za8_u8_m), arm_streaming, arm_shared_za))
293	void svwrite_hor_za8_u8_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint8_t);
294	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za8_s8_m), arm_streaming, arm_shared_za))
295	void svwrite_hor_za8_s8_m(uint64_t, uint32_t, uint64_t, svbool_t, svint8_t);
296	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_u8_m), arm_streaming, arm_shared_za))
297	void svwrite_ver_za128_u8_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint8_t);
298	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_u32_m), arm_streaming, arm_shared_za))
299	void svwrite_ver_za128_u32_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint32_t);
300	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_u64_m), arm_streaming, arm_shared_za))
301	void svwrite_ver_za128_u64_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint64_t);
302	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_u16_m), arm_streaming, arm_shared_za))
303	void svwrite_ver_za128_u16_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint16_t);
304	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_bf16_m), arm_streaming, arm_shared_za))
305	void svwrite_ver_za128_bf16_m(uint64_t, uint32_t, uint64_t, svbool_t, svbfloat16_t);
306	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_s8_m), arm_streaming, arm_shared_za))
307	void svwrite_ver_za128_s8_m(uint64_t, uint32_t, uint64_t, svbool_t, svint8_t);
308	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_f64_m), arm_streaming, arm_shared_za))
309	void svwrite_ver_za128_f64_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat64_t);
310	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_f32_m), arm_streaming, arm_shared_za))
311	void svwrite_ver_za128_f32_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat32_t);
312	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_f16_m), arm_streaming, arm_shared_za))
313	void svwrite_ver_za128_f16_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat16_t);
314	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_s32_m), arm_streaming, arm_shared_za))
315	void svwrite_ver_za128_s32_m(uint64_t, uint32_t, uint64_t, svbool_t, svint32_t);
316	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_s64_m), arm_streaming, arm_shared_za))
317	void svwrite_ver_za128_s64_m(uint64_t, uint32_t, uint64_t, svbool_t, svint64_t);
318	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_s16_m), arm_streaming, arm_shared_za))
319	void svwrite_ver_za128_s16_m(uint64_t, uint32_t, uint64_t, svbool_t, svint16_t);
320	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_u16_m), arm_streaming, arm_shared_za))
321	void svwrite_ver_za16_u16_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint16_t);
322	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_bf16_m), arm_streaming, arm_shared_za))
323	void svwrite_ver_za16_bf16_m(uint64_t, uint32_t, uint64_t, svbool_t, svbfloat16_t);
324	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_f16_m), arm_streaming, arm_shared_za))
325	void svwrite_ver_za16_f16_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat16_t);
326	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_s16_m), arm_streaming, arm_shared_za))
327	void svwrite_ver_za16_s16_m(uint64_t, uint32_t, uint64_t, svbool_t, svint16_t);
328	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_u32_m), arm_streaming, arm_shared_za))
329	void svwrite_ver_za32_u32_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint32_t);
330	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_f32_m), arm_streaming, arm_shared_za))
331	void svwrite_ver_za32_f32_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat32_t);
332	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_s32_m), arm_streaming, arm_shared_za))
333	void svwrite_ver_za32_s32_m(uint64_t, uint32_t, uint64_t, svbool_t, svint32_t);
334	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_u64_m), arm_streaming, arm_shared_za))
335	void svwrite_ver_za64_u64_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint64_t);
336	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_f64_m), arm_streaming, arm_shared_za))
337	void svwrite_ver_za64_f64_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat64_t);
338	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_s64_m), arm_streaming, arm_shared_za))
339	void svwrite_ver_za64_s64_m(uint64_t, uint32_t, uint64_t, svbool_t, svint64_t);
340	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za8_u8_m), arm_streaming, arm_shared_za))
341	void svwrite_ver_za8_u8_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint8_t);
342	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za8_s8_m), arm_streaming, arm_shared_za))
343	void svwrite_ver_za8_s8_m(uint64_t, uint32_t, uint64_t, svbool_t, svint8_t);
344	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svzero_mask_za), arm_streaming_compatible, arm_shared_za))
345	void svzero_mask_za(uint64_t);
346	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svzero_za), arm_streaming_compatible, arm_shared_za))
347	void svzero_za();
348	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddha_za32_u32_m), arm_streaming, arm_shared_za))
349	void svaddha_za32_m(uint64_t, svbool_t, svbool_t, svuint32_t);
350	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddha_za32_s32_m), arm_streaming, arm_shared_za))
351	void svaddha_za32_m(uint64_t, svbool_t, svbool_t, svint32_t);
352	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddva_za32_u32_m), arm_streaming, arm_shared_za))
353	void svaddva_za32_m(uint64_t, svbool_t, svbool_t, svuint32_t);
354	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddva_za32_s32_m), arm_streaming, arm_shared_za))
355	void svaddva_za32_m(uint64_t, svbool_t, svbool_t, svint32_t);
356	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_f16_m), arm_streaming, arm_shared_za))
357	void svmopa_za32_m(uint64_t, svbool_t, svbool_t, svfloat16_t, svfloat16_t);
358	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_bf16_m), arm_streaming, arm_shared_za))
359	void svmopa_za32_m(uint64_t, svbool_t, svbool_t, svbfloat16_t, svbfloat16_t);
360	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_f32_m), arm_streaming, arm_shared_za))
361	void svmopa_za32_m(uint64_t, svbool_t, svbool_t, svfloat32_t, svfloat32_t);
362	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_s8_m), arm_streaming, arm_shared_za))
363	void svmopa_za32_m(uint64_t, svbool_t, svbool_t, svint8_t, svint8_t);
364	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za32_u8_m), arm_streaming, arm_shared_za))
365	void svmopa_za32_m(uint64_t, svbool_t, svbool_t, svuint8_t, svuint8_t);
366	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_f16_m), arm_streaming, arm_shared_za))
367	void svmops_za32_m(uint64_t, svbool_t, svbool_t, svfloat16_t, svfloat16_t);
368	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_bf16_m), arm_streaming, arm_shared_za))
369	void svmops_za32_m(uint64_t, svbool_t, svbool_t, svbfloat16_t, svbfloat16_t);
370	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_f32_m), arm_streaming, arm_shared_za))
371	void svmops_za32_m(uint64_t, svbool_t, svbool_t, svfloat32_t, svfloat32_t);
372	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_s8_m), arm_streaming, arm_shared_za))
373	void svmops_za32_m(uint64_t, svbool_t, svbool_t, svint8_t, svint8_t);
374	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za32_u8_m), arm_streaming, arm_shared_za))
375	void svmops_za32_m(uint64_t, svbool_t, svbool_t, svuint8_t, svuint8_t);
376	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_u8_m), arm_streaming, arm_shared_za, arm_preserves_za))
377	svuint8_t svread_hor_za128_m(svuint8_t, svbool_t, uint64_t, uint32_t, uint64_t);
378	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_u32_m), arm_streaming, arm_shared_za, arm_preserves_za))
379	svuint32_t svread_hor_za128_m(svuint32_t, svbool_t, uint64_t, uint32_t, uint64_t);
380	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_u64_m), arm_streaming, arm_shared_za, arm_preserves_za))
381	svuint64_t svread_hor_za128_m(svuint64_t, svbool_t, uint64_t, uint32_t, uint64_t);
382	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_u16_m), arm_streaming, arm_shared_za, arm_preserves_za))
383	svuint16_t svread_hor_za128_m(svuint16_t, svbool_t, uint64_t, uint32_t, uint64_t);
384	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_bf16_m), arm_streaming, arm_shared_za, arm_preserves_za))
385	svbfloat16_t svread_hor_za128_m(svbfloat16_t, svbool_t, uint64_t, uint32_t, uint64_t);
386	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_s8_m), arm_streaming, arm_shared_za, arm_preserves_za))
387	svint8_t svread_hor_za128_m(svint8_t, svbool_t, uint64_t, uint32_t, uint64_t);
388	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_f64_m), arm_streaming, arm_shared_za, arm_preserves_za))
389	svfloat64_t svread_hor_za128_m(svfloat64_t, svbool_t, uint64_t, uint32_t, uint64_t);
390	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_f32_m), arm_streaming, arm_shared_za, arm_preserves_za))
391	svfloat32_t svread_hor_za128_m(svfloat32_t, svbool_t, uint64_t, uint32_t, uint64_t);
392	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_f16_m), arm_streaming, arm_shared_za, arm_preserves_za))
393	svfloat16_t svread_hor_za128_m(svfloat16_t, svbool_t, uint64_t, uint32_t, uint64_t);
394	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_s32_m), arm_streaming, arm_shared_za, arm_preserves_za))
395	svint32_t svread_hor_za128_m(svint32_t, svbool_t, uint64_t, uint32_t, uint64_t);
396	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_s64_m), arm_streaming, arm_shared_za, arm_preserves_za))
397	svint64_t svread_hor_za128_m(svint64_t, svbool_t, uint64_t, uint32_t, uint64_t);
398	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za128_s16_m), arm_streaming, arm_shared_za, arm_preserves_za))
399	svint16_t svread_hor_za128_m(svint16_t, svbool_t, uint64_t, uint32_t, uint64_t);
400	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_u16_m), arm_streaming, arm_shared_za, arm_preserves_za))
401	svuint16_t svread_hor_za16_m(svuint16_t, svbool_t, uint64_t, uint32_t, uint64_t);
402	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_bf16_m), arm_streaming, arm_shared_za, arm_preserves_za))
403	svbfloat16_t svread_hor_za16_m(svbfloat16_t, svbool_t, uint64_t, uint32_t, uint64_t);
404	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_f16_m), arm_streaming, arm_shared_za, arm_preserves_za))
405	svfloat16_t svread_hor_za16_m(svfloat16_t, svbool_t, uint64_t, uint32_t, uint64_t);
406	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za16_s16_m), arm_streaming, arm_shared_za, arm_preserves_za))
407	svint16_t svread_hor_za16_m(svint16_t, svbool_t, uint64_t, uint32_t, uint64_t);
408	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_u32_m), arm_streaming, arm_shared_za, arm_preserves_za))
409	svuint32_t svread_hor_za32_m(svuint32_t, svbool_t, uint64_t, uint32_t, uint64_t);
410	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_f32_m), arm_streaming, arm_shared_za, arm_preserves_za))
411	svfloat32_t svread_hor_za32_m(svfloat32_t, svbool_t, uint64_t, uint32_t, uint64_t);
412	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za32_s32_m), arm_streaming, arm_shared_za, arm_preserves_za))
413	svint32_t svread_hor_za32_m(svint32_t, svbool_t, uint64_t, uint32_t, uint64_t);
414	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_u64_m), arm_streaming, arm_shared_za, arm_preserves_za))
415	svuint64_t svread_hor_za64_m(svuint64_t, svbool_t, uint64_t, uint32_t, uint64_t);
416	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_f64_m), arm_streaming, arm_shared_za, arm_preserves_za))
417	svfloat64_t svread_hor_za64_m(svfloat64_t, svbool_t, uint64_t, uint32_t, uint64_t);
418	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za64_s64_m), arm_streaming, arm_shared_za, arm_preserves_za))
419	svint64_t svread_hor_za64_m(svint64_t, svbool_t, uint64_t, uint32_t, uint64_t);
420	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za8_u8_m), arm_streaming, arm_shared_za, arm_preserves_za))
421	svuint8_t svread_hor_za8_m(svuint8_t, svbool_t, uint64_t, uint32_t, uint64_t);
422	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_hor_za8_s8_m), arm_streaming, arm_shared_za, arm_preserves_za))
423	svint8_t svread_hor_za8_m(svint8_t, svbool_t, uint64_t, uint32_t, uint64_t);
424	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_u8_m), arm_streaming, arm_shared_za, arm_preserves_za))
425	svuint8_t svread_ver_za128_m(svuint8_t, svbool_t, uint64_t, uint32_t, uint64_t);
426	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_u32_m), arm_streaming, arm_shared_za, arm_preserves_za))
427	svuint32_t svread_ver_za128_m(svuint32_t, svbool_t, uint64_t, uint32_t, uint64_t);
428	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_u64_m), arm_streaming, arm_shared_za, arm_preserves_za))
429	svuint64_t svread_ver_za128_m(svuint64_t, svbool_t, uint64_t, uint32_t, uint64_t);
430	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_u16_m), arm_streaming, arm_shared_za, arm_preserves_za))
431	svuint16_t svread_ver_za128_m(svuint16_t, svbool_t, uint64_t, uint32_t, uint64_t);
432	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_bf16_m), arm_streaming, arm_shared_za, arm_preserves_za))
433	svbfloat16_t svread_ver_za128_m(svbfloat16_t, svbool_t, uint64_t, uint32_t, uint64_t);
434	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_s8_m), arm_streaming, arm_shared_za, arm_preserves_za))
435	svint8_t svread_ver_za128_m(svint8_t, svbool_t, uint64_t, uint32_t, uint64_t);
436	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_f64_m), arm_streaming, arm_shared_za, arm_preserves_za))
437	svfloat64_t svread_ver_za128_m(svfloat64_t, svbool_t, uint64_t, uint32_t, uint64_t);
438	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_f32_m), arm_streaming, arm_shared_za, arm_preserves_za))
439	svfloat32_t svread_ver_za128_m(svfloat32_t, svbool_t, uint64_t, uint32_t, uint64_t);
440	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_f16_m), arm_streaming, arm_shared_za, arm_preserves_za))
441	svfloat16_t svread_ver_za128_m(svfloat16_t, svbool_t, uint64_t, uint32_t, uint64_t);
442	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_s32_m), arm_streaming, arm_shared_za, arm_preserves_za))
443	svint32_t svread_ver_za128_m(svint32_t, svbool_t, uint64_t, uint32_t, uint64_t);
444	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_s64_m), arm_streaming, arm_shared_za, arm_preserves_za))
445	svint64_t svread_ver_za128_m(svint64_t, svbool_t, uint64_t, uint32_t, uint64_t);
446	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za128_s16_m), arm_streaming, arm_shared_za, arm_preserves_za))
447	svint16_t svread_ver_za128_m(svint16_t, svbool_t, uint64_t, uint32_t, uint64_t);
448	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_u16_m), arm_streaming, arm_shared_za, arm_preserves_za))
449	svuint16_t svread_ver_za16_m(svuint16_t, svbool_t, uint64_t, uint32_t, uint64_t);
450	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_bf16_m), arm_streaming, arm_shared_za, arm_preserves_za))
451	svbfloat16_t svread_ver_za16_m(svbfloat16_t, svbool_t, uint64_t, uint32_t, uint64_t);
452	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_f16_m), arm_streaming, arm_shared_za, arm_preserves_za))
453	svfloat16_t svread_ver_za16_m(svfloat16_t, svbool_t, uint64_t, uint32_t, uint64_t);
454	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za16_s16_m), arm_streaming, arm_shared_za, arm_preserves_za))
455	svint16_t svread_ver_za16_m(svint16_t, svbool_t, uint64_t, uint32_t, uint64_t);
456	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_u32_m), arm_streaming, arm_shared_za, arm_preserves_za))
457	svuint32_t svread_ver_za32_m(svuint32_t, svbool_t, uint64_t, uint32_t, uint64_t);
458	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_f32_m), arm_streaming, arm_shared_za, arm_preserves_za))
459	svfloat32_t svread_ver_za32_m(svfloat32_t, svbool_t, uint64_t, uint32_t, uint64_t);
460	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za32_s32_m), arm_streaming, arm_shared_za, arm_preserves_za))
461	svint32_t svread_ver_za32_m(svint32_t, svbool_t, uint64_t, uint32_t, uint64_t);
462	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_u64_m), arm_streaming, arm_shared_za, arm_preserves_za))
463	svuint64_t svread_ver_za64_m(svuint64_t, svbool_t, uint64_t, uint32_t, uint64_t);
464	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_f64_m), arm_streaming, arm_shared_za, arm_preserves_za))
465	svfloat64_t svread_ver_za64_m(svfloat64_t, svbool_t, uint64_t, uint32_t, uint64_t);
466	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za64_s64_m), arm_streaming, arm_shared_za, arm_preserves_za))
467	svint64_t svread_ver_za64_m(svint64_t, svbool_t, uint64_t, uint32_t, uint64_t);
468	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za8_u8_m), arm_streaming, arm_shared_za, arm_preserves_za))
469	svuint8_t svread_ver_za8_m(svuint8_t, svbool_t, uint64_t, uint32_t, uint64_t);
470	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svread_ver_za8_s8_m), arm_streaming, arm_shared_za, arm_preserves_za))
471	svint8_t svread_ver_za8_m(svint8_t, svbool_t, uint64_t, uint32_t, uint64_t);
472	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumopa_za32_s8_m), arm_streaming, arm_shared_za))
473	void svsumopa_za32_m(uint64_t, svbool_t, svbool_t, svint8_t, svuint8_t);
474	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumops_za32_s8_m), arm_streaming, arm_shared_za))
475	void svsumops_za32_m(uint64_t, svbool_t, svbool_t, svint8_t, svuint8_t);
476	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmopa_za32_u8_m), arm_streaming, arm_shared_za))
477	void svusmopa_za32_m(uint64_t, svbool_t, svbool_t, svuint8_t, svint8_t);
478	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmops_za32_u8_m), arm_streaming, arm_shared_za))
479	void svusmops_za32_m(uint64_t, svbool_t, svbool_t, svuint8_t, svint8_t);
480	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_u8_m), arm_streaming, arm_shared_za))
481	void svwrite_hor_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint8_t);
482	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_u32_m), arm_streaming, arm_shared_za))
483	void svwrite_hor_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint32_t);
484	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_u64_m), arm_streaming, arm_shared_za))
485	void svwrite_hor_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint64_t);
486	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_u16_m), arm_streaming, arm_shared_za))
487	void svwrite_hor_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint16_t);
488	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_bf16_m), arm_streaming, arm_shared_za))
489	void svwrite_hor_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svbfloat16_t);
490	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_s8_m), arm_streaming, arm_shared_za))
491	void svwrite_hor_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svint8_t);
492	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_f64_m), arm_streaming, arm_shared_za))
493	void svwrite_hor_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat64_t);
494	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_f32_m), arm_streaming, arm_shared_za))
495	void svwrite_hor_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat32_t);
496	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_f16_m), arm_streaming, arm_shared_za))
497	void svwrite_hor_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat16_t);
498	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_s32_m), arm_streaming, arm_shared_za))
499	void svwrite_hor_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svint32_t);
500	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_s64_m), arm_streaming, arm_shared_za))
501	void svwrite_hor_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svint64_t);
502	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za128_s16_m), arm_streaming, arm_shared_za))
503	void svwrite_hor_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svint16_t);
504	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_u16_m), arm_streaming, arm_shared_za))
505	void svwrite_hor_za16_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint16_t);
506	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_bf16_m), arm_streaming, arm_shared_za))
507	void svwrite_hor_za16_m(uint64_t, uint32_t, uint64_t, svbool_t, svbfloat16_t);
508	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_f16_m), arm_streaming, arm_shared_za))
509	void svwrite_hor_za16_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat16_t);
510	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za16_s16_m), arm_streaming, arm_shared_za))
511	void svwrite_hor_za16_m(uint64_t, uint32_t, uint64_t, svbool_t, svint16_t);
512	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_u32_m), arm_streaming, arm_shared_za))
513	void svwrite_hor_za32_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint32_t);
514	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_f32_m), arm_streaming, arm_shared_za))
515	void svwrite_hor_za32_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat32_t);
516	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za32_s32_m), arm_streaming, arm_shared_za))
517	void svwrite_hor_za32_m(uint64_t, uint32_t, uint64_t, svbool_t, svint32_t);
518	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_u64_m), arm_streaming, arm_shared_za))
519	void svwrite_hor_za64_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint64_t);
520	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_f64_m), arm_streaming, arm_shared_za))
521	void svwrite_hor_za64_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat64_t);
522	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za64_s64_m), arm_streaming, arm_shared_za))
523	void svwrite_hor_za64_m(uint64_t, uint32_t, uint64_t, svbool_t, svint64_t);
524	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za8_u8_m), arm_streaming, arm_shared_za))
525	void svwrite_hor_za8_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint8_t);
526	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_hor_za8_s8_m), arm_streaming, arm_shared_za))
527	void svwrite_hor_za8_m(uint64_t, uint32_t, uint64_t, svbool_t, svint8_t);
528	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_u8_m), arm_streaming, arm_shared_za))
529	void svwrite_ver_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint8_t);
530	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_u32_m), arm_streaming, arm_shared_za))
531	void svwrite_ver_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint32_t);
532	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_u64_m), arm_streaming, arm_shared_za))
533	void svwrite_ver_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint64_t);
534	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_u16_m), arm_streaming, arm_shared_za))
535	void svwrite_ver_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint16_t);
536	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_bf16_m), arm_streaming, arm_shared_za))
537	void svwrite_ver_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svbfloat16_t);
538	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_s8_m), arm_streaming, arm_shared_za))
539	void svwrite_ver_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svint8_t);
540	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_f64_m), arm_streaming, arm_shared_za))
541	void svwrite_ver_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat64_t);
542	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_f32_m), arm_streaming, arm_shared_za))
543	void svwrite_ver_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat32_t);
544	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_f16_m), arm_streaming, arm_shared_za))
545	void svwrite_ver_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat16_t);
546	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_s32_m), arm_streaming, arm_shared_za))
547	void svwrite_ver_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svint32_t);
548	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_s64_m), arm_streaming, arm_shared_za))
549	void svwrite_ver_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svint64_t);
550	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za128_s16_m), arm_streaming, arm_shared_za))
551	void svwrite_ver_za128_m(uint64_t, uint32_t, uint64_t, svbool_t, svint16_t);
552	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_u16_m), arm_streaming, arm_shared_za))
553	void svwrite_ver_za16_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint16_t);
554	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_bf16_m), arm_streaming, arm_shared_za))
555	void svwrite_ver_za16_m(uint64_t, uint32_t, uint64_t, svbool_t, svbfloat16_t);
556	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_f16_m), arm_streaming, arm_shared_za))
557	void svwrite_ver_za16_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat16_t);
558	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za16_s16_m), arm_streaming, arm_shared_za))
559	void svwrite_ver_za16_m(uint64_t, uint32_t, uint64_t, svbool_t, svint16_t);
560	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_u32_m), arm_streaming, arm_shared_za))
561	void svwrite_ver_za32_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint32_t);
562	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_f32_m), arm_streaming, arm_shared_za))
563	void svwrite_ver_za32_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat32_t);
564	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za32_s32_m), arm_streaming, arm_shared_za))
565	void svwrite_ver_za32_m(uint64_t, uint32_t, uint64_t, svbool_t, svint32_t);
566	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_u64_m), arm_streaming, arm_shared_za))
567	void svwrite_ver_za64_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint64_t);
568	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_f64_m), arm_streaming, arm_shared_za))
569	void svwrite_ver_za64_m(uint64_t, uint32_t, uint64_t, svbool_t, svfloat64_t);
570	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za64_s64_m), arm_streaming, arm_shared_za))
571	void svwrite_ver_za64_m(uint64_t, uint32_t, uint64_t, svbool_t, svint64_t);
572	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za8_u8_m), arm_streaming, arm_shared_za))
573	void svwrite_ver_za8_m(uint64_t, uint32_t, uint64_t, svbool_t, svuint8_t);
574	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svwrite_ver_za8_s8_m), arm_streaming, arm_shared_za))
575	void svwrite_ver_za8_m(uint64_t, uint32_t, uint64_t, svbool_t, svint8_t);
576	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za64_f64_m), arm_streaming, arm_shared_za))
577	void svmopa_za64_f64_m(uint64_t, svbool_t, svbool_t, svfloat64_t, svfloat64_t);
578	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za64_f64_m), arm_streaming, arm_shared_za))
579	void svmops_za64_f64_m(uint64_t, svbool_t, svbool_t, svfloat64_t, svfloat64_t);
580	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za64_f64_m), arm_streaming, arm_shared_za))
581	void svmopa_za64_m(uint64_t, svbool_t, svbool_t, svfloat64_t, svfloat64_t);
582	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za64_f64_m), arm_streaming, arm_shared_za))
583	void svmops_za64_m(uint64_t, svbool_t, svbool_t, svfloat64_t, svfloat64_t);
584	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddha_za64_u64_m), arm_streaming, arm_shared_za))
585	void svaddha_za64_u64_m(uint64_t, svbool_t, svbool_t, svuint64_t);
586	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddha_za64_s64_m), arm_streaming, arm_shared_za))
587	void svaddha_za64_s64_m(uint64_t, svbool_t, svbool_t, svint64_t);
588	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddva_za64_u64_m), arm_streaming, arm_shared_za))
589	void svaddva_za64_u64_m(uint64_t, svbool_t, svbool_t, svuint64_t);
590	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddva_za64_s64_m), arm_streaming, arm_shared_za))
591	void svaddva_za64_s64_m(uint64_t, svbool_t, svbool_t, svint64_t);
592	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za64_s16_m), arm_streaming, arm_shared_za))
593	void svmopa_za64_s16_m(uint64_t, svbool_t, svbool_t, svint16_t, svint16_t);
594	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za64_u16_m), arm_streaming, arm_shared_za))
595	void svmopa_za64_u16_m(uint64_t, svbool_t, svbool_t, svuint16_t, svuint16_t);
596	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za64_s16_m), arm_streaming, arm_shared_za))
597	void svmops_za64_s16_m(uint64_t, svbool_t, svbool_t, svint16_t, svint16_t);
598	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za64_u16_m), arm_streaming, arm_shared_za))
599	void svmops_za64_u16_m(uint64_t, svbool_t, svbool_t, svuint16_t, svuint16_t);
600	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumopa_za64_s16_m), arm_streaming, arm_shared_za))
601	void svsumopa_za64_s16_m(uint64_t, svbool_t, svbool_t, svint16_t, svuint16_t);
602	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumops_za64_s16_m), arm_streaming, arm_shared_za))
603	void svsumops_za64_s16_m(uint64_t, svbool_t, svbool_t, svint16_t, svuint16_t);
604	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmopa_za64_u16_m), arm_streaming, arm_shared_za))
605	void svusmopa_za64_u16_m(uint64_t, svbool_t, svbool_t, svuint16_t, svint16_t);
606	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmops_za64_u16_m), arm_streaming, arm_shared_za))
607	void svusmops_za64_u16_m(uint64_t, svbool_t, svbool_t, svuint16_t, svint16_t);
608	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddha_za64_u64_m), arm_streaming, arm_shared_za))
609	void svaddha_za64_m(uint64_t, svbool_t, svbool_t, svuint64_t);
610	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddha_za64_s64_m), arm_streaming, arm_shared_za))
611	void svaddha_za64_m(uint64_t, svbool_t, svbool_t, svint64_t);
612	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddva_za64_u64_m), arm_streaming, arm_shared_za))
613	void svaddva_za64_m(uint64_t, svbool_t, svbool_t, svuint64_t);
614	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svaddva_za64_s64_m), arm_streaming, arm_shared_za))
615	void svaddva_za64_m(uint64_t, svbool_t, svbool_t, svint64_t);
616	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za64_s16_m), arm_streaming, arm_shared_za))
617	void svmopa_za64_m(uint64_t, svbool_t, svbool_t, svint16_t, svint16_t);
618	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmopa_za64_u16_m), arm_streaming, arm_shared_za))
619	void svmopa_za64_m(uint64_t, svbool_t, svbool_t, svuint16_t, svuint16_t);
620	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za64_s16_m), arm_streaming, arm_shared_za))
621	void svmops_za64_m(uint64_t, svbool_t, svbool_t, svint16_t, svint16_t);
622	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svmops_za64_u16_m), arm_streaming, arm_shared_za))
623	void svmops_za64_m(uint64_t, svbool_t, svbool_t, svuint16_t, svuint16_t);
624	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumopa_za64_s16_m), arm_streaming, arm_shared_za))
625	void svsumopa_za64_m(uint64_t, svbool_t, svbool_t, svint16_t, svuint16_t);
626	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svsumops_za64_s16_m), arm_streaming, arm_shared_za))
627	void svsumops_za64_m(uint64_t, svbool_t, svbool_t, svint16_t, svuint16_t);
628	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmopa_za64_u16_m), arm_streaming, arm_shared_za))
629	void svusmopa_za64_m(uint64_t, svbool_t, svbool_t, svuint16_t, svint16_t);
630	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sme_svusmops_za64_u16_m), arm_streaming, arm_shared_za))
631	void svusmops_za64_m(uint64_t, svbool_t, svbool_t, svuint16_t, svint16_t);
632	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svldr_vnum_za), arm_streaming_compatible, arm_shared_za))
633	void svldr_vnum_za(uint32_t, uint64_t, void const *);
634	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sme_svstr_vnum_za), arm_streaming_compatible, arm_shared_za, arm_preserves_za))
635	void svstr_vnum_za(uint32_t, uint64_t, void *);
636	#ifdef __cplusplus
637	} // extern "C"
638	#endif
639
640	#undef __ai
641
642	#endif /* __ARM_SME_H */

lib/include/arm_sve.h+7386-351

...	@@ -35,8 +35,9 @@ typedef __SVUint32_t svuint32_t;	...	@@ -35,8 +35,9 @@ typedef __SVUint32_t svuint32_t;
35	typedef __SVUint64_t svuint64_t;	35	typedef __SVUint64_t svuint64_t;
36	typedef __SVFloat16_t svfloat16_t;	36	typedef __SVFloat16_t svfloat16_t;
37		37
38	typedef __SVBFloat16_t svbfloat16_t;	38	typedef __SVBfloat16_t svbfloat16_t;
39	#include <arm_bf16.h>	39	#include <arm_bf16.h>
		40	#include <arm_vector_types.h>
40	typedef __SVFloat32_t svfloat32_t;	41	typedef __SVFloat32_t svfloat32_t;
41	typedef __SVFloat64_t svfloat64_t;	42	typedef __SVFloat64_t svfloat64_t;
42	typedef __clang_svint8x2_t svint8x2_t;	43	typedef __clang_svint8x2_t svint8x2_t;
...	@@ -124,725 +125,4657 @@ enum svprfop	...	@@ -124,725 +125,4657 @@ enum svprfop
124	#define __aio static __inline__ __attribute__((__always_inline__, __nodebug__, __overloadable__))	125	#define __aio static __inline__ __attribute__((__always_inline__, __nodebug__, __overloadable__))
125		126
126	#define svreinterpret_s8_s8(...) __builtin_sve_reinterpret_s8_s8(__VA_ARGS__)	127	#define svreinterpret_s8_s8(...) __builtin_sve_reinterpret_s8_s8(__VA_ARGS__)
127	#define svreinterpret_s8_s16(...) __builtin_sve_reinterpret_s8_s16(__VA_ARGS__)
128	#define svreinterpret_s8_s32(...) __builtin_sve_reinterpret_s8_s32(__VA_ARGS__)
129	#define svreinterpret_s8_s64(...) __builtin_sve_reinterpret_s8_s64(__VA_ARGS__)
130	#define svreinterpret_s8_u8(...) __builtin_sve_reinterpret_s8_u8(__VA_ARGS__)	128	#define svreinterpret_s8_u8(...) __builtin_sve_reinterpret_s8_u8(__VA_ARGS__)
		129	#define svreinterpret_s8_s16(...) __builtin_sve_reinterpret_s8_s16(__VA_ARGS__)
131	#define svreinterpret_s8_u16(...) __builtin_sve_reinterpret_s8_u16(__VA_ARGS__)	130	#define svreinterpret_s8_u16(...) __builtin_sve_reinterpret_s8_u16(__VA_ARGS__)
		131	#define svreinterpret_s8_s32(...) __builtin_sve_reinterpret_s8_s32(__VA_ARGS__)
132	#define svreinterpret_s8_u32(...) __builtin_sve_reinterpret_s8_u32(__VA_ARGS__)	132	#define svreinterpret_s8_u32(...) __builtin_sve_reinterpret_s8_u32(__VA_ARGS__)
		133	#define svreinterpret_s8_s64(...) __builtin_sve_reinterpret_s8_s64(__VA_ARGS__)
133	#define svreinterpret_s8_u64(...) __builtin_sve_reinterpret_s8_u64(__VA_ARGS__)	134	#define svreinterpret_s8_u64(...) __builtin_sve_reinterpret_s8_u64(__VA_ARGS__)
134	#define svreinterpret_s8_f16(...) __builtin_sve_reinterpret_s8_f16(__VA_ARGS__)	135	#define svreinterpret_s8_f16(...) __builtin_sve_reinterpret_s8_f16(__VA_ARGS__)
135	#define svreinterpret_s8_bf16(...) __builtin_sve_reinterpret_s8_bf16(__VA_ARGS__)	136	#define svreinterpret_s8_bf16(...) __builtin_sve_reinterpret_s8_bf16(__VA_ARGS__)
136	#define svreinterpret_s8_f32(...) __builtin_sve_reinterpret_s8_f32(__VA_ARGS__)	137	#define svreinterpret_s8_f32(...) __builtin_sve_reinterpret_s8_f32(__VA_ARGS__)
137	#define svreinterpret_s8_f64(...) __builtin_sve_reinterpret_s8_f64(__VA_ARGS__)	138	#define svreinterpret_s8_f64(...) __builtin_sve_reinterpret_s8_f64(__VA_ARGS__)
138	#define svreinterpret_s16_s8(...) __builtin_sve_reinterpret_s16_s8(__VA_ARGS__)
139	#define svreinterpret_s16_s16(...) __builtin_sve_reinterpret_s16_s16(__VA_ARGS__)
140	#define svreinterpret_s16_s32(...) __builtin_sve_reinterpret_s16_s32(__VA_ARGS__)
141	#define svreinterpret_s16_s64(...) __builtin_sve_reinterpret_s16_s64(__VA_ARGS__)
142	#define svreinterpret_s16_u8(...) __builtin_sve_reinterpret_s16_u8(__VA_ARGS__)
143	#define svreinterpret_s16_u16(...) __builtin_sve_reinterpret_s16_u16(__VA_ARGS__)
144	#define svreinterpret_s16_u32(...) __builtin_sve_reinterpret_s16_u32(__VA_ARGS__)
145	#define svreinterpret_s16_u64(...) __builtin_sve_reinterpret_s16_u64(__VA_ARGS__)
146	#define svreinterpret_s16_f16(...) __builtin_sve_reinterpret_s16_f16(__VA_ARGS__)
147	#define svreinterpret_s16_bf16(...) __builtin_sve_reinterpret_s16_bf16(__VA_ARGS__)
148	#define svreinterpret_s16_f32(...) __builtin_sve_reinterpret_s16_f32(__VA_ARGS__)
149	#define svreinterpret_s16_f64(...) __builtin_sve_reinterpret_s16_f64(__VA_ARGS__)
150	#define svreinterpret_s32_s8(...) __builtin_sve_reinterpret_s32_s8(__VA_ARGS__)
151	#define svreinterpret_s32_s16(...) __builtin_sve_reinterpret_s32_s16(__VA_ARGS__)
152	#define svreinterpret_s32_s32(...) __builtin_sve_reinterpret_s32_s32(__VA_ARGS__)
153	#define svreinterpret_s32_s64(...) __builtin_sve_reinterpret_s32_s64(__VA_ARGS__)
154	#define svreinterpret_s32_u8(...) __builtin_sve_reinterpret_s32_u8(__VA_ARGS__)
155	#define svreinterpret_s32_u16(...) __builtin_sve_reinterpret_s32_u16(__VA_ARGS__)
156	#define svreinterpret_s32_u32(...) __builtin_sve_reinterpret_s32_u32(__VA_ARGS__)
157	#define svreinterpret_s32_u64(...) __builtin_sve_reinterpret_s32_u64(__VA_ARGS__)
158	#define svreinterpret_s32_f16(...) __builtin_sve_reinterpret_s32_f16(__VA_ARGS__)
159	#define svreinterpret_s32_bf16(...) __builtin_sve_reinterpret_s32_bf16(__VA_ARGS__)
160	#define svreinterpret_s32_f32(...) __builtin_sve_reinterpret_s32_f32(__VA_ARGS__)
161	#define svreinterpret_s32_f64(...) __builtin_sve_reinterpret_s32_f64(__VA_ARGS__)
162	#define svreinterpret_s64_s8(...) __builtin_sve_reinterpret_s64_s8(__VA_ARGS__)
163	#define svreinterpret_s64_s16(...) __builtin_sve_reinterpret_s64_s16(__VA_ARGS__)
164	#define svreinterpret_s64_s32(...) __builtin_sve_reinterpret_s64_s32(__VA_ARGS__)
165	#define svreinterpret_s64_s64(...) __builtin_sve_reinterpret_s64_s64(__VA_ARGS__)
166	#define svreinterpret_s64_u8(...) __builtin_sve_reinterpret_s64_u8(__VA_ARGS__)
167	#define svreinterpret_s64_u16(...) __builtin_sve_reinterpret_s64_u16(__VA_ARGS__)
168	#define svreinterpret_s64_u32(...) __builtin_sve_reinterpret_s64_u32(__VA_ARGS__)
169	#define svreinterpret_s64_u64(...) __builtin_sve_reinterpret_s64_u64(__VA_ARGS__)
170	#define svreinterpret_s64_f16(...) __builtin_sve_reinterpret_s64_f16(__VA_ARGS__)
171	#define svreinterpret_s64_bf16(...) __builtin_sve_reinterpret_s64_bf16(__VA_ARGS__)
172	#define svreinterpret_s64_f32(...) __builtin_sve_reinterpret_s64_f32(__VA_ARGS__)
173	#define svreinterpret_s64_f64(...) __builtin_sve_reinterpret_s64_f64(__VA_ARGS__)
174	#define svreinterpret_u8_s8(...) __builtin_sve_reinterpret_u8_s8(__VA_ARGS__)	139	#define svreinterpret_u8_s8(...) __builtin_sve_reinterpret_u8_s8(__VA_ARGS__)
175	#define svreinterpret_u8_s16(...) __builtin_sve_reinterpret_u8_s16(__VA_ARGS__)
176	#define svreinterpret_u8_s32(...) __builtin_sve_reinterpret_u8_s32(__VA_ARGS__)
177	#define svreinterpret_u8_s64(...) __builtin_sve_reinterpret_u8_s64(__VA_ARGS__)
178	#define svreinterpret_u8_u8(...) __builtin_sve_reinterpret_u8_u8(__VA_ARGS__)	140	#define svreinterpret_u8_u8(...) __builtin_sve_reinterpret_u8_u8(__VA_ARGS__)
		141	#define svreinterpret_u8_s16(...) __builtin_sve_reinterpret_u8_s16(__VA_ARGS__)
179	#define svreinterpret_u8_u16(...) __builtin_sve_reinterpret_u8_u16(__VA_ARGS__)	142	#define svreinterpret_u8_u16(...) __builtin_sve_reinterpret_u8_u16(__VA_ARGS__)
		143	#define svreinterpret_u8_s32(...) __builtin_sve_reinterpret_u8_s32(__VA_ARGS__)
180	#define svreinterpret_u8_u32(...) __builtin_sve_reinterpret_u8_u32(__VA_ARGS__)	144	#define svreinterpret_u8_u32(...) __builtin_sve_reinterpret_u8_u32(__VA_ARGS__)
		145	#define svreinterpret_u8_s64(...) __builtin_sve_reinterpret_u8_s64(__VA_ARGS__)
181	#define svreinterpret_u8_u64(...) __builtin_sve_reinterpret_u8_u64(__VA_ARGS__)	146	#define svreinterpret_u8_u64(...) __builtin_sve_reinterpret_u8_u64(__VA_ARGS__)
182	#define svreinterpret_u8_f16(...) __builtin_sve_reinterpret_u8_f16(__VA_ARGS__)	147	#define svreinterpret_u8_f16(...) __builtin_sve_reinterpret_u8_f16(__VA_ARGS__)
183	#define svreinterpret_u8_bf16(...) __builtin_sve_reinterpret_u8_bf16(__VA_ARGS__)	148	#define svreinterpret_u8_bf16(...) __builtin_sve_reinterpret_u8_bf16(__VA_ARGS__)
184	#define svreinterpret_u8_f32(...) __builtin_sve_reinterpret_u8_f32(__VA_ARGS__)	149	#define svreinterpret_u8_f32(...) __builtin_sve_reinterpret_u8_f32(__VA_ARGS__)
185	#define svreinterpret_u8_f64(...) __builtin_sve_reinterpret_u8_f64(__VA_ARGS__)	150	#define svreinterpret_u8_f64(...) __builtin_sve_reinterpret_u8_f64(__VA_ARGS__)
		151	#define svreinterpret_s16_s8(...) __builtin_sve_reinterpret_s16_s8(__VA_ARGS__)
		152	#define svreinterpret_s16_u8(...) __builtin_sve_reinterpret_s16_u8(__VA_ARGS__)
		153	#define svreinterpret_s16_s16(...) __builtin_sve_reinterpret_s16_s16(__VA_ARGS__)
		154	#define svreinterpret_s16_u16(...) __builtin_sve_reinterpret_s16_u16(__VA_ARGS__)
		155	#define svreinterpret_s16_s32(...) __builtin_sve_reinterpret_s16_s32(__VA_ARGS__)
		156	#define svreinterpret_s16_u32(...) __builtin_sve_reinterpret_s16_u32(__VA_ARGS__)
		157	#define svreinterpret_s16_s64(...) __builtin_sve_reinterpret_s16_s64(__VA_ARGS__)
		158	#define svreinterpret_s16_u64(...) __builtin_sve_reinterpret_s16_u64(__VA_ARGS__)
		159	#define svreinterpret_s16_f16(...) __builtin_sve_reinterpret_s16_f16(__VA_ARGS__)
		160	#define svreinterpret_s16_bf16(...) __builtin_sve_reinterpret_s16_bf16(__VA_ARGS__)
		161	#define svreinterpret_s16_f32(...) __builtin_sve_reinterpret_s16_f32(__VA_ARGS__)
		162	#define svreinterpret_s16_f64(...) __builtin_sve_reinterpret_s16_f64(__VA_ARGS__)
186	#define svreinterpret_u16_s8(...) __builtin_sve_reinterpret_u16_s8(__VA_ARGS__)	163	#define svreinterpret_u16_s8(...) __builtin_sve_reinterpret_u16_s8(__VA_ARGS__)
187	#define svreinterpret_u16_s16(...) __builtin_sve_reinterpret_u16_s16(__VA_ARGS__)
188	#define svreinterpret_u16_s32(...) __builtin_sve_reinterpret_u16_s32(__VA_ARGS__)
189	#define svreinterpret_u16_s64(...) __builtin_sve_reinterpret_u16_s64(__VA_ARGS__)
190	#define svreinterpret_u16_u8(...) __builtin_sve_reinterpret_u16_u8(__VA_ARGS__)	164	#define svreinterpret_u16_u8(...) __builtin_sve_reinterpret_u16_u8(__VA_ARGS__)
		165	#define svreinterpret_u16_s16(...) __builtin_sve_reinterpret_u16_s16(__VA_ARGS__)
191	#define svreinterpret_u16_u16(...) __builtin_sve_reinterpret_u16_u16(__VA_ARGS__)	166	#define svreinterpret_u16_u16(...) __builtin_sve_reinterpret_u16_u16(__VA_ARGS__)
		167	#define svreinterpret_u16_s32(...) __builtin_sve_reinterpret_u16_s32(__VA_ARGS__)
192	#define svreinterpret_u16_u32(...) __builtin_sve_reinterpret_u16_u32(__VA_ARGS__)	168	#define svreinterpret_u16_u32(...) __builtin_sve_reinterpret_u16_u32(__VA_ARGS__)
		169	#define svreinterpret_u16_s64(...) __builtin_sve_reinterpret_u16_s64(__VA_ARGS__)
193	#define svreinterpret_u16_u64(...) __builtin_sve_reinterpret_u16_u64(__VA_ARGS__)	170	#define svreinterpret_u16_u64(...) __builtin_sve_reinterpret_u16_u64(__VA_ARGS__)
194	#define svreinterpret_u16_f16(...) __builtin_sve_reinterpret_u16_f16(__VA_ARGS__)	171	#define svreinterpret_u16_f16(...) __builtin_sve_reinterpret_u16_f16(__VA_ARGS__)
195	#define svreinterpret_u16_bf16(...) __builtin_sve_reinterpret_u16_bf16(__VA_ARGS__)	172	#define svreinterpret_u16_bf16(...) __builtin_sve_reinterpret_u16_bf16(__VA_ARGS__)
196	#define svreinterpret_u16_f32(...) __builtin_sve_reinterpret_u16_f32(__VA_ARGS__)	173	#define svreinterpret_u16_f32(...) __builtin_sve_reinterpret_u16_f32(__VA_ARGS__)
197	#define svreinterpret_u16_f64(...) __builtin_sve_reinterpret_u16_f64(__VA_ARGS__)	174	#define svreinterpret_u16_f64(...) __builtin_sve_reinterpret_u16_f64(__VA_ARGS__)
		175	#define svreinterpret_s32_s8(...) __builtin_sve_reinterpret_s32_s8(__VA_ARGS__)
		176	#define svreinterpret_s32_u8(...) __builtin_sve_reinterpret_s32_u8(__VA_ARGS__)
		177	#define svreinterpret_s32_s16(...) __builtin_sve_reinterpret_s32_s16(__VA_ARGS__)
		178	#define svreinterpret_s32_u16(...) __builtin_sve_reinterpret_s32_u16(__VA_ARGS__)
		179	#define svreinterpret_s32_s32(...) __builtin_sve_reinterpret_s32_s32(__VA_ARGS__)
		180	#define svreinterpret_s32_u32(...) __builtin_sve_reinterpret_s32_u32(__VA_ARGS__)
		181	#define svreinterpret_s32_s64(...) __builtin_sve_reinterpret_s32_s64(__VA_ARGS__)
		182	#define svreinterpret_s32_u64(...) __builtin_sve_reinterpret_s32_u64(__VA_ARGS__)
		183	#define svreinterpret_s32_f16(...) __builtin_sve_reinterpret_s32_f16(__VA_ARGS__)
		184	#define svreinterpret_s32_bf16(...) __builtin_sve_reinterpret_s32_bf16(__VA_ARGS__)
		185	#define svreinterpret_s32_f32(...) __builtin_sve_reinterpret_s32_f32(__VA_ARGS__)
		186	#define svreinterpret_s32_f64(...) __builtin_sve_reinterpret_s32_f64(__VA_ARGS__)
198	#define svreinterpret_u32_s8(...) __builtin_sve_reinterpret_u32_s8(__VA_ARGS__)	187	#define svreinterpret_u32_s8(...) __builtin_sve_reinterpret_u32_s8(__VA_ARGS__)
199	#define svreinterpret_u32_s16(...) __builtin_sve_reinterpret_u32_s16(__VA_ARGS__)
200	#define svreinterpret_u32_s32(...) __builtin_sve_reinterpret_u32_s32(__VA_ARGS__)
201	#define svreinterpret_u32_s64(...) __builtin_sve_reinterpret_u32_s64(__VA_ARGS__)
202	#define svreinterpret_u32_u8(...) __builtin_sve_reinterpret_u32_u8(__VA_ARGS__)	188	#define svreinterpret_u32_u8(...) __builtin_sve_reinterpret_u32_u8(__VA_ARGS__)
		189	#define svreinterpret_u32_s16(...) __builtin_sve_reinterpret_u32_s16(__VA_ARGS__)
203	#define svreinterpret_u32_u16(...) __builtin_sve_reinterpret_u32_u16(__VA_ARGS__)	190	#define svreinterpret_u32_u16(...) __builtin_sve_reinterpret_u32_u16(__VA_ARGS__)
		191	#define svreinterpret_u32_s32(...) __builtin_sve_reinterpret_u32_s32(__VA_ARGS__)
204	#define svreinterpret_u32_u32(...) __builtin_sve_reinterpret_u32_u32(__VA_ARGS__)	192	#define svreinterpret_u32_u32(...) __builtin_sve_reinterpret_u32_u32(__VA_ARGS__)
		193	#define svreinterpret_u32_s64(...) __builtin_sve_reinterpret_u32_s64(__VA_ARGS__)
205	#define svreinterpret_u32_u64(...) __builtin_sve_reinterpret_u32_u64(__VA_ARGS__)	194	#define svreinterpret_u32_u64(...) __builtin_sve_reinterpret_u32_u64(__VA_ARGS__)
206	#define svreinterpret_u32_f16(...) __builtin_sve_reinterpret_u32_f16(__VA_ARGS__)	195	#define svreinterpret_u32_f16(...) __builtin_sve_reinterpret_u32_f16(__VA_ARGS__)
207	#define svreinterpret_u32_bf16(...) __builtin_sve_reinterpret_u32_bf16(__VA_ARGS__)	196	#define svreinterpret_u32_bf16(...) __builtin_sve_reinterpret_u32_bf16(__VA_ARGS__)
208	#define svreinterpret_u32_f32(...) __builtin_sve_reinterpret_u32_f32(__VA_ARGS__)	197	#define svreinterpret_u32_f32(...) __builtin_sve_reinterpret_u32_f32(__VA_ARGS__)
209	#define svreinterpret_u32_f64(...) __builtin_sve_reinterpret_u32_f64(__VA_ARGS__)	198	#define svreinterpret_u32_f64(...) __builtin_sve_reinterpret_u32_f64(__VA_ARGS__)
		199	#define svreinterpret_s64_s8(...) __builtin_sve_reinterpret_s64_s8(__VA_ARGS__)
		200	#define svreinterpret_s64_u8(...) __builtin_sve_reinterpret_s64_u8(__VA_ARGS__)
		201	#define svreinterpret_s64_s16(...) __builtin_sve_reinterpret_s64_s16(__VA_ARGS__)
		202	#define svreinterpret_s64_u16(...) __builtin_sve_reinterpret_s64_u16(__VA_ARGS__)
		203	#define svreinterpret_s64_s32(...) __builtin_sve_reinterpret_s64_s32(__VA_ARGS__)
		204	#define svreinterpret_s64_u32(...) __builtin_sve_reinterpret_s64_u32(__VA_ARGS__)
		205	#define svreinterpret_s64_s64(...) __builtin_sve_reinterpret_s64_s64(__VA_ARGS__)
		206	#define svreinterpret_s64_u64(...) __builtin_sve_reinterpret_s64_u64(__VA_ARGS__)
		207	#define svreinterpret_s64_f16(...) __builtin_sve_reinterpret_s64_f16(__VA_ARGS__)
		208	#define svreinterpret_s64_bf16(...) __builtin_sve_reinterpret_s64_bf16(__VA_ARGS__)
		209	#define svreinterpret_s64_f32(...) __builtin_sve_reinterpret_s64_f32(__VA_ARGS__)
		210	#define svreinterpret_s64_f64(...) __builtin_sve_reinterpret_s64_f64(__VA_ARGS__)
210	#define svreinterpret_u64_s8(...) __builtin_sve_reinterpret_u64_s8(__VA_ARGS__)	211	#define svreinterpret_u64_s8(...) __builtin_sve_reinterpret_u64_s8(__VA_ARGS__)
211	#define svreinterpret_u64_s16(...) __builtin_sve_reinterpret_u64_s16(__VA_ARGS__)
212	#define svreinterpret_u64_s32(...) __builtin_sve_reinterpret_u64_s32(__VA_ARGS__)
213	#define svreinterpret_u64_s64(...) __builtin_sve_reinterpret_u64_s64(__VA_ARGS__)
214	#define svreinterpret_u64_u8(...) __builtin_sve_reinterpret_u64_u8(__VA_ARGS__)	212	#define svreinterpret_u64_u8(...) __builtin_sve_reinterpret_u64_u8(__VA_ARGS__)
		213	#define svreinterpret_u64_s16(...) __builtin_sve_reinterpret_u64_s16(__VA_ARGS__)
215	#define svreinterpret_u64_u16(...) __builtin_sve_reinterpret_u64_u16(__VA_ARGS__)	214	#define svreinterpret_u64_u16(...) __builtin_sve_reinterpret_u64_u16(__VA_ARGS__)
		215	#define svreinterpret_u64_s32(...) __builtin_sve_reinterpret_u64_s32(__VA_ARGS__)
216	#define svreinterpret_u64_u32(...) __builtin_sve_reinterpret_u64_u32(__VA_ARGS__)	216	#define svreinterpret_u64_u32(...) __builtin_sve_reinterpret_u64_u32(__VA_ARGS__)
		217	#define svreinterpret_u64_s64(...) __builtin_sve_reinterpret_u64_s64(__VA_ARGS__)
217	#define svreinterpret_u64_u64(...) __builtin_sve_reinterpret_u64_u64(__VA_ARGS__)	218	#define svreinterpret_u64_u64(...) __builtin_sve_reinterpret_u64_u64(__VA_ARGS__)
218	#define svreinterpret_u64_f16(...) __builtin_sve_reinterpret_u64_f16(__VA_ARGS__)	219	#define svreinterpret_u64_f16(...) __builtin_sve_reinterpret_u64_f16(__VA_ARGS__)
219	#define svreinterpret_u64_bf16(...) __builtin_sve_reinterpret_u64_bf16(__VA_ARGS__)	220	#define svreinterpret_u64_bf16(...) __builtin_sve_reinterpret_u64_bf16(__VA_ARGS__)
220	#define svreinterpret_u64_f32(...) __builtin_sve_reinterpret_u64_f32(__VA_ARGS__)	221	#define svreinterpret_u64_f32(...) __builtin_sve_reinterpret_u64_f32(__VA_ARGS__)
221	#define svreinterpret_u64_f64(...) __builtin_sve_reinterpret_u64_f64(__VA_ARGS__)	222	#define svreinterpret_u64_f64(...) __builtin_sve_reinterpret_u64_f64(__VA_ARGS__)
222	#define svreinterpret_f16_s8(...) __builtin_sve_reinterpret_f16_s8(__VA_ARGS__)	223	#define svreinterpret_f16_s8(...) __builtin_sve_reinterpret_f16_s8(__VA_ARGS__)
223	#define svreinterpret_f16_s16(...) __builtin_sve_reinterpret_f16_s16(__VA_ARGS__)
224	#define svreinterpret_f16_s32(...) __builtin_sve_reinterpret_f16_s32(__VA_ARGS__)
225	#define svreinterpret_f16_s64(...) __builtin_sve_reinterpret_f16_s64(__VA_ARGS__)
226	#define svreinterpret_f16_u8(...) __builtin_sve_reinterpret_f16_u8(__VA_ARGS__)	224	#define svreinterpret_f16_u8(...) __builtin_sve_reinterpret_f16_u8(__VA_ARGS__)
		225	#define svreinterpret_f16_s16(...) __builtin_sve_reinterpret_f16_s16(__VA_ARGS__)
227	#define svreinterpret_f16_u16(...) __builtin_sve_reinterpret_f16_u16(__VA_ARGS__)	226	#define svreinterpret_f16_u16(...) __builtin_sve_reinterpret_f16_u16(__VA_ARGS__)
		227	#define svreinterpret_f16_s32(...) __builtin_sve_reinterpret_f16_s32(__VA_ARGS__)
228	#define svreinterpret_f16_u32(...) __builtin_sve_reinterpret_f16_u32(__VA_ARGS__)	228	#define svreinterpret_f16_u32(...) __builtin_sve_reinterpret_f16_u32(__VA_ARGS__)
		229	#define svreinterpret_f16_s64(...) __builtin_sve_reinterpret_f16_s64(__VA_ARGS__)
229	#define svreinterpret_f16_u64(...) __builtin_sve_reinterpret_f16_u64(__VA_ARGS__)	230	#define svreinterpret_f16_u64(...) __builtin_sve_reinterpret_f16_u64(__VA_ARGS__)
230	#define svreinterpret_f16_f16(...) __builtin_sve_reinterpret_f16_f16(__VA_ARGS__)	231	#define svreinterpret_f16_f16(...) __builtin_sve_reinterpret_f16_f16(__VA_ARGS__)
231	#define svreinterpret_f16_bf16(...) __builtin_sve_reinterpret_f16_bf16(__VA_ARGS__)	232	#define svreinterpret_f16_bf16(...) __builtin_sve_reinterpret_f16_bf16(__VA_ARGS__)
232	#define svreinterpret_f16_f32(...) __builtin_sve_reinterpret_f16_f32(__VA_ARGS__)	233	#define svreinterpret_f16_f32(...) __builtin_sve_reinterpret_f16_f32(__VA_ARGS__)
233	#define svreinterpret_f16_f64(...) __builtin_sve_reinterpret_f16_f64(__VA_ARGS__)	234	#define svreinterpret_f16_f64(...) __builtin_sve_reinterpret_f16_f64(__VA_ARGS__)
234	#define svreinterpret_bf16_s8(...) __builtin_sve_reinterpret_bf16_s8(__VA_ARGS__)	235	#define svreinterpret_bf16_s8(...) __builtin_sve_reinterpret_bf16_s8(__VA_ARGS__)
235	#define svreinterpret_bf16_s16(...) __builtin_sve_reinterpret_bf16_s16(__VA_ARGS__)
236	#define svreinterpret_bf16_s32(...) __builtin_sve_reinterpret_bf16_s32(__VA_ARGS__)
237	#define svreinterpret_bf16_s64(...) __builtin_sve_reinterpret_bf16_s64(__VA_ARGS__)
238	#define svreinterpret_bf16_u8(...) __builtin_sve_reinterpret_bf16_u8(__VA_ARGS__)	236	#define svreinterpret_bf16_u8(...) __builtin_sve_reinterpret_bf16_u8(__VA_ARGS__)
		237	#define svreinterpret_bf16_s16(...) __builtin_sve_reinterpret_bf16_s16(__VA_ARGS__)
239	#define svreinterpret_bf16_u16(...) __builtin_sve_reinterpret_bf16_u16(__VA_ARGS__)	238	#define svreinterpret_bf16_u16(...) __builtin_sve_reinterpret_bf16_u16(__VA_ARGS__)
		239	#define svreinterpret_bf16_s32(...) __builtin_sve_reinterpret_bf16_s32(__VA_ARGS__)
240	#define svreinterpret_bf16_u32(...) __builtin_sve_reinterpret_bf16_u32(__VA_ARGS__)	240	#define svreinterpret_bf16_u32(...) __builtin_sve_reinterpret_bf16_u32(__VA_ARGS__)
		241	#define svreinterpret_bf16_s64(...) __builtin_sve_reinterpret_bf16_s64(__VA_ARGS__)
241	#define svreinterpret_bf16_u64(...) __builtin_sve_reinterpret_bf16_u64(__VA_ARGS__)	242	#define svreinterpret_bf16_u64(...) __builtin_sve_reinterpret_bf16_u64(__VA_ARGS__)
242	#define svreinterpret_bf16_f16(...) __builtin_sve_reinterpret_bf16_f16(__VA_ARGS__)	243	#define svreinterpret_bf16_f16(...) __builtin_sve_reinterpret_bf16_f16(__VA_ARGS__)
243	#define svreinterpret_bf16_bf16(...) __builtin_sve_reinterpret_bf16_bf16(__VA_ARGS__)	244	#define svreinterpret_bf16_bf16(...) __builtin_sve_reinterpret_bf16_bf16(__VA_ARGS__)
244	#define svreinterpret_bf16_f32(...) __builtin_sve_reinterpret_bf16_f32(__VA_ARGS__)	245	#define svreinterpret_bf16_f32(...) __builtin_sve_reinterpret_bf16_f32(__VA_ARGS__)
245	#define svreinterpret_bf16_f64(...) __builtin_sve_reinterpret_bf16_f64(__VA_ARGS__)	246	#define svreinterpret_bf16_f64(...) __builtin_sve_reinterpret_bf16_f64(__VA_ARGS__)
246	#define svreinterpret_f32_s8(...) __builtin_sve_reinterpret_f32_s8(__VA_ARGS__)	247	#define svreinterpret_f32_s8(...) __builtin_sve_reinterpret_f32_s8(__VA_ARGS__)
247	#define svreinterpret_f32_s16(...) __builtin_sve_reinterpret_f32_s16(__VA_ARGS__)
248	#define svreinterpret_f32_s32(...) __builtin_sve_reinterpret_f32_s32(__VA_ARGS__)
249	#define svreinterpret_f32_s64(...) __builtin_sve_reinterpret_f32_s64(__VA_ARGS__)
250	#define svreinterpret_f32_u8(...) __builtin_sve_reinterpret_f32_u8(__VA_ARGS__)	248	#define svreinterpret_f32_u8(...) __builtin_sve_reinterpret_f32_u8(__VA_ARGS__)
		249	#define svreinterpret_f32_s16(...) __builtin_sve_reinterpret_f32_s16(__VA_ARGS__)
251	#define svreinterpret_f32_u16(...) __builtin_sve_reinterpret_f32_u16(__VA_ARGS__)	250	#define svreinterpret_f32_u16(...) __builtin_sve_reinterpret_f32_u16(__VA_ARGS__)
		251	#define svreinterpret_f32_s32(...) __builtin_sve_reinterpret_f32_s32(__VA_ARGS__)
252	#define svreinterpret_f32_u32(...) __builtin_sve_reinterpret_f32_u32(__VA_ARGS__)	252	#define svreinterpret_f32_u32(...) __builtin_sve_reinterpret_f32_u32(__VA_ARGS__)
		253	#define svreinterpret_f32_s64(...) __builtin_sve_reinterpret_f32_s64(__VA_ARGS__)
253	#define svreinterpret_f32_u64(...) __builtin_sve_reinterpret_f32_u64(__VA_ARGS__)	254	#define svreinterpret_f32_u64(...) __builtin_sve_reinterpret_f32_u64(__VA_ARGS__)
254	#define svreinterpret_f32_f16(...) __builtin_sve_reinterpret_f32_f16(__VA_ARGS__)	255	#define svreinterpret_f32_f16(...) __builtin_sve_reinterpret_f32_f16(__VA_ARGS__)
255	#define svreinterpret_f32_bf16(...) __builtin_sve_reinterpret_f32_bf16(__VA_ARGS__)	256	#define svreinterpret_f32_bf16(...) __builtin_sve_reinterpret_f32_bf16(__VA_ARGS__)
256	#define svreinterpret_f32_f32(...) __builtin_sve_reinterpret_f32_f32(__VA_ARGS__)	257	#define svreinterpret_f32_f32(...) __builtin_sve_reinterpret_f32_f32(__VA_ARGS__)
257	#define svreinterpret_f32_f64(...) __builtin_sve_reinterpret_f32_f64(__VA_ARGS__)	258	#define svreinterpret_f32_f64(...) __builtin_sve_reinterpret_f32_f64(__VA_ARGS__)
258	#define svreinterpret_f64_s8(...) __builtin_sve_reinterpret_f64_s8(__VA_ARGS__)	259	#define svreinterpret_f64_s8(...) __builtin_sve_reinterpret_f64_s8(__VA_ARGS__)
259	#define svreinterpret_f64_s16(...) __builtin_sve_reinterpret_f64_s16(__VA_ARGS__)
260	#define svreinterpret_f64_s32(...) __builtin_sve_reinterpret_f64_s32(__VA_ARGS__)
261	#define svreinterpret_f64_s64(...) __builtin_sve_reinterpret_f64_s64(__VA_ARGS__)
262	#define svreinterpret_f64_u8(...) __builtin_sve_reinterpret_f64_u8(__VA_ARGS__)	260	#define svreinterpret_f64_u8(...) __builtin_sve_reinterpret_f64_u8(__VA_ARGS__)
		261	#define svreinterpret_f64_s16(...) __builtin_sve_reinterpret_f64_s16(__VA_ARGS__)
263	#define svreinterpret_f64_u16(...) __builtin_sve_reinterpret_f64_u16(__VA_ARGS__)	262	#define svreinterpret_f64_u16(...) __builtin_sve_reinterpret_f64_u16(__VA_ARGS__)
		263	#define svreinterpret_f64_s32(...) __builtin_sve_reinterpret_f64_s32(__VA_ARGS__)
264	#define svreinterpret_f64_u32(...) __builtin_sve_reinterpret_f64_u32(__VA_ARGS__)	264	#define svreinterpret_f64_u32(...) __builtin_sve_reinterpret_f64_u32(__VA_ARGS__)
		265	#define svreinterpret_f64_s64(...) __builtin_sve_reinterpret_f64_s64(__VA_ARGS__)
265	#define svreinterpret_f64_u64(...) __builtin_sve_reinterpret_f64_u64(__VA_ARGS__)	266	#define svreinterpret_f64_u64(...) __builtin_sve_reinterpret_f64_u64(__VA_ARGS__)
266	#define svreinterpret_f64_f16(...) __builtin_sve_reinterpret_f64_f16(__VA_ARGS__)	267	#define svreinterpret_f64_f16(...) __builtin_sve_reinterpret_f64_f16(__VA_ARGS__)
267	#define svreinterpret_f64_bf16(...) __builtin_sve_reinterpret_f64_bf16(__VA_ARGS__)	268	#define svreinterpret_f64_bf16(...) __builtin_sve_reinterpret_f64_bf16(__VA_ARGS__)
268	#define svreinterpret_f64_f32(...) __builtin_sve_reinterpret_f64_f32(__VA_ARGS__)	269	#define svreinterpret_f64_f32(...) __builtin_sve_reinterpret_f64_f32(__VA_ARGS__)
269	#define svreinterpret_f64_f64(...) __builtin_sve_reinterpret_f64_f64(__VA_ARGS__)	270	#define svreinterpret_f64_f64(...) __builtin_sve_reinterpret_f64_f64(__VA_ARGS__)
270	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svint8_t op) {	271	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svint8_t op) __arm_streaming_compatible {
271	return __builtin_sve_reinterpret_s8_s8(op);	272	return __builtin_sve_reinterpret_s8_s8(op);
272	}	273	}
273		274
274	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svint16_t op) {	275	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svuint8_t op) __arm_streaming_compatible {
275	return __builtin_sve_reinterpret_s8_s16(op);	276	return __builtin_sve_reinterpret_s8_u8(op);
		277	}
		278
		279	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svint16_t op) __arm_streaming_compatible {
		280	return __builtin_sve_reinterpret_s8_s16(op);
		281	}
		282
		283	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svuint16_t op) __arm_streaming_compatible {
		284	return __builtin_sve_reinterpret_s8_u16(op);
		285	}
		286
		287	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svint32_t op) __arm_streaming_compatible {
		288	return __builtin_sve_reinterpret_s8_s32(op);
		289	}
		290
		291	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svuint32_t op) __arm_streaming_compatible {
		292	return __builtin_sve_reinterpret_s8_u32(op);
		293	}
		294
		295	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svint64_t op) __arm_streaming_compatible {
		296	return __builtin_sve_reinterpret_s8_s64(op);
		297	}
		298
		299	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svuint64_t op) __arm_streaming_compatible {
		300	return __builtin_sve_reinterpret_s8_u64(op);
		301	}
		302
		303	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svfloat16_t op) __arm_streaming_compatible {
		304	return __builtin_sve_reinterpret_s8_f16(op);
		305	}
		306
		307	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svbfloat16_t op) __arm_streaming_compatible {
		308	return __builtin_sve_reinterpret_s8_bf16(op);
		309	}
		310
		311	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svfloat32_t op) __arm_streaming_compatible {
		312	return __builtin_sve_reinterpret_s8_f32(op);
		313	}
		314
		315	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svfloat64_t op) __arm_streaming_compatible {
		316	return __builtin_sve_reinterpret_s8_f64(op);
		317	}
		318
		319	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svint8_t op) __arm_streaming_compatible {
		320	return __builtin_sve_reinterpret_u8_s8(op);
		321	}
		322
		323	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svuint8_t op) __arm_streaming_compatible {
		324	return __builtin_sve_reinterpret_u8_u8(op);
		325	}
		326
		327	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svint16_t op) __arm_streaming_compatible {
		328	return __builtin_sve_reinterpret_u8_s16(op);
		329	}
		330
		331	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svuint16_t op) __arm_streaming_compatible {
		332	return __builtin_sve_reinterpret_u8_u16(op);
		333	}
		334
		335	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svint32_t op) __arm_streaming_compatible {
		336	return __builtin_sve_reinterpret_u8_s32(op);
		337	}
		338
		339	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svuint32_t op) __arm_streaming_compatible {
		340	return __builtin_sve_reinterpret_u8_u32(op);
		341	}
		342
		343	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svint64_t op) __arm_streaming_compatible {
		344	return __builtin_sve_reinterpret_u8_s64(op);
		345	}
		346
		347	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svuint64_t op) __arm_streaming_compatible {
		348	return __builtin_sve_reinterpret_u8_u64(op);
		349	}
		350
		351	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svfloat16_t op) __arm_streaming_compatible {
		352	return __builtin_sve_reinterpret_u8_f16(op);
		353	}
		354
		355	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svbfloat16_t op) __arm_streaming_compatible {
		356	return __builtin_sve_reinterpret_u8_bf16(op);
		357	}
		358
		359	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svfloat32_t op) __arm_streaming_compatible {
		360	return __builtin_sve_reinterpret_u8_f32(op);
		361	}
		362
		363	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svfloat64_t op) __arm_streaming_compatible {
		364	return __builtin_sve_reinterpret_u8_f64(op);
		365	}
		366
		367	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svint8_t op) __arm_streaming_compatible {
		368	return __builtin_sve_reinterpret_s16_s8(op);
		369	}
		370
		371	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svuint8_t op) __arm_streaming_compatible {
		372	return __builtin_sve_reinterpret_s16_u8(op);
		373	}
		374
		375	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svint16_t op) __arm_streaming_compatible {
		376	return __builtin_sve_reinterpret_s16_s16(op);
		377	}
		378
		379	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svuint16_t op) __arm_streaming_compatible {
		380	return __builtin_sve_reinterpret_s16_u16(op);
		381	}
		382
		383	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svint32_t op) __arm_streaming_compatible {
		384	return __builtin_sve_reinterpret_s16_s32(op);
		385	}
		386
		387	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svuint32_t op) __arm_streaming_compatible {
		388	return __builtin_sve_reinterpret_s16_u32(op);
		389	}
		390
		391	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svint64_t op) __arm_streaming_compatible {
		392	return __builtin_sve_reinterpret_s16_s64(op);
		393	}
		394
		395	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svuint64_t op) __arm_streaming_compatible {
		396	return __builtin_sve_reinterpret_s16_u64(op);
		397	}
		398
		399	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svfloat16_t op) __arm_streaming_compatible {
		400	return __builtin_sve_reinterpret_s16_f16(op);
		401	}
		402
		403	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svbfloat16_t op) __arm_streaming_compatible {
		404	return __builtin_sve_reinterpret_s16_bf16(op);
		405	}
		406
		407	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svfloat32_t op) __arm_streaming_compatible {
		408	return __builtin_sve_reinterpret_s16_f32(op);
		409	}
		410
		411	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svfloat64_t op) __arm_streaming_compatible {
		412	return __builtin_sve_reinterpret_s16_f64(op);
		413	}
		414
		415	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svint8_t op) __arm_streaming_compatible {
		416	return __builtin_sve_reinterpret_u16_s8(op);
		417	}
		418
		419	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svuint8_t op) __arm_streaming_compatible {
		420	return __builtin_sve_reinterpret_u16_u8(op);
		421	}
		422
		423	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svint16_t op) __arm_streaming_compatible {
		424	return __builtin_sve_reinterpret_u16_s16(op);
		425	}
		426
		427	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svuint16_t op) __arm_streaming_compatible {
		428	return __builtin_sve_reinterpret_u16_u16(op);
		429	}
		430
		431	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svint32_t op) __arm_streaming_compatible {
		432	return __builtin_sve_reinterpret_u16_s32(op);
		433	}
		434
		435	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svuint32_t op) __arm_streaming_compatible {
		436	return __builtin_sve_reinterpret_u16_u32(op);
		437	}
		438
		439	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svint64_t op) __arm_streaming_compatible {
		440	return __builtin_sve_reinterpret_u16_s64(op);
		441	}
		442
		443	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svuint64_t op) __arm_streaming_compatible {
		444	return __builtin_sve_reinterpret_u16_u64(op);
		445	}
		446
		447	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svfloat16_t op) __arm_streaming_compatible {
		448	return __builtin_sve_reinterpret_u16_f16(op);
		449	}
		450
		451	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svbfloat16_t op) __arm_streaming_compatible {
		452	return __builtin_sve_reinterpret_u16_bf16(op);
		453	}
		454
		455	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svfloat32_t op) __arm_streaming_compatible {
		456	return __builtin_sve_reinterpret_u16_f32(op);
		457	}
		458
		459	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svfloat64_t op) __arm_streaming_compatible {
		460	return __builtin_sve_reinterpret_u16_f64(op);
		461	}
		462
		463	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svint8_t op) __arm_streaming_compatible {
		464	return __builtin_sve_reinterpret_s32_s8(op);
		465	}
		466
		467	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svuint8_t op) __arm_streaming_compatible {
		468	return __builtin_sve_reinterpret_s32_u8(op);
		469	}
		470
		471	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svint16_t op) __arm_streaming_compatible {
		472	return __builtin_sve_reinterpret_s32_s16(op);
		473	}
		474
		475	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svuint16_t op) __arm_streaming_compatible {
		476	return __builtin_sve_reinterpret_s32_u16(op);
		477	}
		478
		479	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svint32_t op) __arm_streaming_compatible {
		480	return __builtin_sve_reinterpret_s32_s32(op);
		481	}
		482
		483	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svuint32_t op) __arm_streaming_compatible {
		484	return __builtin_sve_reinterpret_s32_u32(op);
		485	}
		486
		487	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svint64_t op) __arm_streaming_compatible {
		488	return __builtin_sve_reinterpret_s32_s64(op);
		489	}
		490
		491	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svuint64_t op) __arm_streaming_compatible {
		492	return __builtin_sve_reinterpret_s32_u64(op);
		493	}
		494
		495	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svfloat16_t op) __arm_streaming_compatible {
		496	return __builtin_sve_reinterpret_s32_f16(op);
		497	}
		498
		499	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svbfloat16_t op) __arm_streaming_compatible {
		500	return __builtin_sve_reinterpret_s32_bf16(op);
		501	}
		502
		503	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svfloat32_t op) __arm_streaming_compatible {
		504	return __builtin_sve_reinterpret_s32_f32(op);
		505	}
		506
		507	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svfloat64_t op) __arm_streaming_compatible {
		508	return __builtin_sve_reinterpret_s32_f64(op);
		509	}
		510
		511	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svint8_t op) __arm_streaming_compatible {
		512	return __builtin_sve_reinterpret_u32_s8(op);
		513	}
		514
		515	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svuint8_t op) __arm_streaming_compatible {
		516	return __builtin_sve_reinterpret_u32_u8(op);
		517	}
		518
		519	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svint16_t op) __arm_streaming_compatible {
		520	return __builtin_sve_reinterpret_u32_s16(op);
		521	}
		522
		523	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svuint16_t op) __arm_streaming_compatible {
		524	return __builtin_sve_reinterpret_u32_u16(op);
		525	}
		526
		527	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svint32_t op) __arm_streaming_compatible {
		528	return __builtin_sve_reinterpret_u32_s32(op);
		529	}
		530
		531	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svuint32_t op) __arm_streaming_compatible {
		532	return __builtin_sve_reinterpret_u32_u32(op);
		533	}
		534
		535	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svint64_t op) __arm_streaming_compatible {
		536	return __builtin_sve_reinterpret_u32_s64(op);
		537	}
		538
		539	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svuint64_t op) __arm_streaming_compatible {
		540	return __builtin_sve_reinterpret_u32_u64(op);
		541	}
		542
		543	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svfloat16_t op) __arm_streaming_compatible {
		544	return __builtin_sve_reinterpret_u32_f16(op);
		545	}
		546
		547	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svbfloat16_t op) __arm_streaming_compatible {
		548	return __builtin_sve_reinterpret_u32_bf16(op);
		549	}
		550
		551	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svfloat32_t op) __arm_streaming_compatible {
		552	return __builtin_sve_reinterpret_u32_f32(op);
		553	}
		554
		555	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svfloat64_t op) __arm_streaming_compatible {
		556	return __builtin_sve_reinterpret_u32_f64(op);
		557	}
		558
		559	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svint8_t op) __arm_streaming_compatible {
		560	return __builtin_sve_reinterpret_s64_s8(op);
		561	}
		562
		563	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svuint8_t op) __arm_streaming_compatible {
		564	return __builtin_sve_reinterpret_s64_u8(op);
		565	}
		566
		567	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svint16_t op) __arm_streaming_compatible {
		568	return __builtin_sve_reinterpret_s64_s16(op);
		569	}
		570
		571	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svuint16_t op) __arm_streaming_compatible {
		572	return __builtin_sve_reinterpret_s64_u16(op);
		573	}
		574
		575	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svint32_t op) __arm_streaming_compatible {
		576	return __builtin_sve_reinterpret_s64_s32(op);
		577	}
		578
		579	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svuint32_t op) __arm_streaming_compatible {
		580	return __builtin_sve_reinterpret_s64_u32(op);
		581	}
		582
		583	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svint64_t op) __arm_streaming_compatible {
		584	return __builtin_sve_reinterpret_s64_s64(op);
		585	}
		586
		587	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svuint64_t op) __arm_streaming_compatible {
		588	return __builtin_sve_reinterpret_s64_u64(op);
		589	}
		590
		591	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svfloat16_t op) __arm_streaming_compatible {
		592	return __builtin_sve_reinterpret_s64_f16(op);
		593	}
		594
		595	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svbfloat16_t op) __arm_streaming_compatible {
		596	return __builtin_sve_reinterpret_s64_bf16(op);
		597	}
		598
		599	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svfloat32_t op) __arm_streaming_compatible {
		600	return __builtin_sve_reinterpret_s64_f32(op);
		601	}
		602
		603	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svfloat64_t op) __arm_streaming_compatible {
		604	return __builtin_sve_reinterpret_s64_f64(op);
		605	}
		606
		607	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svint8_t op) __arm_streaming_compatible {
		608	return __builtin_sve_reinterpret_u64_s8(op);
		609	}
		610
		611	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svuint8_t op) __arm_streaming_compatible {
		612	return __builtin_sve_reinterpret_u64_u8(op);
		613	}
		614
		615	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svint16_t op) __arm_streaming_compatible {
		616	return __builtin_sve_reinterpret_u64_s16(op);
		617	}
		618
		619	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svuint16_t op) __arm_streaming_compatible {
		620	return __builtin_sve_reinterpret_u64_u16(op);
		621	}
		622
		623	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svint32_t op) __arm_streaming_compatible {
		624	return __builtin_sve_reinterpret_u64_s32(op);
		625	}
		626
		627	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svuint32_t op) __arm_streaming_compatible {
		628	return __builtin_sve_reinterpret_u64_u32(op);
		629	}
		630
		631	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svint64_t op) __arm_streaming_compatible {
		632	return __builtin_sve_reinterpret_u64_s64(op);
		633	}
		634
		635	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svuint64_t op) __arm_streaming_compatible {
		636	return __builtin_sve_reinterpret_u64_u64(op);
		637	}
		638
		639	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svfloat16_t op) __arm_streaming_compatible {
		640	return __builtin_sve_reinterpret_u64_f16(op);
		641	}
		642
		643	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svbfloat16_t op) __arm_streaming_compatible {
		644	return __builtin_sve_reinterpret_u64_bf16(op);
		645	}
		646
		647	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svfloat32_t op) __arm_streaming_compatible {
		648	return __builtin_sve_reinterpret_u64_f32(op);
		649	}
		650
		651	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svfloat64_t op) __arm_streaming_compatible {
		652	return __builtin_sve_reinterpret_u64_f64(op);
		653	}
		654
		655	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svint8_t op) __arm_streaming_compatible {
		656	return __builtin_sve_reinterpret_f16_s8(op);
		657	}
		658
		659	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svuint8_t op) __arm_streaming_compatible {
		660	return __builtin_sve_reinterpret_f16_u8(op);
		661	}
		662
		663	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svint16_t op) __arm_streaming_compatible {
		664	return __builtin_sve_reinterpret_f16_s16(op);
		665	}
		666
		667	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svuint16_t op) __arm_streaming_compatible {
		668	return __builtin_sve_reinterpret_f16_u16(op);
		669	}
		670
		671	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svint32_t op) __arm_streaming_compatible {
		672	return __builtin_sve_reinterpret_f16_s32(op);
		673	}
		674
		675	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svuint32_t op) __arm_streaming_compatible {
		676	return __builtin_sve_reinterpret_f16_u32(op);
		677	}
		678
		679	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svint64_t op) __arm_streaming_compatible {
		680	return __builtin_sve_reinterpret_f16_s64(op);
		681	}
		682
		683	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svuint64_t op) __arm_streaming_compatible {
		684	return __builtin_sve_reinterpret_f16_u64(op);
		685	}
		686
		687	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svfloat16_t op) __arm_streaming_compatible {
		688	return __builtin_sve_reinterpret_f16_f16(op);
		689	}
		690
		691	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svbfloat16_t op) __arm_streaming_compatible {
		692	return __builtin_sve_reinterpret_f16_bf16(op);
		693	}
		694
		695	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svfloat32_t op) __arm_streaming_compatible {
		696	return __builtin_sve_reinterpret_f16_f32(op);
		697	}
		698
		699	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svfloat64_t op) __arm_streaming_compatible {
		700	return __builtin_sve_reinterpret_f16_f64(op);
		701	}
		702
		703	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svint8_t op) __arm_streaming_compatible {
		704	return __builtin_sve_reinterpret_bf16_s8(op);
		705	}
		706
		707	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svuint8_t op) __arm_streaming_compatible {
		708	return __builtin_sve_reinterpret_bf16_u8(op);
		709	}
		710
		711	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svint16_t op) __arm_streaming_compatible {
		712	return __builtin_sve_reinterpret_bf16_s16(op);
		713	}
		714
		715	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svuint16_t op) __arm_streaming_compatible {
		716	return __builtin_sve_reinterpret_bf16_u16(op);
		717	}
		718
		719	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svint32_t op) __arm_streaming_compatible {
		720	return __builtin_sve_reinterpret_bf16_s32(op);
		721	}
		722
		723	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svuint32_t op) __arm_streaming_compatible {
		724	return __builtin_sve_reinterpret_bf16_u32(op);
		725	}
		726
		727	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svint64_t op) __arm_streaming_compatible {
		728	return __builtin_sve_reinterpret_bf16_s64(op);
		729	}
		730
		731	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svuint64_t op) __arm_streaming_compatible {
		732	return __builtin_sve_reinterpret_bf16_u64(op);
		733	}
		734
		735	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svfloat16_t op) __arm_streaming_compatible {
		736	return __builtin_sve_reinterpret_bf16_f16(op);
		737	}
		738
		739	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svbfloat16_t op) __arm_streaming_compatible {
		740	return __builtin_sve_reinterpret_bf16_bf16(op);
		741	}
		742
		743	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svfloat32_t op) __arm_streaming_compatible {
		744	return __builtin_sve_reinterpret_bf16_f32(op);
		745	}
		746
		747	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svfloat64_t op) __arm_streaming_compatible {
		748	return __builtin_sve_reinterpret_bf16_f64(op);
		749	}
		750
		751	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svint8_t op) __arm_streaming_compatible {
		752	return __builtin_sve_reinterpret_f32_s8(op);
		753	}
		754
		755	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svuint8_t op) __arm_streaming_compatible {
		756	return __builtin_sve_reinterpret_f32_u8(op);
		757	}
		758
		759	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svint16_t op) __arm_streaming_compatible {
		760	return __builtin_sve_reinterpret_f32_s16(op);
		761	}
		762
		763	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svuint16_t op) __arm_streaming_compatible {
		764	return __builtin_sve_reinterpret_f32_u16(op);
		765	}
		766
		767	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svint32_t op) __arm_streaming_compatible {
		768	return __builtin_sve_reinterpret_f32_s32(op);
		769	}
		770
		771	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svuint32_t op) __arm_streaming_compatible {
		772	return __builtin_sve_reinterpret_f32_u32(op);
		773	}
		774
		775	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svint64_t op) __arm_streaming_compatible {
		776	return __builtin_sve_reinterpret_f32_s64(op);
		777	}
		778
		779	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svuint64_t op) __arm_streaming_compatible {
		780	return __builtin_sve_reinterpret_f32_u64(op);
		781	}
		782
		783	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svfloat16_t op) __arm_streaming_compatible {
		784	return __builtin_sve_reinterpret_f32_f16(op);
		785	}
		786
		787	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svbfloat16_t op) __arm_streaming_compatible {
		788	return __builtin_sve_reinterpret_f32_bf16(op);
		789	}
		790
		791	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svfloat32_t op) __arm_streaming_compatible {
		792	return __builtin_sve_reinterpret_f32_f32(op);
		793	}
		794
		795	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svfloat64_t op) __arm_streaming_compatible {
		796	return __builtin_sve_reinterpret_f32_f64(op);
		797	}
		798
		799	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svint8_t op) __arm_streaming_compatible {
		800	return __builtin_sve_reinterpret_f64_s8(op);
		801	}
		802
		803	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svuint8_t op) __arm_streaming_compatible {
		804	return __builtin_sve_reinterpret_f64_u8(op);
		805	}
		806
		807	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svint16_t op) __arm_streaming_compatible {
		808	return __builtin_sve_reinterpret_f64_s16(op);
		809	}
		810
		811	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svuint16_t op) __arm_streaming_compatible {
		812	return __builtin_sve_reinterpret_f64_u16(op);
		813	}
		814
		815	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svint32_t op) __arm_streaming_compatible {
		816	return __builtin_sve_reinterpret_f64_s32(op);
		817	}
		818
		819	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svuint32_t op) __arm_streaming_compatible {
		820	return __builtin_sve_reinterpret_f64_u32(op);
		821	}
		822
		823	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svint64_t op) __arm_streaming_compatible {
		824	return __builtin_sve_reinterpret_f64_s64(op);
		825	}
		826
		827	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svuint64_t op) __arm_streaming_compatible {
		828	return __builtin_sve_reinterpret_f64_u64(op);
		829	}
		830
		831	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svfloat16_t op) __arm_streaming_compatible {
		832	return __builtin_sve_reinterpret_f64_f16(op);
		833	}
		834
		835	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svbfloat16_t op) __arm_streaming_compatible {
		836	return __builtin_sve_reinterpret_f64_bf16(op);
		837	}
		838
		839	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svfloat32_t op) __arm_streaming_compatible {
		840	return __builtin_sve_reinterpret_f64_f32(op);
		841	}
		842
		843	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svfloat64_t op) __arm_streaming_compatible {
		844	return __builtin_sve_reinterpret_f64_f64(op);
		845	}
		846
		847	#define svreinterpret_s8_s8_x2(...) __builtin_sve_reinterpret_s8_s8_x2(__VA_ARGS__)
		848	#define svreinterpret_s8_u8_x2(...) __builtin_sve_reinterpret_s8_u8_x2(__VA_ARGS__)
		849	#define svreinterpret_s8_s16_x2(...) __builtin_sve_reinterpret_s8_s16_x2(__VA_ARGS__)
		850	#define svreinterpret_s8_u16_x2(...) __builtin_sve_reinterpret_s8_u16_x2(__VA_ARGS__)
		851	#define svreinterpret_s8_s32_x2(...) __builtin_sve_reinterpret_s8_s32_x2(__VA_ARGS__)
		852	#define svreinterpret_s8_u32_x2(...) __builtin_sve_reinterpret_s8_u32_x2(__VA_ARGS__)
		853	#define svreinterpret_s8_s64_x2(...) __builtin_sve_reinterpret_s8_s64_x2(__VA_ARGS__)
		854	#define svreinterpret_s8_u64_x2(...) __builtin_sve_reinterpret_s8_u64_x2(__VA_ARGS__)
		855	#define svreinterpret_s8_f16_x2(...) __builtin_sve_reinterpret_s8_f16_x2(__VA_ARGS__)
		856	#define svreinterpret_s8_bf16_x2(...) __builtin_sve_reinterpret_s8_bf16_x2(__VA_ARGS__)
		857	#define svreinterpret_s8_f32_x2(...) __builtin_sve_reinterpret_s8_f32_x2(__VA_ARGS__)
		858	#define svreinterpret_s8_f64_x2(...) __builtin_sve_reinterpret_s8_f64_x2(__VA_ARGS__)
		859	#define svreinterpret_u8_s8_x2(...) __builtin_sve_reinterpret_u8_s8_x2(__VA_ARGS__)
		860	#define svreinterpret_u8_u8_x2(...) __builtin_sve_reinterpret_u8_u8_x2(__VA_ARGS__)
		861	#define svreinterpret_u8_s16_x2(...) __builtin_sve_reinterpret_u8_s16_x2(__VA_ARGS__)
		862	#define svreinterpret_u8_u16_x2(...) __builtin_sve_reinterpret_u8_u16_x2(__VA_ARGS__)
		863	#define svreinterpret_u8_s32_x2(...) __builtin_sve_reinterpret_u8_s32_x2(__VA_ARGS__)
		864	#define svreinterpret_u8_u32_x2(...) __builtin_sve_reinterpret_u8_u32_x2(__VA_ARGS__)
		865	#define svreinterpret_u8_s64_x2(...) __builtin_sve_reinterpret_u8_s64_x2(__VA_ARGS__)
		866	#define svreinterpret_u8_u64_x2(...) __builtin_sve_reinterpret_u8_u64_x2(__VA_ARGS__)
		867	#define svreinterpret_u8_f16_x2(...) __builtin_sve_reinterpret_u8_f16_x2(__VA_ARGS__)
		868	#define svreinterpret_u8_bf16_x2(...) __builtin_sve_reinterpret_u8_bf16_x2(__VA_ARGS__)
		869	#define svreinterpret_u8_f32_x2(...) __builtin_sve_reinterpret_u8_f32_x2(__VA_ARGS__)
		870	#define svreinterpret_u8_f64_x2(...) __builtin_sve_reinterpret_u8_f64_x2(__VA_ARGS__)
		871	#define svreinterpret_s16_s8_x2(...) __builtin_sve_reinterpret_s16_s8_x2(__VA_ARGS__)
		872	#define svreinterpret_s16_u8_x2(...) __builtin_sve_reinterpret_s16_u8_x2(__VA_ARGS__)
		873	#define svreinterpret_s16_s16_x2(...) __builtin_sve_reinterpret_s16_s16_x2(__VA_ARGS__)
		874	#define svreinterpret_s16_u16_x2(...) __builtin_sve_reinterpret_s16_u16_x2(__VA_ARGS__)
		875	#define svreinterpret_s16_s32_x2(...) __builtin_sve_reinterpret_s16_s32_x2(__VA_ARGS__)
		876	#define svreinterpret_s16_u32_x2(...) __builtin_sve_reinterpret_s16_u32_x2(__VA_ARGS__)
		877	#define svreinterpret_s16_s64_x2(...) __builtin_sve_reinterpret_s16_s64_x2(__VA_ARGS__)
		878	#define svreinterpret_s16_u64_x2(...) __builtin_sve_reinterpret_s16_u64_x2(__VA_ARGS__)
		879	#define svreinterpret_s16_f16_x2(...) __builtin_sve_reinterpret_s16_f16_x2(__VA_ARGS__)
		880	#define svreinterpret_s16_bf16_x2(...) __builtin_sve_reinterpret_s16_bf16_x2(__VA_ARGS__)
		881	#define svreinterpret_s16_f32_x2(...) __builtin_sve_reinterpret_s16_f32_x2(__VA_ARGS__)
		882	#define svreinterpret_s16_f64_x2(...) __builtin_sve_reinterpret_s16_f64_x2(__VA_ARGS__)
		883	#define svreinterpret_u16_s8_x2(...) __builtin_sve_reinterpret_u16_s8_x2(__VA_ARGS__)
		884	#define svreinterpret_u16_u8_x2(...) __builtin_sve_reinterpret_u16_u8_x2(__VA_ARGS__)
		885	#define svreinterpret_u16_s16_x2(...) __builtin_sve_reinterpret_u16_s16_x2(__VA_ARGS__)
		886	#define svreinterpret_u16_u16_x2(...) __builtin_sve_reinterpret_u16_u16_x2(__VA_ARGS__)
		887	#define svreinterpret_u16_s32_x2(...) __builtin_sve_reinterpret_u16_s32_x2(__VA_ARGS__)
		888	#define svreinterpret_u16_u32_x2(...) __builtin_sve_reinterpret_u16_u32_x2(__VA_ARGS__)
		889	#define svreinterpret_u16_s64_x2(...) __builtin_sve_reinterpret_u16_s64_x2(__VA_ARGS__)
		890	#define svreinterpret_u16_u64_x2(...) __builtin_sve_reinterpret_u16_u64_x2(__VA_ARGS__)
		891	#define svreinterpret_u16_f16_x2(...) __builtin_sve_reinterpret_u16_f16_x2(__VA_ARGS__)
		892	#define svreinterpret_u16_bf16_x2(...) __builtin_sve_reinterpret_u16_bf16_x2(__VA_ARGS__)
		893	#define svreinterpret_u16_f32_x2(...) __builtin_sve_reinterpret_u16_f32_x2(__VA_ARGS__)
		894	#define svreinterpret_u16_f64_x2(...) __builtin_sve_reinterpret_u16_f64_x2(__VA_ARGS__)
		895	#define svreinterpret_s32_s8_x2(...) __builtin_sve_reinterpret_s32_s8_x2(__VA_ARGS__)
		896	#define svreinterpret_s32_u8_x2(...) __builtin_sve_reinterpret_s32_u8_x2(__VA_ARGS__)
		897	#define svreinterpret_s32_s16_x2(...) __builtin_sve_reinterpret_s32_s16_x2(__VA_ARGS__)
		898	#define svreinterpret_s32_u16_x2(...) __builtin_sve_reinterpret_s32_u16_x2(__VA_ARGS__)
		899	#define svreinterpret_s32_s32_x2(...) __builtin_sve_reinterpret_s32_s32_x2(__VA_ARGS__)
		900	#define svreinterpret_s32_u32_x2(...) __builtin_sve_reinterpret_s32_u32_x2(__VA_ARGS__)
		901	#define svreinterpret_s32_s64_x2(...) __builtin_sve_reinterpret_s32_s64_x2(__VA_ARGS__)
		902	#define svreinterpret_s32_u64_x2(...) __builtin_sve_reinterpret_s32_u64_x2(__VA_ARGS__)
		903	#define svreinterpret_s32_f16_x2(...) __builtin_sve_reinterpret_s32_f16_x2(__VA_ARGS__)
		904	#define svreinterpret_s32_bf16_x2(...) __builtin_sve_reinterpret_s32_bf16_x2(__VA_ARGS__)
		905	#define svreinterpret_s32_f32_x2(...) __builtin_sve_reinterpret_s32_f32_x2(__VA_ARGS__)
		906	#define svreinterpret_s32_f64_x2(...) __builtin_sve_reinterpret_s32_f64_x2(__VA_ARGS__)
		907	#define svreinterpret_u32_s8_x2(...) __builtin_sve_reinterpret_u32_s8_x2(__VA_ARGS__)
		908	#define svreinterpret_u32_u8_x2(...) __builtin_sve_reinterpret_u32_u8_x2(__VA_ARGS__)
		909	#define svreinterpret_u32_s16_x2(...) __builtin_sve_reinterpret_u32_s16_x2(__VA_ARGS__)
		910	#define svreinterpret_u32_u16_x2(...) __builtin_sve_reinterpret_u32_u16_x2(__VA_ARGS__)
		911	#define svreinterpret_u32_s32_x2(...) __builtin_sve_reinterpret_u32_s32_x2(__VA_ARGS__)
		912	#define svreinterpret_u32_u32_x2(...) __builtin_sve_reinterpret_u32_u32_x2(__VA_ARGS__)
		913	#define svreinterpret_u32_s64_x2(...) __builtin_sve_reinterpret_u32_s64_x2(__VA_ARGS__)
		914	#define svreinterpret_u32_u64_x2(...) __builtin_sve_reinterpret_u32_u64_x2(__VA_ARGS__)
		915	#define svreinterpret_u32_f16_x2(...) __builtin_sve_reinterpret_u32_f16_x2(__VA_ARGS__)
		916	#define svreinterpret_u32_bf16_x2(...) __builtin_sve_reinterpret_u32_bf16_x2(__VA_ARGS__)
		917	#define svreinterpret_u32_f32_x2(...) __builtin_sve_reinterpret_u32_f32_x2(__VA_ARGS__)
		918	#define svreinterpret_u32_f64_x2(...) __builtin_sve_reinterpret_u32_f64_x2(__VA_ARGS__)
		919	#define svreinterpret_s64_s8_x2(...) __builtin_sve_reinterpret_s64_s8_x2(__VA_ARGS__)
		920	#define svreinterpret_s64_u8_x2(...) __builtin_sve_reinterpret_s64_u8_x2(__VA_ARGS__)
		921	#define svreinterpret_s64_s16_x2(...) __builtin_sve_reinterpret_s64_s16_x2(__VA_ARGS__)
		922	#define svreinterpret_s64_u16_x2(...) __builtin_sve_reinterpret_s64_u16_x2(__VA_ARGS__)
		923	#define svreinterpret_s64_s32_x2(...) __builtin_sve_reinterpret_s64_s32_x2(__VA_ARGS__)
		924	#define svreinterpret_s64_u32_x2(...) __builtin_sve_reinterpret_s64_u32_x2(__VA_ARGS__)
		925	#define svreinterpret_s64_s64_x2(...) __builtin_sve_reinterpret_s64_s64_x2(__VA_ARGS__)
		926	#define svreinterpret_s64_u64_x2(...) __builtin_sve_reinterpret_s64_u64_x2(__VA_ARGS__)
		927	#define svreinterpret_s64_f16_x2(...) __builtin_sve_reinterpret_s64_f16_x2(__VA_ARGS__)
		928	#define svreinterpret_s64_bf16_x2(...) __builtin_sve_reinterpret_s64_bf16_x2(__VA_ARGS__)
		929	#define svreinterpret_s64_f32_x2(...) __builtin_sve_reinterpret_s64_f32_x2(__VA_ARGS__)
		930	#define svreinterpret_s64_f64_x2(...) __builtin_sve_reinterpret_s64_f64_x2(__VA_ARGS__)
		931	#define svreinterpret_u64_s8_x2(...) __builtin_sve_reinterpret_u64_s8_x2(__VA_ARGS__)
		932	#define svreinterpret_u64_u8_x2(...) __builtin_sve_reinterpret_u64_u8_x2(__VA_ARGS__)
		933	#define svreinterpret_u64_s16_x2(...) __builtin_sve_reinterpret_u64_s16_x2(__VA_ARGS__)
		934	#define svreinterpret_u64_u16_x2(...) __builtin_sve_reinterpret_u64_u16_x2(__VA_ARGS__)
		935	#define svreinterpret_u64_s32_x2(...) __builtin_sve_reinterpret_u64_s32_x2(__VA_ARGS__)
		936	#define svreinterpret_u64_u32_x2(...) __builtin_sve_reinterpret_u64_u32_x2(__VA_ARGS__)
		937	#define svreinterpret_u64_s64_x2(...) __builtin_sve_reinterpret_u64_s64_x2(__VA_ARGS__)
		938	#define svreinterpret_u64_u64_x2(...) __builtin_sve_reinterpret_u64_u64_x2(__VA_ARGS__)
		939	#define svreinterpret_u64_f16_x2(...) __builtin_sve_reinterpret_u64_f16_x2(__VA_ARGS__)
		940	#define svreinterpret_u64_bf16_x2(...) __builtin_sve_reinterpret_u64_bf16_x2(__VA_ARGS__)
		941	#define svreinterpret_u64_f32_x2(...) __builtin_sve_reinterpret_u64_f32_x2(__VA_ARGS__)
		942	#define svreinterpret_u64_f64_x2(...) __builtin_sve_reinterpret_u64_f64_x2(__VA_ARGS__)
		943	#define svreinterpret_f16_s8_x2(...) __builtin_sve_reinterpret_f16_s8_x2(__VA_ARGS__)
		944	#define svreinterpret_f16_u8_x2(...) __builtin_sve_reinterpret_f16_u8_x2(__VA_ARGS__)
		945	#define svreinterpret_f16_s16_x2(...) __builtin_sve_reinterpret_f16_s16_x2(__VA_ARGS__)
		946	#define svreinterpret_f16_u16_x2(...) __builtin_sve_reinterpret_f16_u16_x2(__VA_ARGS__)
		947	#define svreinterpret_f16_s32_x2(...) __builtin_sve_reinterpret_f16_s32_x2(__VA_ARGS__)
		948	#define svreinterpret_f16_u32_x2(...) __builtin_sve_reinterpret_f16_u32_x2(__VA_ARGS__)
		949	#define svreinterpret_f16_s64_x2(...) __builtin_sve_reinterpret_f16_s64_x2(__VA_ARGS__)
		950	#define svreinterpret_f16_u64_x2(...) __builtin_sve_reinterpret_f16_u64_x2(__VA_ARGS__)
		951	#define svreinterpret_f16_f16_x2(...) __builtin_sve_reinterpret_f16_f16_x2(__VA_ARGS__)
		952	#define svreinterpret_f16_bf16_x2(...) __builtin_sve_reinterpret_f16_bf16_x2(__VA_ARGS__)
		953	#define svreinterpret_f16_f32_x2(...) __builtin_sve_reinterpret_f16_f32_x2(__VA_ARGS__)
		954	#define svreinterpret_f16_f64_x2(...) __builtin_sve_reinterpret_f16_f64_x2(__VA_ARGS__)
		955	#define svreinterpret_bf16_s8_x2(...) __builtin_sve_reinterpret_bf16_s8_x2(__VA_ARGS__)
		956	#define svreinterpret_bf16_u8_x2(...) __builtin_sve_reinterpret_bf16_u8_x2(__VA_ARGS__)
		957	#define svreinterpret_bf16_s16_x2(...) __builtin_sve_reinterpret_bf16_s16_x2(__VA_ARGS__)
		958	#define svreinterpret_bf16_u16_x2(...) __builtin_sve_reinterpret_bf16_u16_x2(__VA_ARGS__)
		959	#define svreinterpret_bf16_s32_x2(...) __builtin_sve_reinterpret_bf16_s32_x2(__VA_ARGS__)
		960	#define svreinterpret_bf16_u32_x2(...) __builtin_sve_reinterpret_bf16_u32_x2(__VA_ARGS__)
		961	#define svreinterpret_bf16_s64_x2(...) __builtin_sve_reinterpret_bf16_s64_x2(__VA_ARGS__)
		962	#define svreinterpret_bf16_u64_x2(...) __builtin_sve_reinterpret_bf16_u64_x2(__VA_ARGS__)
		963	#define svreinterpret_bf16_f16_x2(...) __builtin_sve_reinterpret_bf16_f16_x2(__VA_ARGS__)
		964	#define svreinterpret_bf16_bf16_x2(...) __builtin_sve_reinterpret_bf16_bf16_x2(__VA_ARGS__)
		965	#define svreinterpret_bf16_f32_x2(...) __builtin_sve_reinterpret_bf16_f32_x2(__VA_ARGS__)
		966	#define svreinterpret_bf16_f64_x2(...) __builtin_sve_reinterpret_bf16_f64_x2(__VA_ARGS__)
		967	#define svreinterpret_f32_s8_x2(...) __builtin_sve_reinterpret_f32_s8_x2(__VA_ARGS__)
		968	#define svreinterpret_f32_u8_x2(...) __builtin_sve_reinterpret_f32_u8_x2(__VA_ARGS__)
		969	#define svreinterpret_f32_s16_x2(...) __builtin_sve_reinterpret_f32_s16_x2(__VA_ARGS__)
		970	#define svreinterpret_f32_u16_x2(...) __builtin_sve_reinterpret_f32_u16_x2(__VA_ARGS__)
		971	#define svreinterpret_f32_s32_x2(...) __builtin_sve_reinterpret_f32_s32_x2(__VA_ARGS__)
		972	#define svreinterpret_f32_u32_x2(...) __builtin_sve_reinterpret_f32_u32_x2(__VA_ARGS__)
		973	#define svreinterpret_f32_s64_x2(...) __builtin_sve_reinterpret_f32_s64_x2(__VA_ARGS__)
		974	#define svreinterpret_f32_u64_x2(...) __builtin_sve_reinterpret_f32_u64_x2(__VA_ARGS__)
		975	#define svreinterpret_f32_f16_x2(...) __builtin_sve_reinterpret_f32_f16_x2(__VA_ARGS__)
		976	#define svreinterpret_f32_bf16_x2(...) __builtin_sve_reinterpret_f32_bf16_x2(__VA_ARGS__)
		977	#define svreinterpret_f32_f32_x2(...) __builtin_sve_reinterpret_f32_f32_x2(__VA_ARGS__)
		978	#define svreinterpret_f32_f64_x2(...) __builtin_sve_reinterpret_f32_f64_x2(__VA_ARGS__)
		979	#define svreinterpret_f64_s8_x2(...) __builtin_sve_reinterpret_f64_s8_x2(__VA_ARGS__)
		980	#define svreinterpret_f64_u8_x2(...) __builtin_sve_reinterpret_f64_u8_x2(__VA_ARGS__)
		981	#define svreinterpret_f64_s16_x2(...) __builtin_sve_reinterpret_f64_s16_x2(__VA_ARGS__)
		982	#define svreinterpret_f64_u16_x2(...) __builtin_sve_reinterpret_f64_u16_x2(__VA_ARGS__)
		983	#define svreinterpret_f64_s32_x2(...) __builtin_sve_reinterpret_f64_s32_x2(__VA_ARGS__)
		984	#define svreinterpret_f64_u32_x2(...) __builtin_sve_reinterpret_f64_u32_x2(__VA_ARGS__)
		985	#define svreinterpret_f64_s64_x2(...) __builtin_sve_reinterpret_f64_s64_x2(__VA_ARGS__)
		986	#define svreinterpret_f64_u64_x2(...) __builtin_sve_reinterpret_f64_u64_x2(__VA_ARGS__)
		987	#define svreinterpret_f64_f16_x2(...) __builtin_sve_reinterpret_f64_f16_x2(__VA_ARGS__)
		988	#define svreinterpret_f64_bf16_x2(...) __builtin_sve_reinterpret_f64_bf16_x2(__VA_ARGS__)
		989	#define svreinterpret_f64_f32_x2(...) __builtin_sve_reinterpret_f64_f32_x2(__VA_ARGS__)
		990	#define svreinterpret_f64_f64_x2(...) __builtin_sve_reinterpret_f64_f64_x2(__VA_ARGS__)
		991	__aio __attribute__((target("sve"))) svint8x2_t svreinterpret_s8(svint8x2_t op) __arm_streaming_compatible {
		992	return __builtin_sve_reinterpret_s8_s8_x2(op);
		993	}
		994
		995	__aio __attribute__((target("sve"))) svint8x2_t svreinterpret_s8(svuint8x2_t op) __arm_streaming_compatible {
		996	return __builtin_sve_reinterpret_s8_u8_x2(op);
		997	}
		998
		999	__aio __attribute__((target("sve"))) svint8x2_t svreinterpret_s8(svint16x2_t op) __arm_streaming_compatible {
		1000	return __builtin_sve_reinterpret_s8_s16_x2(op);
		1001	}
		1002
		1003	__aio __attribute__((target("sve"))) svint8x2_t svreinterpret_s8(svuint16x2_t op) __arm_streaming_compatible {
		1004	return __builtin_sve_reinterpret_s8_u16_x2(op);
		1005	}
		1006
		1007	__aio __attribute__((target("sve"))) svint8x2_t svreinterpret_s8(svint32x2_t op) __arm_streaming_compatible {
		1008	return __builtin_sve_reinterpret_s8_s32_x2(op);
		1009	}
		1010
		1011	__aio __attribute__((target("sve"))) svint8x2_t svreinterpret_s8(svuint32x2_t op) __arm_streaming_compatible {
		1012	return __builtin_sve_reinterpret_s8_u32_x2(op);
		1013	}
		1014
		1015	__aio __attribute__((target("sve"))) svint8x2_t svreinterpret_s8(svint64x2_t op) __arm_streaming_compatible {
		1016	return __builtin_sve_reinterpret_s8_s64_x2(op);
		1017	}
		1018
		1019	__aio __attribute__((target("sve"))) svint8x2_t svreinterpret_s8(svuint64x2_t op) __arm_streaming_compatible {
		1020	return __builtin_sve_reinterpret_s8_u64_x2(op);
		1021	}
		1022
		1023	__aio __attribute__((target("sve"))) svint8x2_t svreinterpret_s8(svfloat16x2_t op) __arm_streaming_compatible {
		1024	return __builtin_sve_reinterpret_s8_f16_x2(op);
		1025	}
		1026
		1027	__aio __attribute__((target("sve"))) svint8x2_t svreinterpret_s8(svbfloat16x2_t op) __arm_streaming_compatible {
		1028	return __builtin_sve_reinterpret_s8_bf16_x2(op);
		1029	}
		1030
		1031	__aio __attribute__((target("sve"))) svint8x2_t svreinterpret_s8(svfloat32x2_t op) __arm_streaming_compatible {
		1032	return __builtin_sve_reinterpret_s8_f32_x2(op);
		1033	}
		1034
		1035	__aio __attribute__((target("sve"))) svint8x2_t svreinterpret_s8(svfloat64x2_t op) __arm_streaming_compatible {
		1036	return __builtin_sve_reinterpret_s8_f64_x2(op);
		1037	}
		1038
		1039	__aio __attribute__((target("sve"))) svuint8x2_t svreinterpret_u8(svint8x2_t op) __arm_streaming_compatible {
		1040	return __builtin_sve_reinterpret_u8_s8_x2(op);
		1041	}
		1042
		1043	__aio __attribute__((target("sve"))) svuint8x2_t svreinterpret_u8(svuint8x2_t op) __arm_streaming_compatible {
		1044	return __builtin_sve_reinterpret_u8_u8_x2(op);
		1045	}
		1046
		1047	__aio __attribute__((target("sve"))) svuint8x2_t svreinterpret_u8(svint16x2_t op) __arm_streaming_compatible {
		1048	return __builtin_sve_reinterpret_u8_s16_x2(op);
		1049	}
		1050
		1051	__aio __attribute__((target("sve"))) svuint8x2_t svreinterpret_u8(svuint16x2_t op) __arm_streaming_compatible {
		1052	return __builtin_sve_reinterpret_u8_u16_x2(op);
		1053	}
		1054
		1055	__aio __attribute__((target("sve"))) svuint8x2_t svreinterpret_u8(svint32x2_t op) __arm_streaming_compatible {
		1056	return __builtin_sve_reinterpret_u8_s32_x2(op);
		1057	}
		1058
		1059	__aio __attribute__((target("sve"))) svuint8x2_t svreinterpret_u8(svuint32x2_t op) __arm_streaming_compatible {
		1060	return __builtin_sve_reinterpret_u8_u32_x2(op);
		1061	}
		1062
		1063	__aio __attribute__((target("sve"))) svuint8x2_t svreinterpret_u8(svint64x2_t op) __arm_streaming_compatible {
		1064	return __builtin_sve_reinterpret_u8_s64_x2(op);
		1065	}
		1066
		1067	__aio __attribute__((target("sve"))) svuint8x2_t svreinterpret_u8(svuint64x2_t op) __arm_streaming_compatible {
		1068	return __builtin_sve_reinterpret_u8_u64_x2(op);
		1069	}
		1070
		1071	__aio __attribute__((target("sve"))) svuint8x2_t svreinterpret_u8(svfloat16x2_t op) __arm_streaming_compatible {
		1072	return __builtin_sve_reinterpret_u8_f16_x2(op);
		1073	}
		1074
		1075	__aio __attribute__((target("sve"))) svuint8x2_t svreinterpret_u8(svbfloat16x2_t op) __arm_streaming_compatible {
		1076	return __builtin_sve_reinterpret_u8_bf16_x2(op);
		1077	}
		1078
		1079	__aio __attribute__((target("sve"))) svuint8x2_t svreinterpret_u8(svfloat32x2_t op) __arm_streaming_compatible {
		1080	return __builtin_sve_reinterpret_u8_f32_x2(op);
		1081	}
		1082
		1083	__aio __attribute__((target("sve"))) svuint8x2_t svreinterpret_u8(svfloat64x2_t op) __arm_streaming_compatible {
		1084	return __builtin_sve_reinterpret_u8_f64_x2(op);
		1085	}
		1086
		1087	__aio __attribute__((target("sve"))) svint16x2_t svreinterpret_s16(svint8x2_t op) __arm_streaming_compatible {
		1088	return __builtin_sve_reinterpret_s16_s8_x2(op);
		1089	}
		1090
		1091	__aio __attribute__((target("sve"))) svint16x2_t svreinterpret_s16(svuint8x2_t op) __arm_streaming_compatible {
		1092	return __builtin_sve_reinterpret_s16_u8_x2(op);
		1093	}
		1094
		1095	__aio __attribute__((target("sve"))) svint16x2_t svreinterpret_s16(svint16x2_t op) __arm_streaming_compatible {
		1096	return __builtin_sve_reinterpret_s16_s16_x2(op);
		1097	}
		1098
		1099	__aio __attribute__((target("sve"))) svint16x2_t svreinterpret_s16(svuint16x2_t op) __arm_streaming_compatible {
		1100	return __builtin_sve_reinterpret_s16_u16_x2(op);
		1101	}
		1102
		1103	__aio __attribute__((target("sve"))) svint16x2_t svreinterpret_s16(svint32x2_t op) __arm_streaming_compatible {
		1104	return __builtin_sve_reinterpret_s16_s32_x2(op);
		1105	}
		1106
		1107	__aio __attribute__((target("sve"))) svint16x2_t svreinterpret_s16(svuint32x2_t op) __arm_streaming_compatible {
		1108	return __builtin_sve_reinterpret_s16_u32_x2(op);
		1109	}
		1110
		1111	__aio __attribute__((target("sve"))) svint16x2_t svreinterpret_s16(svint64x2_t op) __arm_streaming_compatible {
		1112	return __builtin_sve_reinterpret_s16_s64_x2(op);
		1113	}
		1114
		1115	__aio __attribute__((target("sve"))) svint16x2_t svreinterpret_s16(svuint64x2_t op) __arm_streaming_compatible {
		1116	return __builtin_sve_reinterpret_s16_u64_x2(op);
		1117	}
		1118
		1119	__aio __attribute__((target("sve"))) svint16x2_t svreinterpret_s16(svfloat16x2_t op) __arm_streaming_compatible {
		1120	return __builtin_sve_reinterpret_s16_f16_x2(op);
		1121	}
		1122
		1123	__aio __attribute__((target("sve"))) svint16x2_t svreinterpret_s16(svbfloat16x2_t op) __arm_streaming_compatible {
		1124	return __builtin_sve_reinterpret_s16_bf16_x2(op);
		1125	}
		1126
		1127	__aio __attribute__((target("sve"))) svint16x2_t svreinterpret_s16(svfloat32x2_t op) __arm_streaming_compatible {
		1128	return __builtin_sve_reinterpret_s16_f32_x2(op);
		1129	}
		1130
		1131	__aio __attribute__((target("sve"))) svint16x2_t svreinterpret_s16(svfloat64x2_t op) __arm_streaming_compatible {
		1132	return __builtin_sve_reinterpret_s16_f64_x2(op);
		1133	}
		1134
		1135	__aio __attribute__((target("sve"))) svuint16x2_t svreinterpret_u16(svint8x2_t op) __arm_streaming_compatible {
		1136	return __builtin_sve_reinterpret_u16_s8_x2(op);
		1137	}
		1138
		1139	__aio __attribute__((target("sve"))) svuint16x2_t svreinterpret_u16(svuint8x2_t op) __arm_streaming_compatible {
		1140	return __builtin_sve_reinterpret_u16_u8_x2(op);
		1141	}
		1142
		1143	__aio __attribute__((target("sve"))) svuint16x2_t svreinterpret_u16(svint16x2_t op) __arm_streaming_compatible {
		1144	return __builtin_sve_reinterpret_u16_s16_x2(op);
		1145	}
		1146
		1147	__aio __attribute__((target("sve"))) svuint16x2_t svreinterpret_u16(svuint16x2_t op) __arm_streaming_compatible {
		1148	return __builtin_sve_reinterpret_u16_u16_x2(op);
		1149	}
		1150
		1151	__aio __attribute__((target("sve"))) svuint16x2_t svreinterpret_u16(svint32x2_t op) __arm_streaming_compatible {
		1152	return __builtin_sve_reinterpret_u16_s32_x2(op);
		1153	}
		1154
		1155	__aio __attribute__((target("sve"))) svuint16x2_t svreinterpret_u16(svuint32x2_t op) __arm_streaming_compatible {
		1156	return __builtin_sve_reinterpret_u16_u32_x2(op);
		1157	}
		1158
		1159	__aio __attribute__((target("sve"))) svuint16x2_t svreinterpret_u16(svint64x2_t op) __arm_streaming_compatible {
		1160	return __builtin_sve_reinterpret_u16_s64_x2(op);
		1161	}
		1162
		1163	__aio __attribute__((target("sve"))) svuint16x2_t svreinterpret_u16(svuint64x2_t op) __arm_streaming_compatible {
		1164	return __builtin_sve_reinterpret_u16_u64_x2(op);
		1165	}
		1166
		1167	__aio __attribute__((target("sve"))) svuint16x2_t svreinterpret_u16(svfloat16x2_t op) __arm_streaming_compatible {
		1168	return __builtin_sve_reinterpret_u16_f16_x2(op);
		1169	}
		1170
		1171	__aio __attribute__((target("sve"))) svuint16x2_t svreinterpret_u16(svbfloat16x2_t op) __arm_streaming_compatible {
		1172	return __builtin_sve_reinterpret_u16_bf16_x2(op);
		1173	}
		1174
		1175	__aio __attribute__((target("sve"))) svuint16x2_t svreinterpret_u16(svfloat32x2_t op) __arm_streaming_compatible {
		1176	return __builtin_sve_reinterpret_u16_f32_x2(op);
		1177	}
		1178
		1179	__aio __attribute__((target("sve"))) svuint16x2_t svreinterpret_u16(svfloat64x2_t op) __arm_streaming_compatible {
		1180	return __builtin_sve_reinterpret_u16_f64_x2(op);
		1181	}
		1182
		1183	__aio __attribute__((target("sve"))) svint32x2_t svreinterpret_s32(svint8x2_t op) __arm_streaming_compatible {
		1184	return __builtin_sve_reinterpret_s32_s8_x2(op);
		1185	}
		1186
		1187	__aio __attribute__((target("sve"))) svint32x2_t svreinterpret_s32(svuint8x2_t op) __arm_streaming_compatible {
		1188	return __builtin_sve_reinterpret_s32_u8_x2(op);
		1189	}
		1190
		1191	__aio __attribute__((target("sve"))) svint32x2_t svreinterpret_s32(svint16x2_t op) __arm_streaming_compatible {
		1192	return __builtin_sve_reinterpret_s32_s16_x2(op);
		1193	}
		1194
		1195	__aio __attribute__((target("sve"))) svint32x2_t svreinterpret_s32(svuint16x2_t op) __arm_streaming_compatible {
		1196	return __builtin_sve_reinterpret_s32_u16_x2(op);
		1197	}
		1198
		1199	__aio __attribute__((target("sve"))) svint32x2_t svreinterpret_s32(svint32x2_t op) __arm_streaming_compatible {
		1200	return __builtin_sve_reinterpret_s32_s32_x2(op);
		1201	}
		1202
		1203	__aio __attribute__((target("sve"))) svint32x2_t svreinterpret_s32(svuint32x2_t op) __arm_streaming_compatible {
		1204	return __builtin_sve_reinterpret_s32_u32_x2(op);
		1205	}
		1206
		1207	__aio __attribute__((target("sve"))) svint32x2_t svreinterpret_s32(svint64x2_t op) __arm_streaming_compatible {
		1208	return __builtin_sve_reinterpret_s32_s64_x2(op);
		1209	}
		1210
		1211	__aio __attribute__((target("sve"))) svint32x2_t svreinterpret_s32(svuint64x2_t op) __arm_streaming_compatible {
		1212	return __builtin_sve_reinterpret_s32_u64_x2(op);
		1213	}
		1214
		1215	__aio __attribute__((target("sve"))) svint32x2_t svreinterpret_s32(svfloat16x2_t op) __arm_streaming_compatible {
		1216	return __builtin_sve_reinterpret_s32_f16_x2(op);
		1217	}
		1218
		1219	__aio __attribute__((target("sve"))) svint32x2_t svreinterpret_s32(svbfloat16x2_t op) __arm_streaming_compatible {
		1220	return __builtin_sve_reinterpret_s32_bf16_x2(op);
		1221	}
		1222
		1223	__aio __attribute__((target("sve"))) svint32x2_t svreinterpret_s32(svfloat32x2_t op) __arm_streaming_compatible {
		1224	return __builtin_sve_reinterpret_s32_f32_x2(op);
		1225	}
		1226
		1227	__aio __attribute__((target("sve"))) svint32x2_t svreinterpret_s32(svfloat64x2_t op) __arm_streaming_compatible {
		1228	return __builtin_sve_reinterpret_s32_f64_x2(op);
		1229	}
		1230
		1231	__aio __attribute__((target("sve"))) svuint32x2_t svreinterpret_u32(svint8x2_t op) __arm_streaming_compatible {
		1232	return __builtin_sve_reinterpret_u32_s8_x2(op);
		1233	}
		1234
		1235	__aio __attribute__((target("sve"))) svuint32x2_t svreinterpret_u32(svuint8x2_t op) __arm_streaming_compatible {
		1236	return __builtin_sve_reinterpret_u32_u8_x2(op);
		1237	}
		1238
		1239	__aio __attribute__((target("sve"))) svuint32x2_t svreinterpret_u32(svint16x2_t op) __arm_streaming_compatible {
		1240	return __builtin_sve_reinterpret_u32_s16_x2(op);
		1241	}
		1242
		1243	__aio __attribute__((target("sve"))) svuint32x2_t svreinterpret_u32(svuint16x2_t op) __arm_streaming_compatible {
		1244	return __builtin_sve_reinterpret_u32_u16_x2(op);
		1245	}
		1246
		1247	__aio __attribute__((target("sve"))) svuint32x2_t svreinterpret_u32(svint32x2_t op) __arm_streaming_compatible {
		1248	return __builtin_sve_reinterpret_u32_s32_x2(op);
		1249	}
		1250
		1251	__aio __attribute__((target("sve"))) svuint32x2_t svreinterpret_u32(svuint32x2_t op) __arm_streaming_compatible {
		1252	return __builtin_sve_reinterpret_u32_u32_x2(op);
		1253	}
		1254
		1255	__aio __attribute__((target("sve"))) svuint32x2_t svreinterpret_u32(svint64x2_t op) __arm_streaming_compatible {
		1256	return __builtin_sve_reinterpret_u32_s64_x2(op);
		1257	}
		1258
		1259	__aio __attribute__((target("sve"))) svuint32x2_t svreinterpret_u32(svuint64x2_t op) __arm_streaming_compatible {
		1260	return __builtin_sve_reinterpret_u32_u64_x2(op);
		1261	}
		1262
		1263	__aio __attribute__((target("sve"))) svuint32x2_t svreinterpret_u32(svfloat16x2_t op) __arm_streaming_compatible {
		1264	return __builtin_sve_reinterpret_u32_f16_x2(op);
		1265	}
		1266
		1267	__aio __attribute__((target("sve"))) svuint32x2_t svreinterpret_u32(svbfloat16x2_t op) __arm_streaming_compatible {
		1268	return __builtin_sve_reinterpret_u32_bf16_x2(op);
		1269	}
		1270
		1271	__aio __attribute__((target("sve"))) svuint32x2_t svreinterpret_u32(svfloat32x2_t op) __arm_streaming_compatible {
		1272	return __builtin_sve_reinterpret_u32_f32_x2(op);
		1273	}
		1274
		1275	__aio __attribute__((target("sve"))) svuint32x2_t svreinterpret_u32(svfloat64x2_t op) __arm_streaming_compatible {
		1276	return __builtin_sve_reinterpret_u32_f64_x2(op);
		1277	}
		1278
		1279	__aio __attribute__((target("sve"))) svint64x2_t svreinterpret_s64(svint8x2_t op) __arm_streaming_compatible {
		1280	return __builtin_sve_reinterpret_s64_s8_x2(op);
		1281	}
		1282
		1283	__aio __attribute__((target("sve"))) svint64x2_t svreinterpret_s64(svuint8x2_t op) __arm_streaming_compatible {
		1284	return __builtin_sve_reinterpret_s64_u8_x2(op);
		1285	}
		1286
		1287	__aio __attribute__((target("sve"))) svint64x2_t svreinterpret_s64(svint16x2_t op) __arm_streaming_compatible {
		1288	return __builtin_sve_reinterpret_s64_s16_x2(op);
		1289	}
		1290
		1291	__aio __attribute__((target("sve"))) svint64x2_t svreinterpret_s64(svuint16x2_t op) __arm_streaming_compatible {
		1292	return __builtin_sve_reinterpret_s64_u16_x2(op);
		1293	}
		1294
		1295	__aio __attribute__((target("sve"))) svint64x2_t svreinterpret_s64(svint32x2_t op) __arm_streaming_compatible {
		1296	return __builtin_sve_reinterpret_s64_s32_x2(op);
		1297	}
		1298
		1299	__aio __attribute__((target("sve"))) svint64x2_t svreinterpret_s64(svuint32x2_t op) __arm_streaming_compatible {
		1300	return __builtin_sve_reinterpret_s64_u32_x2(op);
		1301	}
		1302
		1303	__aio __attribute__((target("sve"))) svint64x2_t svreinterpret_s64(svint64x2_t op) __arm_streaming_compatible {
		1304	return __builtin_sve_reinterpret_s64_s64_x2(op);
		1305	}
		1306
		1307	__aio __attribute__((target("sve"))) svint64x2_t svreinterpret_s64(svuint64x2_t op) __arm_streaming_compatible {
		1308	return __builtin_sve_reinterpret_s64_u64_x2(op);
		1309	}
		1310
		1311	__aio __attribute__((target("sve"))) svint64x2_t svreinterpret_s64(svfloat16x2_t op) __arm_streaming_compatible {
		1312	return __builtin_sve_reinterpret_s64_f16_x2(op);
		1313	}
		1314
		1315	__aio __attribute__((target("sve"))) svint64x2_t svreinterpret_s64(svbfloat16x2_t op) __arm_streaming_compatible {
		1316	return __builtin_sve_reinterpret_s64_bf16_x2(op);
		1317	}
		1318
		1319	__aio __attribute__((target("sve"))) svint64x2_t svreinterpret_s64(svfloat32x2_t op) __arm_streaming_compatible {
		1320	return __builtin_sve_reinterpret_s64_f32_x2(op);
		1321	}
		1322
		1323	__aio __attribute__((target("sve"))) svint64x2_t svreinterpret_s64(svfloat64x2_t op) __arm_streaming_compatible {
		1324	return __builtin_sve_reinterpret_s64_f64_x2(op);
		1325	}
		1326
		1327	__aio __attribute__((target("sve"))) svuint64x2_t svreinterpret_u64(svint8x2_t op) __arm_streaming_compatible {
		1328	return __builtin_sve_reinterpret_u64_s8_x2(op);
		1329	}
		1330
		1331	__aio __attribute__((target("sve"))) svuint64x2_t svreinterpret_u64(svuint8x2_t op) __arm_streaming_compatible {
		1332	return __builtin_sve_reinterpret_u64_u8_x2(op);
		1333	}
		1334
		1335	__aio __attribute__((target("sve"))) svuint64x2_t svreinterpret_u64(svint16x2_t op) __arm_streaming_compatible {
		1336	return __builtin_sve_reinterpret_u64_s16_x2(op);
		1337	}
		1338
		1339	__aio __attribute__((target("sve"))) svuint64x2_t svreinterpret_u64(svuint16x2_t op) __arm_streaming_compatible {
		1340	return __builtin_sve_reinterpret_u64_u16_x2(op);
		1341	}
		1342
		1343	__aio __attribute__((target("sve"))) svuint64x2_t svreinterpret_u64(svint32x2_t op) __arm_streaming_compatible {
		1344	return __builtin_sve_reinterpret_u64_s32_x2(op);
		1345	}
		1346
		1347	__aio __attribute__((target("sve"))) svuint64x2_t svreinterpret_u64(svuint32x2_t op) __arm_streaming_compatible {
		1348	return __builtin_sve_reinterpret_u64_u32_x2(op);
		1349	}
		1350
		1351	__aio __attribute__((target("sve"))) svuint64x2_t svreinterpret_u64(svint64x2_t op) __arm_streaming_compatible {
		1352	return __builtin_sve_reinterpret_u64_s64_x2(op);
		1353	}
		1354
		1355	__aio __attribute__((target("sve"))) svuint64x2_t svreinterpret_u64(svuint64x2_t op) __arm_streaming_compatible {
		1356	return __builtin_sve_reinterpret_u64_u64_x2(op);
		1357	}
		1358
		1359	__aio __attribute__((target("sve"))) svuint64x2_t svreinterpret_u64(svfloat16x2_t op) __arm_streaming_compatible {
		1360	return __builtin_sve_reinterpret_u64_f16_x2(op);
		1361	}
		1362
		1363	__aio __attribute__((target("sve"))) svuint64x2_t svreinterpret_u64(svbfloat16x2_t op) __arm_streaming_compatible {
		1364	return __builtin_sve_reinterpret_u64_bf16_x2(op);
		1365	}
		1366
		1367	__aio __attribute__((target("sve"))) svuint64x2_t svreinterpret_u64(svfloat32x2_t op) __arm_streaming_compatible {
		1368	return __builtin_sve_reinterpret_u64_f32_x2(op);
		1369	}
		1370
		1371	__aio __attribute__((target("sve"))) svuint64x2_t svreinterpret_u64(svfloat64x2_t op) __arm_streaming_compatible {
		1372	return __builtin_sve_reinterpret_u64_f64_x2(op);
		1373	}
		1374
		1375	__aio __attribute__((target("sve"))) svfloat16x2_t svreinterpret_f16(svint8x2_t op) __arm_streaming_compatible {
		1376	return __builtin_sve_reinterpret_f16_s8_x2(op);
		1377	}
		1378
		1379	__aio __attribute__((target("sve"))) svfloat16x2_t svreinterpret_f16(svuint8x2_t op) __arm_streaming_compatible {
		1380	return __builtin_sve_reinterpret_f16_u8_x2(op);
		1381	}
		1382
		1383	__aio __attribute__((target("sve"))) svfloat16x2_t svreinterpret_f16(svint16x2_t op) __arm_streaming_compatible {
		1384	return __builtin_sve_reinterpret_f16_s16_x2(op);
		1385	}
		1386
		1387	__aio __attribute__((target("sve"))) svfloat16x2_t svreinterpret_f16(svuint16x2_t op) __arm_streaming_compatible {
		1388	return __builtin_sve_reinterpret_f16_u16_x2(op);
		1389	}
		1390
		1391	__aio __attribute__((target("sve"))) svfloat16x2_t svreinterpret_f16(svint32x2_t op) __arm_streaming_compatible {
		1392	return __builtin_sve_reinterpret_f16_s32_x2(op);
		1393	}
		1394
		1395	__aio __attribute__((target("sve"))) svfloat16x2_t svreinterpret_f16(svuint32x2_t op) __arm_streaming_compatible {
		1396	return __builtin_sve_reinterpret_f16_u32_x2(op);
		1397	}
		1398
		1399	__aio __attribute__((target("sve"))) svfloat16x2_t svreinterpret_f16(svint64x2_t op) __arm_streaming_compatible {
		1400	return __builtin_sve_reinterpret_f16_s64_x2(op);
		1401	}
		1402
		1403	__aio __attribute__((target("sve"))) svfloat16x2_t svreinterpret_f16(svuint64x2_t op) __arm_streaming_compatible {
		1404	return __builtin_sve_reinterpret_f16_u64_x2(op);
		1405	}
		1406
		1407	__aio __attribute__((target("sve"))) svfloat16x2_t svreinterpret_f16(svfloat16x2_t op) __arm_streaming_compatible {
		1408	return __builtin_sve_reinterpret_f16_f16_x2(op);
		1409	}
		1410
		1411	__aio __attribute__((target("sve"))) svfloat16x2_t svreinterpret_f16(svbfloat16x2_t op) __arm_streaming_compatible {
		1412	return __builtin_sve_reinterpret_f16_bf16_x2(op);
		1413	}
		1414
		1415	__aio __attribute__((target("sve"))) svfloat16x2_t svreinterpret_f16(svfloat32x2_t op) __arm_streaming_compatible {
		1416	return __builtin_sve_reinterpret_f16_f32_x2(op);
		1417	}
		1418
		1419	__aio __attribute__((target("sve"))) svfloat16x2_t svreinterpret_f16(svfloat64x2_t op) __arm_streaming_compatible {
		1420	return __builtin_sve_reinterpret_f16_f64_x2(op);
		1421	}
		1422
		1423	__aio __attribute__((target("sve"))) svbfloat16x2_t svreinterpret_bf16(svint8x2_t op) __arm_streaming_compatible {
		1424	return __builtin_sve_reinterpret_bf16_s8_x2(op);
		1425	}
		1426
		1427	__aio __attribute__((target("sve"))) svbfloat16x2_t svreinterpret_bf16(svuint8x2_t op) __arm_streaming_compatible {
		1428	return __builtin_sve_reinterpret_bf16_u8_x2(op);
		1429	}
		1430
		1431	__aio __attribute__((target("sve"))) svbfloat16x2_t svreinterpret_bf16(svint16x2_t op) __arm_streaming_compatible {
		1432	return __builtin_sve_reinterpret_bf16_s16_x2(op);
		1433	}
		1434
		1435	__aio __attribute__((target("sve"))) svbfloat16x2_t svreinterpret_bf16(svuint16x2_t op) __arm_streaming_compatible {
		1436	return __builtin_sve_reinterpret_bf16_u16_x2(op);
		1437	}
		1438
		1439	__aio __attribute__((target("sve"))) svbfloat16x2_t svreinterpret_bf16(svint32x2_t op) __arm_streaming_compatible {
		1440	return __builtin_sve_reinterpret_bf16_s32_x2(op);
		1441	}
		1442
		1443	__aio __attribute__((target("sve"))) svbfloat16x2_t svreinterpret_bf16(svuint32x2_t op) __arm_streaming_compatible {
		1444	return __builtin_sve_reinterpret_bf16_u32_x2(op);
		1445	}
		1446
		1447	__aio __attribute__((target("sve"))) svbfloat16x2_t svreinterpret_bf16(svint64x2_t op) __arm_streaming_compatible {
		1448	return __builtin_sve_reinterpret_bf16_s64_x2(op);
		1449	}
		1450
		1451	__aio __attribute__((target("sve"))) svbfloat16x2_t svreinterpret_bf16(svuint64x2_t op) __arm_streaming_compatible {
		1452	return __builtin_sve_reinterpret_bf16_u64_x2(op);
		1453	}
		1454
		1455	__aio __attribute__((target("sve"))) svbfloat16x2_t svreinterpret_bf16(svfloat16x2_t op) __arm_streaming_compatible {
		1456	return __builtin_sve_reinterpret_bf16_f16_x2(op);
		1457	}
		1458
		1459	__aio __attribute__((target("sve"))) svbfloat16x2_t svreinterpret_bf16(svbfloat16x2_t op) __arm_streaming_compatible {
		1460	return __builtin_sve_reinterpret_bf16_bf16_x2(op);
		1461	}
		1462
		1463	__aio __attribute__((target("sve"))) svbfloat16x2_t svreinterpret_bf16(svfloat32x2_t op) __arm_streaming_compatible {
		1464	return __builtin_sve_reinterpret_bf16_f32_x2(op);
		1465	}
		1466
		1467	__aio __attribute__((target("sve"))) svbfloat16x2_t svreinterpret_bf16(svfloat64x2_t op) __arm_streaming_compatible {
		1468	return __builtin_sve_reinterpret_bf16_f64_x2(op);
		1469	}
		1470
		1471	__aio __attribute__((target("sve"))) svfloat32x2_t svreinterpret_f32(svint8x2_t op) __arm_streaming_compatible {
		1472	return __builtin_sve_reinterpret_f32_s8_x2(op);
		1473	}
		1474
		1475	__aio __attribute__((target("sve"))) svfloat32x2_t svreinterpret_f32(svuint8x2_t op) __arm_streaming_compatible {
		1476	return __builtin_sve_reinterpret_f32_u8_x2(op);
		1477	}
		1478
		1479	__aio __attribute__((target("sve"))) svfloat32x2_t svreinterpret_f32(svint16x2_t op) __arm_streaming_compatible {
		1480	return __builtin_sve_reinterpret_f32_s16_x2(op);
		1481	}
		1482
		1483	__aio __attribute__((target("sve"))) svfloat32x2_t svreinterpret_f32(svuint16x2_t op) __arm_streaming_compatible {
		1484	return __builtin_sve_reinterpret_f32_u16_x2(op);
		1485	}
		1486
		1487	__aio __attribute__((target("sve"))) svfloat32x2_t svreinterpret_f32(svint32x2_t op) __arm_streaming_compatible {
		1488	return __builtin_sve_reinterpret_f32_s32_x2(op);
		1489	}
		1490
		1491	__aio __attribute__((target("sve"))) svfloat32x2_t svreinterpret_f32(svuint32x2_t op) __arm_streaming_compatible {
		1492	return __builtin_sve_reinterpret_f32_u32_x2(op);
		1493	}
		1494
		1495	__aio __attribute__((target("sve"))) svfloat32x2_t svreinterpret_f32(svint64x2_t op) __arm_streaming_compatible {
		1496	return __builtin_sve_reinterpret_f32_s64_x2(op);
		1497	}
		1498
		1499	__aio __attribute__((target("sve"))) svfloat32x2_t svreinterpret_f32(svuint64x2_t op) __arm_streaming_compatible {
		1500	return __builtin_sve_reinterpret_f32_u64_x2(op);
		1501	}
		1502
		1503	__aio __attribute__((target("sve"))) svfloat32x2_t svreinterpret_f32(svfloat16x2_t op) __arm_streaming_compatible {
		1504	return __builtin_sve_reinterpret_f32_f16_x2(op);
		1505	}
		1506
		1507	__aio __attribute__((target("sve"))) svfloat32x2_t svreinterpret_f32(svbfloat16x2_t op) __arm_streaming_compatible {
		1508	return __builtin_sve_reinterpret_f32_bf16_x2(op);
		1509	}
		1510
		1511	__aio __attribute__((target("sve"))) svfloat32x2_t svreinterpret_f32(svfloat32x2_t op) __arm_streaming_compatible {
		1512	return __builtin_sve_reinterpret_f32_f32_x2(op);
		1513	}
		1514
		1515	__aio __attribute__((target("sve"))) svfloat32x2_t svreinterpret_f32(svfloat64x2_t op) __arm_streaming_compatible {
		1516	return __builtin_sve_reinterpret_f32_f64_x2(op);
		1517	}
		1518
		1519	__aio __attribute__((target("sve"))) svfloat64x2_t svreinterpret_f64(svint8x2_t op) __arm_streaming_compatible {
		1520	return __builtin_sve_reinterpret_f64_s8_x2(op);
		1521	}
		1522
		1523	__aio __attribute__((target("sve"))) svfloat64x2_t svreinterpret_f64(svuint8x2_t op) __arm_streaming_compatible {
		1524	return __builtin_sve_reinterpret_f64_u8_x2(op);
		1525	}
		1526
		1527	__aio __attribute__((target("sve"))) svfloat64x2_t svreinterpret_f64(svint16x2_t op) __arm_streaming_compatible {
		1528	return __builtin_sve_reinterpret_f64_s16_x2(op);
		1529	}
		1530
		1531	__aio __attribute__((target("sve"))) svfloat64x2_t svreinterpret_f64(svuint16x2_t op) __arm_streaming_compatible {
		1532	return __builtin_sve_reinterpret_f64_u16_x2(op);
		1533	}
		1534
		1535	__aio __attribute__((target("sve"))) svfloat64x2_t svreinterpret_f64(svint32x2_t op) __arm_streaming_compatible {
		1536	return __builtin_sve_reinterpret_f64_s32_x2(op);
		1537	}
		1538
		1539	__aio __attribute__((target("sve"))) svfloat64x2_t svreinterpret_f64(svuint32x2_t op) __arm_streaming_compatible {
		1540	return __builtin_sve_reinterpret_f64_u32_x2(op);
		1541	}
		1542
		1543	__aio __attribute__((target("sve"))) svfloat64x2_t svreinterpret_f64(svint64x2_t op) __arm_streaming_compatible {
		1544	return __builtin_sve_reinterpret_f64_s64_x2(op);
		1545	}
		1546
		1547	__aio __attribute__((target("sve"))) svfloat64x2_t svreinterpret_f64(svuint64x2_t op) __arm_streaming_compatible {
		1548	return __builtin_sve_reinterpret_f64_u64_x2(op);
		1549	}
		1550
		1551	__aio __attribute__((target("sve"))) svfloat64x2_t svreinterpret_f64(svfloat16x2_t op) __arm_streaming_compatible {
		1552	return __builtin_sve_reinterpret_f64_f16_x2(op);
		1553	}
		1554
		1555	__aio __attribute__((target("sve"))) svfloat64x2_t svreinterpret_f64(svbfloat16x2_t op) __arm_streaming_compatible {
		1556	return __builtin_sve_reinterpret_f64_bf16_x2(op);
		1557	}
		1558
		1559	__aio __attribute__((target("sve"))) svfloat64x2_t svreinterpret_f64(svfloat32x2_t op) __arm_streaming_compatible {
		1560	return __builtin_sve_reinterpret_f64_f32_x2(op);
		1561	}
		1562
		1563	__aio __attribute__((target("sve"))) svfloat64x2_t svreinterpret_f64(svfloat64x2_t op) __arm_streaming_compatible {
		1564	return __builtin_sve_reinterpret_f64_f64_x2(op);
		1565	}
		1566
		1567	#define svreinterpret_s8_s8_x3(...) __builtin_sve_reinterpret_s8_s8_x3(__VA_ARGS__)
		1568	#define svreinterpret_s8_u8_x3(...) __builtin_sve_reinterpret_s8_u8_x3(__VA_ARGS__)
		1569	#define svreinterpret_s8_s16_x3(...) __builtin_sve_reinterpret_s8_s16_x3(__VA_ARGS__)
		1570	#define svreinterpret_s8_u16_x3(...) __builtin_sve_reinterpret_s8_u16_x3(__VA_ARGS__)
		1571	#define svreinterpret_s8_s32_x3(...) __builtin_sve_reinterpret_s8_s32_x3(__VA_ARGS__)
		1572	#define svreinterpret_s8_u32_x3(...) __builtin_sve_reinterpret_s8_u32_x3(__VA_ARGS__)
		1573	#define svreinterpret_s8_s64_x3(...) __builtin_sve_reinterpret_s8_s64_x3(__VA_ARGS__)
		1574	#define svreinterpret_s8_u64_x3(...) __builtin_sve_reinterpret_s8_u64_x3(__VA_ARGS__)
		1575	#define svreinterpret_s8_f16_x3(...) __builtin_sve_reinterpret_s8_f16_x3(__VA_ARGS__)
		1576	#define svreinterpret_s8_bf16_x3(...) __builtin_sve_reinterpret_s8_bf16_x3(__VA_ARGS__)
		1577	#define svreinterpret_s8_f32_x3(...) __builtin_sve_reinterpret_s8_f32_x3(__VA_ARGS__)
		1578	#define svreinterpret_s8_f64_x3(...) __builtin_sve_reinterpret_s8_f64_x3(__VA_ARGS__)
		1579	#define svreinterpret_u8_s8_x3(...) __builtin_sve_reinterpret_u8_s8_x3(__VA_ARGS__)
		1580	#define svreinterpret_u8_u8_x3(...) __builtin_sve_reinterpret_u8_u8_x3(__VA_ARGS__)
		1581	#define svreinterpret_u8_s16_x3(...) __builtin_sve_reinterpret_u8_s16_x3(__VA_ARGS__)
		1582	#define svreinterpret_u8_u16_x3(...) __builtin_sve_reinterpret_u8_u16_x3(__VA_ARGS__)
		1583	#define svreinterpret_u8_s32_x3(...) __builtin_sve_reinterpret_u8_s32_x3(__VA_ARGS__)
		1584	#define svreinterpret_u8_u32_x3(...) __builtin_sve_reinterpret_u8_u32_x3(__VA_ARGS__)
		1585	#define svreinterpret_u8_s64_x3(...) __builtin_sve_reinterpret_u8_s64_x3(__VA_ARGS__)
		1586	#define svreinterpret_u8_u64_x3(...) __builtin_sve_reinterpret_u8_u64_x3(__VA_ARGS__)
		1587	#define svreinterpret_u8_f16_x3(...) __builtin_sve_reinterpret_u8_f16_x3(__VA_ARGS__)
		1588	#define svreinterpret_u8_bf16_x3(...) __builtin_sve_reinterpret_u8_bf16_x3(__VA_ARGS__)
		1589	#define svreinterpret_u8_f32_x3(...) __builtin_sve_reinterpret_u8_f32_x3(__VA_ARGS__)
		1590	#define svreinterpret_u8_f64_x3(...) __builtin_sve_reinterpret_u8_f64_x3(__VA_ARGS__)
		1591	#define svreinterpret_s16_s8_x3(...) __builtin_sve_reinterpret_s16_s8_x3(__VA_ARGS__)
		1592	#define svreinterpret_s16_u8_x3(...) __builtin_sve_reinterpret_s16_u8_x3(__VA_ARGS__)
		1593	#define svreinterpret_s16_s16_x3(...) __builtin_sve_reinterpret_s16_s16_x3(__VA_ARGS__)
		1594	#define svreinterpret_s16_u16_x3(...) __builtin_sve_reinterpret_s16_u16_x3(__VA_ARGS__)
		1595	#define svreinterpret_s16_s32_x3(...) __builtin_sve_reinterpret_s16_s32_x3(__VA_ARGS__)
		1596	#define svreinterpret_s16_u32_x3(...) __builtin_sve_reinterpret_s16_u32_x3(__VA_ARGS__)
		1597	#define svreinterpret_s16_s64_x3(...) __builtin_sve_reinterpret_s16_s64_x3(__VA_ARGS__)
		1598	#define svreinterpret_s16_u64_x3(...) __builtin_sve_reinterpret_s16_u64_x3(__VA_ARGS__)
		1599	#define svreinterpret_s16_f16_x3(...) __builtin_sve_reinterpret_s16_f16_x3(__VA_ARGS__)
		1600	#define svreinterpret_s16_bf16_x3(...) __builtin_sve_reinterpret_s16_bf16_x3(__VA_ARGS__)
		1601	#define svreinterpret_s16_f32_x3(...) __builtin_sve_reinterpret_s16_f32_x3(__VA_ARGS__)
		1602	#define svreinterpret_s16_f64_x3(...) __builtin_sve_reinterpret_s16_f64_x3(__VA_ARGS__)
		1603	#define svreinterpret_u16_s8_x3(...) __builtin_sve_reinterpret_u16_s8_x3(__VA_ARGS__)
		1604	#define svreinterpret_u16_u8_x3(...) __builtin_sve_reinterpret_u16_u8_x3(__VA_ARGS__)
		1605	#define svreinterpret_u16_s16_x3(...) __builtin_sve_reinterpret_u16_s16_x3(__VA_ARGS__)
		1606	#define svreinterpret_u16_u16_x3(...) __builtin_sve_reinterpret_u16_u16_x3(__VA_ARGS__)
		1607	#define svreinterpret_u16_s32_x3(...) __builtin_sve_reinterpret_u16_s32_x3(__VA_ARGS__)
		1608	#define svreinterpret_u16_u32_x3(...) __builtin_sve_reinterpret_u16_u32_x3(__VA_ARGS__)
		1609	#define svreinterpret_u16_s64_x3(...) __builtin_sve_reinterpret_u16_s64_x3(__VA_ARGS__)
		1610	#define svreinterpret_u16_u64_x3(...) __builtin_sve_reinterpret_u16_u64_x3(__VA_ARGS__)
		1611	#define svreinterpret_u16_f16_x3(...) __builtin_sve_reinterpret_u16_f16_x3(__VA_ARGS__)
		1612	#define svreinterpret_u16_bf16_x3(...) __builtin_sve_reinterpret_u16_bf16_x3(__VA_ARGS__)
		1613	#define svreinterpret_u16_f32_x3(...) __builtin_sve_reinterpret_u16_f32_x3(__VA_ARGS__)
		1614	#define svreinterpret_u16_f64_x3(...) __builtin_sve_reinterpret_u16_f64_x3(__VA_ARGS__)
		1615	#define svreinterpret_s32_s8_x3(...) __builtin_sve_reinterpret_s32_s8_x3(__VA_ARGS__)
		1616	#define svreinterpret_s32_u8_x3(...) __builtin_sve_reinterpret_s32_u8_x3(__VA_ARGS__)
		1617	#define svreinterpret_s32_s16_x3(...) __builtin_sve_reinterpret_s32_s16_x3(__VA_ARGS__)
		1618	#define svreinterpret_s32_u16_x3(...) __builtin_sve_reinterpret_s32_u16_x3(__VA_ARGS__)
		1619	#define svreinterpret_s32_s32_x3(...) __builtin_sve_reinterpret_s32_s32_x3(__VA_ARGS__)
		1620	#define svreinterpret_s32_u32_x3(...) __builtin_sve_reinterpret_s32_u32_x3(__VA_ARGS__)
		1621	#define svreinterpret_s32_s64_x3(...) __builtin_sve_reinterpret_s32_s64_x3(__VA_ARGS__)
		1622	#define svreinterpret_s32_u64_x3(...) __builtin_sve_reinterpret_s32_u64_x3(__VA_ARGS__)
		1623	#define svreinterpret_s32_f16_x3(...) __builtin_sve_reinterpret_s32_f16_x3(__VA_ARGS__)
		1624	#define svreinterpret_s32_bf16_x3(...) __builtin_sve_reinterpret_s32_bf16_x3(__VA_ARGS__)
		1625	#define svreinterpret_s32_f32_x3(...) __builtin_sve_reinterpret_s32_f32_x3(__VA_ARGS__)
		1626	#define svreinterpret_s32_f64_x3(...) __builtin_sve_reinterpret_s32_f64_x3(__VA_ARGS__)
		1627	#define svreinterpret_u32_s8_x3(...) __builtin_sve_reinterpret_u32_s8_x3(__VA_ARGS__)
		1628	#define svreinterpret_u32_u8_x3(...) __builtin_sve_reinterpret_u32_u8_x3(__VA_ARGS__)
		1629	#define svreinterpret_u32_s16_x3(...) __builtin_sve_reinterpret_u32_s16_x3(__VA_ARGS__)
		1630	#define svreinterpret_u32_u16_x3(...) __builtin_sve_reinterpret_u32_u16_x3(__VA_ARGS__)
		1631	#define svreinterpret_u32_s32_x3(...) __builtin_sve_reinterpret_u32_s32_x3(__VA_ARGS__)
		1632	#define svreinterpret_u32_u32_x3(...) __builtin_sve_reinterpret_u32_u32_x3(__VA_ARGS__)
		1633	#define svreinterpret_u32_s64_x3(...) __builtin_sve_reinterpret_u32_s64_x3(__VA_ARGS__)
		1634	#define svreinterpret_u32_u64_x3(...) __builtin_sve_reinterpret_u32_u64_x3(__VA_ARGS__)
		1635	#define svreinterpret_u32_f16_x3(...) __builtin_sve_reinterpret_u32_f16_x3(__VA_ARGS__)
		1636	#define svreinterpret_u32_bf16_x3(...) __builtin_sve_reinterpret_u32_bf16_x3(__VA_ARGS__)
		1637	#define svreinterpret_u32_f32_x3(...) __builtin_sve_reinterpret_u32_f32_x3(__VA_ARGS__)
		1638	#define svreinterpret_u32_f64_x3(...) __builtin_sve_reinterpret_u32_f64_x3(__VA_ARGS__)
		1639	#define svreinterpret_s64_s8_x3(...) __builtin_sve_reinterpret_s64_s8_x3(__VA_ARGS__)
		1640	#define svreinterpret_s64_u8_x3(...) __builtin_sve_reinterpret_s64_u8_x3(__VA_ARGS__)
		1641	#define svreinterpret_s64_s16_x3(...) __builtin_sve_reinterpret_s64_s16_x3(__VA_ARGS__)
		1642	#define svreinterpret_s64_u16_x3(...) __builtin_sve_reinterpret_s64_u16_x3(__VA_ARGS__)
		1643	#define svreinterpret_s64_s32_x3(...) __builtin_sve_reinterpret_s64_s32_x3(__VA_ARGS__)
		1644	#define svreinterpret_s64_u32_x3(...) __builtin_sve_reinterpret_s64_u32_x3(__VA_ARGS__)
		1645	#define svreinterpret_s64_s64_x3(...) __builtin_sve_reinterpret_s64_s64_x3(__VA_ARGS__)
		1646	#define svreinterpret_s64_u64_x3(...) __builtin_sve_reinterpret_s64_u64_x3(__VA_ARGS__)
		1647	#define svreinterpret_s64_f16_x3(...) __builtin_sve_reinterpret_s64_f16_x3(__VA_ARGS__)
		1648	#define svreinterpret_s64_bf16_x3(...) __builtin_sve_reinterpret_s64_bf16_x3(__VA_ARGS__)
		1649	#define svreinterpret_s64_f32_x3(...) __builtin_sve_reinterpret_s64_f32_x3(__VA_ARGS__)
		1650	#define svreinterpret_s64_f64_x3(...) __builtin_sve_reinterpret_s64_f64_x3(__VA_ARGS__)
		1651	#define svreinterpret_u64_s8_x3(...) __builtin_sve_reinterpret_u64_s8_x3(__VA_ARGS__)
		1652	#define svreinterpret_u64_u8_x3(...) __builtin_sve_reinterpret_u64_u8_x3(__VA_ARGS__)
		1653	#define svreinterpret_u64_s16_x3(...) __builtin_sve_reinterpret_u64_s16_x3(__VA_ARGS__)
		1654	#define svreinterpret_u64_u16_x3(...) __builtin_sve_reinterpret_u64_u16_x3(__VA_ARGS__)
		1655	#define svreinterpret_u64_s32_x3(...) __builtin_sve_reinterpret_u64_s32_x3(__VA_ARGS__)
		1656	#define svreinterpret_u64_u32_x3(...) __builtin_sve_reinterpret_u64_u32_x3(__VA_ARGS__)
		1657	#define svreinterpret_u64_s64_x3(...) __builtin_sve_reinterpret_u64_s64_x3(__VA_ARGS__)
		1658	#define svreinterpret_u64_u64_x3(...) __builtin_sve_reinterpret_u64_u64_x3(__VA_ARGS__)
		1659	#define svreinterpret_u64_f16_x3(...) __builtin_sve_reinterpret_u64_f16_x3(__VA_ARGS__)
		1660	#define svreinterpret_u64_bf16_x3(...) __builtin_sve_reinterpret_u64_bf16_x3(__VA_ARGS__)
		1661	#define svreinterpret_u64_f32_x3(...) __builtin_sve_reinterpret_u64_f32_x3(__VA_ARGS__)
		1662	#define svreinterpret_u64_f64_x3(...) __builtin_sve_reinterpret_u64_f64_x3(__VA_ARGS__)
		1663	#define svreinterpret_f16_s8_x3(...) __builtin_sve_reinterpret_f16_s8_x3(__VA_ARGS__)
		1664	#define svreinterpret_f16_u8_x3(...) __builtin_sve_reinterpret_f16_u8_x3(__VA_ARGS__)
		1665	#define svreinterpret_f16_s16_x3(...) __builtin_sve_reinterpret_f16_s16_x3(__VA_ARGS__)
		1666	#define svreinterpret_f16_u16_x3(...) __builtin_sve_reinterpret_f16_u16_x3(__VA_ARGS__)
		1667	#define svreinterpret_f16_s32_x3(...) __builtin_sve_reinterpret_f16_s32_x3(__VA_ARGS__)
		1668	#define svreinterpret_f16_u32_x3(...) __builtin_sve_reinterpret_f16_u32_x3(__VA_ARGS__)
		1669	#define svreinterpret_f16_s64_x3(...) __builtin_sve_reinterpret_f16_s64_x3(__VA_ARGS__)
		1670	#define svreinterpret_f16_u64_x3(...) __builtin_sve_reinterpret_f16_u64_x3(__VA_ARGS__)
		1671	#define svreinterpret_f16_f16_x3(...) __builtin_sve_reinterpret_f16_f16_x3(__VA_ARGS__)
		1672	#define svreinterpret_f16_bf16_x3(...) __builtin_sve_reinterpret_f16_bf16_x3(__VA_ARGS__)
		1673	#define svreinterpret_f16_f32_x3(...) __builtin_sve_reinterpret_f16_f32_x3(__VA_ARGS__)
		1674	#define svreinterpret_f16_f64_x3(...) __builtin_sve_reinterpret_f16_f64_x3(__VA_ARGS__)
		1675	#define svreinterpret_bf16_s8_x3(...) __builtin_sve_reinterpret_bf16_s8_x3(__VA_ARGS__)
		1676	#define svreinterpret_bf16_u8_x3(...) __builtin_sve_reinterpret_bf16_u8_x3(__VA_ARGS__)
		1677	#define svreinterpret_bf16_s16_x3(...) __builtin_sve_reinterpret_bf16_s16_x3(__VA_ARGS__)
		1678	#define svreinterpret_bf16_u16_x3(...) __builtin_sve_reinterpret_bf16_u16_x3(__VA_ARGS__)
		1679	#define svreinterpret_bf16_s32_x3(...) __builtin_sve_reinterpret_bf16_s32_x3(__VA_ARGS__)
		1680	#define svreinterpret_bf16_u32_x3(...) __builtin_sve_reinterpret_bf16_u32_x3(__VA_ARGS__)
		1681	#define svreinterpret_bf16_s64_x3(...) __builtin_sve_reinterpret_bf16_s64_x3(__VA_ARGS__)
		1682	#define svreinterpret_bf16_u64_x3(...) __builtin_sve_reinterpret_bf16_u64_x3(__VA_ARGS__)
		1683	#define svreinterpret_bf16_f16_x3(...) __builtin_sve_reinterpret_bf16_f16_x3(__VA_ARGS__)
		1684	#define svreinterpret_bf16_bf16_x3(...) __builtin_sve_reinterpret_bf16_bf16_x3(__VA_ARGS__)
		1685	#define svreinterpret_bf16_f32_x3(...) __builtin_sve_reinterpret_bf16_f32_x3(__VA_ARGS__)
		1686	#define svreinterpret_bf16_f64_x3(...) __builtin_sve_reinterpret_bf16_f64_x3(__VA_ARGS__)
		1687	#define svreinterpret_f32_s8_x3(...) __builtin_sve_reinterpret_f32_s8_x3(__VA_ARGS__)
		1688	#define svreinterpret_f32_u8_x3(...) __builtin_sve_reinterpret_f32_u8_x3(__VA_ARGS__)
		1689	#define svreinterpret_f32_s16_x3(...) __builtin_sve_reinterpret_f32_s16_x3(__VA_ARGS__)
		1690	#define svreinterpret_f32_u16_x3(...) __builtin_sve_reinterpret_f32_u16_x3(__VA_ARGS__)
		1691	#define svreinterpret_f32_s32_x3(...) __builtin_sve_reinterpret_f32_s32_x3(__VA_ARGS__)
		1692	#define svreinterpret_f32_u32_x3(...) __builtin_sve_reinterpret_f32_u32_x3(__VA_ARGS__)
		1693	#define svreinterpret_f32_s64_x3(...) __builtin_sve_reinterpret_f32_s64_x3(__VA_ARGS__)
		1694	#define svreinterpret_f32_u64_x3(...) __builtin_sve_reinterpret_f32_u64_x3(__VA_ARGS__)
		1695	#define svreinterpret_f32_f16_x3(...) __builtin_sve_reinterpret_f32_f16_x3(__VA_ARGS__)
		1696	#define svreinterpret_f32_bf16_x3(...) __builtin_sve_reinterpret_f32_bf16_x3(__VA_ARGS__)
		1697	#define svreinterpret_f32_f32_x3(...) __builtin_sve_reinterpret_f32_f32_x3(__VA_ARGS__)
		1698	#define svreinterpret_f32_f64_x3(...) __builtin_sve_reinterpret_f32_f64_x3(__VA_ARGS__)
		1699	#define svreinterpret_f64_s8_x3(...) __builtin_sve_reinterpret_f64_s8_x3(__VA_ARGS__)
		1700	#define svreinterpret_f64_u8_x3(...) __builtin_sve_reinterpret_f64_u8_x3(__VA_ARGS__)
		1701	#define svreinterpret_f64_s16_x3(...) __builtin_sve_reinterpret_f64_s16_x3(__VA_ARGS__)
		1702	#define svreinterpret_f64_u16_x3(...) __builtin_sve_reinterpret_f64_u16_x3(__VA_ARGS__)
		1703	#define svreinterpret_f64_s32_x3(...) __builtin_sve_reinterpret_f64_s32_x3(__VA_ARGS__)
		1704	#define svreinterpret_f64_u32_x3(...) __builtin_sve_reinterpret_f64_u32_x3(__VA_ARGS__)
		1705	#define svreinterpret_f64_s64_x3(...) __builtin_sve_reinterpret_f64_s64_x3(__VA_ARGS__)
		1706	#define svreinterpret_f64_u64_x3(...) __builtin_sve_reinterpret_f64_u64_x3(__VA_ARGS__)
		1707	#define svreinterpret_f64_f16_x3(...) __builtin_sve_reinterpret_f64_f16_x3(__VA_ARGS__)
		1708	#define svreinterpret_f64_bf16_x3(...) __builtin_sve_reinterpret_f64_bf16_x3(__VA_ARGS__)
		1709	#define svreinterpret_f64_f32_x3(...) __builtin_sve_reinterpret_f64_f32_x3(__VA_ARGS__)
		1710	#define svreinterpret_f64_f64_x3(...) __builtin_sve_reinterpret_f64_f64_x3(__VA_ARGS__)
		1711	__aio __attribute__((target("sve"))) svint8x3_t svreinterpret_s8(svint8x3_t op) __arm_streaming_compatible {
		1712	return __builtin_sve_reinterpret_s8_s8_x3(op);
		1713	}
		1714
		1715	__aio __attribute__((target("sve"))) svint8x3_t svreinterpret_s8(svuint8x3_t op) __arm_streaming_compatible {
		1716	return __builtin_sve_reinterpret_s8_u8_x3(op);
		1717	}
		1718
		1719	__aio __attribute__((target("sve"))) svint8x3_t svreinterpret_s8(svint16x3_t op) __arm_streaming_compatible {
		1720	return __builtin_sve_reinterpret_s8_s16_x3(op);
		1721	}
		1722
		1723	__aio __attribute__((target("sve"))) svint8x3_t svreinterpret_s8(svuint16x3_t op) __arm_streaming_compatible {
		1724	return __builtin_sve_reinterpret_s8_u16_x3(op);
		1725	}
		1726
		1727	__aio __attribute__((target("sve"))) svint8x3_t svreinterpret_s8(svint32x3_t op) __arm_streaming_compatible {
		1728	return __builtin_sve_reinterpret_s8_s32_x3(op);
		1729	}
		1730
		1731	__aio __attribute__((target("sve"))) svint8x3_t svreinterpret_s8(svuint32x3_t op) __arm_streaming_compatible {
		1732	return __builtin_sve_reinterpret_s8_u32_x3(op);
		1733	}
		1734
		1735	__aio __attribute__((target("sve"))) svint8x3_t svreinterpret_s8(svint64x3_t op) __arm_streaming_compatible {
		1736	return __builtin_sve_reinterpret_s8_s64_x3(op);
		1737	}
		1738
		1739	__aio __attribute__((target("sve"))) svint8x3_t svreinterpret_s8(svuint64x3_t op) __arm_streaming_compatible {
		1740	return __builtin_sve_reinterpret_s8_u64_x3(op);
		1741	}
		1742
		1743	__aio __attribute__((target("sve"))) svint8x3_t svreinterpret_s8(svfloat16x3_t op) __arm_streaming_compatible {
		1744	return __builtin_sve_reinterpret_s8_f16_x3(op);
		1745	}
		1746
		1747	__aio __attribute__((target("sve"))) svint8x3_t svreinterpret_s8(svbfloat16x3_t op) __arm_streaming_compatible {
		1748	return __builtin_sve_reinterpret_s8_bf16_x3(op);
		1749	}
		1750
		1751	__aio __attribute__((target("sve"))) svint8x3_t svreinterpret_s8(svfloat32x3_t op) __arm_streaming_compatible {
		1752	return __builtin_sve_reinterpret_s8_f32_x3(op);
		1753	}
		1754
		1755	__aio __attribute__((target("sve"))) svint8x3_t svreinterpret_s8(svfloat64x3_t op) __arm_streaming_compatible {
		1756	return __builtin_sve_reinterpret_s8_f64_x3(op);
		1757	}
		1758
		1759	__aio __attribute__((target("sve"))) svuint8x3_t svreinterpret_u8(svint8x3_t op) __arm_streaming_compatible {
		1760	return __builtin_sve_reinterpret_u8_s8_x3(op);
		1761	}
		1762
		1763	__aio __attribute__((target("sve"))) svuint8x3_t svreinterpret_u8(svuint8x3_t op) __arm_streaming_compatible {
		1764	return __builtin_sve_reinterpret_u8_u8_x3(op);
		1765	}
		1766
		1767	__aio __attribute__((target("sve"))) svuint8x3_t svreinterpret_u8(svint16x3_t op) __arm_streaming_compatible {
		1768	return __builtin_sve_reinterpret_u8_s16_x3(op);
		1769	}
		1770
		1771	__aio __attribute__((target("sve"))) svuint8x3_t svreinterpret_u8(svuint16x3_t op) __arm_streaming_compatible {
		1772	return __builtin_sve_reinterpret_u8_u16_x3(op);
		1773	}
		1774
		1775	__aio __attribute__((target("sve"))) svuint8x3_t svreinterpret_u8(svint32x3_t op) __arm_streaming_compatible {
		1776	return __builtin_sve_reinterpret_u8_s32_x3(op);
		1777	}
		1778
		1779	__aio __attribute__((target("sve"))) svuint8x3_t svreinterpret_u8(svuint32x3_t op) __arm_streaming_compatible {
		1780	return __builtin_sve_reinterpret_u8_u32_x3(op);
		1781	}
		1782
		1783	__aio __attribute__((target("sve"))) svuint8x3_t svreinterpret_u8(svint64x3_t op) __arm_streaming_compatible {
		1784	return __builtin_sve_reinterpret_u8_s64_x3(op);
		1785	}
		1786
		1787	__aio __attribute__((target("sve"))) svuint8x3_t svreinterpret_u8(svuint64x3_t op) __arm_streaming_compatible {
		1788	return __builtin_sve_reinterpret_u8_u64_x3(op);
		1789	}
		1790
		1791	__aio __attribute__((target("sve"))) svuint8x3_t svreinterpret_u8(svfloat16x3_t op) __arm_streaming_compatible {
		1792	return __builtin_sve_reinterpret_u8_f16_x3(op);
		1793	}
		1794
		1795	__aio __attribute__((target("sve"))) svuint8x3_t svreinterpret_u8(svbfloat16x3_t op) __arm_streaming_compatible {
		1796	return __builtin_sve_reinterpret_u8_bf16_x3(op);
		1797	}
		1798
		1799	__aio __attribute__((target("sve"))) svuint8x3_t svreinterpret_u8(svfloat32x3_t op) __arm_streaming_compatible {
		1800	return __builtin_sve_reinterpret_u8_f32_x3(op);
		1801	}
		1802
		1803	__aio __attribute__((target("sve"))) svuint8x3_t svreinterpret_u8(svfloat64x3_t op) __arm_streaming_compatible {
		1804	return __builtin_sve_reinterpret_u8_f64_x3(op);
		1805	}
		1806
		1807	__aio __attribute__((target("sve"))) svint16x3_t svreinterpret_s16(svint8x3_t op) __arm_streaming_compatible {
		1808	return __builtin_sve_reinterpret_s16_s8_x3(op);
		1809	}
		1810
		1811	__aio __attribute__((target("sve"))) svint16x3_t svreinterpret_s16(svuint8x3_t op) __arm_streaming_compatible {
		1812	return __builtin_sve_reinterpret_s16_u8_x3(op);
		1813	}
		1814
		1815	__aio __attribute__((target("sve"))) svint16x3_t svreinterpret_s16(svint16x3_t op) __arm_streaming_compatible {
		1816	return __builtin_sve_reinterpret_s16_s16_x3(op);
		1817	}
		1818
		1819	__aio __attribute__((target("sve"))) svint16x3_t svreinterpret_s16(svuint16x3_t op) __arm_streaming_compatible {
		1820	return __builtin_sve_reinterpret_s16_u16_x3(op);
		1821	}
		1822
		1823	__aio __attribute__((target("sve"))) svint16x3_t svreinterpret_s16(svint32x3_t op) __arm_streaming_compatible {
		1824	return __builtin_sve_reinterpret_s16_s32_x3(op);
		1825	}
		1826
		1827	__aio __attribute__((target("sve"))) svint16x3_t svreinterpret_s16(svuint32x3_t op) __arm_streaming_compatible {
		1828	return __builtin_sve_reinterpret_s16_u32_x3(op);
		1829	}
		1830
		1831	__aio __attribute__((target("sve"))) svint16x3_t svreinterpret_s16(svint64x3_t op) __arm_streaming_compatible {
		1832	return __builtin_sve_reinterpret_s16_s64_x3(op);
		1833	}
		1834
		1835	__aio __attribute__((target("sve"))) svint16x3_t svreinterpret_s16(svuint64x3_t op) __arm_streaming_compatible {
		1836	return __builtin_sve_reinterpret_s16_u64_x3(op);
		1837	}
		1838
		1839	__aio __attribute__((target("sve"))) svint16x3_t svreinterpret_s16(svfloat16x3_t op) __arm_streaming_compatible {
		1840	return __builtin_sve_reinterpret_s16_f16_x3(op);
		1841	}
		1842
		1843	__aio __attribute__((target("sve"))) svint16x3_t svreinterpret_s16(svbfloat16x3_t op) __arm_streaming_compatible {
		1844	return __builtin_sve_reinterpret_s16_bf16_x3(op);
		1845	}
		1846
		1847	__aio __attribute__((target("sve"))) svint16x3_t svreinterpret_s16(svfloat32x3_t op) __arm_streaming_compatible {
		1848	return __builtin_sve_reinterpret_s16_f32_x3(op);
		1849	}
		1850
		1851	__aio __attribute__((target("sve"))) svint16x3_t svreinterpret_s16(svfloat64x3_t op) __arm_streaming_compatible {
		1852	return __builtin_sve_reinterpret_s16_f64_x3(op);
		1853	}
		1854
		1855	__aio __attribute__((target("sve"))) svuint16x3_t svreinterpret_u16(svint8x3_t op) __arm_streaming_compatible {
		1856	return __builtin_sve_reinterpret_u16_s8_x3(op);
		1857	}
		1858
		1859	__aio __attribute__((target("sve"))) svuint16x3_t svreinterpret_u16(svuint8x3_t op) __arm_streaming_compatible {
		1860	return __builtin_sve_reinterpret_u16_u8_x3(op);
		1861	}
		1862
		1863	__aio __attribute__((target("sve"))) svuint16x3_t svreinterpret_u16(svint16x3_t op) __arm_streaming_compatible {
		1864	return __builtin_sve_reinterpret_u16_s16_x3(op);
		1865	}
		1866
		1867	__aio __attribute__((target("sve"))) svuint16x3_t svreinterpret_u16(svuint16x3_t op) __arm_streaming_compatible {
		1868	return __builtin_sve_reinterpret_u16_u16_x3(op);
		1869	}
		1870
		1871	__aio __attribute__((target("sve"))) svuint16x3_t svreinterpret_u16(svint32x3_t op) __arm_streaming_compatible {
		1872	return __builtin_sve_reinterpret_u16_s32_x3(op);
		1873	}
		1874
		1875	__aio __attribute__((target("sve"))) svuint16x3_t svreinterpret_u16(svuint32x3_t op) __arm_streaming_compatible {
		1876	return __builtin_sve_reinterpret_u16_u32_x3(op);
		1877	}
		1878
		1879	__aio __attribute__((target("sve"))) svuint16x3_t svreinterpret_u16(svint64x3_t op) __arm_streaming_compatible {
		1880	return __builtin_sve_reinterpret_u16_s64_x3(op);
		1881	}
		1882
		1883	__aio __attribute__((target("sve"))) svuint16x3_t svreinterpret_u16(svuint64x3_t op) __arm_streaming_compatible {
		1884	return __builtin_sve_reinterpret_u16_u64_x3(op);
		1885	}
		1886
		1887	__aio __attribute__((target("sve"))) svuint16x3_t svreinterpret_u16(svfloat16x3_t op) __arm_streaming_compatible {
		1888	return __builtin_sve_reinterpret_u16_f16_x3(op);
		1889	}
		1890
		1891	__aio __attribute__((target("sve"))) svuint16x3_t svreinterpret_u16(svbfloat16x3_t op) __arm_streaming_compatible {
		1892	return __builtin_sve_reinterpret_u16_bf16_x3(op);
		1893	}
		1894
		1895	__aio __attribute__((target("sve"))) svuint16x3_t svreinterpret_u16(svfloat32x3_t op) __arm_streaming_compatible {
		1896	return __builtin_sve_reinterpret_u16_f32_x3(op);
		1897	}
		1898
		1899	__aio __attribute__((target("sve"))) svuint16x3_t svreinterpret_u16(svfloat64x3_t op) __arm_streaming_compatible {
		1900	return __builtin_sve_reinterpret_u16_f64_x3(op);
		1901	}
		1902
		1903	__aio __attribute__((target("sve"))) svint32x3_t svreinterpret_s32(svint8x3_t op) __arm_streaming_compatible {
		1904	return __builtin_sve_reinterpret_s32_s8_x3(op);
		1905	}
		1906
		1907	__aio __attribute__((target("sve"))) svint32x3_t svreinterpret_s32(svuint8x3_t op) __arm_streaming_compatible {
		1908	return __builtin_sve_reinterpret_s32_u8_x3(op);
		1909	}
		1910
		1911	__aio __attribute__((target("sve"))) svint32x3_t svreinterpret_s32(svint16x3_t op) __arm_streaming_compatible {
		1912	return __builtin_sve_reinterpret_s32_s16_x3(op);
		1913	}
		1914
		1915	__aio __attribute__((target("sve"))) svint32x3_t svreinterpret_s32(svuint16x3_t op) __arm_streaming_compatible {
		1916	return __builtin_sve_reinterpret_s32_u16_x3(op);
		1917	}
		1918
		1919	__aio __attribute__((target("sve"))) svint32x3_t svreinterpret_s32(svint32x3_t op) __arm_streaming_compatible {
		1920	return __builtin_sve_reinterpret_s32_s32_x3(op);
		1921	}
		1922
		1923	__aio __attribute__((target("sve"))) svint32x3_t svreinterpret_s32(svuint32x3_t op) __arm_streaming_compatible {
		1924	return __builtin_sve_reinterpret_s32_u32_x3(op);
		1925	}
		1926
		1927	__aio __attribute__((target("sve"))) svint32x3_t svreinterpret_s32(svint64x3_t op) __arm_streaming_compatible {
		1928	return __builtin_sve_reinterpret_s32_s64_x3(op);
		1929	}
		1930
		1931	__aio __attribute__((target("sve"))) svint32x3_t svreinterpret_s32(svuint64x3_t op) __arm_streaming_compatible {
		1932	return __builtin_sve_reinterpret_s32_u64_x3(op);
		1933	}
		1934
		1935	__aio __attribute__((target("sve"))) svint32x3_t svreinterpret_s32(svfloat16x3_t op) __arm_streaming_compatible {
		1936	return __builtin_sve_reinterpret_s32_f16_x3(op);
		1937	}
		1938
		1939	__aio __attribute__((target("sve"))) svint32x3_t svreinterpret_s32(svbfloat16x3_t op) __arm_streaming_compatible {
		1940	return __builtin_sve_reinterpret_s32_bf16_x3(op);
		1941	}
		1942
		1943	__aio __attribute__((target("sve"))) svint32x3_t svreinterpret_s32(svfloat32x3_t op) __arm_streaming_compatible {
		1944	return __builtin_sve_reinterpret_s32_f32_x3(op);
		1945	}
		1946
		1947	__aio __attribute__((target("sve"))) svint32x3_t svreinterpret_s32(svfloat64x3_t op) __arm_streaming_compatible {
		1948	return __builtin_sve_reinterpret_s32_f64_x3(op);
		1949	}
		1950
		1951	__aio __attribute__((target("sve"))) svuint32x3_t svreinterpret_u32(svint8x3_t op) __arm_streaming_compatible {
		1952	return __builtin_sve_reinterpret_u32_s8_x3(op);
		1953	}
		1954
		1955	__aio __attribute__((target("sve"))) svuint32x3_t svreinterpret_u32(svuint8x3_t op) __arm_streaming_compatible {
		1956	return __builtin_sve_reinterpret_u32_u8_x3(op);
		1957	}
		1958
		1959	__aio __attribute__((target("sve"))) svuint32x3_t svreinterpret_u32(svint16x3_t op) __arm_streaming_compatible {
		1960	return __builtin_sve_reinterpret_u32_s16_x3(op);
		1961	}
		1962
		1963	__aio __attribute__((target("sve"))) svuint32x3_t svreinterpret_u32(svuint16x3_t op) __arm_streaming_compatible {
		1964	return __builtin_sve_reinterpret_u32_u16_x3(op);
		1965	}
		1966
		1967	__aio __attribute__((target("sve"))) svuint32x3_t svreinterpret_u32(svint32x3_t op) __arm_streaming_compatible {
		1968	return __builtin_sve_reinterpret_u32_s32_x3(op);
		1969	}
		1970
		1971	__aio __attribute__((target("sve"))) svuint32x3_t svreinterpret_u32(svuint32x3_t op) __arm_streaming_compatible {
		1972	return __builtin_sve_reinterpret_u32_u32_x3(op);
		1973	}
		1974
		1975	__aio __attribute__((target("sve"))) svuint32x3_t svreinterpret_u32(svint64x3_t op) __arm_streaming_compatible {
		1976	return __builtin_sve_reinterpret_u32_s64_x3(op);
		1977	}
		1978
		1979	__aio __attribute__((target("sve"))) svuint32x3_t svreinterpret_u32(svuint64x3_t op) __arm_streaming_compatible {
		1980	return __builtin_sve_reinterpret_u32_u64_x3(op);
		1981	}
		1982
		1983	__aio __attribute__((target("sve"))) svuint32x3_t svreinterpret_u32(svfloat16x3_t op) __arm_streaming_compatible {
		1984	return __builtin_sve_reinterpret_u32_f16_x3(op);
		1985	}
		1986
		1987	__aio __attribute__((target("sve"))) svuint32x3_t svreinterpret_u32(svbfloat16x3_t op) __arm_streaming_compatible {
		1988	return __builtin_sve_reinterpret_u32_bf16_x3(op);
		1989	}
		1990
		1991	__aio __attribute__((target("sve"))) svuint32x3_t svreinterpret_u32(svfloat32x3_t op) __arm_streaming_compatible {
		1992	return __builtin_sve_reinterpret_u32_f32_x3(op);
		1993	}
		1994
		1995	__aio __attribute__((target("sve"))) svuint32x3_t svreinterpret_u32(svfloat64x3_t op) __arm_streaming_compatible {
		1996	return __builtin_sve_reinterpret_u32_f64_x3(op);
		1997	}
		1998
		1999	__aio __attribute__((target("sve"))) svint64x3_t svreinterpret_s64(svint8x3_t op) __arm_streaming_compatible {
		2000	return __builtin_sve_reinterpret_s64_s8_x3(op);
		2001	}
		2002
		2003	__aio __attribute__((target("sve"))) svint64x3_t svreinterpret_s64(svuint8x3_t op) __arm_streaming_compatible {
		2004	return __builtin_sve_reinterpret_s64_u8_x3(op);
		2005	}
		2006
		2007	__aio __attribute__((target("sve"))) svint64x3_t svreinterpret_s64(svint16x3_t op) __arm_streaming_compatible {
		2008	return __builtin_sve_reinterpret_s64_s16_x3(op);
		2009	}
		2010
		2011	__aio __attribute__((target("sve"))) svint64x3_t svreinterpret_s64(svuint16x3_t op) __arm_streaming_compatible {
		2012	return __builtin_sve_reinterpret_s64_u16_x3(op);
		2013	}
		2014
		2015	__aio __attribute__((target("sve"))) svint64x3_t svreinterpret_s64(svint32x3_t op) __arm_streaming_compatible {
		2016	return __builtin_sve_reinterpret_s64_s32_x3(op);
		2017	}
		2018
		2019	__aio __attribute__((target("sve"))) svint64x3_t svreinterpret_s64(svuint32x3_t op) __arm_streaming_compatible {
		2020	return __builtin_sve_reinterpret_s64_u32_x3(op);
		2021	}
		2022
		2023	__aio __attribute__((target("sve"))) svint64x3_t svreinterpret_s64(svint64x3_t op) __arm_streaming_compatible {
		2024	return __builtin_sve_reinterpret_s64_s64_x3(op);
		2025	}
		2026
		2027	__aio __attribute__((target("sve"))) svint64x3_t svreinterpret_s64(svuint64x3_t op) __arm_streaming_compatible {
		2028	return __builtin_sve_reinterpret_s64_u64_x3(op);
		2029	}
		2030
		2031	__aio __attribute__((target("sve"))) svint64x3_t svreinterpret_s64(svfloat16x3_t op) __arm_streaming_compatible {
		2032	return __builtin_sve_reinterpret_s64_f16_x3(op);
		2033	}
		2034
		2035	__aio __attribute__((target("sve"))) svint64x3_t svreinterpret_s64(svbfloat16x3_t op) __arm_streaming_compatible {
		2036	return __builtin_sve_reinterpret_s64_bf16_x3(op);
		2037	}
		2038
		2039	__aio __attribute__((target("sve"))) svint64x3_t svreinterpret_s64(svfloat32x3_t op) __arm_streaming_compatible {
		2040	return __builtin_sve_reinterpret_s64_f32_x3(op);
		2041	}
		2042
		2043	__aio __attribute__((target("sve"))) svint64x3_t svreinterpret_s64(svfloat64x3_t op) __arm_streaming_compatible {
		2044	return __builtin_sve_reinterpret_s64_f64_x3(op);
		2045	}
		2046
		2047	__aio __attribute__((target("sve"))) svuint64x3_t svreinterpret_u64(svint8x3_t op) __arm_streaming_compatible {
		2048	return __builtin_sve_reinterpret_u64_s8_x3(op);
		2049	}
		2050
		2051	__aio __attribute__((target("sve"))) svuint64x3_t svreinterpret_u64(svuint8x3_t op) __arm_streaming_compatible {
		2052	return __builtin_sve_reinterpret_u64_u8_x3(op);
		2053	}
		2054
		2055	__aio __attribute__((target("sve"))) svuint64x3_t svreinterpret_u64(svint16x3_t op) __arm_streaming_compatible {
		2056	return __builtin_sve_reinterpret_u64_s16_x3(op);
		2057	}
		2058
		2059	__aio __attribute__((target("sve"))) svuint64x3_t svreinterpret_u64(svuint16x3_t op) __arm_streaming_compatible {
		2060	return __builtin_sve_reinterpret_u64_u16_x3(op);
		2061	}
		2062
		2063	__aio __attribute__((target("sve"))) svuint64x3_t svreinterpret_u64(svint32x3_t op) __arm_streaming_compatible {
		2064	return __builtin_sve_reinterpret_u64_s32_x3(op);
		2065	}
		2066
		2067	__aio __attribute__((target("sve"))) svuint64x3_t svreinterpret_u64(svuint32x3_t op) __arm_streaming_compatible {
		2068	return __builtin_sve_reinterpret_u64_u32_x3(op);
276	}	2069	}
277		2070
278	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svint32_t op) {	2071	__aio __attribute__((target("sve"))) svuint64x3_t svreinterpret_u64(svint64x3_t op) __arm_streaming_compatible {
279	return __builtin_sve_reinterpret_s8_s32(op);	2072	return __builtin_sve_reinterpret_u64_s64_x3(op);
280	}	2073	}
281		2074
282	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svint64_t op) {	2075	__aio __attribute__((target("sve"))) svuint64x3_t svreinterpret_u64(svuint64x3_t op) __arm_streaming_compatible {
283	return __builtin_sve_reinterpret_s8_s64(op);	2076	return __builtin_sve_reinterpret_u64_u64_x3(op);
284	}	2077	}
285		2078
286	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svuint8_t op) {	2079	__aio __attribute__((target("sve"))) svuint64x3_t svreinterpret_u64(svfloat16x3_t op) __arm_streaming_compatible {
287	return __builtin_sve_reinterpret_s8_u8(op);	2080	return __builtin_sve_reinterpret_u64_f16_x3(op);
288	}	2081	}
289		2082
290	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svuint16_t op) {	2083	__aio __attribute__((target("sve"))) svuint64x3_t svreinterpret_u64(svbfloat16x3_t op) __arm_streaming_compatible {
291	return __builtin_sve_reinterpret_s8_u16(op);	2084	return __builtin_sve_reinterpret_u64_bf16_x3(op);
292	}	2085	}
293		2086
294	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svuint32_t op) {	2087	__aio __attribute__((target("sve"))) svuint64x3_t svreinterpret_u64(svfloat32x3_t op) __arm_streaming_compatible {
295	return __builtin_sve_reinterpret_s8_u32(op);	2088	return __builtin_sve_reinterpret_u64_f32_x3(op);
296	}	2089	}
297		2090
298	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svuint64_t op) {	2091	__aio __attribute__((target("sve"))) svuint64x3_t svreinterpret_u64(svfloat64x3_t op) __arm_streaming_compatible {
299	return __builtin_sve_reinterpret_s8_u64(op);	2092	return __builtin_sve_reinterpret_u64_f64_x3(op);
300	}	2093	}
301		2094
302	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svfloat16_t op) {	2095	__aio __attribute__((target("sve"))) svfloat16x3_t svreinterpret_f16(svint8x3_t op) __arm_streaming_compatible {
303	return __builtin_sve_reinterpret_s8_f16(op);	2096	return __builtin_sve_reinterpret_f16_s8_x3(op);
304	}	2097	}
305		2098
306	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svbfloat16_t op) {	2099	__aio __attribute__((target("sve"))) svfloat16x3_t svreinterpret_f16(svuint8x3_t op) __arm_streaming_compatible {
307	return __builtin_sve_reinterpret_s8_bf16(op);	2100	return __builtin_sve_reinterpret_f16_u8_x3(op);
308	}	2101	}
309		2102
310	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svfloat32_t op) {	2103	__aio __attribute__((target("sve"))) svfloat16x3_t svreinterpret_f16(svint16x3_t op) __arm_streaming_compatible {
311	return __builtin_sve_reinterpret_s8_f32(op);	2104	return __builtin_sve_reinterpret_f16_s16_x3(op);
312	}	2105	}
313		2106
314	__aio __attribute__((target("sve"))) svint8_t svreinterpret_s8(svfloat64_t op) {	2107	__aio __attribute__((target("sve"))) svfloat16x3_t svreinterpret_f16(svuint16x3_t op) __arm_streaming_compatible {
315	return __builtin_sve_reinterpret_s8_f64(op);	2108	return __builtin_sve_reinterpret_f16_u16_x3(op);
316	}	2109	}
317		2110
318	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svint8_t op) {	2111	__aio __attribute__((target("sve"))) svfloat16x3_t svreinterpret_f16(svint32x3_t op) __arm_streaming_compatible {
319	return __builtin_sve_reinterpret_s16_s8(op);	2112	return __builtin_sve_reinterpret_f16_s32_x3(op);
320	}	2113	}
321		2114
322	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svint16_t op) {	2115	__aio __attribute__((target("sve"))) svfloat16x3_t svreinterpret_f16(svuint32x3_t op) __arm_streaming_compatible {
323	return __builtin_sve_reinterpret_s16_s16(op);	2116	return __builtin_sve_reinterpret_f16_u32_x3(op);
324	}	2117	}
325		2118
326	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svint32_t op) {	2119	__aio __attribute__((target("sve"))) svfloat16x3_t svreinterpret_f16(svint64x3_t op) __arm_streaming_compatible {
327	return __builtin_sve_reinterpret_s16_s32(op);	2120	return __builtin_sve_reinterpret_f16_s64_x3(op);
328	}	2121	}
329		2122
330	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svint64_t op) {	2123	__aio __attribute__((target("sve"))) svfloat16x3_t svreinterpret_f16(svuint64x3_t op) __arm_streaming_compatible {
331	return __builtin_sve_reinterpret_s16_s64(op);	2124	return __builtin_sve_reinterpret_f16_u64_x3(op);
332	}	2125	}
333		2126
334	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svuint8_t op) {	2127	__aio __attribute__((target("sve"))) svfloat16x3_t svreinterpret_f16(svfloat16x3_t op) __arm_streaming_compatible {
335	return __builtin_sve_reinterpret_s16_u8(op);	2128	return __builtin_sve_reinterpret_f16_f16_x3(op);
336	}	2129	}
337		2130
338	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svuint16_t op) {	2131	__aio __attribute__((target("sve"))) svfloat16x3_t svreinterpret_f16(svbfloat16x3_t op) __arm_streaming_compatible {
339	return __builtin_sve_reinterpret_s16_u16(op);	2132	return __builtin_sve_reinterpret_f16_bf16_x3(op);
340	}	2133	}
341		2134
342	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svuint32_t op) {	2135	__aio __attribute__((target("sve"))) svfloat16x3_t svreinterpret_f16(svfloat32x3_t op) __arm_streaming_compatible {
343	return __builtin_sve_reinterpret_s16_u32(op);	2136	return __builtin_sve_reinterpret_f16_f32_x3(op);
344	}	2137	}
345		2138
346	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svuint64_t op) {	2139	__aio __attribute__((target("sve"))) svfloat16x3_t svreinterpret_f16(svfloat64x3_t op) __arm_streaming_compatible {
347	return __builtin_sve_reinterpret_s16_u64(op);	2140	return __builtin_sve_reinterpret_f16_f64_x3(op);
348	}	2141	}
349		2142
350	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svfloat16_t op) {	2143	__aio __attribute__((target("sve"))) svbfloat16x3_t svreinterpret_bf16(svint8x3_t op) __arm_streaming_compatible {
351	return __builtin_sve_reinterpret_s16_f16(op);	2144	return __builtin_sve_reinterpret_bf16_s8_x3(op);
352	}	2145	}
353		2146
354	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svbfloat16_t op) {	2147	__aio __attribute__((target("sve"))) svbfloat16x3_t svreinterpret_bf16(svuint8x3_t op) __arm_streaming_compatible {
355	return __builtin_sve_reinterpret_s16_bf16(op);	2148	return __builtin_sve_reinterpret_bf16_u8_x3(op);
356	}	2149	}
357		2150
358	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svfloat32_t op) {	2151	__aio __attribute__((target("sve"))) svbfloat16x3_t svreinterpret_bf16(svint16x3_t op) __arm_streaming_compatible {
359	return __builtin_sve_reinterpret_s16_f32(op);	2152	return __builtin_sve_reinterpret_bf16_s16_x3(op);
360	}	2153	}
361		2154
362	__aio __attribute__((target("sve"))) svint16_t svreinterpret_s16(svfloat64_t op) {	2155	__aio __attribute__((target("sve"))) svbfloat16x3_t svreinterpret_bf16(svuint16x3_t op) __arm_streaming_compatible {
363	return __builtin_sve_reinterpret_s16_f64(op);	2156	return __builtin_sve_reinterpret_bf16_u16_x3(op);
364	}	2157	}
365		2158
366	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svint8_t op) {	2159	__aio __attribute__((target("sve"))) svbfloat16x3_t svreinterpret_bf16(svint32x3_t op) __arm_streaming_compatible {
367	return __builtin_sve_reinterpret_s32_s8(op);	2160	return __builtin_sve_reinterpret_bf16_s32_x3(op);
368	}	2161	}
369		2162
370	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svint16_t op) {	2163	__aio __attribute__((target("sve"))) svbfloat16x3_t svreinterpret_bf16(svuint32x3_t op) __arm_streaming_compatible {
371	return __builtin_sve_reinterpret_s32_s16(op);	2164	return __builtin_sve_reinterpret_bf16_u32_x3(op);
372	}	2165	}
373		2166
374	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svint32_t op) {	2167	__aio __attribute__((target("sve"))) svbfloat16x3_t svreinterpret_bf16(svint64x3_t op) __arm_streaming_compatible {
375	return __builtin_sve_reinterpret_s32_s32(op);	2168	return __builtin_sve_reinterpret_bf16_s64_x3(op);
376	}	2169	}
377		2170
378	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svint64_t op) {	2171	__aio __attribute__((target("sve"))) svbfloat16x3_t svreinterpret_bf16(svuint64x3_t op) __arm_streaming_compatible {
379	return __builtin_sve_reinterpret_s32_s64(op);	2172	return __builtin_sve_reinterpret_bf16_u64_x3(op);
380	}	2173	}
381		2174
382	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svuint8_t op) {	2175	__aio __attribute__((target("sve"))) svbfloat16x3_t svreinterpret_bf16(svfloat16x3_t op) __arm_streaming_compatible {
383	return __builtin_sve_reinterpret_s32_u8(op);	2176	return __builtin_sve_reinterpret_bf16_f16_x3(op);
384	}	2177	}
385		2178
386	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svuint16_t op) {	2179	__aio __attribute__((target("sve"))) svbfloat16x3_t svreinterpret_bf16(svbfloat16x3_t op) __arm_streaming_compatible {
387	return __builtin_sve_reinterpret_s32_u16(op);	2180	return __builtin_sve_reinterpret_bf16_bf16_x3(op);
388	}	2181	}
389		2182
390	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svuint32_t op) {	2183	__aio __attribute__((target("sve"))) svbfloat16x3_t svreinterpret_bf16(svfloat32x3_t op) __arm_streaming_compatible {
391	return __builtin_sve_reinterpret_s32_u32(op);	2184	return __builtin_sve_reinterpret_bf16_f32_x3(op);
392	}	2185	}
393		2186
394	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svuint64_t op) {	2187	__aio __attribute__((target("sve"))) svbfloat16x3_t svreinterpret_bf16(svfloat64x3_t op) __arm_streaming_compatible {
395	return __builtin_sve_reinterpret_s32_u64(op);	2188	return __builtin_sve_reinterpret_bf16_f64_x3(op);
396	}	2189	}
397		2190
398	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svfloat16_t op) {	2191	__aio __attribute__((target("sve"))) svfloat32x3_t svreinterpret_f32(svint8x3_t op) __arm_streaming_compatible {
399	return __builtin_sve_reinterpret_s32_f16(op);	2192	return __builtin_sve_reinterpret_f32_s8_x3(op);
400	}	2193	}
401		2194
402	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svbfloat16_t op) {	2195	__aio __attribute__((target("sve"))) svfloat32x3_t svreinterpret_f32(svuint8x3_t op) __arm_streaming_compatible {
403	return __builtin_sve_reinterpret_s32_bf16(op);	2196	return __builtin_sve_reinterpret_f32_u8_x3(op);
404	}	2197	}
405		2198
406	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svfloat32_t op) {	2199	__aio __attribute__((target("sve"))) svfloat32x3_t svreinterpret_f32(svint16x3_t op) __arm_streaming_compatible {
407	return __builtin_sve_reinterpret_s32_f32(op);	2200	return __builtin_sve_reinterpret_f32_s16_x3(op);
408	}	2201	}
409		2202
410	__aio __attribute__((target("sve"))) svint32_t svreinterpret_s32(svfloat64_t op) {	2203	__aio __attribute__((target("sve"))) svfloat32x3_t svreinterpret_f32(svuint16x3_t op) __arm_streaming_compatible {
411	return __builtin_sve_reinterpret_s32_f64(op);	2204	return __builtin_sve_reinterpret_f32_u16_x3(op);
412	}	2205	}
413		2206
414	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svint8_t op) {	2207	__aio __attribute__((target("sve"))) svfloat32x3_t svreinterpret_f32(svint32x3_t op) __arm_streaming_compatible {
415	return __builtin_sve_reinterpret_s64_s8(op);	2208	return __builtin_sve_reinterpret_f32_s32_x3(op);
416	}	2209	}
417		2210
418	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svint16_t op) {	2211	__aio __attribute__((target("sve"))) svfloat32x3_t svreinterpret_f32(svuint32x3_t op) __arm_streaming_compatible {
419	return __builtin_sve_reinterpret_s64_s16(op);	2212	return __builtin_sve_reinterpret_f32_u32_x3(op);
420	}	2213	}
421		2214
422	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svint32_t op) {	2215	__aio __attribute__((target("sve"))) svfloat32x3_t svreinterpret_f32(svint64x3_t op) __arm_streaming_compatible {
423	return __builtin_sve_reinterpret_s64_s32(op);	2216	return __builtin_sve_reinterpret_f32_s64_x3(op);
424	}	2217	}
425		2218
426	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svint64_t op) {	2219	__aio __attribute__((target("sve"))) svfloat32x3_t svreinterpret_f32(svuint64x3_t op) __arm_streaming_compatible {
427	return __builtin_sve_reinterpret_s64_s64(op);	2220	return __builtin_sve_reinterpret_f32_u64_x3(op);
428	}	2221	}
429		2222
430	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svuint8_t op) {	2223	__aio __attribute__((target("sve"))) svfloat32x3_t svreinterpret_f32(svfloat16x3_t op) __arm_streaming_compatible {
431	return __builtin_sve_reinterpret_s64_u8(op);	2224	return __builtin_sve_reinterpret_f32_f16_x3(op);
432	}	2225	}
433		2226
434	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svuint16_t op) {	2227	__aio __attribute__((target("sve"))) svfloat32x3_t svreinterpret_f32(svbfloat16x3_t op) __arm_streaming_compatible {
435	return __builtin_sve_reinterpret_s64_u16(op);	2228	return __builtin_sve_reinterpret_f32_bf16_x3(op);
436	}	2229	}
437		2230
438	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svuint32_t op) {	2231	__aio __attribute__((target("sve"))) svfloat32x3_t svreinterpret_f32(svfloat32x3_t op) __arm_streaming_compatible {
439	return __builtin_sve_reinterpret_s64_u32(op);	2232	return __builtin_sve_reinterpret_f32_f32_x3(op);
440	}	2233	}
441		2234
442	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svuint64_t op) {	2235	__aio __attribute__((target("sve"))) svfloat32x3_t svreinterpret_f32(svfloat64x3_t op) __arm_streaming_compatible {
443	return __builtin_sve_reinterpret_s64_u64(op);	2236	return __builtin_sve_reinterpret_f32_f64_x3(op);
444	}	2237	}
445		2238
446	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svfloat16_t op) {	2239	__aio __attribute__((target("sve"))) svfloat64x3_t svreinterpret_f64(svint8x3_t op) __arm_streaming_compatible {
447	return __builtin_sve_reinterpret_s64_f16(op);	2240	return __builtin_sve_reinterpret_f64_s8_x3(op);
448	}	2241	}
449		2242
450	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svbfloat16_t op) {	2243	__aio __attribute__((target("sve"))) svfloat64x3_t svreinterpret_f64(svuint8x3_t op) __arm_streaming_compatible {
451	return __builtin_sve_reinterpret_s64_bf16(op);	2244	return __builtin_sve_reinterpret_f64_u8_x3(op);
452	}	2245	}
453		2246
454	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svfloat32_t op) {	2247	__aio __attribute__((target("sve"))) svfloat64x3_t svreinterpret_f64(svint16x3_t op) __arm_streaming_compatible {
455	return __builtin_sve_reinterpret_s64_f32(op);	2248	return __builtin_sve_reinterpret_f64_s16_x3(op);
456	}	2249	}
457		2250
458	__aio __attribute__((target("sve"))) svint64_t svreinterpret_s64(svfloat64_t op) {	2251	__aio __attribute__((target("sve"))) svfloat64x3_t svreinterpret_f64(svuint16x3_t op) __arm_streaming_compatible {
459	return __builtin_sve_reinterpret_s64_f64(op);	2252	return __builtin_sve_reinterpret_f64_u16_x3(op);
460	}	2253	}
461		2254
462	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svint8_t op) {	2255	__aio __attribute__((target("sve"))) svfloat64x3_t svreinterpret_f64(svint32x3_t op) __arm_streaming_compatible {
463	return __builtin_sve_reinterpret_u8_s8(op);	2256	return __builtin_sve_reinterpret_f64_s32_x3(op);
464	}	2257	}
465		2258
466	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svint16_t op) {	2259	__aio __attribute__((target("sve"))) svfloat64x3_t svreinterpret_f64(svuint32x3_t op) __arm_streaming_compatible {
467	return __builtin_sve_reinterpret_u8_s16(op);	2260	return __builtin_sve_reinterpret_f64_u32_x3(op);
468	}	2261	}
469		2262
470	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svint32_t op) {	2263	__aio __attribute__((target("sve"))) svfloat64x3_t svreinterpret_f64(svint64x3_t op) __arm_streaming_compatible {
471	return __builtin_sve_reinterpret_u8_s32(op);	2264	return __builtin_sve_reinterpret_f64_s64_x3(op);
472	}	2265	}
473		2266
474	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svint64_t op) {	2267	__aio __attribute__((target("sve"))) svfloat64x3_t svreinterpret_f64(svuint64x3_t op) __arm_streaming_compatible {
475	return __builtin_sve_reinterpret_u8_s64(op);	2268	return __builtin_sve_reinterpret_f64_u64_x3(op);
476	}	2269	}
477		2270
478	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svuint8_t op) {	2271	__aio __attribute__((target("sve"))) svfloat64x3_t svreinterpret_f64(svfloat16x3_t op) __arm_streaming_compatible {
479	return __builtin_sve_reinterpret_u8_u8(op);	2272	return __builtin_sve_reinterpret_f64_f16_x3(op);
480	}	2273	}
481		2274
482	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svuint16_t op) {	2275	__aio __attribute__((target("sve"))) svfloat64x3_t svreinterpret_f64(svbfloat16x3_t op) __arm_streaming_compatible {
483	return __builtin_sve_reinterpret_u8_u16(op);	2276	return __builtin_sve_reinterpret_f64_bf16_x3(op);
484	}	2277	}
485		2278
486	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svuint32_t op) {	2279	__aio __attribute__((target("sve"))) svfloat64x3_t svreinterpret_f64(svfloat32x3_t op) __arm_streaming_compatible {
487	return __builtin_sve_reinterpret_u8_u32(op);	2280	return __builtin_sve_reinterpret_f64_f32_x3(op);
488	}	2281	}
489		2282
490	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svuint64_t op) {	2283	__aio __attribute__((target("sve"))) svfloat64x3_t svreinterpret_f64(svfloat64x3_t op) __arm_streaming_compatible {
491	return __builtin_sve_reinterpret_u8_u64(op);	2284	return __builtin_sve_reinterpret_f64_f64_x3(op);
492	}	2285	}
493		2286
494	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svfloat16_t op) {	2287	#define svreinterpret_s8_s8_x4(...) __builtin_sve_reinterpret_s8_s8_x4(__VA_ARGS__)
495	return __builtin_sve_reinterpret_u8_f16(op);	2288	#define svreinterpret_s8_u8_x4(...) __builtin_sve_reinterpret_s8_u8_x4(__VA_ARGS__)
		2289	#define svreinterpret_s8_s16_x4(...) __builtin_sve_reinterpret_s8_s16_x4(__VA_ARGS__)
		2290	#define svreinterpret_s8_u16_x4(...) __builtin_sve_reinterpret_s8_u16_x4(__VA_ARGS__)
		2291	#define svreinterpret_s8_s32_x4(...) __builtin_sve_reinterpret_s8_s32_x4(__VA_ARGS__)
		2292	#define svreinterpret_s8_u32_x4(...) __builtin_sve_reinterpret_s8_u32_x4(__VA_ARGS__)
		2293	#define svreinterpret_s8_s64_x4(...) __builtin_sve_reinterpret_s8_s64_x4(__VA_ARGS__)
		2294	#define svreinterpret_s8_u64_x4(...) __builtin_sve_reinterpret_s8_u64_x4(__VA_ARGS__)
		2295	#define svreinterpret_s8_f16_x4(...) __builtin_sve_reinterpret_s8_f16_x4(__VA_ARGS__)
		2296	#define svreinterpret_s8_bf16_x4(...) __builtin_sve_reinterpret_s8_bf16_x4(__VA_ARGS__)
		2297	#define svreinterpret_s8_f32_x4(...) __builtin_sve_reinterpret_s8_f32_x4(__VA_ARGS__)
		2298	#define svreinterpret_s8_f64_x4(...) __builtin_sve_reinterpret_s8_f64_x4(__VA_ARGS__)
		2299	#define svreinterpret_u8_s8_x4(...) __builtin_sve_reinterpret_u8_s8_x4(__VA_ARGS__)
		2300	#define svreinterpret_u8_u8_x4(...) __builtin_sve_reinterpret_u8_u8_x4(__VA_ARGS__)
		2301	#define svreinterpret_u8_s16_x4(...) __builtin_sve_reinterpret_u8_s16_x4(__VA_ARGS__)
		2302	#define svreinterpret_u8_u16_x4(...) __builtin_sve_reinterpret_u8_u16_x4(__VA_ARGS__)
		2303	#define svreinterpret_u8_s32_x4(...) __builtin_sve_reinterpret_u8_s32_x4(__VA_ARGS__)
		2304	#define svreinterpret_u8_u32_x4(...) __builtin_sve_reinterpret_u8_u32_x4(__VA_ARGS__)
		2305	#define svreinterpret_u8_s64_x4(...) __builtin_sve_reinterpret_u8_s64_x4(__VA_ARGS__)
		2306	#define svreinterpret_u8_u64_x4(...) __builtin_sve_reinterpret_u8_u64_x4(__VA_ARGS__)
		2307	#define svreinterpret_u8_f16_x4(...) __builtin_sve_reinterpret_u8_f16_x4(__VA_ARGS__)
		2308	#define svreinterpret_u8_bf16_x4(...) __builtin_sve_reinterpret_u8_bf16_x4(__VA_ARGS__)
		2309	#define svreinterpret_u8_f32_x4(...) __builtin_sve_reinterpret_u8_f32_x4(__VA_ARGS__)
		2310	#define svreinterpret_u8_f64_x4(...) __builtin_sve_reinterpret_u8_f64_x4(__VA_ARGS__)
		2311	#define svreinterpret_s16_s8_x4(...) __builtin_sve_reinterpret_s16_s8_x4(__VA_ARGS__)
		2312	#define svreinterpret_s16_u8_x4(...) __builtin_sve_reinterpret_s16_u8_x4(__VA_ARGS__)
		2313	#define svreinterpret_s16_s16_x4(...) __builtin_sve_reinterpret_s16_s16_x4(__VA_ARGS__)
		2314	#define svreinterpret_s16_u16_x4(...) __builtin_sve_reinterpret_s16_u16_x4(__VA_ARGS__)
		2315	#define svreinterpret_s16_s32_x4(...) __builtin_sve_reinterpret_s16_s32_x4(__VA_ARGS__)
		2316	#define svreinterpret_s16_u32_x4(...) __builtin_sve_reinterpret_s16_u32_x4(__VA_ARGS__)
		2317	#define svreinterpret_s16_s64_x4(...) __builtin_sve_reinterpret_s16_s64_x4(__VA_ARGS__)
		2318	#define svreinterpret_s16_u64_x4(...) __builtin_sve_reinterpret_s16_u64_x4(__VA_ARGS__)
		2319	#define svreinterpret_s16_f16_x4(...) __builtin_sve_reinterpret_s16_f16_x4(__VA_ARGS__)
		2320	#define svreinterpret_s16_bf16_x4(...) __builtin_sve_reinterpret_s16_bf16_x4(__VA_ARGS__)
		2321	#define svreinterpret_s16_f32_x4(...) __builtin_sve_reinterpret_s16_f32_x4(__VA_ARGS__)
		2322	#define svreinterpret_s16_f64_x4(...) __builtin_sve_reinterpret_s16_f64_x4(__VA_ARGS__)
		2323	#define svreinterpret_u16_s8_x4(...) __builtin_sve_reinterpret_u16_s8_x4(__VA_ARGS__)
		2324	#define svreinterpret_u16_u8_x4(...) __builtin_sve_reinterpret_u16_u8_x4(__VA_ARGS__)
		2325	#define svreinterpret_u16_s16_x4(...) __builtin_sve_reinterpret_u16_s16_x4(__VA_ARGS__)
		2326	#define svreinterpret_u16_u16_x4(...) __builtin_sve_reinterpret_u16_u16_x4(__VA_ARGS__)
		2327	#define svreinterpret_u16_s32_x4(...) __builtin_sve_reinterpret_u16_s32_x4(__VA_ARGS__)
		2328	#define svreinterpret_u16_u32_x4(...) __builtin_sve_reinterpret_u16_u32_x4(__VA_ARGS__)
		2329	#define svreinterpret_u16_s64_x4(...) __builtin_sve_reinterpret_u16_s64_x4(__VA_ARGS__)
		2330	#define svreinterpret_u16_u64_x4(...) __builtin_sve_reinterpret_u16_u64_x4(__VA_ARGS__)
		2331	#define svreinterpret_u16_f16_x4(...) __builtin_sve_reinterpret_u16_f16_x4(__VA_ARGS__)
		2332	#define svreinterpret_u16_bf16_x4(...) __builtin_sve_reinterpret_u16_bf16_x4(__VA_ARGS__)
		2333	#define svreinterpret_u16_f32_x4(...) __builtin_sve_reinterpret_u16_f32_x4(__VA_ARGS__)
		2334	#define svreinterpret_u16_f64_x4(...) __builtin_sve_reinterpret_u16_f64_x4(__VA_ARGS__)
		2335	#define svreinterpret_s32_s8_x4(...) __builtin_sve_reinterpret_s32_s8_x4(__VA_ARGS__)
		2336	#define svreinterpret_s32_u8_x4(...) __builtin_sve_reinterpret_s32_u8_x4(__VA_ARGS__)
		2337	#define svreinterpret_s32_s16_x4(...) __builtin_sve_reinterpret_s32_s16_x4(__VA_ARGS__)
		2338	#define svreinterpret_s32_u16_x4(...) __builtin_sve_reinterpret_s32_u16_x4(__VA_ARGS__)
		2339	#define svreinterpret_s32_s32_x4(...) __builtin_sve_reinterpret_s32_s32_x4(__VA_ARGS__)
		2340	#define svreinterpret_s32_u32_x4(...) __builtin_sve_reinterpret_s32_u32_x4(__VA_ARGS__)
		2341	#define svreinterpret_s32_s64_x4(...) __builtin_sve_reinterpret_s32_s64_x4(__VA_ARGS__)
		2342	#define svreinterpret_s32_u64_x4(...) __builtin_sve_reinterpret_s32_u64_x4(__VA_ARGS__)
		2343	#define svreinterpret_s32_f16_x4(...) __builtin_sve_reinterpret_s32_f16_x4(__VA_ARGS__)
		2344	#define svreinterpret_s32_bf16_x4(...) __builtin_sve_reinterpret_s32_bf16_x4(__VA_ARGS__)
		2345	#define svreinterpret_s32_f32_x4(...) __builtin_sve_reinterpret_s32_f32_x4(__VA_ARGS__)
		2346	#define svreinterpret_s32_f64_x4(...) __builtin_sve_reinterpret_s32_f64_x4(__VA_ARGS__)
		2347	#define svreinterpret_u32_s8_x4(...) __builtin_sve_reinterpret_u32_s8_x4(__VA_ARGS__)
		2348	#define svreinterpret_u32_u8_x4(...) __builtin_sve_reinterpret_u32_u8_x4(__VA_ARGS__)
		2349	#define svreinterpret_u32_s16_x4(...) __builtin_sve_reinterpret_u32_s16_x4(__VA_ARGS__)
		2350	#define svreinterpret_u32_u16_x4(...) __builtin_sve_reinterpret_u32_u16_x4(__VA_ARGS__)
		2351	#define svreinterpret_u32_s32_x4(...) __builtin_sve_reinterpret_u32_s32_x4(__VA_ARGS__)
		2352	#define svreinterpret_u32_u32_x4(...) __builtin_sve_reinterpret_u32_u32_x4(__VA_ARGS__)
		2353	#define svreinterpret_u32_s64_x4(...) __builtin_sve_reinterpret_u32_s64_x4(__VA_ARGS__)
		2354	#define svreinterpret_u32_u64_x4(...) __builtin_sve_reinterpret_u32_u64_x4(__VA_ARGS__)
		2355	#define svreinterpret_u32_f16_x4(...) __builtin_sve_reinterpret_u32_f16_x4(__VA_ARGS__)
		2356	#define svreinterpret_u32_bf16_x4(...) __builtin_sve_reinterpret_u32_bf16_x4(__VA_ARGS__)
		2357	#define svreinterpret_u32_f32_x4(...) __builtin_sve_reinterpret_u32_f32_x4(__VA_ARGS__)
		2358	#define svreinterpret_u32_f64_x4(...) __builtin_sve_reinterpret_u32_f64_x4(__VA_ARGS__)
		2359	#define svreinterpret_s64_s8_x4(...) __builtin_sve_reinterpret_s64_s8_x4(__VA_ARGS__)
		2360	#define svreinterpret_s64_u8_x4(...) __builtin_sve_reinterpret_s64_u8_x4(__VA_ARGS__)
		2361	#define svreinterpret_s64_s16_x4(...) __builtin_sve_reinterpret_s64_s16_x4(__VA_ARGS__)
		2362	#define svreinterpret_s64_u16_x4(...) __builtin_sve_reinterpret_s64_u16_x4(__VA_ARGS__)
		2363	#define svreinterpret_s64_s32_x4(...) __builtin_sve_reinterpret_s64_s32_x4(__VA_ARGS__)
		2364	#define svreinterpret_s64_u32_x4(...) __builtin_sve_reinterpret_s64_u32_x4(__VA_ARGS__)
		2365	#define svreinterpret_s64_s64_x4(...) __builtin_sve_reinterpret_s64_s64_x4(__VA_ARGS__)
		2366	#define svreinterpret_s64_u64_x4(...) __builtin_sve_reinterpret_s64_u64_x4(__VA_ARGS__)
		2367	#define svreinterpret_s64_f16_x4(...) __builtin_sve_reinterpret_s64_f16_x4(__VA_ARGS__)
		2368	#define svreinterpret_s64_bf16_x4(...) __builtin_sve_reinterpret_s64_bf16_x4(__VA_ARGS__)
		2369	#define svreinterpret_s64_f32_x4(...) __builtin_sve_reinterpret_s64_f32_x4(__VA_ARGS__)
		2370	#define svreinterpret_s64_f64_x4(...) __builtin_sve_reinterpret_s64_f64_x4(__VA_ARGS__)
		2371	#define svreinterpret_u64_s8_x4(...) __builtin_sve_reinterpret_u64_s8_x4(__VA_ARGS__)
		2372	#define svreinterpret_u64_u8_x4(...) __builtin_sve_reinterpret_u64_u8_x4(__VA_ARGS__)
		2373	#define svreinterpret_u64_s16_x4(...) __builtin_sve_reinterpret_u64_s16_x4(__VA_ARGS__)
		2374	#define svreinterpret_u64_u16_x4(...) __builtin_sve_reinterpret_u64_u16_x4(__VA_ARGS__)
		2375	#define svreinterpret_u64_s32_x4(...) __builtin_sve_reinterpret_u64_s32_x4(__VA_ARGS__)
		2376	#define svreinterpret_u64_u32_x4(...) __builtin_sve_reinterpret_u64_u32_x4(__VA_ARGS__)
		2377	#define svreinterpret_u64_s64_x4(...) __builtin_sve_reinterpret_u64_s64_x4(__VA_ARGS__)
		2378	#define svreinterpret_u64_u64_x4(...) __builtin_sve_reinterpret_u64_u64_x4(__VA_ARGS__)
		2379	#define svreinterpret_u64_f16_x4(...) __builtin_sve_reinterpret_u64_f16_x4(__VA_ARGS__)
		2380	#define svreinterpret_u64_bf16_x4(...) __builtin_sve_reinterpret_u64_bf16_x4(__VA_ARGS__)
		2381	#define svreinterpret_u64_f32_x4(...) __builtin_sve_reinterpret_u64_f32_x4(__VA_ARGS__)
		2382	#define svreinterpret_u64_f64_x4(...) __builtin_sve_reinterpret_u64_f64_x4(__VA_ARGS__)
		2383	#define svreinterpret_f16_s8_x4(...) __builtin_sve_reinterpret_f16_s8_x4(__VA_ARGS__)
		2384	#define svreinterpret_f16_u8_x4(...) __builtin_sve_reinterpret_f16_u8_x4(__VA_ARGS__)
		2385	#define svreinterpret_f16_s16_x4(...) __builtin_sve_reinterpret_f16_s16_x4(__VA_ARGS__)
		2386	#define svreinterpret_f16_u16_x4(...) __builtin_sve_reinterpret_f16_u16_x4(__VA_ARGS__)
		2387	#define svreinterpret_f16_s32_x4(...) __builtin_sve_reinterpret_f16_s32_x4(__VA_ARGS__)
		2388	#define svreinterpret_f16_u32_x4(...) __builtin_sve_reinterpret_f16_u32_x4(__VA_ARGS__)
		2389	#define svreinterpret_f16_s64_x4(...) __builtin_sve_reinterpret_f16_s64_x4(__VA_ARGS__)
		2390	#define svreinterpret_f16_u64_x4(...) __builtin_sve_reinterpret_f16_u64_x4(__VA_ARGS__)
		2391	#define svreinterpret_f16_f16_x4(...) __builtin_sve_reinterpret_f16_f16_x4(__VA_ARGS__)
		2392	#define svreinterpret_f16_bf16_x4(...) __builtin_sve_reinterpret_f16_bf16_x4(__VA_ARGS__)
		2393	#define svreinterpret_f16_f32_x4(...) __builtin_sve_reinterpret_f16_f32_x4(__VA_ARGS__)
		2394	#define svreinterpret_f16_f64_x4(...) __builtin_sve_reinterpret_f16_f64_x4(__VA_ARGS__)
		2395	#define svreinterpret_bf16_s8_x4(...) __builtin_sve_reinterpret_bf16_s8_x4(__VA_ARGS__)
		2396	#define svreinterpret_bf16_u8_x4(...) __builtin_sve_reinterpret_bf16_u8_x4(__VA_ARGS__)
		2397	#define svreinterpret_bf16_s16_x4(...) __builtin_sve_reinterpret_bf16_s16_x4(__VA_ARGS__)
		2398	#define svreinterpret_bf16_u16_x4(...) __builtin_sve_reinterpret_bf16_u16_x4(__VA_ARGS__)
		2399	#define svreinterpret_bf16_s32_x4(...) __builtin_sve_reinterpret_bf16_s32_x4(__VA_ARGS__)
		2400	#define svreinterpret_bf16_u32_x4(...) __builtin_sve_reinterpret_bf16_u32_x4(__VA_ARGS__)
		2401	#define svreinterpret_bf16_s64_x4(...) __builtin_sve_reinterpret_bf16_s64_x4(__VA_ARGS__)
		2402	#define svreinterpret_bf16_u64_x4(...) __builtin_sve_reinterpret_bf16_u64_x4(__VA_ARGS__)
		2403	#define svreinterpret_bf16_f16_x4(...) __builtin_sve_reinterpret_bf16_f16_x4(__VA_ARGS__)
		2404	#define svreinterpret_bf16_bf16_x4(...) __builtin_sve_reinterpret_bf16_bf16_x4(__VA_ARGS__)
		2405	#define svreinterpret_bf16_f32_x4(...) __builtin_sve_reinterpret_bf16_f32_x4(__VA_ARGS__)
		2406	#define svreinterpret_bf16_f64_x4(...) __builtin_sve_reinterpret_bf16_f64_x4(__VA_ARGS__)
		2407	#define svreinterpret_f32_s8_x4(...) __builtin_sve_reinterpret_f32_s8_x4(__VA_ARGS__)
		2408	#define svreinterpret_f32_u8_x4(...) __builtin_sve_reinterpret_f32_u8_x4(__VA_ARGS__)
		2409	#define svreinterpret_f32_s16_x4(...) __builtin_sve_reinterpret_f32_s16_x4(__VA_ARGS__)
		2410	#define svreinterpret_f32_u16_x4(...) __builtin_sve_reinterpret_f32_u16_x4(__VA_ARGS__)
		2411	#define svreinterpret_f32_s32_x4(...) __builtin_sve_reinterpret_f32_s32_x4(__VA_ARGS__)
		2412	#define svreinterpret_f32_u32_x4(...) __builtin_sve_reinterpret_f32_u32_x4(__VA_ARGS__)
		2413	#define svreinterpret_f32_s64_x4(...) __builtin_sve_reinterpret_f32_s64_x4(__VA_ARGS__)
		2414	#define svreinterpret_f32_u64_x4(...) __builtin_sve_reinterpret_f32_u64_x4(__VA_ARGS__)
		2415	#define svreinterpret_f32_f16_x4(...) __builtin_sve_reinterpret_f32_f16_x4(__VA_ARGS__)
		2416	#define svreinterpret_f32_bf16_x4(...) __builtin_sve_reinterpret_f32_bf16_x4(__VA_ARGS__)
		2417	#define svreinterpret_f32_f32_x4(...) __builtin_sve_reinterpret_f32_f32_x4(__VA_ARGS__)
		2418	#define svreinterpret_f32_f64_x4(...) __builtin_sve_reinterpret_f32_f64_x4(__VA_ARGS__)
		2419	#define svreinterpret_f64_s8_x4(...) __builtin_sve_reinterpret_f64_s8_x4(__VA_ARGS__)
		2420	#define svreinterpret_f64_u8_x4(...) __builtin_sve_reinterpret_f64_u8_x4(__VA_ARGS__)
		2421	#define svreinterpret_f64_s16_x4(...) __builtin_sve_reinterpret_f64_s16_x4(__VA_ARGS__)
		2422	#define svreinterpret_f64_u16_x4(...) __builtin_sve_reinterpret_f64_u16_x4(__VA_ARGS__)
		2423	#define svreinterpret_f64_s32_x4(...) __builtin_sve_reinterpret_f64_s32_x4(__VA_ARGS__)
		2424	#define svreinterpret_f64_u32_x4(...) __builtin_sve_reinterpret_f64_u32_x4(__VA_ARGS__)
		2425	#define svreinterpret_f64_s64_x4(...) __builtin_sve_reinterpret_f64_s64_x4(__VA_ARGS__)
		2426	#define svreinterpret_f64_u64_x4(...) __builtin_sve_reinterpret_f64_u64_x4(__VA_ARGS__)
		2427	#define svreinterpret_f64_f16_x4(...) __builtin_sve_reinterpret_f64_f16_x4(__VA_ARGS__)
		2428	#define svreinterpret_f64_bf16_x4(...) __builtin_sve_reinterpret_f64_bf16_x4(__VA_ARGS__)
		2429	#define svreinterpret_f64_f32_x4(...) __builtin_sve_reinterpret_f64_f32_x4(__VA_ARGS__)
		2430	#define svreinterpret_f64_f64_x4(...) __builtin_sve_reinterpret_f64_f64_x4(__VA_ARGS__)
		2431	__aio __attribute__((target("sve"))) svint8x4_t svreinterpret_s8(svint8x4_t op) __arm_streaming_compatible {
		2432	return __builtin_sve_reinterpret_s8_s8_x4(op);
496	}	2433	}
497		2434
498	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svbfloat16_t op) {	2435	__aio __attribute__((target("sve"))) svint8x4_t svreinterpret_s8(svuint8x4_t op) __arm_streaming_compatible {
499	return __builtin_sve_reinterpret_u8_bf16(op);	2436	return __builtin_sve_reinterpret_s8_u8_x4(op);
500	}	2437	}
501		2438
502	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svfloat32_t op) {	2439	__aio __attribute__((target("sve"))) svint8x4_t svreinterpret_s8(svint16x4_t op) __arm_streaming_compatible {
503	return __builtin_sve_reinterpret_u8_f32(op);	2440	return __builtin_sve_reinterpret_s8_s16_x4(op);
504	}	2441	}
505		2442
506	__aio __attribute__((target("sve"))) svuint8_t svreinterpret_u8(svfloat64_t op) {	2443	__aio __attribute__((target("sve"))) svint8x4_t svreinterpret_s8(svuint16x4_t op) __arm_streaming_compatible {
507	return __builtin_sve_reinterpret_u8_f64(op);	2444	return __builtin_sve_reinterpret_s8_u16_x4(op);
508	}	2445	}
509		2446
510	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svint8_t op) {	2447	__aio __attribute__((target("sve"))) svint8x4_t svreinterpret_s8(svint32x4_t op) __arm_streaming_compatible {
511	return __builtin_sve_reinterpret_u16_s8(op);	2448	return __builtin_sve_reinterpret_s8_s32_x4(op);
512	}	2449	}
513		2450
514	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svint16_t op) {	2451	__aio __attribute__((target("sve"))) svint8x4_t svreinterpret_s8(svuint32x4_t op) __arm_streaming_compatible {
515	return __builtin_sve_reinterpret_u16_s16(op);	2452	return __builtin_sve_reinterpret_s8_u32_x4(op);
516	}	2453	}
517		2454
518	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svint32_t op) {	2455	__aio __attribute__((target("sve"))) svint8x4_t svreinterpret_s8(svint64x4_t op) __arm_streaming_compatible {
519	return __builtin_sve_reinterpret_u16_s32(op);	2456	return __builtin_sve_reinterpret_s8_s64_x4(op);
520	}	2457	}
521		2458
522	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svint64_t op) {	2459	__aio __attribute__((target("sve"))) svint8x4_t svreinterpret_s8(svuint64x4_t op) __arm_streaming_compatible {
523	return __builtin_sve_reinterpret_u16_s64(op);	2460	return __builtin_sve_reinterpret_s8_u64_x4(op);
524	}	2461	}
525		2462
526	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svuint8_t op) {	2463	__aio __attribute__((target("sve"))) svint8x4_t svreinterpret_s8(svfloat16x4_t op) __arm_streaming_compatible {
527	return __builtin_sve_reinterpret_u16_u8(op);	2464	return __builtin_sve_reinterpret_s8_f16_x4(op);
528	}	2465	}
529		2466
530	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svuint16_t op) {	2467	__aio __attribute__((target("sve"))) svint8x4_t svreinterpret_s8(svbfloat16x4_t op) __arm_streaming_compatible {
531	return __builtin_sve_reinterpret_u16_u16(op);	2468	return __builtin_sve_reinterpret_s8_bf16_x4(op);
532	}	2469	}
533		2470
534	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svuint32_t op) {	2471	__aio __attribute__((target("sve"))) svint8x4_t svreinterpret_s8(svfloat32x4_t op) __arm_streaming_compatible {
535	return __builtin_sve_reinterpret_u16_u32(op);	2472	return __builtin_sve_reinterpret_s8_f32_x4(op);
536	}	2473	}
537		2474
538	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svuint64_t op) {	2475	__aio __attribute__((target("sve"))) svint8x4_t svreinterpret_s8(svfloat64x4_t op) __arm_streaming_compatible {
539	return __builtin_sve_reinterpret_u16_u64(op);	2476	return __builtin_sve_reinterpret_s8_f64_x4(op);
540	}	2477	}
541		2478
542	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svfloat16_t op) {	2479	__aio __attribute__((target("sve"))) svuint8x4_t svreinterpret_u8(svint8x4_t op) __arm_streaming_compatible {
543	return __builtin_sve_reinterpret_u16_f16(op);	2480	return __builtin_sve_reinterpret_u8_s8_x4(op);
544	}	2481	}
545		2482
546	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svbfloat16_t op) {	2483	__aio __attribute__((target("sve"))) svuint8x4_t svreinterpret_u8(svuint8x4_t op) __arm_streaming_compatible {
547	return __builtin_sve_reinterpret_u16_bf16(op);	2484	return __builtin_sve_reinterpret_u8_u8_x4(op);
548	}	2485	}
549		2486
550	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svfloat32_t op) {	2487	__aio __attribute__((target("sve"))) svuint8x4_t svreinterpret_u8(svint16x4_t op) __arm_streaming_compatible {
551	return __builtin_sve_reinterpret_u16_f32(op);	2488	return __builtin_sve_reinterpret_u8_s16_x4(op);
552	}	2489	}
553		2490
554	__aio __attribute__((target("sve"))) svuint16_t svreinterpret_u16(svfloat64_t op) {	2491	__aio __attribute__((target("sve"))) svuint8x4_t svreinterpret_u8(svuint16x4_t op) __arm_streaming_compatible {
555	return __builtin_sve_reinterpret_u16_f64(op);	2492	return __builtin_sve_reinterpret_u8_u16_x4(op);
556	}	2493	}
557		2494
558	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svint8_t op) {	2495	__aio __attribute__((target("sve"))) svuint8x4_t svreinterpret_u8(svint32x4_t op) __arm_streaming_compatible {
559	return __builtin_sve_reinterpret_u32_s8(op);	2496	return __builtin_sve_reinterpret_u8_s32_x4(op);
560	}	2497	}
561		2498
562	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svint16_t op) {	2499	__aio __attribute__((target("sve"))) svuint8x4_t svreinterpret_u8(svuint32x4_t op) __arm_streaming_compatible {
563	return __builtin_sve_reinterpret_u32_s16(op);	2500	return __builtin_sve_reinterpret_u8_u32_x4(op);
564	}	2501	}
565		2502
566	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svint32_t op) {	2503	__aio __attribute__((target("sve"))) svuint8x4_t svreinterpret_u8(svint64x4_t op) __arm_streaming_compatible {
567	return __builtin_sve_reinterpret_u32_s32(op);	2504	return __builtin_sve_reinterpret_u8_s64_x4(op);
568	}	2505	}
569		2506
570	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svint64_t op) {	2507	__aio __attribute__((target("sve"))) svuint8x4_t svreinterpret_u8(svuint64x4_t op) __arm_streaming_compatible {
571	return __builtin_sve_reinterpret_u32_s64(op);	2508	return __builtin_sve_reinterpret_u8_u64_x4(op);
572	}	2509	}
573		2510
574	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svuint8_t op) {	2511	__aio __attribute__((target("sve"))) svuint8x4_t svreinterpret_u8(svfloat16x4_t op) __arm_streaming_compatible {
575	return __builtin_sve_reinterpret_u32_u8(op);	2512	return __builtin_sve_reinterpret_u8_f16_x4(op);
576	}	2513	}
577		2514
578	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svuint16_t op) {	2515	__aio __attribute__((target("sve"))) svuint8x4_t svreinterpret_u8(svbfloat16x4_t op) __arm_streaming_compatible {
579	return __builtin_sve_reinterpret_u32_u16(op);	2516	return __builtin_sve_reinterpret_u8_bf16_x4(op);
580	}	2517	}
581		2518
582	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svuint32_t op) {	2519	__aio __attribute__((target("sve"))) svuint8x4_t svreinterpret_u8(svfloat32x4_t op) __arm_streaming_compatible {
583	return __builtin_sve_reinterpret_u32_u32(op);	2520	return __builtin_sve_reinterpret_u8_f32_x4(op);
584	}	2521	}
585		2522
586	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svuint64_t op) {	2523	__aio __attribute__((target("sve"))) svuint8x4_t svreinterpret_u8(svfloat64x4_t op) __arm_streaming_compatible {
587	return __builtin_sve_reinterpret_u32_u64(op);	2524	return __builtin_sve_reinterpret_u8_f64_x4(op);
588	}	2525	}
589		2526
590	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svfloat16_t op) {	2527	__aio __attribute__((target("sve"))) svint16x4_t svreinterpret_s16(svint8x4_t op) __arm_streaming_compatible {
591	return __builtin_sve_reinterpret_u32_f16(op);	2528	return __builtin_sve_reinterpret_s16_s8_x4(op);
592	}	2529	}
593		2530
594	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svbfloat16_t op) {	2531	__aio __attribute__((target("sve"))) svint16x4_t svreinterpret_s16(svuint8x4_t op) __arm_streaming_compatible {
595	return __builtin_sve_reinterpret_u32_bf16(op);	2532	return __builtin_sve_reinterpret_s16_u8_x4(op);
596	}	2533	}
597		2534
598	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svfloat32_t op) {	2535	__aio __attribute__((target("sve"))) svint16x4_t svreinterpret_s16(svint16x4_t op) __arm_streaming_compatible {
599	return __builtin_sve_reinterpret_u32_f32(op);	2536	return __builtin_sve_reinterpret_s16_s16_x4(op);
600	}	2537	}
601		2538
602	__aio __attribute__((target("sve"))) svuint32_t svreinterpret_u32(svfloat64_t op) {	2539	__aio __attribute__((target("sve"))) svint16x4_t svreinterpret_s16(svuint16x4_t op) __arm_streaming_compatible {
603	return __builtin_sve_reinterpret_u32_f64(op);	2540	return __builtin_sve_reinterpret_s16_u16_x4(op);
604	}	2541	}
605		2542
606	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svint8_t op) {	2543	__aio __attribute__((target("sve"))) svint16x4_t svreinterpret_s16(svint32x4_t op) __arm_streaming_compatible {
607	return __builtin_sve_reinterpret_u64_s8(op);	2544	return __builtin_sve_reinterpret_s16_s32_x4(op);
608	}	2545	}
609		2546
610	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svint16_t op) {	2547	__aio __attribute__((target("sve"))) svint16x4_t svreinterpret_s16(svuint32x4_t op) __arm_streaming_compatible {
611	return __builtin_sve_reinterpret_u64_s16(op);	2548	return __builtin_sve_reinterpret_s16_u32_x4(op);
612	}	2549	}
613		2550
614	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svint32_t op) {	2551	__aio __attribute__((target("sve"))) svint16x4_t svreinterpret_s16(svint64x4_t op) __arm_streaming_compatible {
615	return __builtin_sve_reinterpret_u64_s32(op);	2552	return __builtin_sve_reinterpret_s16_s64_x4(op);
616	}	2553	}
617		2554
618	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svint64_t op) {	2555	__aio __attribute__((target("sve"))) svint16x4_t svreinterpret_s16(svuint64x4_t op) __arm_streaming_compatible {
619	return __builtin_sve_reinterpret_u64_s64(op);	2556	return __builtin_sve_reinterpret_s16_u64_x4(op);
620	}	2557	}
621		2558
622	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svuint8_t op) {	2559	__aio __attribute__((target("sve"))) svint16x4_t svreinterpret_s16(svfloat16x4_t op) __arm_streaming_compatible {
623	return __builtin_sve_reinterpret_u64_u8(op);	2560	return __builtin_sve_reinterpret_s16_f16_x4(op);
624	}	2561	}
625		2562
626	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svuint16_t op) {	2563	__aio __attribute__((target("sve"))) svint16x4_t svreinterpret_s16(svbfloat16x4_t op) __arm_streaming_compatible {
627	return __builtin_sve_reinterpret_u64_u16(op);	2564	return __builtin_sve_reinterpret_s16_bf16_x4(op);
628	}	2565	}
629		2566
630	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svuint32_t op) {	2567	__aio __attribute__((target("sve"))) svint16x4_t svreinterpret_s16(svfloat32x4_t op) __arm_streaming_compatible {
631	return __builtin_sve_reinterpret_u64_u32(op);	2568	return __builtin_sve_reinterpret_s16_f32_x4(op);
632	}	2569	}
633		2570
634	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svuint64_t op) {	2571	__aio __attribute__((target("sve"))) svint16x4_t svreinterpret_s16(svfloat64x4_t op) __arm_streaming_compatible {
635	return __builtin_sve_reinterpret_u64_u64(op);	2572	return __builtin_sve_reinterpret_s16_f64_x4(op);
636	}	2573	}
637		2574
638	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svfloat16_t op) {	2575	__aio __attribute__((target("sve"))) svuint16x4_t svreinterpret_u16(svint8x4_t op) __arm_streaming_compatible {
639	return __builtin_sve_reinterpret_u64_f16(op);	2576	return __builtin_sve_reinterpret_u16_s8_x4(op);
640	}	2577	}
641		2578
642	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svbfloat16_t op) {	2579	__aio __attribute__((target("sve"))) svuint16x4_t svreinterpret_u16(svuint8x4_t op) __arm_streaming_compatible {
643	return __builtin_sve_reinterpret_u64_bf16(op);	2580	return __builtin_sve_reinterpret_u16_u8_x4(op);
644	}	2581	}
645		2582
646	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svfloat32_t op) {	2583	__aio __attribute__((target("sve"))) svuint16x4_t svreinterpret_u16(svint16x4_t op) __arm_streaming_compatible {
647	return __builtin_sve_reinterpret_u64_f32(op);	2584	return __builtin_sve_reinterpret_u16_s16_x4(op);
648	}	2585	}
649		2586
650	__aio __attribute__((target("sve"))) svuint64_t svreinterpret_u64(svfloat64_t op) {	2587	__aio __attribute__((target("sve"))) svuint16x4_t svreinterpret_u16(svuint16x4_t op) __arm_streaming_compatible {
651	return __builtin_sve_reinterpret_u64_f64(op);	2588	return __builtin_sve_reinterpret_u16_u16_x4(op);
652	}	2589	}
653		2590
654	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svint8_t op) {	2591	__aio __attribute__((target("sve"))) svuint16x4_t svreinterpret_u16(svint32x4_t op) __arm_streaming_compatible {
655	return __builtin_sve_reinterpret_f16_s8(op);	2592	return __builtin_sve_reinterpret_u16_s32_x4(op);
656	}	2593	}
657		2594
658	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svint16_t op) {	2595	__aio __attribute__((target("sve"))) svuint16x4_t svreinterpret_u16(svuint32x4_t op) __arm_streaming_compatible {
659	return __builtin_sve_reinterpret_f16_s16(op);	2596	return __builtin_sve_reinterpret_u16_u32_x4(op);
660	}	2597	}
661		2598
662	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svint32_t op) {	2599	__aio __attribute__((target("sve"))) svuint16x4_t svreinterpret_u16(svint64x4_t op) __arm_streaming_compatible {
663	return __builtin_sve_reinterpret_f16_s32(op);	2600	return __builtin_sve_reinterpret_u16_s64_x4(op);
664	}	2601	}
665		2602
666	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svint64_t op) {	2603	__aio __attribute__((target("sve"))) svuint16x4_t svreinterpret_u16(svuint64x4_t op) __arm_streaming_compatible {
667	return __builtin_sve_reinterpret_f16_s64(op);	2604	return __builtin_sve_reinterpret_u16_u64_x4(op);
668	}	2605	}
669		2606
670	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svuint8_t op) {	2607	__aio __attribute__((target("sve"))) svuint16x4_t svreinterpret_u16(svfloat16x4_t op) __arm_streaming_compatible {
671	return __builtin_sve_reinterpret_f16_u8(op);	2608	return __builtin_sve_reinterpret_u16_f16_x4(op);
672	}	2609	}
673		2610
674	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svuint16_t op) {	2611	__aio __attribute__((target("sve"))) svuint16x4_t svreinterpret_u16(svbfloat16x4_t op) __arm_streaming_compatible {
675	return __builtin_sve_reinterpret_f16_u16(op);	2612	return __builtin_sve_reinterpret_u16_bf16_x4(op);
676	}	2613	}
677		2614
678	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svuint32_t op) {	2615	__aio __attribute__((target("sve"))) svuint16x4_t svreinterpret_u16(svfloat32x4_t op) __arm_streaming_compatible {
679	return __builtin_sve_reinterpret_f16_u32(op);	2616	return __builtin_sve_reinterpret_u16_f32_x4(op);
680	}	2617	}
681		2618
682	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svuint64_t op) {	2619	__aio __attribute__((target("sve"))) svuint16x4_t svreinterpret_u16(svfloat64x4_t op) __arm_streaming_compatible {
683	return __builtin_sve_reinterpret_f16_u64(op);	2620	return __builtin_sve_reinterpret_u16_f64_x4(op);
684	}	2621	}
685		2622
686	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svfloat16_t op) {	2623	__aio __attribute__((target("sve"))) svint32x4_t svreinterpret_s32(svint8x4_t op) __arm_streaming_compatible {
687	return __builtin_sve_reinterpret_f16_f16(op);	2624	return __builtin_sve_reinterpret_s32_s8_x4(op);
688	}	2625	}
689		2626
690	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svbfloat16_t op) {	2627	__aio __attribute__((target("sve"))) svint32x4_t svreinterpret_s32(svuint8x4_t op) __arm_streaming_compatible {
691	return __builtin_sve_reinterpret_f16_bf16(op);	2628	return __builtin_sve_reinterpret_s32_u8_x4(op);
692	}	2629	}
693		2630
694	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svfloat32_t op) {	2631	__aio __attribute__((target("sve"))) svint32x4_t svreinterpret_s32(svint16x4_t op) __arm_streaming_compatible {
695	return __builtin_sve_reinterpret_f16_f32(op);	2632	return __builtin_sve_reinterpret_s32_s16_x4(op);
696	}	2633	}
697		2634
698	__aio __attribute__((target("sve"))) svfloat16_t svreinterpret_f16(svfloat64_t op) {	2635	__aio __attribute__((target("sve"))) svint32x4_t svreinterpret_s32(svuint16x4_t op) __arm_streaming_compatible {
699	return __builtin_sve_reinterpret_f16_f64(op);	2636	return __builtin_sve_reinterpret_s32_u16_x4(op);
700	}	2637	}
701		2638
702	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svint8_t op) {	2639	__aio __attribute__((target("sve"))) svint32x4_t svreinterpret_s32(svint32x4_t op) __arm_streaming_compatible {
703	return __builtin_sve_reinterpret_bf16_s8(op);	2640	return __builtin_sve_reinterpret_s32_s32_x4(op);
704	}	2641	}
705		2642
706	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svint16_t op) {	2643	__aio __attribute__((target("sve"))) svint32x4_t svreinterpret_s32(svuint32x4_t op) __arm_streaming_compatible {
707	return __builtin_sve_reinterpret_bf16_s16(op);	2644	return __builtin_sve_reinterpret_s32_u32_x4(op);
708	}	2645	}
709		2646
710	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svint32_t op) {	2647	__aio __attribute__((target("sve"))) svint32x4_t svreinterpret_s32(svint64x4_t op) __arm_streaming_compatible {
711	return __builtin_sve_reinterpret_bf16_s32(op);	2648	return __builtin_sve_reinterpret_s32_s64_x4(op);
712	}	2649	}
713		2650
714	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svint64_t op) {	2651	__aio __attribute__((target("sve"))) svint32x4_t svreinterpret_s32(svuint64x4_t op) __arm_streaming_compatible {
715	return __builtin_sve_reinterpret_bf16_s64(op);	2652	return __builtin_sve_reinterpret_s32_u64_x4(op);
716	}	2653	}
717		2654
718	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svuint8_t op) {	2655	__aio __attribute__((target("sve"))) svint32x4_t svreinterpret_s32(svfloat16x4_t op) __arm_streaming_compatible {
719	return __builtin_sve_reinterpret_bf16_u8(op);	2656	return __builtin_sve_reinterpret_s32_f16_x4(op);
720	}	2657	}
721		2658
722	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svuint16_t op) {	2659	__aio __attribute__((target("sve"))) svint32x4_t svreinterpret_s32(svbfloat16x4_t op) __arm_streaming_compatible {
723	return __builtin_sve_reinterpret_bf16_u16(op);	2660	return __builtin_sve_reinterpret_s32_bf16_x4(op);
724	}	2661	}
725		2662
726	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svuint32_t op) {	2663	__aio __attribute__((target("sve"))) svint32x4_t svreinterpret_s32(svfloat32x4_t op) __arm_streaming_compatible {
727	return __builtin_sve_reinterpret_bf16_u32(op);	2664	return __builtin_sve_reinterpret_s32_f32_x4(op);
728	}	2665	}
729		2666
730	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svuint64_t op) {	2667	__aio __attribute__((target("sve"))) svint32x4_t svreinterpret_s32(svfloat64x4_t op) __arm_streaming_compatible {
731	return __builtin_sve_reinterpret_bf16_u64(op);	2668	return __builtin_sve_reinterpret_s32_f64_x4(op);
732	}	2669	}
733		2670
734	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svfloat16_t op) {	2671	__aio __attribute__((target("sve"))) svuint32x4_t svreinterpret_u32(svint8x4_t op) __arm_streaming_compatible {
735	return __builtin_sve_reinterpret_bf16_f16(op);	2672	return __builtin_sve_reinterpret_u32_s8_x4(op);
736	}	2673	}
737		2674
738	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svbfloat16_t op) {	2675	__aio __attribute__((target("sve"))) svuint32x4_t svreinterpret_u32(svuint8x4_t op) __arm_streaming_compatible {
739	return __builtin_sve_reinterpret_bf16_bf16(op);	2676	return __builtin_sve_reinterpret_u32_u8_x4(op);
740	}	2677	}
741		2678
742	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svfloat32_t op) {	2679	__aio __attribute__((target("sve"))) svuint32x4_t svreinterpret_u32(svint16x4_t op) __arm_streaming_compatible {
743	return __builtin_sve_reinterpret_bf16_f32(op);	2680	return __builtin_sve_reinterpret_u32_s16_x4(op);
744	}	2681	}
745		2682
746	__aio __attribute__((target("sve"))) svbfloat16_t svreinterpret_bf16(svfloat64_t op) {	2683	__aio __attribute__((target("sve"))) svuint32x4_t svreinterpret_u32(svuint16x4_t op) __arm_streaming_compatible {
747	return __builtin_sve_reinterpret_bf16_f64(op);	2684	return __builtin_sve_reinterpret_u32_u16_x4(op);
748	}	2685	}
749		2686
750	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svint8_t op) {	2687	__aio __attribute__((target("sve"))) svuint32x4_t svreinterpret_u32(svint32x4_t op) __arm_streaming_compatible {
751	return __builtin_sve_reinterpret_f32_s8(op);	2688	return __builtin_sve_reinterpret_u32_s32_x4(op);
752	}	2689	}
753		2690
754	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svint16_t op) {	2691	__aio __attribute__((target("sve"))) svuint32x4_t svreinterpret_u32(svuint32x4_t op) __arm_streaming_compatible {
755	return __builtin_sve_reinterpret_f32_s16(op);	2692	return __builtin_sve_reinterpret_u32_u32_x4(op);
756	}	2693	}
757		2694
758	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svint32_t op) {	2695	__aio __attribute__((target("sve"))) svuint32x4_t svreinterpret_u32(svint64x4_t op) __arm_streaming_compatible {
759	return __builtin_sve_reinterpret_f32_s32(op);	2696	return __builtin_sve_reinterpret_u32_s64_x4(op);
760	}	2697	}
761		2698
762	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svint64_t op) {	2699	__aio __attribute__((target("sve"))) svuint32x4_t svreinterpret_u32(svuint64x4_t op) __arm_streaming_compatible {
763	return __builtin_sve_reinterpret_f32_s64(op);	2700	return __builtin_sve_reinterpret_u32_u64_x4(op);
764	}	2701	}
765		2702
766	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svuint8_t op) {	2703	__aio __attribute__((target("sve"))) svuint32x4_t svreinterpret_u32(svfloat16x4_t op) __arm_streaming_compatible {
767	return __builtin_sve_reinterpret_f32_u8(op);	2704	return __builtin_sve_reinterpret_u32_f16_x4(op);
768	}	2705	}
769		2706
770	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svuint16_t op) {	2707	__aio __attribute__((target("sve"))) svuint32x4_t svreinterpret_u32(svbfloat16x4_t op) __arm_streaming_compatible {
771	return __builtin_sve_reinterpret_f32_u16(op);	2708	return __builtin_sve_reinterpret_u32_bf16_x4(op);
772	}	2709	}
773		2710
774	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svuint32_t op) {	2711	__aio __attribute__((target("sve"))) svuint32x4_t svreinterpret_u32(svfloat32x4_t op) __arm_streaming_compatible {
775	return __builtin_sve_reinterpret_f32_u32(op);	2712	return __builtin_sve_reinterpret_u32_f32_x4(op);
776	}	2713	}
777		2714
778	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svuint64_t op) {	2715	__aio __attribute__((target("sve"))) svuint32x4_t svreinterpret_u32(svfloat64x4_t op) __arm_streaming_compatible {
779	return __builtin_sve_reinterpret_f32_u64(op);	2716	return __builtin_sve_reinterpret_u32_f64_x4(op);
780	}	2717	}
781		2718
782	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svfloat16_t op) {	2719	__aio __attribute__((target("sve"))) svint64x4_t svreinterpret_s64(svint8x4_t op) __arm_streaming_compatible {
783	return __builtin_sve_reinterpret_f32_f16(op);	2720	return __builtin_sve_reinterpret_s64_s8_x4(op);
784	}	2721	}
785		2722
786	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svbfloat16_t op) {	2723	__aio __attribute__((target("sve"))) svint64x4_t svreinterpret_s64(svuint8x4_t op) __arm_streaming_compatible {
787	return __builtin_sve_reinterpret_f32_bf16(op);	2724	return __builtin_sve_reinterpret_s64_u8_x4(op);
788	}	2725	}
789		2726
790	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svfloat32_t op) {	2727	__aio __attribute__((target("sve"))) svint64x4_t svreinterpret_s64(svint16x4_t op) __arm_streaming_compatible {
791	return __builtin_sve_reinterpret_f32_f32(op);	2728	return __builtin_sve_reinterpret_s64_s16_x4(op);
792	}	2729	}
793		2730
794	__aio __attribute__((target("sve"))) svfloat32_t svreinterpret_f32(svfloat64_t op) {	2731	__aio __attribute__((target("sve"))) svint64x4_t svreinterpret_s64(svuint16x4_t op) __arm_streaming_compatible {
795	return __builtin_sve_reinterpret_f32_f64(op);	2732	return __builtin_sve_reinterpret_s64_u16_x4(op);
796	}	2733	}
797		2734
798	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svint8_t op) {	2735	__aio __attribute__((target("sve"))) svint64x4_t svreinterpret_s64(svint32x4_t op) __arm_streaming_compatible {
799	return __builtin_sve_reinterpret_f64_s8(op);	2736	return __builtin_sve_reinterpret_s64_s32_x4(op);
800	}	2737	}
801		2738
802	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svint16_t op) {	2739	__aio __attribute__((target("sve"))) svint64x4_t svreinterpret_s64(svuint32x4_t op) __arm_streaming_compatible {
803	return __builtin_sve_reinterpret_f64_s16(op);	2740	return __builtin_sve_reinterpret_s64_u32_x4(op);
804	}	2741	}
805		2742
806	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svint32_t op) {	2743	__aio __attribute__((target("sve"))) svint64x4_t svreinterpret_s64(svint64x4_t op) __arm_streaming_compatible {
807	return __builtin_sve_reinterpret_f64_s32(op);	2744	return __builtin_sve_reinterpret_s64_s64_x4(op);
808	}	2745	}
809		2746
810	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svint64_t op) {	2747	__aio __attribute__((target("sve"))) svint64x4_t svreinterpret_s64(svuint64x4_t op) __arm_streaming_compatible {
811	return __builtin_sve_reinterpret_f64_s64(op);	2748	return __builtin_sve_reinterpret_s64_u64_x4(op);
812	}	2749	}
813		2750
814	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svuint8_t op) {	2751	__aio __attribute__((target("sve"))) svint64x4_t svreinterpret_s64(svfloat16x4_t op) __arm_streaming_compatible {
815	return __builtin_sve_reinterpret_f64_u8(op);	2752	return __builtin_sve_reinterpret_s64_f16_x4(op);
816	}	2753	}
817		2754
818	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svuint16_t op) {	2755	__aio __attribute__((target("sve"))) svint64x4_t svreinterpret_s64(svbfloat16x4_t op) __arm_streaming_compatible {
819	return __builtin_sve_reinterpret_f64_u16(op);	2756	return __builtin_sve_reinterpret_s64_bf16_x4(op);
820	}	2757	}
821		2758
822	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svuint32_t op) {	2759	__aio __attribute__((target("sve"))) svint64x4_t svreinterpret_s64(svfloat32x4_t op) __arm_streaming_compatible {
823	return __builtin_sve_reinterpret_f64_u32(op);	2760	return __builtin_sve_reinterpret_s64_f32_x4(op);
824	}	2761	}
825		2762
826	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svuint64_t op) {	2763	__aio __attribute__((target("sve"))) svint64x4_t svreinterpret_s64(svfloat64x4_t op) __arm_streaming_compatible {
827	return __builtin_sve_reinterpret_f64_u64(op);	2764	return __builtin_sve_reinterpret_s64_f64_x4(op);
828	}	2765	}
829		2766
830	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svfloat16_t op) {	2767	__aio __attribute__((target("sve"))) svuint64x4_t svreinterpret_u64(svint8x4_t op) __arm_streaming_compatible {
831	return __builtin_sve_reinterpret_f64_f16(op);	2768	return __builtin_sve_reinterpret_u64_s8_x4(op);
832	}	2769	}
833		2770
834	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svbfloat16_t op) {	2771	__aio __attribute__((target("sve"))) svuint64x4_t svreinterpret_u64(svuint8x4_t op) __arm_streaming_compatible {
835	return __builtin_sve_reinterpret_f64_bf16(op);	2772	return __builtin_sve_reinterpret_u64_u8_x4(op);
836	}	2773	}
837		2774
838	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svfloat32_t op) {	2775	__aio __attribute__((target("sve"))) svuint64x4_t svreinterpret_u64(svint16x4_t op) __arm_streaming_compatible {
839	return __builtin_sve_reinterpret_f64_f32(op);	2776	return __builtin_sve_reinterpret_u64_s16_x4(op);
840	}	2777	}
841		2778
842	__aio __attribute__((target("sve"))) svfloat64_t svreinterpret_f64(svfloat64_t op) {	2779	__aio __attribute__((target("sve"))) svuint64x4_t svreinterpret_u64(svuint16x4_t op) __arm_streaming_compatible {
843	return __builtin_sve_reinterpret_f64_f64(op);	2780	return __builtin_sve_reinterpret_u64_u16_x4(op);
		2781	}
		2782
		2783	__aio __attribute__((target("sve"))) svuint64x4_t svreinterpret_u64(svint32x4_t op) __arm_streaming_compatible {
		2784	return __builtin_sve_reinterpret_u64_s32_x4(op);
		2785	}
		2786
		2787	__aio __attribute__((target("sve"))) svuint64x4_t svreinterpret_u64(svuint32x4_t op) __arm_streaming_compatible {
		2788	return __builtin_sve_reinterpret_u64_u32_x4(op);
		2789	}
		2790
		2791	__aio __attribute__((target("sve"))) svuint64x4_t svreinterpret_u64(svint64x4_t op) __arm_streaming_compatible {
		2792	return __builtin_sve_reinterpret_u64_s64_x4(op);
		2793	}
		2794
		2795	__aio __attribute__((target("sve"))) svuint64x4_t svreinterpret_u64(svuint64x4_t op) __arm_streaming_compatible {
		2796	return __builtin_sve_reinterpret_u64_u64_x4(op);
		2797	}
		2798
		2799	__aio __attribute__((target("sve"))) svuint64x4_t svreinterpret_u64(svfloat16x4_t op) __arm_streaming_compatible {
		2800	return __builtin_sve_reinterpret_u64_f16_x4(op);
		2801	}
		2802
		2803	__aio __attribute__((target("sve"))) svuint64x4_t svreinterpret_u64(svbfloat16x4_t op) __arm_streaming_compatible {
		2804	return __builtin_sve_reinterpret_u64_bf16_x4(op);
		2805	}
		2806
		2807	__aio __attribute__((target("sve"))) svuint64x4_t svreinterpret_u64(svfloat32x4_t op) __arm_streaming_compatible {
		2808	return __builtin_sve_reinterpret_u64_f32_x4(op);
		2809	}
		2810
		2811	__aio __attribute__((target("sve"))) svuint64x4_t svreinterpret_u64(svfloat64x4_t op) __arm_streaming_compatible {
		2812	return __builtin_sve_reinterpret_u64_f64_x4(op);
		2813	}
		2814
		2815	__aio __attribute__((target("sve"))) svfloat16x4_t svreinterpret_f16(svint8x4_t op) __arm_streaming_compatible {
		2816	return __builtin_sve_reinterpret_f16_s8_x4(op);
		2817	}
		2818
		2819	__aio __attribute__((target("sve"))) svfloat16x4_t svreinterpret_f16(svuint8x4_t op) __arm_streaming_compatible {
		2820	return __builtin_sve_reinterpret_f16_u8_x4(op);
		2821	}
		2822
		2823	__aio __attribute__((target("sve"))) svfloat16x4_t svreinterpret_f16(svint16x4_t op) __arm_streaming_compatible {
		2824	return __builtin_sve_reinterpret_f16_s16_x4(op);
		2825	}
		2826
		2827	__aio __attribute__((target("sve"))) svfloat16x4_t svreinterpret_f16(svuint16x4_t op) __arm_streaming_compatible {
		2828	return __builtin_sve_reinterpret_f16_u16_x4(op);
		2829	}
		2830
		2831	__aio __attribute__((target("sve"))) svfloat16x4_t svreinterpret_f16(svint32x4_t op) __arm_streaming_compatible {
		2832	return __builtin_sve_reinterpret_f16_s32_x4(op);
		2833	}
		2834
		2835	__aio __attribute__((target("sve"))) svfloat16x4_t svreinterpret_f16(svuint32x4_t op) __arm_streaming_compatible {
		2836	return __builtin_sve_reinterpret_f16_u32_x4(op);
		2837	}
		2838
		2839	__aio __attribute__((target("sve"))) svfloat16x4_t svreinterpret_f16(svint64x4_t op) __arm_streaming_compatible {
		2840	return __builtin_sve_reinterpret_f16_s64_x4(op);
		2841	}
		2842
		2843	__aio __attribute__((target("sve"))) svfloat16x4_t svreinterpret_f16(svuint64x4_t op) __arm_streaming_compatible {
		2844	return __builtin_sve_reinterpret_f16_u64_x4(op);
		2845	}
		2846
		2847	__aio __attribute__((target("sve"))) svfloat16x4_t svreinterpret_f16(svfloat16x4_t op) __arm_streaming_compatible {
		2848	return __builtin_sve_reinterpret_f16_f16_x4(op);
		2849	}
		2850
		2851	__aio __attribute__((target("sve"))) svfloat16x4_t svreinterpret_f16(svbfloat16x4_t op) __arm_streaming_compatible {
		2852	return __builtin_sve_reinterpret_f16_bf16_x4(op);
		2853	}
		2854
		2855	__aio __attribute__((target("sve"))) svfloat16x4_t svreinterpret_f16(svfloat32x4_t op) __arm_streaming_compatible {
		2856	return __builtin_sve_reinterpret_f16_f32_x4(op);
		2857	}
		2858
		2859	__aio __attribute__((target("sve"))) svfloat16x4_t svreinterpret_f16(svfloat64x4_t op) __arm_streaming_compatible {
		2860	return __builtin_sve_reinterpret_f16_f64_x4(op);
		2861	}
		2862
		2863	__aio __attribute__((target("sve"))) svbfloat16x4_t svreinterpret_bf16(svint8x4_t op) __arm_streaming_compatible {
		2864	return __builtin_sve_reinterpret_bf16_s8_x4(op);
		2865	}
		2866
		2867	__aio __attribute__((target("sve"))) svbfloat16x4_t svreinterpret_bf16(svuint8x4_t op) __arm_streaming_compatible {
		2868	return __builtin_sve_reinterpret_bf16_u8_x4(op);
		2869	}
		2870
		2871	__aio __attribute__((target("sve"))) svbfloat16x4_t svreinterpret_bf16(svint16x4_t op) __arm_streaming_compatible {
		2872	return __builtin_sve_reinterpret_bf16_s16_x4(op);
		2873	}
		2874
		2875	__aio __attribute__((target("sve"))) svbfloat16x4_t svreinterpret_bf16(svuint16x4_t op) __arm_streaming_compatible {
		2876	return __builtin_sve_reinterpret_bf16_u16_x4(op);
		2877	}
		2878
		2879	__aio __attribute__((target("sve"))) svbfloat16x4_t svreinterpret_bf16(svint32x4_t op) __arm_streaming_compatible {
		2880	return __builtin_sve_reinterpret_bf16_s32_x4(op);
		2881	}
		2882
		2883	__aio __attribute__((target("sve"))) svbfloat16x4_t svreinterpret_bf16(svuint32x4_t op) __arm_streaming_compatible {
		2884	return __builtin_sve_reinterpret_bf16_u32_x4(op);
		2885	}
		2886
		2887	__aio __attribute__((target("sve"))) svbfloat16x4_t svreinterpret_bf16(svint64x4_t op) __arm_streaming_compatible {
		2888	return __builtin_sve_reinterpret_bf16_s64_x4(op);
		2889	}
		2890
		2891	__aio __attribute__((target("sve"))) svbfloat16x4_t svreinterpret_bf16(svuint64x4_t op) __arm_streaming_compatible {
		2892	return __builtin_sve_reinterpret_bf16_u64_x4(op);
		2893	}
		2894
		2895	__aio __attribute__((target("sve"))) svbfloat16x4_t svreinterpret_bf16(svfloat16x4_t op) __arm_streaming_compatible {
		2896	return __builtin_sve_reinterpret_bf16_f16_x4(op);
		2897	}
		2898
		2899	__aio __attribute__((target("sve"))) svbfloat16x4_t svreinterpret_bf16(svbfloat16x4_t op) __arm_streaming_compatible {
		2900	return __builtin_sve_reinterpret_bf16_bf16_x4(op);
		2901	}
		2902
		2903	__aio __attribute__((target("sve"))) svbfloat16x4_t svreinterpret_bf16(svfloat32x4_t op) __arm_streaming_compatible {
		2904	return __builtin_sve_reinterpret_bf16_f32_x4(op);
		2905	}
		2906
		2907	__aio __attribute__((target("sve"))) svbfloat16x4_t svreinterpret_bf16(svfloat64x4_t op) __arm_streaming_compatible {
		2908	return __builtin_sve_reinterpret_bf16_f64_x4(op);
		2909	}
		2910
		2911	__aio __attribute__((target("sve"))) svfloat32x4_t svreinterpret_f32(svint8x4_t op) __arm_streaming_compatible {
		2912	return __builtin_sve_reinterpret_f32_s8_x4(op);
		2913	}
		2914
		2915	__aio __attribute__((target("sve"))) svfloat32x4_t svreinterpret_f32(svuint8x4_t op) __arm_streaming_compatible {
		2916	return __builtin_sve_reinterpret_f32_u8_x4(op);
		2917	}
		2918
		2919	__aio __attribute__((target("sve"))) svfloat32x4_t svreinterpret_f32(svint16x4_t op) __arm_streaming_compatible {
		2920	return __builtin_sve_reinterpret_f32_s16_x4(op);
		2921	}
		2922
		2923	__aio __attribute__((target("sve"))) svfloat32x4_t svreinterpret_f32(svuint16x4_t op) __arm_streaming_compatible {
		2924	return __builtin_sve_reinterpret_f32_u16_x4(op);
		2925	}
		2926
		2927	__aio __attribute__((target("sve"))) svfloat32x4_t svreinterpret_f32(svint32x4_t op) __arm_streaming_compatible {
		2928	return __builtin_sve_reinterpret_f32_s32_x4(op);
		2929	}
		2930
		2931	__aio __attribute__((target("sve"))) svfloat32x4_t svreinterpret_f32(svuint32x4_t op) __arm_streaming_compatible {
		2932	return __builtin_sve_reinterpret_f32_u32_x4(op);
		2933	}
		2934
		2935	__aio __attribute__((target("sve"))) svfloat32x4_t svreinterpret_f32(svint64x4_t op) __arm_streaming_compatible {
		2936	return __builtin_sve_reinterpret_f32_s64_x4(op);
		2937	}
		2938
		2939	__aio __attribute__((target("sve"))) svfloat32x4_t svreinterpret_f32(svuint64x4_t op) __arm_streaming_compatible {
		2940	return __builtin_sve_reinterpret_f32_u64_x4(op);
		2941	}
		2942
		2943	__aio __attribute__((target("sve"))) svfloat32x4_t svreinterpret_f32(svfloat16x4_t op) __arm_streaming_compatible {
		2944	return __builtin_sve_reinterpret_f32_f16_x4(op);
		2945	}
		2946
		2947	__aio __attribute__((target("sve"))) svfloat32x4_t svreinterpret_f32(svbfloat16x4_t op) __arm_streaming_compatible {
		2948	return __builtin_sve_reinterpret_f32_bf16_x4(op);
		2949	}
		2950
		2951	__aio __attribute__((target("sve"))) svfloat32x4_t svreinterpret_f32(svfloat32x4_t op) __arm_streaming_compatible {
		2952	return __builtin_sve_reinterpret_f32_f32_x4(op);
		2953	}
		2954
		2955	__aio __attribute__((target("sve"))) svfloat32x4_t svreinterpret_f32(svfloat64x4_t op) __arm_streaming_compatible {
		2956	return __builtin_sve_reinterpret_f32_f64_x4(op);
		2957	}
		2958
		2959	__aio __attribute__((target("sve"))) svfloat64x4_t svreinterpret_f64(svint8x4_t op) __arm_streaming_compatible {
		2960	return __builtin_sve_reinterpret_f64_s8_x4(op);
		2961	}
		2962
		2963	__aio __attribute__((target("sve"))) svfloat64x4_t svreinterpret_f64(svuint8x4_t op) __arm_streaming_compatible {
		2964	return __builtin_sve_reinterpret_f64_u8_x4(op);
		2965	}
		2966
		2967	__aio __attribute__((target("sve"))) svfloat64x4_t svreinterpret_f64(svint16x4_t op) __arm_streaming_compatible {
		2968	return __builtin_sve_reinterpret_f64_s16_x4(op);
		2969	}
		2970
		2971	__aio __attribute__((target("sve"))) svfloat64x4_t svreinterpret_f64(svuint16x4_t op) __arm_streaming_compatible {
		2972	return __builtin_sve_reinterpret_f64_u16_x4(op);
		2973	}
		2974
		2975	__aio __attribute__((target("sve"))) svfloat64x4_t svreinterpret_f64(svint32x4_t op) __arm_streaming_compatible {
		2976	return __builtin_sve_reinterpret_f64_s32_x4(op);
		2977	}
		2978
		2979	__aio __attribute__((target("sve"))) svfloat64x4_t svreinterpret_f64(svuint32x4_t op) __arm_streaming_compatible {
		2980	return __builtin_sve_reinterpret_f64_u32_x4(op);
		2981	}
		2982
		2983	__aio __attribute__((target("sve"))) svfloat64x4_t svreinterpret_f64(svint64x4_t op) __arm_streaming_compatible {
		2984	return __builtin_sve_reinterpret_f64_s64_x4(op);
		2985	}
		2986
		2987	__aio __attribute__((target("sve"))) svfloat64x4_t svreinterpret_f64(svuint64x4_t op) __arm_streaming_compatible {
		2988	return __builtin_sve_reinterpret_f64_u64_x4(op);
		2989	}
		2990
		2991	__aio __attribute__((target("sve"))) svfloat64x4_t svreinterpret_f64(svfloat16x4_t op) __arm_streaming_compatible {
		2992	return __builtin_sve_reinterpret_f64_f16_x4(op);
		2993	}
		2994
		2995	__aio __attribute__((target("sve"))) svfloat64x4_t svreinterpret_f64(svbfloat16x4_t op) __arm_streaming_compatible {
		2996	return __builtin_sve_reinterpret_f64_bf16_x4(op);
		2997	}
		2998
		2999	__aio __attribute__((target("sve"))) svfloat64x4_t svreinterpret_f64(svfloat32x4_t op) __arm_streaming_compatible {
		3000	return __builtin_sve_reinterpret_f64_f32_x4(op);
		3001	}
		3002
		3003	__aio __attribute__((target("sve"))) svfloat64x4_t svreinterpret_f64(svfloat64x4_t op) __arm_streaming_compatible {
		3004	return __builtin_sve_reinterpret_f64_f64_x4(op);
844	}	3005	}
845		3006
		3007	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_n_bf16_m)))
		3008	svbfloat16_t svadd_n_bf16_m(svbool_t, svbfloat16_t, bfloat16_t);
		3009	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_n_bf16_x)))
		3010	svbfloat16_t svadd_n_bf16_x(svbool_t, svbfloat16_t, bfloat16_t);
		3011	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_n_bf16_z)))
		3012	svbfloat16_t svadd_n_bf16_z(svbool_t, svbfloat16_t, bfloat16_t);
		3013	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_bf16_m)))
		3014	svbfloat16_t svadd_bf16_m(svbool_t, svbfloat16_t, svbfloat16_t);
		3015	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_bf16_x)))
		3016	svbfloat16_t svadd_bf16_x(svbool_t, svbfloat16_t, svbfloat16_t);
		3017	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_bf16_z)))
		3018	svbfloat16_t svadd_bf16_z(svbool_t, svbfloat16_t, svbfloat16_t);
		3019	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_bf16)))
		3020	svbfloat16_t svclamp_bf16(svbfloat16_t, svbfloat16_t, svbfloat16_t);
		3021	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_n_bf16_m)))
		3022	svbfloat16_t svmax_n_bf16_m(svbool_t, svbfloat16_t, bfloat16_t);
		3023	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_n_bf16_x)))
		3024	svbfloat16_t svmax_n_bf16_x(svbool_t, svbfloat16_t, bfloat16_t);
		3025	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_n_bf16_z)))
		3026	svbfloat16_t svmax_n_bf16_z(svbool_t, svbfloat16_t, bfloat16_t);
		3027	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_bf16_m)))
		3028	svbfloat16_t svmax_bf16_m(svbool_t, svbfloat16_t, svbfloat16_t);
		3029	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_bf16_x)))
		3030	svbfloat16_t svmax_bf16_x(svbool_t, svbfloat16_t, svbfloat16_t);
		3031	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_bf16_z)))
		3032	svbfloat16_t svmax_bf16_z(svbool_t, svbfloat16_t, svbfloat16_t);
		3033	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_n_bf16_m)))
		3034	svbfloat16_t svmaxnm_n_bf16_m(svbool_t, svbfloat16_t, bfloat16_t);
		3035	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_n_bf16_x)))
		3036	svbfloat16_t svmaxnm_n_bf16_x(svbool_t, svbfloat16_t, bfloat16_t);
		3037	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_n_bf16_z)))
		3038	svbfloat16_t svmaxnm_n_bf16_z(svbool_t, svbfloat16_t, bfloat16_t);
		3039	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_bf16_m)))
		3040	svbfloat16_t svmaxnm_bf16_m(svbool_t, svbfloat16_t, svbfloat16_t);
		3041	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_bf16_x)))
		3042	svbfloat16_t svmaxnm_bf16_x(svbool_t, svbfloat16_t, svbfloat16_t);
		3043	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_bf16_z)))
		3044	svbfloat16_t svmaxnm_bf16_z(svbool_t, svbfloat16_t, svbfloat16_t);
		3045	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_n_bf16_m)))
		3046	svbfloat16_t svmin_n_bf16_m(svbool_t, svbfloat16_t, bfloat16_t);
		3047	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_n_bf16_x)))
		3048	svbfloat16_t svmin_n_bf16_x(svbool_t, svbfloat16_t, bfloat16_t);
		3049	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_n_bf16_z)))
		3050	svbfloat16_t svmin_n_bf16_z(svbool_t, svbfloat16_t, bfloat16_t);
		3051	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_bf16_m)))
		3052	svbfloat16_t svmin_bf16_m(svbool_t, svbfloat16_t, svbfloat16_t);
		3053	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_bf16_x)))
		3054	svbfloat16_t svmin_bf16_x(svbool_t, svbfloat16_t, svbfloat16_t);
		3055	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_bf16_z)))
		3056	svbfloat16_t svmin_bf16_z(svbool_t, svbfloat16_t, svbfloat16_t);
		3057	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_n_bf16_m)))
		3058	svbfloat16_t svminnm_n_bf16_m(svbool_t, svbfloat16_t, bfloat16_t);
		3059	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_n_bf16_x)))
		3060	svbfloat16_t svminnm_n_bf16_x(svbool_t, svbfloat16_t, bfloat16_t);
		3061	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_n_bf16_z)))
		3062	svbfloat16_t svminnm_n_bf16_z(svbool_t, svbfloat16_t, bfloat16_t);
		3063	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_bf16_m)))
		3064	svbfloat16_t svminnm_bf16_m(svbool_t, svbfloat16_t, svbfloat16_t);
		3065	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_bf16_x)))
		3066	svbfloat16_t svminnm_bf16_x(svbool_t, svbfloat16_t, svbfloat16_t);
		3067	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_bf16_z)))
		3068	svbfloat16_t svminnm_bf16_z(svbool_t, svbfloat16_t, svbfloat16_t);
		3069	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmla_n_bf16_m)))
		3070	svbfloat16_t svmla_n_bf16_m(svbool_t, svbfloat16_t, svbfloat16_t, bfloat16_t);
		3071	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmla_n_bf16_x)))
		3072	svbfloat16_t svmla_n_bf16_x(svbool_t, svbfloat16_t, svbfloat16_t, bfloat16_t);
		3073	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmla_n_bf16_z)))
		3074	svbfloat16_t svmla_n_bf16_z(svbool_t, svbfloat16_t, svbfloat16_t, bfloat16_t);
		3075	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmla_bf16_m)))
		3076	svbfloat16_t svmla_bf16_m(svbool_t, svbfloat16_t, svbfloat16_t, svbfloat16_t);
		3077	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmla_bf16_x)))
		3078	svbfloat16_t svmla_bf16_x(svbool_t, svbfloat16_t, svbfloat16_t, svbfloat16_t);
		3079	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmla_bf16_z)))
		3080	svbfloat16_t svmla_bf16_z(svbool_t, svbfloat16_t, svbfloat16_t, svbfloat16_t);
		3081	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmla_lane_bf16)))
		3082	svbfloat16_t svmla_lane_bf16(svbfloat16_t, svbfloat16_t, svbfloat16_t, uint64_t);
		3083	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmls_n_bf16_m)))
		3084	svbfloat16_t svmls_n_bf16_m(svbool_t, svbfloat16_t, svbfloat16_t, bfloat16_t);
		3085	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmls_n_bf16_x)))
		3086	svbfloat16_t svmls_n_bf16_x(svbool_t, svbfloat16_t, svbfloat16_t, bfloat16_t);
		3087	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmls_n_bf16_z)))
		3088	svbfloat16_t svmls_n_bf16_z(svbool_t, svbfloat16_t, svbfloat16_t, bfloat16_t);
		3089	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmls_bf16_m)))
		3090	svbfloat16_t svmls_bf16_m(svbool_t, svbfloat16_t, svbfloat16_t, svbfloat16_t);
		3091	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmls_bf16_x)))
		3092	svbfloat16_t svmls_bf16_x(svbool_t, svbfloat16_t, svbfloat16_t, svbfloat16_t);
		3093	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmls_bf16_z)))
		3094	svbfloat16_t svmls_bf16_z(svbool_t, svbfloat16_t, svbfloat16_t, svbfloat16_t);
		3095	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmls_lane_bf16)))
		3096	svbfloat16_t svmls_lane_bf16(svbfloat16_t, svbfloat16_t, svbfloat16_t, uint64_t);
		3097	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmul_n_bf16_m)))
		3098	svbfloat16_t svmul_n_bf16_m(svbool_t, svbfloat16_t, bfloat16_t);
		3099	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmul_n_bf16_x)))
		3100	svbfloat16_t svmul_n_bf16_x(svbool_t, svbfloat16_t, bfloat16_t);
		3101	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmul_n_bf16_z)))
		3102	svbfloat16_t svmul_n_bf16_z(svbool_t, svbfloat16_t, bfloat16_t);
		3103	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmul_bf16_m)))
		3104	svbfloat16_t svmul_bf16_m(svbool_t, svbfloat16_t, svbfloat16_t);
		3105	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmul_bf16_x)))
		3106	svbfloat16_t svmul_bf16_x(svbool_t, svbfloat16_t, svbfloat16_t);
		3107	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmul_bf16_z)))
		3108	svbfloat16_t svmul_bf16_z(svbool_t, svbfloat16_t, svbfloat16_t);
		3109	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmul_lane_bf16)))
		3110	svbfloat16_t svmul_lane_bf16(svbfloat16_t, svbfloat16_t, uint64_t);
		3111	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsub_n_bf16_m)))
		3112	svbfloat16_t svsub_n_bf16_m(svbool_t, svbfloat16_t, bfloat16_t);
		3113	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsub_n_bf16_x)))
		3114	svbfloat16_t svsub_n_bf16_x(svbool_t, svbfloat16_t, bfloat16_t);
		3115	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsub_n_bf16_z)))
		3116	svbfloat16_t svsub_n_bf16_z(svbool_t, svbfloat16_t, bfloat16_t);
		3117	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsub_bf16_m)))
		3118	svbfloat16_t svsub_bf16_m(svbool_t, svbfloat16_t, svbfloat16_t);
		3119	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsub_bf16_x)))
		3120	svbfloat16_t svsub_bf16_x(svbool_t, svbfloat16_t, svbfloat16_t);
		3121	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsub_bf16_z)))
		3122	svbfloat16_t svsub_bf16_z(svbool_t, svbfloat16_t, svbfloat16_t);
		3123	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_n_bf16_m)))
		3124	svbfloat16_t svadd_m(svbool_t, svbfloat16_t, bfloat16_t);
		3125	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_n_bf16_x)))
		3126	svbfloat16_t svadd_x(svbool_t, svbfloat16_t, bfloat16_t);
		3127	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_n_bf16_z)))
		3128	svbfloat16_t svadd_z(svbool_t, svbfloat16_t, bfloat16_t);
		3129	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_bf16_m)))
		3130	svbfloat16_t svadd_m(svbool_t, svbfloat16_t, svbfloat16_t);
		3131	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_bf16_x)))
		3132	svbfloat16_t svadd_x(svbool_t, svbfloat16_t, svbfloat16_t);
		3133	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_bf16_z)))
		3134	svbfloat16_t svadd_z(svbool_t, svbfloat16_t, svbfloat16_t);
		3135	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_bf16)))
		3136	svbfloat16_t svclamp(svbfloat16_t, svbfloat16_t, svbfloat16_t);
		3137	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_n_bf16_m)))
		3138	svbfloat16_t svmax_m(svbool_t, svbfloat16_t, bfloat16_t);
		3139	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_n_bf16_x)))
		3140	svbfloat16_t svmax_x(svbool_t, svbfloat16_t, bfloat16_t);
		3141	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_n_bf16_z)))
		3142	svbfloat16_t svmax_z(svbool_t, svbfloat16_t, bfloat16_t);
		3143	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_bf16_m)))
		3144	svbfloat16_t svmax_m(svbool_t, svbfloat16_t, svbfloat16_t);
		3145	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_bf16_x)))
		3146	svbfloat16_t svmax_x(svbool_t, svbfloat16_t, svbfloat16_t);
		3147	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_bf16_z)))
		3148	svbfloat16_t svmax_z(svbool_t, svbfloat16_t, svbfloat16_t);
		3149	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_n_bf16_m)))
		3150	svbfloat16_t svmaxnm_m(svbool_t, svbfloat16_t, bfloat16_t);
		3151	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_n_bf16_x)))
		3152	svbfloat16_t svmaxnm_x(svbool_t, svbfloat16_t, bfloat16_t);
		3153	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_n_bf16_z)))
		3154	svbfloat16_t svmaxnm_z(svbool_t, svbfloat16_t, bfloat16_t);
		3155	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_bf16_m)))
		3156	svbfloat16_t svmaxnm_m(svbool_t, svbfloat16_t, svbfloat16_t);
		3157	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_bf16_x)))
		3158	svbfloat16_t svmaxnm_x(svbool_t, svbfloat16_t, svbfloat16_t);
		3159	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_bf16_z)))
		3160	svbfloat16_t svmaxnm_z(svbool_t, svbfloat16_t, svbfloat16_t);
		3161	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_n_bf16_m)))
		3162	svbfloat16_t svmin_m(svbool_t, svbfloat16_t, bfloat16_t);
		3163	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_n_bf16_x)))
		3164	svbfloat16_t svmin_x(svbool_t, svbfloat16_t, bfloat16_t);
		3165	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_n_bf16_z)))
		3166	svbfloat16_t svmin_z(svbool_t, svbfloat16_t, bfloat16_t);
		3167	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_bf16_m)))
		3168	svbfloat16_t svmin_m(svbool_t, svbfloat16_t, svbfloat16_t);
		3169	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_bf16_x)))
		3170	svbfloat16_t svmin_x(svbool_t, svbfloat16_t, svbfloat16_t);
		3171	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_bf16_z)))
		3172	svbfloat16_t svmin_z(svbool_t, svbfloat16_t, svbfloat16_t);
		3173	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_n_bf16_m)))
		3174	svbfloat16_t svminnm_m(svbool_t, svbfloat16_t, bfloat16_t);
		3175	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_n_bf16_x)))
		3176	svbfloat16_t svminnm_x(svbool_t, svbfloat16_t, bfloat16_t);
		3177	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_n_bf16_z)))
		3178	svbfloat16_t svminnm_z(svbool_t, svbfloat16_t, bfloat16_t);
		3179	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_bf16_m)))
		3180	svbfloat16_t svminnm_m(svbool_t, svbfloat16_t, svbfloat16_t);
		3181	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_bf16_x)))
		3182	svbfloat16_t svminnm_x(svbool_t, svbfloat16_t, svbfloat16_t);
		3183	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_bf16_z)))
		3184	svbfloat16_t svminnm_z(svbool_t, svbfloat16_t, svbfloat16_t);
		3185	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmla_n_bf16_m)))
		3186	svbfloat16_t svmla_m(svbool_t, svbfloat16_t, svbfloat16_t, bfloat16_t);
		3187	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmla_n_bf16_x)))
		3188	svbfloat16_t svmla_x(svbool_t, svbfloat16_t, svbfloat16_t, bfloat16_t);
		3189	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmla_n_bf16_z)))
		3190	svbfloat16_t svmla_z(svbool_t, svbfloat16_t, svbfloat16_t, bfloat16_t);
		3191	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmla_bf16_m)))
		3192	svbfloat16_t svmla_m(svbool_t, svbfloat16_t, svbfloat16_t, svbfloat16_t);
		3193	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmla_bf16_x)))
		3194	svbfloat16_t svmla_x(svbool_t, svbfloat16_t, svbfloat16_t, svbfloat16_t);
		3195	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmla_bf16_z)))
		3196	svbfloat16_t svmla_z(svbool_t, svbfloat16_t, svbfloat16_t, svbfloat16_t);
		3197	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmla_lane_bf16)))
		3198	svbfloat16_t svmla_lane(svbfloat16_t, svbfloat16_t, svbfloat16_t, uint64_t);
		3199	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmls_n_bf16_m)))
		3200	svbfloat16_t svmls_m(svbool_t, svbfloat16_t, svbfloat16_t, bfloat16_t);
		3201	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmls_n_bf16_x)))
		3202	svbfloat16_t svmls_x(svbool_t, svbfloat16_t, svbfloat16_t, bfloat16_t);
		3203	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmls_n_bf16_z)))
		3204	svbfloat16_t svmls_z(svbool_t, svbfloat16_t, svbfloat16_t, bfloat16_t);
		3205	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmls_bf16_m)))
		3206	svbfloat16_t svmls_m(svbool_t, svbfloat16_t, svbfloat16_t, svbfloat16_t);
		3207	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmls_bf16_x)))
		3208	svbfloat16_t svmls_x(svbool_t, svbfloat16_t, svbfloat16_t, svbfloat16_t);
		3209	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmls_bf16_z)))
		3210	svbfloat16_t svmls_z(svbool_t, svbfloat16_t, svbfloat16_t, svbfloat16_t);
		3211	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmls_lane_bf16)))
		3212	svbfloat16_t svmls_lane(svbfloat16_t, svbfloat16_t, svbfloat16_t, uint64_t);
		3213	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmul_n_bf16_m)))
		3214	svbfloat16_t svmul_m(svbool_t, svbfloat16_t, bfloat16_t);
		3215	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmul_n_bf16_x)))
		3216	svbfloat16_t svmul_x(svbool_t, svbfloat16_t, bfloat16_t);
		3217	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmul_n_bf16_z)))
		3218	svbfloat16_t svmul_z(svbool_t, svbfloat16_t, bfloat16_t);
		3219	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmul_bf16_m)))
		3220	svbfloat16_t svmul_m(svbool_t, svbfloat16_t, svbfloat16_t);
		3221	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmul_bf16_x)))
		3222	svbfloat16_t svmul_x(svbool_t, svbfloat16_t, svbfloat16_t);
		3223	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmul_bf16_z)))
		3224	svbfloat16_t svmul_z(svbool_t, svbfloat16_t, svbfloat16_t);
		3225	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmul_lane_bf16)))
		3226	svbfloat16_t svmul_lane(svbfloat16_t, svbfloat16_t, uint64_t);
		3227	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsub_n_bf16_m)))
		3228	svbfloat16_t svsub_m(svbool_t, svbfloat16_t, bfloat16_t);
		3229	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsub_n_bf16_x)))
		3230	svbfloat16_t svsub_x(svbool_t, svbfloat16_t, bfloat16_t);
		3231	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsub_n_bf16_z)))
		3232	svbfloat16_t svsub_z(svbool_t, svbfloat16_t, bfloat16_t);
		3233	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsub_bf16_m)))
		3234	svbfloat16_t svsub_m(svbool_t, svbfloat16_t, svbfloat16_t);
		3235	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsub_bf16_x)))
		3236	svbfloat16_t svsub_x(svbool_t, svbfloat16_t, svbfloat16_t);
		3237	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsub_bf16_z)))
		3238	svbfloat16_t svsub_z(svbool_t, svbfloat16_t, svbfloat16_t);
		3239	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_u8_x2)))
		3240	svuint8x2_t svadd_single_u8_x2(svuint8x2_t, svuint8_t);
		3241	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_u32_x2)))
		3242	svuint32x2_t svadd_single_u32_x2(svuint32x2_t, svuint32_t);
		3243	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_u64_x2)))
		3244	svuint64x2_t svadd_single_u64_x2(svuint64x2_t, svuint64_t);
		3245	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_u16_x2)))
		3246	svuint16x2_t svadd_single_u16_x2(svuint16x2_t, svuint16_t);
		3247	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_s8_x2)))
		3248	svint8x2_t svadd_single_s8_x2(svint8x2_t, svint8_t);
		3249	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_s32_x2)))
		3250	svint32x2_t svadd_single_s32_x2(svint32x2_t, svint32_t);
		3251	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_s64_x2)))
		3252	svint64x2_t svadd_single_s64_x2(svint64x2_t, svint64_t);
		3253	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_s16_x2)))
		3254	svint16x2_t svadd_single_s16_x2(svint16x2_t, svint16_t);
		3255	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_u8_x4)))
		3256	svuint8x4_t svadd_single_u8_x4(svuint8x4_t, svuint8_t);
		3257	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_u32_x4)))
		3258	svuint32x4_t svadd_single_u32_x4(svuint32x4_t, svuint32_t);
		3259	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_u64_x4)))
		3260	svuint64x4_t svadd_single_u64_x4(svuint64x4_t, svuint64_t);
		3261	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_u16_x4)))
		3262	svuint16x4_t svadd_single_u16_x4(svuint16x4_t, svuint16_t);
		3263	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_s8_x4)))
		3264	svint8x4_t svadd_single_s8_x4(svint8x4_t, svint8_t);
		3265	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_s32_x4)))
		3266	svint32x4_t svadd_single_s32_x4(svint32x4_t, svint32_t);
		3267	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_s64_x4)))
		3268	svint64x4_t svadd_single_s64_x4(svint64x4_t, svint64_t);
		3269	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_s16_x4)))
		3270	svint16x4_t svadd_single_s16_x4(svint16x4_t, svint16_t);
		3271	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_f64_x2)))
		3272	svfloat64x2_t svclamp_single_f64_x2(svfloat64x2_t, svfloat64_t, svfloat64_t);
		3273	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_f32_x2)))
		3274	svfloat32x2_t svclamp_single_f32_x2(svfloat32x2_t, svfloat32_t, svfloat32_t);
		3275	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_f16_x2)))
		3276	svfloat16x2_t svclamp_single_f16_x2(svfloat16x2_t, svfloat16_t, svfloat16_t);
		3277	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_s8_x2)))
		3278	svint8x2_t svclamp_single_s8_x2(svint8x2_t, svint8_t, svint8_t);
		3279	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_s32_x2)))
		3280	svint32x2_t svclamp_single_s32_x2(svint32x2_t, svint32_t, svint32_t);
		3281	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_s64_x2)))
		3282	svint64x2_t svclamp_single_s64_x2(svint64x2_t, svint64_t, svint64_t);
		3283	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_s16_x2)))
		3284	svint16x2_t svclamp_single_s16_x2(svint16x2_t, svint16_t, svint16_t);
		3285	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_u8_x2)))
		3286	svuint8x2_t svclamp_single_u8_x2(svuint8x2_t, svuint8_t, svuint8_t);
		3287	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_u32_x2)))
		3288	svuint32x2_t svclamp_single_u32_x2(svuint32x2_t, svuint32_t, svuint32_t);
		3289	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_u64_x2)))
		3290	svuint64x2_t svclamp_single_u64_x2(svuint64x2_t, svuint64_t, svuint64_t);
		3291	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_u16_x2)))
		3292	svuint16x2_t svclamp_single_u16_x2(svuint16x2_t, svuint16_t, svuint16_t);
		3293	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_f64_x4)))
		3294	svfloat64x4_t svclamp_single_f64_x4(svfloat64x4_t, svfloat64_t, svfloat64_t);
		3295	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_f32_x4)))
		3296	svfloat32x4_t svclamp_single_f32_x4(svfloat32x4_t, svfloat32_t, svfloat32_t);
		3297	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_f16_x4)))
		3298	svfloat16x4_t svclamp_single_f16_x4(svfloat16x4_t, svfloat16_t, svfloat16_t);
		3299	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_s8_x4)))
		3300	svint8x4_t svclamp_single_s8_x4(svint8x4_t, svint8_t, svint8_t);
		3301	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_s32_x4)))
		3302	svint32x4_t svclamp_single_s32_x4(svint32x4_t, svint32_t, svint32_t);
		3303	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_s64_x4)))
		3304	svint64x4_t svclamp_single_s64_x4(svint64x4_t, svint64_t, svint64_t);
		3305	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_s16_x4)))
		3306	svint16x4_t svclamp_single_s16_x4(svint16x4_t, svint16_t, svint16_t);
		3307	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_u8_x4)))
		3308	svuint8x4_t svclamp_single_u8_x4(svuint8x4_t, svuint8_t, svuint8_t);
		3309	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_u32_x4)))
		3310	svuint32x4_t svclamp_single_u32_x4(svuint32x4_t, svuint32_t, svuint32_t);
		3311	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_u64_x4)))
		3312	svuint64x4_t svclamp_single_u64_x4(svuint64x4_t, svuint64_t, svuint64_t);
		3313	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_u16_x4)))
		3314	svuint16x4_t svclamp_single_u16_x4(svuint16x4_t, svuint16_t, svuint16_t);
		3315	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_bf16_f32_x2)))
		3316	svbfloat16_t svcvt_bf16_f32_x2(svfloat32x2_t);
		3317	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_f16_f32_x2)))
		3318	svfloat16_t svcvt_f16_f32_x2(svfloat32x2_t);
		3319	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_s32_f32_x2)))
		3320	svint32x2_t svcvt_s32_f32_x2(svfloat32x2_t);
		3321	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_u32_f32_x2)))
		3322	svuint32x2_t svcvt_u32_f32_x2(svfloat32x2_t);
		3323	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_s32_f32_x4)))
		3324	svint32x4_t svcvt_s32_f32_x4(svfloat32x4_t);
		3325	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_u32_f32_x4)))
		3326	svuint32x4_t svcvt_u32_f32_x4(svfloat32x4_t);
		3327	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_f32_s32_x2)))
		3328	svfloat32x2_t svcvt_f32_s32_x2(svint32x2_t);
		3329	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_f32_s32_x4)))
		3330	svfloat32x4_t svcvt_f32_s32_x4(svint32x4_t);
		3331	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_f32_u32_x2)))
		3332	svfloat32x2_t svcvt_f32_u32_x2(svuint32x2_t);
		3333	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_f32_u32_x4)))
		3334	svfloat32x4_t svcvt_f32_u32_x4(svuint32x4_t);
		3335	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvtn_bf16_f32_x2)))
		3336	svbfloat16_t svcvtn_bf16_f32_x2(svfloat32x2_t);
		3337	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvtn_f16_f32_x2)))
		3338	svfloat16_t svcvtn_f16_f32_x2(svfloat32x2_t);
		3339	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_f64_x2)))
		3340	svfloat64x2_t svmax_single_f64_x2(svfloat64x2_t, svfloat64_t);
		3341	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_f32_x2)))
		3342	svfloat32x2_t svmax_single_f32_x2(svfloat32x2_t, svfloat32_t);
		3343	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_f16_x2)))
		3344	svfloat16x2_t svmax_single_f16_x2(svfloat16x2_t, svfloat16_t);
		3345	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_s8_x2)))
		3346	svint8x2_t svmax_single_s8_x2(svint8x2_t, svint8_t);
		3347	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_s32_x2)))
		3348	svint32x2_t svmax_single_s32_x2(svint32x2_t, svint32_t);
		3349	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_s64_x2)))
		3350	svint64x2_t svmax_single_s64_x2(svint64x2_t, svint64_t);
		3351	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_s16_x2)))
		3352	svint16x2_t svmax_single_s16_x2(svint16x2_t, svint16_t);
		3353	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_u8_x2)))
		3354	svuint8x2_t svmax_single_u8_x2(svuint8x2_t, svuint8_t);
		3355	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_u32_x2)))
		3356	svuint32x2_t svmax_single_u32_x2(svuint32x2_t, svuint32_t);
		3357	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_u64_x2)))
		3358	svuint64x2_t svmax_single_u64_x2(svuint64x2_t, svuint64_t);
		3359	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_u16_x2)))
		3360	svuint16x2_t svmax_single_u16_x2(svuint16x2_t, svuint16_t);
		3361	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_f64_x4)))
		3362	svfloat64x4_t svmax_single_f64_x4(svfloat64x4_t, svfloat64_t);
		3363	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_f32_x4)))
		3364	svfloat32x4_t svmax_single_f32_x4(svfloat32x4_t, svfloat32_t);
		3365	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_f16_x4)))
		3366	svfloat16x4_t svmax_single_f16_x4(svfloat16x4_t, svfloat16_t);
		3367	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_s8_x4)))
		3368	svint8x4_t svmax_single_s8_x4(svint8x4_t, svint8_t);
		3369	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_s32_x4)))
		3370	svint32x4_t svmax_single_s32_x4(svint32x4_t, svint32_t);
		3371	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_s64_x4)))
		3372	svint64x4_t svmax_single_s64_x4(svint64x4_t, svint64_t);
		3373	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_s16_x4)))
		3374	svint16x4_t svmax_single_s16_x4(svint16x4_t, svint16_t);
		3375	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_u8_x4)))
		3376	svuint8x4_t svmax_single_u8_x4(svuint8x4_t, svuint8_t);
		3377	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_u32_x4)))
		3378	svuint32x4_t svmax_single_u32_x4(svuint32x4_t, svuint32_t);
		3379	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_u64_x4)))
		3380	svuint64x4_t svmax_single_u64_x4(svuint64x4_t, svuint64_t);
		3381	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_u16_x4)))
		3382	svuint16x4_t svmax_single_u16_x4(svuint16x4_t, svuint16_t);
		3383	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_f64_x2)))
		3384	svfloat64x2_t svmax_f64_x2(svfloat64x2_t, svfloat64x2_t);
		3385	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_f32_x2)))
		3386	svfloat32x2_t svmax_f32_x2(svfloat32x2_t, svfloat32x2_t);
		3387	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_f16_x2)))
		3388	svfloat16x2_t svmax_f16_x2(svfloat16x2_t, svfloat16x2_t);
		3389	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_s8_x2)))
		3390	svint8x2_t svmax_s8_x2(svint8x2_t, svint8x2_t);
		3391	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_s32_x2)))
		3392	svint32x2_t svmax_s32_x2(svint32x2_t, svint32x2_t);
		3393	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_s64_x2)))
		3394	svint64x2_t svmax_s64_x2(svint64x2_t, svint64x2_t);
		3395	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_s16_x2)))
		3396	svint16x2_t svmax_s16_x2(svint16x2_t, svint16x2_t);
		3397	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_u8_x2)))
		3398	svuint8x2_t svmax_u8_x2(svuint8x2_t, svuint8x2_t);
		3399	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_u32_x2)))
		3400	svuint32x2_t svmax_u32_x2(svuint32x2_t, svuint32x2_t);
		3401	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_u64_x2)))
		3402	svuint64x2_t svmax_u64_x2(svuint64x2_t, svuint64x2_t);
		3403	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_u16_x2)))
		3404	svuint16x2_t svmax_u16_x2(svuint16x2_t, svuint16x2_t);
		3405	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_f64_x4)))
		3406	svfloat64x4_t svmax_f64_x4(svfloat64x4_t, svfloat64x4_t);
		3407	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_f32_x4)))
		3408	svfloat32x4_t svmax_f32_x4(svfloat32x4_t, svfloat32x4_t);
		3409	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_f16_x4)))
		3410	svfloat16x4_t svmax_f16_x4(svfloat16x4_t, svfloat16x4_t);
		3411	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_s8_x4)))
		3412	svint8x4_t svmax_s8_x4(svint8x4_t, svint8x4_t);
		3413	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_s32_x4)))
		3414	svint32x4_t svmax_s32_x4(svint32x4_t, svint32x4_t);
		3415	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_s64_x4)))
		3416	svint64x4_t svmax_s64_x4(svint64x4_t, svint64x4_t);
		3417	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_s16_x4)))
		3418	svint16x4_t svmax_s16_x4(svint16x4_t, svint16x4_t);
		3419	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_u8_x4)))
		3420	svuint8x4_t svmax_u8_x4(svuint8x4_t, svuint8x4_t);
		3421	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_u32_x4)))
		3422	svuint32x4_t svmax_u32_x4(svuint32x4_t, svuint32x4_t);
		3423	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_u64_x4)))
		3424	svuint64x4_t svmax_u64_x4(svuint64x4_t, svuint64x4_t);
		3425	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_u16_x4)))
		3426	svuint16x4_t svmax_u16_x4(svuint16x4_t, svuint16x4_t);
		3427	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_single_f64_x2)))
		3428	svfloat64x2_t svmaxnm_single_f64_x2(svfloat64x2_t, svfloat64_t);
		3429	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_single_f32_x2)))
		3430	svfloat32x2_t svmaxnm_single_f32_x2(svfloat32x2_t, svfloat32_t);
		3431	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_single_f16_x2)))
		3432	svfloat16x2_t svmaxnm_single_f16_x2(svfloat16x2_t, svfloat16_t);
		3433	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_single_f64_x4)))
		3434	svfloat64x4_t svmaxnm_single_f64_x4(svfloat64x4_t, svfloat64_t);
		3435	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_single_f32_x4)))
		3436	svfloat32x4_t svmaxnm_single_f32_x4(svfloat32x4_t, svfloat32_t);
		3437	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_single_f16_x4)))
		3438	svfloat16x4_t svmaxnm_single_f16_x4(svfloat16x4_t, svfloat16_t);
		3439	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_f64_x2)))
		3440	svfloat64x2_t svmaxnm_f64_x2(svfloat64x2_t, svfloat64x2_t);
		3441	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_f32_x2)))
		3442	svfloat32x2_t svmaxnm_f32_x2(svfloat32x2_t, svfloat32x2_t);
		3443	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_f16_x2)))
		3444	svfloat16x2_t svmaxnm_f16_x2(svfloat16x2_t, svfloat16x2_t);
		3445	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_f64_x4)))
		3446	svfloat64x4_t svmaxnm_f64_x4(svfloat64x4_t, svfloat64x4_t);
		3447	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_f32_x4)))
		3448	svfloat32x4_t svmaxnm_f32_x4(svfloat32x4_t, svfloat32x4_t);
		3449	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_f16_x4)))
		3450	svfloat16x4_t svmaxnm_f16_x4(svfloat16x4_t, svfloat16x4_t);
		3451	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_f64_x2)))
		3452	svfloat64x2_t svmin_single_f64_x2(svfloat64x2_t, svfloat64_t);
		3453	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_f32_x2)))
		3454	svfloat32x2_t svmin_single_f32_x2(svfloat32x2_t, svfloat32_t);
		3455	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_f16_x2)))
		3456	svfloat16x2_t svmin_single_f16_x2(svfloat16x2_t, svfloat16_t);
		3457	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_s8_x2)))
		3458	svint8x2_t svmin_single_s8_x2(svint8x2_t, svint8_t);
		3459	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_s32_x2)))
		3460	svint32x2_t svmin_single_s32_x2(svint32x2_t, svint32_t);
		3461	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_s64_x2)))
		3462	svint64x2_t svmin_single_s64_x2(svint64x2_t, svint64_t);
		3463	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_s16_x2)))
		3464	svint16x2_t svmin_single_s16_x2(svint16x2_t, svint16_t);
		3465	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_u8_x2)))
		3466	svuint8x2_t svmin_single_u8_x2(svuint8x2_t, svuint8_t);
		3467	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_u32_x2)))
		3468	svuint32x2_t svmin_single_u32_x2(svuint32x2_t, svuint32_t);
		3469	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_u64_x2)))
		3470	svuint64x2_t svmin_single_u64_x2(svuint64x2_t, svuint64_t);
		3471	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_u16_x2)))
		3472	svuint16x2_t svmin_single_u16_x2(svuint16x2_t, svuint16_t);
		3473	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_f64_x4)))
		3474	svfloat64x4_t svmin_single_f64_x4(svfloat64x4_t, svfloat64_t);
		3475	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_f32_x4)))
		3476	svfloat32x4_t svmin_single_f32_x4(svfloat32x4_t, svfloat32_t);
		3477	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_f16_x4)))
		3478	svfloat16x4_t svmin_single_f16_x4(svfloat16x4_t, svfloat16_t);
		3479	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_s8_x4)))
		3480	svint8x4_t svmin_single_s8_x4(svint8x4_t, svint8_t);
		3481	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_s32_x4)))
		3482	svint32x4_t svmin_single_s32_x4(svint32x4_t, svint32_t);
		3483	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_s64_x4)))
		3484	svint64x4_t svmin_single_s64_x4(svint64x4_t, svint64_t);
		3485	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_s16_x4)))
		3486	svint16x4_t svmin_single_s16_x4(svint16x4_t, svint16_t);
		3487	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_u8_x4)))
		3488	svuint8x4_t svmin_single_u8_x4(svuint8x4_t, svuint8_t);
		3489	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_u32_x4)))
		3490	svuint32x4_t svmin_single_u32_x4(svuint32x4_t, svuint32_t);
		3491	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_u64_x4)))
		3492	svuint64x4_t svmin_single_u64_x4(svuint64x4_t, svuint64_t);
		3493	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_u16_x4)))
		3494	svuint16x4_t svmin_single_u16_x4(svuint16x4_t, svuint16_t);
		3495	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_f64_x2)))
		3496	svfloat64x2_t svmin_f64_x2(svfloat64x2_t, svfloat64x2_t);
		3497	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_f32_x2)))
		3498	svfloat32x2_t svmin_f32_x2(svfloat32x2_t, svfloat32x2_t);
		3499	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_f16_x2)))
		3500	svfloat16x2_t svmin_f16_x2(svfloat16x2_t, svfloat16x2_t);
		3501	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_s8_x2)))
		3502	svint8x2_t svmin_s8_x2(svint8x2_t, svint8x2_t);
		3503	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_s32_x2)))
		3504	svint32x2_t svmin_s32_x2(svint32x2_t, svint32x2_t);
		3505	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_s64_x2)))
		3506	svint64x2_t svmin_s64_x2(svint64x2_t, svint64x2_t);
		3507	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_s16_x2)))
		3508	svint16x2_t svmin_s16_x2(svint16x2_t, svint16x2_t);
		3509	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_u8_x2)))
		3510	svuint8x2_t svmin_u8_x2(svuint8x2_t, svuint8x2_t);
		3511	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_u32_x2)))
		3512	svuint32x2_t svmin_u32_x2(svuint32x2_t, svuint32x2_t);
		3513	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_u64_x2)))
		3514	svuint64x2_t svmin_u64_x2(svuint64x2_t, svuint64x2_t);
		3515	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_u16_x2)))
		3516	svuint16x2_t svmin_u16_x2(svuint16x2_t, svuint16x2_t);
		3517	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_f64_x4)))
		3518	svfloat64x4_t svmin_f64_x4(svfloat64x4_t, svfloat64x4_t);
		3519	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_f32_x4)))
		3520	svfloat32x4_t svmin_f32_x4(svfloat32x4_t, svfloat32x4_t);
		3521	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_f16_x4)))
		3522	svfloat16x4_t svmin_f16_x4(svfloat16x4_t, svfloat16x4_t);
		3523	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_s8_x4)))
		3524	svint8x4_t svmin_s8_x4(svint8x4_t, svint8x4_t);
		3525	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_s32_x4)))
		3526	svint32x4_t svmin_s32_x4(svint32x4_t, svint32x4_t);
		3527	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_s64_x4)))
		3528	svint64x4_t svmin_s64_x4(svint64x4_t, svint64x4_t);
		3529	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_s16_x4)))
		3530	svint16x4_t svmin_s16_x4(svint16x4_t, svint16x4_t);
		3531	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_u8_x4)))
		3532	svuint8x4_t svmin_u8_x4(svuint8x4_t, svuint8x4_t);
		3533	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_u32_x4)))
		3534	svuint32x4_t svmin_u32_x4(svuint32x4_t, svuint32x4_t);
		3535	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_u64_x4)))
		3536	svuint64x4_t svmin_u64_x4(svuint64x4_t, svuint64x4_t);
		3537	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_u16_x4)))
		3538	svuint16x4_t svmin_u16_x4(svuint16x4_t, svuint16x4_t);
		3539	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_single_f64_x2)))
		3540	svfloat64x2_t svminnm_single_f64_x2(svfloat64x2_t, svfloat64_t);
		3541	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_single_f32_x2)))
		3542	svfloat32x2_t svminnm_single_f32_x2(svfloat32x2_t, svfloat32_t);
		3543	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_single_f16_x2)))
		3544	svfloat16x2_t svminnm_single_f16_x2(svfloat16x2_t, svfloat16_t);
		3545	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_single_f64_x4)))
		3546	svfloat64x4_t svminnm_single_f64_x4(svfloat64x4_t, svfloat64_t);
		3547	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_single_f32_x4)))
		3548	svfloat32x4_t svminnm_single_f32_x4(svfloat32x4_t, svfloat32_t);
		3549	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_single_f16_x4)))
		3550	svfloat16x4_t svminnm_single_f16_x4(svfloat16x4_t, svfloat16_t);
		3551	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_f64_x2)))
		3552	svfloat64x2_t svminnm_f64_x2(svfloat64x2_t, svfloat64x2_t);
		3553	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_f32_x2)))
		3554	svfloat32x2_t svminnm_f32_x2(svfloat32x2_t, svfloat32x2_t);
		3555	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_f16_x2)))
		3556	svfloat16x2_t svminnm_f16_x2(svfloat16x2_t, svfloat16x2_t);
		3557	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_f64_x4)))
		3558	svfloat64x4_t svminnm_f64_x4(svfloat64x4_t, svfloat64x4_t);
		3559	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_f32_x4)))
		3560	svfloat32x4_t svminnm_f32_x4(svfloat32x4_t, svfloat32x4_t);
		3561	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_f16_x4)))
		3562	svfloat16x4_t svminnm_f16_x4(svfloat16x4_t, svfloat16x4_t);
		3563	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_s16_s32_x2)))
		3564	svint16_t svqcvt_s16_s32_x2(svint32x2_t);
		3565	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_s16_s64_x4)))
		3566	svint16_t svqcvt_s16_s64_x4(svint64x4_t);
		3567	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_s8_s32_x4)))
		3568	svint8_t svqcvt_s8_s32_x4(svint32x4_t);
		3569	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_u16_s32_x2)))
		3570	svuint16_t svqcvt_u16_s32_x2(svint32x2_t);
		3571	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_u16_u32_x2)))
		3572	svuint16_t svqcvt_u16_u32_x2(svuint32x2_t);
		3573	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_u16_s64_x4)))
		3574	svuint16_t svqcvt_u16_s64_x4(svint64x4_t);
		3575	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_u16_u64_x4)))
		3576	svuint16_t svqcvt_u16_u64_x4(svuint64x4_t);
		3577	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_u8_s32_x4)))
		3578	svuint8_t svqcvt_u8_s32_x4(svint32x4_t);
		3579	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_u8_u32_x4)))
		3580	svuint8_t svqcvt_u8_u32_x4(svuint32x4_t);
		3581	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_s16_s64_x4)))
		3582	svint16_t svqcvtn_s16_s64_x4(svint64x4_t);
		3583	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_s8_s32_x4)))
		3584	svint8_t svqcvtn_s8_s32_x4(svint32x4_t);
		3585	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_u16_s64_x4)))
		3586	svuint16_t svqcvtn_u16_s64_x4(svint64x4_t);
		3587	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_u16_u64_x4)))
		3588	svuint16_t svqcvtn_u16_u64_x4(svuint64x4_t);
		3589	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_u8_s32_x4)))
		3590	svuint8_t svqcvtn_u8_s32_x4(svint32x4_t);
		3591	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_u8_u32_x4)))
		3592	svuint8_t svqcvtn_u8_u32_x4(svuint32x4_t);
		3593	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_single_s8_x2)))
		3594	svint8x2_t svqdmulh_single_s8_x2(svint8x2_t, svint8_t);
		3595	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_single_s32_x2)))
		3596	svint32x2_t svqdmulh_single_s32_x2(svint32x2_t, svint32_t);
		3597	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_single_s64_x2)))
		3598	svint64x2_t svqdmulh_single_s64_x2(svint64x2_t, svint64_t);
		3599	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_single_s16_x2)))
		3600	svint16x2_t svqdmulh_single_s16_x2(svint16x2_t, svint16_t);
		3601	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_single_s8_x4)))
		3602	svint8x4_t svqdmulh_single_s8_x4(svint8x4_t, svint8_t);
		3603	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_single_s32_x4)))
		3604	svint32x4_t svqdmulh_single_s32_x4(svint32x4_t, svint32_t);
		3605	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_single_s64_x4)))
		3606	svint64x4_t svqdmulh_single_s64_x4(svint64x4_t, svint64_t);
		3607	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_single_s16_x4)))
		3608	svint16x4_t svqdmulh_single_s16_x4(svint16x4_t, svint16_t);
		3609	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_s8_x2)))
		3610	svint8x2_t svqdmulh_s8_x2(svint8x2_t, svint8x2_t);
		3611	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_s32_x2)))
		3612	svint32x2_t svqdmulh_s32_x2(svint32x2_t, svint32x2_t);
		3613	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_s64_x2)))
		3614	svint64x2_t svqdmulh_s64_x2(svint64x2_t, svint64x2_t);
		3615	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_s16_x2)))
		3616	svint16x2_t svqdmulh_s16_x2(svint16x2_t, svint16x2_t);
		3617	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_s8_x4)))
		3618	svint8x4_t svqdmulh_s8_x4(svint8x4_t, svint8x4_t);
		3619	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_s32_x4)))
		3620	svint32x4_t svqdmulh_s32_x4(svint32x4_t, svint32x4_t);
		3621	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_s64_x4)))
		3622	svint64x4_t svqdmulh_s64_x4(svint64x4_t, svint64x4_t);
		3623	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_s16_x4)))
		3624	svint16x4_t svqdmulh_s16_x4(svint16x4_t, svint16x4_t);
		3625	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshr_n_s16_s32_x2)))
		3626	svint16_t svqrshr_n_s16_s32_x2(svint32x2_t, uint64_t);
		3627	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshr_n_u16_u32_x2)))
		3628	svuint16_t svqrshr_n_u16_u32_x2(svuint32x2_t, uint64_t);
		3629	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshr_n_s8_s32_x4)))
		3630	svint8_t svqrshr_n_s8_s32_x4(svint32x4_t, uint64_t);
		3631	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshr_n_s16_s64_x4)))
		3632	svint16_t svqrshr_n_s16_s64_x4(svint64x4_t, uint64_t);
		3633	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshr_n_u8_u32_x4)))
		3634	svuint8_t svqrshr_n_u8_u32_x4(svuint32x4_t, uint64_t);
		3635	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshr_n_u16_u64_x4)))
		3636	svuint16_t svqrshr_n_u16_u64_x4(svuint64x4_t, uint64_t);
		3637	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrn_n_s8_s32_x4)))
		3638	svint8_t svqrshrn_n_s8_s32_x4(svint32x4_t, uint64_t);
		3639	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrn_n_s16_s64_x4)))
		3640	svint16_t svqrshrn_n_s16_s64_x4(svint64x4_t, uint64_t);
		3641	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrn_n_u8_u32_x4)))
		3642	svuint8_t svqrshrn_n_u8_u32_x4(svuint32x4_t, uint64_t);
		3643	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrn_n_u16_u64_x4)))
		3644	svuint16_t svqrshrn_n_u16_u64_x4(svuint64x4_t, uint64_t);
		3645	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshru_n_u16_s32_x2)))
		3646	svuint16_t svqrshru_n_u16_s32_x2(svint32x2_t, uint64_t);
		3647	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshru_n_u8_s32_x4)))
		3648	svuint8_t svqrshru_n_u8_s32_x4(svint32x4_t, uint64_t);
		3649	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshru_n_u16_s64_x4)))
		3650	svuint16_t svqrshru_n_u16_s64_x4(svint64x4_t, uint64_t);
		3651	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrun_n_u8_s32_x4)))
		3652	svuint8_t svqrshrun_n_u8_s32_x4(svint32x4_t, uint64_t);
		3653	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrun_n_u16_s64_x4)))
		3654	svuint16_t svqrshrun_n_u16_s64_x4(svint64x4_t, uint64_t);
		3655	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svreinterpret_b)))
		3656	svbool_t svreinterpret_b(svcount_t);
		3657	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svreinterpret_c)))
		3658	svcount_t svreinterpret_c(svbool_t);
		3659	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrinta_f32_x2)))
		3660	svfloat32x2_t svrinta_f32_x2(svfloat32x2_t);
		3661	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrinta_f32_x4)))
		3662	svfloat32x4_t svrinta_f32_x4(svfloat32x4_t);
		3663	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrintm_f32_x2)))
		3664	svfloat32x2_t svrintm_f32_x2(svfloat32x2_t);
		3665	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrintm_f32_x4)))
		3666	svfloat32x4_t svrintm_f32_x4(svfloat32x4_t);
		3667	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrintn_f32_x2)))
		3668	svfloat32x2_t svrintn_f32_x2(svfloat32x2_t);
		3669	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrintn_f32_x4)))
		3670	svfloat32x4_t svrintn_f32_x4(svfloat32x4_t);
		3671	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrintp_f32_x2)))
		3672	svfloat32x2_t svrintp_f32_x2(svfloat32x2_t);
		3673	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrintp_f32_x4)))
		3674	svfloat32x4_t svrintp_f32_x4(svfloat32x4_t);
		3675	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_s8_x2)))
		3676	svint8x2_t svrshl_single_s8_x2(svint8x2_t, svint8_t);
		3677	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_s32_x2)))
		3678	svint32x2_t svrshl_single_s32_x2(svint32x2_t, svint32_t);
		3679	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_s64_x2)))
		3680	svint64x2_t svrshl_single_s64_x2(svint64x2_t, svint64_t);
		3681	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_s16_x2)))
		3682	svint16x2_t svrshl_single_s16_x2(svint16x2_t, svint16_t);
		3683	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_u8_x2)))
		3684	svuint8x2_t svrshl_single_u8_x2(svuint8x2_t, svuint8_t);
		3685	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_u32_x2)))
		3686	svuint32x2_t svrshl_single_u32_x2(svuint32x2_t, svuint32_t);
		3687	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_u64_x2)))
		3688	svuint64x2_t svrshl_single_u64_x2(svuint64x2_t, svuint64_t);
		3689	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_u16_x2)))
		3690	svuint16x2_t svrshl_single_u16_x2(svuint16x2_t, svuint16_t);
		3691	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_s8_x4)))
		3692	svint8x4_t svrshl_single_s8_x4(svint8x4_t, svint8_t);
		3693	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_s32_x4)))
		3694	svint32x4_t svrshl_single_s32_x4(svint32x4_t, svint32_t);
		3695	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_s64_x4)))
		3696	svint64x4_t svrshl_single_s64_x4(svint64x4_t, svint64_t);
		3697	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_s16_x4)))
		3698	svint16x4_t svrshl_single_s16_x4(svint16x4_t, svint16_t);
		3699	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_u8_x4)))
		3700	svuint8x4_t svrshl_single_u8_x4(svuint8x4_t, svuint8_t);
		3701	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_u32_x4)))
		3702	svuint32x4_t svrshl_single_u32_x4(svuint32x4_t, svuint32_t);
		3703	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_u64_x4)))
		3704	svuint64x4_t svrshl_single_u64_x4(svuint64x4_t, svuint64_t);
		3705	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_u16_x4)))
		3706	svuint16x4_t svrshl_single_u16_x4(svuint16x4_t, svuint16_t);
		3707	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_s8_x2)))
		3708	svint8x2_t svrshl_s8_x2(svint8x2_t, svint8x2_t);
		3709	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_s32_x2)))
		3710	svint32x2_t svrshl_s32_x2(svint32x2_t, svint32x2_t);
		3711	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_s64_x2)))
		3712	svint64x2_t svrshl_s64_x2(svint64x2_t, svint64x2_t);
		3713	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_s16_x2)))
		3714	svint16x2_t svrshl_s16_x2(svint16x2_t, svint16x2_t);
		3715	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_u8_x2)))
		3716	svuint8x2_t svrshl_u8_x2(svuint8x2_t, svuint8x2_t);
		3717	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_u32_x2)))
		3718	svuint32x2_t svrshl_u32_x2(svuint32x2_t, svuint32x2_t);
		3719	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_u64_x2)))
		3720	svuint64x2_t svrshl_u64_x2(svuint64x2_t, svuint64x2_t);
		3721	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_u16_x2)))
		3722	svuint16x2_t svrshl_u16_x2(svuint16x2_t, svuint16x2_t);
		3723	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_s8_x4)))
		3724	svint8x4_t svrshl_s8_x4(svint8x4_t, svint8x4_t);
		3725	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_s32_x4)))
		3726	svint32x4_t svrshl_s32_x4(svint32x4_t, svint32x4_t);
		3727	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_s64_x4)))
		3728	svint64x4_t svrshl_s64_x4(svint64x4_t, svint64x4_t);
		3729	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_s16_x4)))
		3730	svint16x4_t svrshl_s16_x4(svint16x4_t, svint16x4_t);
		3731	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_u8_x4)))
		3732	svuint8x4_t svrshl_u8_x4(svuint8x4_t, svuint8x4_t);
		3733	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_u32_x4)))
		3734	svuint32x4_t svrshl_u32_x4(svuint32x4_t, svuint32x4_t);
		3735	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_u64_x4)))
		3736	svuint64x4_t svrshl_u64_x4(svuint64x4_t, svuint64x4_t);
		3737	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_u16_x4)))
		3738	svuint16x4_t svrshl_u16_x4(svuint16x4_t, svuint16x4_t);
		3739	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_u8_x2)))
		3740	svuint8x2_t svsel_u8_x2(svcount_t, svuint8x2_t, svuint8x2_t);
		3741	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_u32_x2)))
		3742	svuint32x2_t svsel_u32_x2(svcount_t, svuint32x2_t, svuint32x2_t);
		3743	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_u64_x2)))
		3744	svuint64x2_t svsel_u64_x2(svcount_t, svuint64x2_t, svuint64x2_t);
		3745	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_u16_x2)))
		3746	svuint16x2_t svsel_u16_x2(svcount_t, svuint16x2_t, svuint16x2_t);
		3747	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_bf16_x2)))
		3748	svbfloat16x2_t svsel_bf16_x2(svcount_t, svbfloat16x2_t, svbfloat16x2_t);
		3749	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_s8_x2)))
		3750	svint8x2_t svsel_s8_x2(svcount_t, svint8x2_t, svint8x2_t);
		3751	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_f64_x2)))
		3752	svfloat64x2_t svsel_f64_x2(svcount_t, svfloat64x2_t, svfloat64x2_t);
		3753	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_f32_x2)))
		3754	svfloat32x2_t svsel_f32_x2(svcount_t, svfloat32x2_t, svfloat32x2_t);
		3755	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_f16_x2)))
		3756	svfloat16x2_t svsel_f16_x2(svcount_t, svfloat16x2_t, svfloat16x2_t);
		3757	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_s32_x2)))
		3758	svint32x2_t svsel_s32_x2(svcount_t, svint32x2_t, svint32x2_t);
		3759	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_s64_x2)))
		3760	svint64x2_t svsel_s64_x2(svcount_t, svint64x2_t, svint64x2_t);
		3761	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_s16_x2)))
		3762	svint16x2_t svsel_s16_x2(svcount_t, svint16x2_t, svint16x2_t);
		3763	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_u8_x4)))
		3764	svuint8x4_t svsel_u8_x4(svcount_t, svuint8x4_t, svuint8x4_t);
		3765	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_u32_x4)))
		3766	svuint32x4_t svsel_u32_x4(svcount_t, svuint32x4_t, svuint32x4_t);
		3767	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_u64_x4)))
		3768	svuint64x4_t svsel_u64_x4(svcount_t, svuint64x4_t, svuint64x4_t);
		3769	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_u16_x4)))
		3770	svuint16x4_t svsel_u16_x4(svcount_t, svuint16x4_t, svuint16x4_t);
		3771	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_bf16_x4)))
		3772	svbfloat16x4_t svsel_bf16_x4(svcount_t, svbfloat16x4_t, svbfloat16x4_t);
		3773	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_s8_x4)))
		3774	svint8x4_t svsel_s8_x4(svcount_t, svint8x4_t, svint8x4_t);
		3775	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_f64_x4)))
		3776	svfloat64x4_t svsel_f64_x4(svcount_t, svfloat64x4_t, svfloat64x4_t);
		3777	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_f32_x4)))
		3778	svfloat32x4_t svsel_f32_x4(svcount_t, svfloat32x4_t, svfloat32x4_t);
		3779	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_f16_x4)))
		3780	svfloat16x4_t svsel_f16_x4(svcount_t, svfloat16x4_t, svfloat16x4_t);
		3781	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_s32_x4)))
		3782	svint32x4_t svsel_s32_x4(svcount_t, svint32x4_t, svint32x4_t);
		3783	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_s64_x4)))
		3784	svint64x4_t svsel_s64_x4(svcount_t, svint64x4_t, svint64x4_t);
		3785	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_s16_x4)))
		3786	svint16x4_t svsel_s16_x4(svcount_t, svint16x4_t, svint16x4_t);
		3787	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_s32_s16_x2)))
		3788	svint32x2_t svunpk_s32_s16_x2(svint16_t);
		3789	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_s64_s32_x2)))
		3790	svint64x2_t svunpk_s64_s32_x2(svint32_t);
		3791	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_s16_s8_x2)))
		3792	svint16x2_t svunpk_s16_s8_x2(svint8_t);
		3793	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_u32_u16_x2)))
		3794	svuint32x2_t svunpk_u32_u16_x2(svuint16_t);
		3795	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_u64_u32_x2)))
		3796	svuint64x2_t svunpk_u64_u32_x2(svuint32_t);
		3797	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_u16_u8_x2)))
		3798	svuint16x2_t svunpk_u16_u8_x2(svuint8_t);
		3799	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_s32_s16_x4)))
		3800	svint32x4_t svunpk_s32_s16_x4(svint16x2_t);
		3801	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_s64_s32_x4)))
		3802	svint64x4_t svunpk_s64_s32_x4(svint32x2_t);
		3803	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_s16_s8_x4)))
		3804	svint16x4_t svunpk_s16_s8_x4(svint8x2_t);
		3805	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_u32_u16_x4)))
		3806	svuint32x4_t svunpk_u32_u16_x4(svuint16x2_t);
		3807	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_u64_u32_x4)))
		3808	svuint64x4_t svunpk_u64_u32_x4(svuint32x2_t);
		3809	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_u16_u8_x4)))
		3810	svuint16x4_t svunpk_u16_u8_x4(svuint8x2_t);
		3811	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_u8_x2)))
		3812	svuint8x2_t svuzp_u8_x2(svuint8x2_t);
		3813	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_u32_x2)))
		3814	svuint32x2_t svuzp_u32_x2(svuint32x2_t);
		3815	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_u64_x2)))
		3816	svuint64x2_t svuzp_u64_x2(svuint64x2_t);
		3817	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_u16_x2)))
		3818	svuint16x2_t svuzp_u16_x2(svuint16x2_t);
		3819	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_bf16_x2)))
		3820	svbfloat16x2_t svuzp_bf16_x2(svbfloat16x2_t);
		3821	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_s8_x2)))
		3822	svint8x2_t svuzp_s8_x2(svint8x2_t);
		3823	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_f64_x2)))
		3824	svfloat64x2_t svuzp_f64_x2(svfloat64x2_t);
		3825	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_f32_x2)))
		3826	svfloat32x2_t svuzp_f32_x2(svfloat32x2_t);
		3827	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_f16_x2)))
		3828	svfloat16x2_t svuzp_f16_x2(svfloat16x2_t);
		3829	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_s32_x2)))
		3830	svint32x2_t svuzp_s32_x2(svint32x2_t);
		3831	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_s64_x2)))
		3832	svint64x2_t svuzp_s64_x2(svint64x2_t);
		3833	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_s16_x2)))
		3834	svint16x2_t svuzp_s16_x2(svint16x2_t);
		3835	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_u8_x4)))
		3836	svuint8x4_t svuzp_u8_x4(svuint8x4_t);
		3837	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_u32_x4)))
		3838	svuint32x4_t svuzp_u32_x4(svuint32x4_t);
		3839	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_u64_x4)))
		3840	svuint64x4_t svuzp_u64_x4(svuint64x4_t);
		3841	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_u16_x4)))
		3842	svuint16x4_t svuzp_u16_x4(svuint16x4_t);
		3843	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_bf16_x4)))
		3844	svbfloat16x4_t svuzp_bf16_x4(svbfloat16x4_t);
		3845	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_s8_x4)))
		3846	svint8x4_t svuzp_s8_x4(svint8x4_t);
		3847	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_f64_x4)))
		3848	svfloat64x4_t svuzp_f64_x4(svfloat64x4_t);
		3849	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_f32_x4)))
		3850	svfloat32x4_t svuzp_f32_x4(svfloat32x4_t);
		3851	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_f16_x4)))
		3852	svfloat16x4_t svuzp_f16_x4(svfloat16x4_t);
		3853	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_s32_x4)))
		3854	svint32x4_t svuzp_s32_x4(svint32x4_t);
		3855	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_s64_x4)))
		3856	svint64x4_t svuzp_s64_x4(svint64x4_t);
		3857	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_s16_x4)))
		3858	svint16x4_t svuzp_s16_x4(svint16x4_t);
		3859	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_u8_x2)))
		3860	svuint8x2_t svuzpq_u8_x2(svuint8x2_t);
		3861	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_u32_x2)))
		3862	svuint32x2_t svuzpq_u32_x2(svuint32x2_t);
		3863	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_u64_x2)))
		3864	svuint64x2_t svuzpq_u64_x2(svuint64x2_t);
		3865	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_u16_x2)))
		3866	svuint16x2_t svuzpq_u16_x2(svuint16x2_t);
		3867	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_bf16_x2)))
		3868	svbfloat16x2_t svuzpq_bf16_x2(svbfloat16x2_t);
		3869	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_s8_x2)))
		3870	svint8x2_t svuzpq_s8_x2(svint8x2_t);
		3871	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_f64_x2)))
		3872	svfloat64x2_t svuzpq_f64_x2(svfloat64x2_t);
		3873	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_f32_x2)))
		3874	svfloat32x2_t svuzpq_f32_x2(svfloat32x2_t);
		3875	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_f16_x2)))
		3876	svfloat16x2_t svuzpq_f16_x2(svfloat16x2_t);
		3877	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_s32_x2)))
		3878	svint32x2_t svuzpq_s32_x2(svint32x2_t);
		3879	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_s64_x2)))
		3880	svint64x2_t svuzpq_s64_x2(svint64x2_t);
		3881	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_s16_x2)))
		3882	svint16x2_t svuzpq_s16_x2(svint16x2_t);
		3883	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_u8_x4)))
		3884	svuint8x4_t svuzpq_u8_x4(svuint8x4_t);
		3885	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_u32_x4)))
		3886	svuint32x4_t svuzpq_u32_x4(svuint32x4_t);
		3887	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_u64_x4)))
		3888	svuint64x4_t svuzpq_u64_x4(svuint64x4_t);
		3889	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_u16_x4)))
		3890	svuint16x4_t svuzpq_u16_x4(svuint16x4_t);
		3891	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_bf16_x4)))
		3892	svbfloat16x4_t svuzpq_bf16_x4(svbfloat16x4_t);
		3893	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_s8_x4)))
		3894	svint8x4_t svuzpq_s8_x4(svint8x4_t);
		3895	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_f64_x4)))
		3896	svfloat64x4_t svuzpq_f64_x4(svfloat64x4_t);
		3897	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_f32_x4)))
		3898	svfloat32x4_t svuzpq_f32_x4(svfloat32x4_t);
		3899	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_f16_x4)))
		3900	svfloat16x4_t svuzpq_f16_x4(svfloat16x4_t);
		3901	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_s32_x4)))
		3902	svint32x4_t svuzpq_s32_x4(svint32x4_t);
		3903	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_s64_x4)))
		3904	svint64x4_t svuzpq_s64_x4(svint64x4_t);
		3905	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_s16_x4)))
		3906	svint16x4_t svuzpq_s16_x4(svint16x4_t);
		3907	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_u8_x2)))
		3908	svuint8x2_t svzip_u8_x2(svuint8x2_t);
		3909	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_u32_x2)))
		3910	svuint32x2_t svzip_u32_x2(svuint32x2_t);
		3911	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_u64_x2)))
		3912	svuint64x2_t svzip_u64_x2(svuint64x2_t);
		3913	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_u16_x2)))
		3914	svuint16x2_t svzip_u16_x2(svuint16x2_t);
		3915	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_bf16_x2)))
		3916	svbfloat16x2_t svzip_bf16_x2(svbfloat16x2_t);
		3917	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_s8_x2)))
		3918	svint8x2_t svzip_s8_x2(svint8x2_t);
		3919	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_f64_x2)))
		3920	svfloat64x2_t svzip_f64_x2(svfloat64x2_t);
		3921	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_f32_x2)))
		3922	svfloat32x2_t svzip_f32_x2(svfloat32x2_t);
		3923	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_f16_x2)))
		3924	svfloat16x2_t svzip_f16_x2(svfloat16x2_t);
		3925	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_s32_x2)))
		3926	svint32x2_t svzip_s32_x2(svint32x2_t);
		3927	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_s64_x2)))
		3928	svint64x2_t svzip_s64_x2(svint64x2_t);
		3929	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_s16_x2)))
		3930	svint16x2_t svzip_s16_x2(svint16x2_t);
		3931	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_u8_x4)))
		3932	svuint8x4_t svzip_u8_x4(svuint8x4_t);
		3933	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_u32_x4)))
		3934	svuint32x4_t svzip_u32_x4(svuint32x4_t);
		3935	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_u64_x4)))
		3936	svuint64x4_t svzip_u64_x4(svuint64x4_t);
		3937	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_u16_x4)))
		3938	svuint16x4_t svzip_u16_x4(svuint16x4_t);
		3939	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_bf16_x4)))
		3940	svbfloat16x4_t svzip_bf16_x4(svbfloat16x4_t);
		3941	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_s8_x4)))
		3942	svint8x4_t svzip_s8_x4(svint8x4_t);
		3943	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_f64_x4)))
		3944	svfloat64x4_t svzip_f64_x4(svfloat64x4_t);
		3945	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_f32_x4)))
		3946	svfloat32x4_t svzip_f32_x4(svfloat32x4_t);
		3947	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_f16_x4)))
		3948	svfloat16x4_t svzip_f16_x4(svfloat16x4_t);
		3949	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_s32_x4)))
		3950	svint32x4_t svzip_s32_x4(svint32x4_t);
		3951	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_s64_x4)))
		3952	svint64x4_t svzip_s64_x4(svint64x4_t);
		3953	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_s16_x4)))
		3954	svint16x4_t svzip_s16_x4(svint16x4_t);
		3955	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_u8_x2)))
		3956	svuint8x2_t svzipq_u8_x2(svuint8x2_t);
		3957	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_u32_x2)))
		3958	svuint32x2_t svzipq_u32_x2(svuint32x2_t);
		3959	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_u64_x2)))
		3960	svuint64x2_t svzipq_u64_x2(svuint64x2_t);
		3961	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_u16_x2)))
		3962	svuint16x2_t svzipq_u16_x2(svuint16x2_t);
		3963	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_bf16_x2)))
		3964	svbfloat16x2_t svzipq_bf16_x2(svbfloat16x2_t);
		3965	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_s8_x2)))
		3966	svint8x2_t svzipq_s8_x2(svint8x2_t);
		3967	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_f64_x2)))
		3968	svfloat64x2_t svzipq_f64_x2(svfloat64x2_t);
		3969	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_f32_x2)))
		3970	svfloat32x2_t svzipq_f32_x2(svfloat32x2_t);
		3971	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_f16_x2)))
		3972	svfloat16x2_t svzipq_f16_x2(svfloat16x2_t);
		3973	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_s32_x2)))
		3974	svint32x2_t svzipq_s32_x2(svint32x2_t);
		3975	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_s64_x2)))
		3976	svint64x2_t svzipq_s64_x2(svint64x2_t);
		3977	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_s16_x2)))
		3978	svint16x2_t svzipq_s16_x2(svint16x2_t);
		3979	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_u8_x4)))
		3980	svuint8x4_t svzipq_u8_x4(svuint8x4_t);
		3981	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_u32_x4)))
		3982	svuint32x4_t svzipq_u32_x4(svuint32x4_t);
		3983	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_u64_x4)))
		3984	svuint64x4_t svzipq_u64_x4(svuint64x4_t);
		3985	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_u16_x4)))
		3986	svuint16x4_t svzipq_u16_x4(svuint16x4_t);
		3987	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_bf16_x4)))
		3988	svbfloat16x4_t svzipq_bf16_x4(svbfloat16x4_t);
		3989	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_s8_x4)))
		3990	svint8x4_t svzipq_s8_x4(svint8x4_t);
		3991	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_f64_x4)))
		3992	svfloat64x4_t svzipq_f64_x4(svfloat64x4_t);
		3993	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_f32_x4)))
		3994	svfloat32x4_t svzipq_f32_x4(svfloat32x4_t);
		3995	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_f16_x4)))
		3996	svfloat16x4_t svzipq_f16_x4(svfloat16x4_t);
		3997	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_s32_x4)))
		3998	svint32x4_t svzipq_s32_x4(svint32x4_t);
		3999	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_s64_x4)))
		4000	svint64x4_t svzipq_s64_x4(svint64x4_t);
		4001	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_s16_x4)))
		4002	svint16x4_t svzipq_s16_x4(svint16x4_t);
		4003	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_u8_x2)))
		4004	svuint8x2_t svadd(svuint8x2_t, svuint8_t);
		4005	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_u32_x2)))
		4006	svuint32x2_t svadd(svuint32x2_t, svuint32_t);
		4007	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_u64_x2)))
		4008	svuint64x2_t svadd(svuint64x2_t, svuint64_t);
		4009	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_u16_x2)))
		4010	svuint16x2_t svadd(svuint16x2_t, svuint16_t);
		4011	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_s8_x2)))
		4012	svint8x2_t svadd(svint8x2_t, svint8_t);
		4013	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_s32_x2)))
		4014	svint32x2_t svadd(svint32x2_t, svint32_t);
		4015	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_s64_x2)))
		4016	svint64x2_t svadd(svint64x2_t, svint64_t);
		4017	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_s16_x2)))
		4018	svint16x2_t svadd(svint16x2_t, svint16_t);
		4019	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_u8_x4)))
		4020	svuint8x4_t svadd(svuint8x4_t, svuint8_t);
		4021	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_u32_x4)))
		4022	svuint32x4_t svadd(svuint32x4_t, svuint32_t);
		4023	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_u64_x4)))
		4024	svuint64x4_t svadd(svuint64x4_t, svuint64_t);
		4025	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_u16_x4)))
		4026	svuint16x4_t svadd(svuint16x4_t, svuint16_t);
		4027	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_s8_x4)))
		4028	svint8x4_t svadd(svint8x4_t, svint8_t);
		4029	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_s32_x4)))
		4030	svint32x4_t svadd(svint32x4_t, svint32_t);
		4031	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_s64_x4)))
		4032	svint64x4_t svadd(svint64x4_t, svint64_t);
		4033	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svadd_single_s16_x4)))
		4034	svint16x4_t svadd(svint16x4_t, svint16_t);
		4035	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_f64_x2)))
		4036	svfloat64x2_t svclamp(svfloat64x2_t, svfloat64_t, svfloat64_t);
		4037	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_f32_x2)))
		4038	svfloat32x2_t svclamp(svfloat32x2_t, svfloat32_t, svfloat32_t);
		4039	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_f16_x2)))
		4040	svfloat16x2_t svclamp(svfloat16x2_t, svfloat16_t, svfloat16_t);
		4041	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_s8_x2)))
		4042	svint8x2_t svclamp(svint8x2_t, svint8_t, svint8_t);
		4043	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_s32_x2)))
		4044	svint32x2_t svclamp(svint32x2_t, svint32_t, svint32_t);
		4045	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_s64_x2)))
		4046	svint64x2_t svclamp(svint64x2_t, svint64_t, svint64_t);
		4047	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_s16_x2)))
		4048	svint16x2_t svclamp(svint16x2_t, svint16_t, svint16_t);
		4049	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_u8_x2)))
		4050	svuint8x2_t svclamp(svuint8x2_t, svuint8_t, svuint8_t);
		4051	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_u32_x2)))
		4052	svuint32x2_t svclamp(svuint32x2_t, svuint32_t, svuint32_t);
		4053	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_u64_x2)))
		4054	svuint64x2_t svclamp(svuint64x2_t, svuint64_t, svuint64_t);
		4055	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_u16_x2)))
		4056	svuint16x2_t svclamp(svuint16x2_t, svuint16_t, svuint16_t);
		4057	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_f64_x4)))
		4058	svfloat64x4_t svclamp(svfloat64x4_t, svfloat64_t, svfloat64_t);
		4059	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_f32_x4)))
		4060	svfloat32x4_t svclamp(svfloat32x4_t, svfloat32_t, svfloat32_t);
		4061	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_f16_x4)))
		4062	svfloat16x4_t svclamp(svfloat16x4_t, svfloat16_t, svfloat16_t);
		4063	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_s8_x4)))
		4064	svint8x4_t svclamp(svint8x4_t, svint8_t, svint8_t);
		4065	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_s32_x4)))
		4066	svint32x4_t svclamp(svint32x4_t, svint32_t, svint32_t);
		4067	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_s64_x4)))
		4068	svint64x4_t svclamp(svint64x4_t, svint64_t, svint64_t);
		4069	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_s16_x4)))
		4070	svint16x4_t svclamp(svint16x4_t, svint16_t, svint16_t);
		4071	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_u8_x4)))
		4072	svuint8x4_t svclamp(svuint8x4_t, svuint8_t, svuint8_t);
		4073	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_u32_x4)))
		4074	svuint32x4_t svclamp(svuint32x4_t, svuint32_t, svuint32_t);
		4075	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_u64_x4)))
		4076	svuint64x4_t svclamp(svuint64x4_t, svuint64_t, svuint64_t);
		4077	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_single_u16_x4)))
		4078	svuint16x4_t svclamp(svuint16x4_t, svuint16_t, svuint16_t);
		4079	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_bf16_f32_x2)))
		4080	svbfloat16_t svcvt_bf16(svfloat32x2_t);
		4081	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_f16_f32_x2)))
		4082	svfloat16_t svcvt_f16(svfloat32x2_t);
		4083	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_s32_f32_x2)))
		4084	svint32x2_t svcvt_s32(svfloat32x2_t);
		4085	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_u32_f32_x2)))
		4086	svuint32x2_t svcvt_u32(svfloat32x2_t);
		4087	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_s32_f32_x4)))
		4088	svint32x4_t svcvt_s32(svfloat32x4_t);
		4089	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_u32_f32_x4)))
		4090	svuint32x4_t svcvt_u32(svfloat32x4_t);
		4091	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_f32_s32_x2)))
		4092	svfloat32x2_t svcvt_f32(svint32x2_t);
		4093	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_f32_s32_x4)))
		4094	svfloat32x4_t svcvt_f32(svint32x4_t);
		4095	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_f32_u32_x2)))
		4096	svfloat32x2_t svcvt_f32(svuint32x2_t);
		4097	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvt_f32_u32_x4)))
		4098	svfloat32x4_t svcvt_f32(svuint32x4_t);
		4099	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvtn_bf16_f32_x2)))
		4100	svbfloat16_t svcvtn_bf16(svfloat32x2_t);
		4101	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcvtn_f16_f32_x2)))
		4102	svfloat16_t svcvtn_f16(svfloat32x2_t);
		4103	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_f64_x2)))
		4104	svfloat64x2_t svmax(svfloat64x2_t, svfloat64_t);
		4105	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_f32_x2)))
		4106	svfloat32x2_t svmax(svfloat32x2_t, svfloat32_t);
		4107	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_f16_x2)))
		4108	svfloat16x2_t svmax(svfloat16x2_t, svfloat16_t);
		4109	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_s8_x2)))
		4110	svint8x2_t svmax(svint8x2_t, svint8_t);
		4111	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_s32_x2)))
		4112	svint32x2_t svmax(svint32x2_t, svint32_t);
		4113	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_s64_x2)))
		4114	svint64x2_t svmax(svint64x2_t, svint64_t);
		4115	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_s16_x2)))
		4116	svint16x2_t svmax(svint16x2_t, svint16_t);
		4117	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_u8_x2)))
		4118	svuint8x2_t svmax(svuint8x2_t, svuint8_t);
		4119	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_u32_x2)))
		4120	svuint32x2_t svmax(svuint32x2_t, svuint32_t);
		4121	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_u64_x2)))
		4122	svuint64x2_t svmax(svuint64x2_t, svuint64_t);
		4123	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_u16_x2)))
		4124	svuint16x2_t svmax(svuint16x2_t, svuint16_t);
		4125	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_f64_x4)))
		4126	svfloat64x4_t svmax(svfloat64x4_t, svfloat64_t);
		4127	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_f32_x4)))
		4128	svfloat32x4_t svmax(svfloat32x4_t, svfloat32_t);
		4129	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_f16_x4)))
		4130	svfloat16x4_t svmax(svfloat16x4_t, svfloat16_t);
		4131	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_s8_x4)))
		4132	svint8x4_t svmax(svint8x4_t, svint8_t);
		4133	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_s32_x4)))
		4134	svint32x4_t svmax(svint32x4_t, svint32_t);
		4135	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_s64_x4)))
		4136	svint64x4_t svmax(svint64x4_t, svint64_t);
		4137	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_s16_x4)))
		4138	svint16x4_t svmax(svint16x4_t, svint16_t);
		4139	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_u8_x4)))
		4140	svuint8x4_t svmax(svuint8x4_t, svuint8_t);
		4141	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_u32_x4)))
		4142	svuint32x4_t svmax(svuint32x4_t, svuint32_t);
		4143	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_u64_x4)))
		4144	svuint64x4_t svmax(svuint64x4_t, svuint64_t);
		4145	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_single_u16_x4)))
		4146	svuint16x4_t svmax(svuint16x4_t, svuint16_t);
		4147	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_f64_x2)))
		4148	svfloat64x2_t svmax(svfloat64x2_t, svfloat64x2_t);
		4149	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_f32_x2)))
		4150	svfloat32x2_t svmax(svfloat32x2_t, svfloat32x2_t);
		4151	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_f16_x2)))
		4152	svfloat16x2_t svmax(svfloat16x2_t, svfloat16x2_t);
		4153	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_s8_x2)))
		4154	svint8x2_t svmax(svint8x2_t, svint8x2_t);
		4155	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_s32_x2)))
		4156	svint32x2_t svmax(svint32x2_t, svint32x2_t);
		4157	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_s64_x2)))
		4158	svint64x2_t svmax(svint64x2_t, svint64x2_t);
		4159	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_s16_x2)))
		4160	svint16x2_t svmax(svint16x2_t, svint16x2_t);
		4161	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_u8_x2)))
		4162	svuint8x2_t svmax(svuint8x2_t, svuint8x2_t);
		4163	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_u32_x2)))
		4164	svuint32x2_t svmax(svuint32x2_t, svuint32x2_t);
		4165	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_u64_x2)))
		4166	svuint64x2_t svmax(svuint64x2_t, svuint64x2_t);
		4167	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_u16_x2)))
		4168	svuint16x2_t svmax(svuint16x2_t, svuint16x2_t);
		4169	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_f64_x4)))
		4170	svfloat64x4_t svmax(svfloat64x4_t, svfloat64x4_t);
		4171	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_f32_x4)))
		4172	svfloat32x4_t svmax(svfloat32x4_t, svfloat32x4_t);
		4173	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_f16_x4)))
		4174	svfloat16x4_t svmax(svfloat16x4_t, svfloat16x4_t);
		4175	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_s8_x4)))
		4176	svint8x4_t svmax(svint8x4_t, svint8x4_t);
		4177	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_s32_x4)))
		4178	svint32x4_t svmax(svint32x4_t, svint32x4_t);
		4179	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_s64_x4)))
		4180	svint64x4_t svmax(svint64x4_t, svint64x4_t);
		4181	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_s16_x4)))
		4182	svint16x4_t svmax(svint16x4_t, svint16x4_t);
		4183	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_u8_x4)))
		4184	svuint8x4_t svmax(svuint8x4_t, svuint8x4_t);
		4185	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_u32_x4)))
		4186	svuint32x4_t svmax(svuint32x4_t, svuint32x4_t);
		4187	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_u64_x4)))
		4188	svuint64x4_t svmax(svuint64x4_t, svuint64x4_t);
		4189	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmax_u16_x4)))
		4190	svuint16x4_t svmax(svuint16x4_t, svuint16x4_t);
		4191	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_single_f64_x2)))
		4192	svfloat64x2_t svmaxnm(svfloat64x2_t, svfloat64_t);
		4193	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_single_f32_x2)))
		4194	svfloat32x2_t svmaxnm(svfloat32x2_t, svfloat32_t);
		4195	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_single_f16_x2)))
		4196	svfloat16x2_t svmaxnm(svfloat16x2_t, svfloat16_t);
		4197	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_single_f64_x4)))
		4198	svfloat64x4_t svmaxnm(svfloat64x4_t, svfloat64_t);
		4199	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_single_f32_x4)))
		4200	svfloat32x4_t svmaxnm(svfloat32x4_t, svfloat32_t);
		4201	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_single_f16_x4)))
		4202	svfloat16x4_t svmaxnm(svfloat16x4_t, svfloat16_t);
		4203	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_f64_x2)))
		4204	svfloat64x2_t svmaxnm(svfloat64x2_t, svfloat64x2_t);
		4205	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_f32_x2)))
		4206	svfloat32x2_t svmaxnm(svfloat32x2_t, svfloat32x2_t);
		4207	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_f16_x2)))
		4208	svfloat16x2_t svmaxnm(svfloat16x2_t, svfloat16x2_t);
		4209	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_f64_x4)))
		4210	svfloat64x4_t svmaxnm(svfloat64x4_t, svfloat64x4_t);
		4211	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_f32_x4)))
		4212	svfloat32x4_t svmaxnm(svfloat32x4_t, svfloat32x4_t);
		4213	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnm_f16_x4)))
		4214	svfloat16x4_t svmaxnm(svfloat16x4_t, svfloat16x4_t);
		4215	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_f64_x2)))
		4216	svfloat64x2_t svmin(svfloat64x2_t, svfloat64_t);
		4217	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_f32_x2)))
		4218	svfloat32x2_t svmin(svfloat32x2_t, svfloat32_t);
		4219	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_f16_x2)))
		4220	svfloat16x2_t svmin(svfloat16x2_t, svfloat16_t);
		4221	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_s8_x2)))
		4222	svint8x2_t svmin(svint8x2_t, svint8_t);
		4223	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_s32_x2)))
		4224	svint32x2_t svmin(svint32x2_t, svint32_t);
		4225	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_s64_x2)))
		4226	svint64x2_t svmin(svint64x2_t, svint64_t);
		4227	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_s16_x2)))
		4228	svint16x2_t svmin(svint16x2_t, svint16_t);
		4229	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_u8_x2)))
		4230	svuint8x2_t svmin(svuint8x2_t, svuint8_t);
		4231	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_u32_x2)))
		4232	svuint32x2_t svmin(svuint32x2_t, svuint32_t);
		4233	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_u64_x2)))
		4234	svuint64x2_t svmin(svuint64x2_t, svuint64_t);
		4235	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_u16_x2)))
		4236	svuint16x2_t svmin(svuint16x2_t, svuint16_t);
		4237	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_f64_x4)))
		4238	svfloat64x4_t svmin(svfloat64x4_t, svfloat64_t);
		4239	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_f32_x4)))
		4240	svfloat32x4_t svmin(svfloat32x4_t, svfloat32_t);
		4241	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_f16_x4)))
		4242	svfloat16x4_t svmin(svfloat16x4_t, svfloat16_t);
		4243	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_s8_x4)))
		4244	svint8x4_t svmin(svint8x4_t, svint8_t);
		4245	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_s32_x4)))
		4246	svint32x4_t svmin(svint32x4_t, svint32_t);
		4247	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_s64_x4)))
		4248	svint64x4_t svmin(svint64x4_t, svint64_t);
		4249	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_s16_x4)))
		4250	svint16x4_t svmin(svint16x4_t, svint16_t);
		4251	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_u8_x4)))
		4252	svuint8x4_t svmin(svuint8x4_t, svuint8_t);
		4253	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_u32_x4)))
		4254	svuint32x4_t svmin(svuint32x4_t, svuint32_t);
		4255	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_u64_x4)))
		4256	svuint64x4_t svmin(svuint64x4_t, svuint64_t);
		4257	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_single_u16_x4)))
		4258	svuint16x4_t svmin(svuint16x4_t, svuint16_t);
		4259	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_f64_x2)))
		4260	svfloat64x2_t svmin(svfloat64x2_t, svfloat64x2_t);
		4261	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_f32_x2)))
		4262	svfloat32x2_t svmin(svfloat32x2_t, svfloat32x2_t);
		4263	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_f16_x2)))
		4264	svfloat16x2_t svmin(svfloat16x2_t, svfloat16x2_t);
		4265	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_s8_x2)))
		4266	svint8x2_t svmin(svint8x2_t, svint8x2_t);
		4267	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_s32_x2)))
		4268	svint32x2_t svmin(svint32x2_t, svint32x2_t);
		4269	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_s64_x2)))
		4270	svint64x2_t svmin(svint64x2_t, svint64x2_t);
		4271	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_s16_x2)))
		4272	svint16x2_t svmin(svint16x2_t, svint16x2_t);
		4273	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_u8_x2)))
		4274	svuint8x2_t svmin(svuint8x2_t, svuint8x2_t);
		4275	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_u32_x2)))
		4276	svuint32x2_t svmin(svuint32x2_t, svuint32x2_t);
		4277	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_u64_x2)))
		4278	svuint64x2_t svmin(svuint64x2_t, svuint64x2_t);
		4279	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_u16_x2)))
		4280	svuint16x2_t svmin(svuint16x2_t, svuint16x2_t);
		4281	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_f64_x4)))
		4282	svfloat64x4_t svmin(svfloat64x4_t, svfloat64x4_t);
		4283	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_f32_x4)))
		4284	svfloat32x4_t svmin(svfloat32x4_t, svfloat32x4_t);
		4285	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_f16_x4)))
		4286	svfloat16x4_t svmin(svfloat16x4_t, svfloat16x4_t);
		4287	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_s8_x4)))
		4288	svint8x4_t svmin(svint8x4_t, svint8x4_t);
		4289	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_s32_x4)))
		4290	svint32x4_t svmin(svint32x4_t, svint32x4_t);
		4291	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_s64_x4)))
		4292	svint64x4_t svmin(svint64x4_t, svint64x4_t);
		4293	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_s16_x4)))
		4294	svint16x4_t svmin(svint16x4_t, svint16x4_t);
		4295	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_u8_x4)))
		4296	svuint8x4_t svmin(svuint8x4_t, svuint8x4_t);
		4297	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_u32_x4)))
		4298	svuint32x4_t svmin(svuint32x4_t, svuint32x4_t);
		4299	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_u64_x4)))
		4300	svuint64x4_t svmin(svuint64x4_t, svuint64x4_t);
		4301	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmin_u16_x4)))
		4302	svuint16x4_t svmin(svuint16x4_t, svuint16x4_t);
		4303	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_single_f64_x2)))
		4304	svfloat64x2_t svminnm(svfloat64x2_t, svfloat64_t);
		4305	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_single_f32_x2)))
		4306	svfloat32x2_t svminnm(svfloat32x2_t, svfloat32_t);
		4307	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_single_f16_x2)))
		4308	svfloat16x2_t svminnm(svfloat16x2_t, svfloat16_t);
		4309	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_single_f64_x4)))
		4310	svfloat64x4_t svminnm(svfloat64x4_t, svfloat64_t);
		4311	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_single_f32_x4)))
		4312	svfloat32x4_t svminnm(svfloat32x4_t, svfloat32_t);
		4313	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_single_f16_x4)))
		4314	svfloat16x4_t svminnm(svfloat16x4_t, svfloat16_t);
		4315	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_f64_x2)))
		4316	svfloat64x2_t svminnm(svfloat64x2_t, svfloat64x2_t);
		4317	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_f32_x2)))
		4318	svfloat32x2_t svminnm(svfloat32x2_t, svfloat32x2_t);
		4319	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_f16_x2)))
		4320	svfloat16x2_t svminnm(svfloat16x2_t, svfloat16x2_t);
		4321	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_f64_x4)))
		4322	svfloat64x4_t svminnm(svfloat64x4_t, svfloat64x4_t);
		4323	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_f32_x4)))
		4324	svfloat32x4_t svminnm(svfloat32x4_t, svfloat32x4_t);
		4325	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnm_f16_x4)))
		4326	svfloat16x4_t svminnm(svfloat16x4_t, svfloat16x4_t);
		4327	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_s16_s32_x2)))
		4328	svint16_t svqcvt_s16(svint32x2_t);
		4329	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_s16_s64_x4)))
		4330	svint16_t svqcvt_s16(svint64x4_t);
		4331	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_s8_s32_x4)))
		4332	svint8_t svqcvt_s8(svint32x4_t);
		4333	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_u16_s32_x2)))
		4334	svuint16_t svqcvt_u16(svint32x2_t);
		4335	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_u16_u32_x2)))
		4336	svuint16_t svqcvt_u16(svuint32x2_t);
		4337	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_u16_s64_x4)))
		4338	svuint16_t svqcvt_u16(svint64x4_t);
		4339	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_u16_u64_x4)))
		4340	svuint16_t svqcvt_u16(svuint64x4_t);
		4341	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_u8_s32_x4)))
		4342	svuint8_t svqcvt_u8(svint32x4_t);
		4343	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvt_u8_u32_x4)))
		4344	svuint8_t svqcvt_u8(svuint32x4_t);
		4345	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_s16_s64_x4)))
		4346	svint16_t svqcvtn_s16(svint64x4_t);
		4347	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_s8_s32_x4)))
		4348	svint8_t svqcvtn_s8(svint32x4_t);
		4349	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_u16_s64_x4)))
		4350	svuint16_t svqcvtn_u16(svint64x4_t);
		4351	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_u16_u64_x4)))
		4352	svuint16_t svqcvtn_u16(svuint64x4_t);
		4353	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_u8_s32_x4)))
		4354	svuint8_t svqcvtn_u8(svint32x4_t);
		4355	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_u8_u32_x4)))
		4356	svuint8_t svqcvtn_u8(svuint32x4_t);
		4357	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_single_s8_x2)))
		4358	svint8x2_t svqdmulh(svint8x2_t, svint8_t);
		4359	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_single_s32_x2)))
		4360	svint32x2_t svqdmulh(svint32x2_t, svint32_t);
		4361	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_single_s64_x2)))
		4362	svint64x2_t svqdmulh(svint64x2_t, svint64_t);
		4363	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_single_s16_x2)))
		4364	svint16x2_t svqdmulh(svint16x2_t, svint16_t);
		4365	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_single_s8_x4)))
		4366	svint8x4_t svqdmulh(svint8x4_t, svint8_t);
		4367	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_single_s32_x4)))
		4368	svint32x4_t svqdmulh(svint32x4_t, svint32_t);
		4369	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_single_s64_x4)))
		4370	svint64x4_t svqdmulh(svint64x4_t, svint64_t);
		4371	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_single_s16_x4)))
		4372	svint16x4_t svqdmulh(svint16x4_t, svint16_t);
		4373	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_s8_x2)))
		4374	svint8x2_t svqdmulh(svint8x2_t, svint8x2_t);
		4375	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_s32_x2)))
		4376	svint32x2_t svqdmulh(svint32x2_t, svint32x2_t);
		4377	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_s64_x2)))
		4378	svint64x2_t svqdmulh(svint64x2_t, svint64x2_t);
		4379	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_s16_x2)))
		4380	svint16x2_t svqdmulh(svint16x2_t, svint16x2_t);
		4381	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_s8_x4)))
		4382	svint8x4_t svqdmulh(svint8x4_t, svint8x4_t);
		4383	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_s32_x4)))
		4384	svint32x4_t svqdmulh(svint32x4_t, svint32x4_t);
		4385	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_s64_x4)))
		4386	svint64x4_t svqdmulh(svint64x4_t, svint64x4_t);
		4387	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqdmulh_s16_x4)))
		4388	svint16x4_t svqdmulh(svint16x4_t, svint16x4_t);
		4389	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshr_n_s16_s32_x2)))
		4390	svint16_t svqrshr_s16(svint32x2_t, uint64_t);
		4391	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshr_n_u16_u32_x2)))
		4392	svuint16_t svqrshr_u16(svuint32x2_t, uint64_t);
		4393	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshr_n_s8_s32_x4)))
		4394	svint8_t svqrshr_s8(svint32x4_t, uint64_t);
		4395	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshr_n_s16_s64_x4)))
		4396	svint16_t svqrshr_s16(svint64x4_t, uint64_t);
		4397	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshr_n_u8_u32_x4)))
		4398	svuint8_t svqrshr_u8(svuint32x4_t, uint64_t);
		4399	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshr_n_u16_u64_x4)))
		4400	svuint16_t svqrshr_u16(svuint64x4_t, uint64_t);
		4401	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrn_n_s8_s32_x4)))
		4402	svint8_t svqrshrn_s8(svint32x4_t, uint64_t);
		4403	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrn_n_s16_s64_x4)))
		4404	svint16_t svqrshrn_s16(svint64x4_t, uint64_t);
		4405	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrn_n_u8_u32_x4)))
		4406	svuint8_t svqrshrn_u8(svuint32x4_t, uint64_t);
		4407	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrn_n_u16_u64_x4)))
		4408	svuint16_t svqrshrn_u16(svuint64x4_t, uint64_t);
		4409	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshru_n_u16_s32_x2)))
		4410	svuint16_t svqrshru_u16(svint32x2_t, uint64_t);
		4411	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshru_n_u8_s32_x4)))
		4412	svuint8_t svqrshru_u8(svint32x4_t, uint64_t);
		4413	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshru_n_u16_s64_x4)))
		4414	svuint16_t svqrshru_u16(svint64x4_t, uint64_t);
		4415	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrun_n_u8_s32_x4)))
		4416	svuint8_t svqrshrun_u8(svint32x4_t, uint64_t);
		4417	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrun_n_u16_s64_x4)))
		4418	svuint16_t svqrshrun_u16(svint64x4_t, uint64_t);
		4419	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svreinterpret_b)))
		4420	svbool_t svreinterpret(svcount_t);
		4421	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svreinterpret_c)))
		4422	svcount_t svreinterpret(svbool_t);
		4423	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrinta_f32_x2)))
		4424	svfloat32x2_t svrinta(svfloat32x2_t);
		4425	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrinta_f32_x4)))
		4426	svfloat32x4_t svrinta(svfloat32x4_t);
		4427	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrintm_f32_x2)))
		4428	svfloat32x2_t svrintm(svfloat32x2_t);
		4429	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrintm_f32_x4)))
		4430	svfloat32x4_t svrintm(svfloat32x4_t);
		4431	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrintn_f32_x2)))
		4432	svfloat32x2_t svrintn(svfloat32x2_t);
		4433	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrintn_f32_x4)))
		4434	svfloat32x4_t svrintn(svfloat32x4_t);
		4435	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrintp_f32_x2)))
		4436	svfloat32x2_t svrintp(svfloat32x2_t);
		4437	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrintp_f32_x4)))
		4438	svfloat32x4_t svrintp(svfloat32x4_t);
		4439	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_s8_x2)))
		4440	svint8x2_t svrshl(svint8x2_t, svint8_t);
		4441	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_s32_x2)))
		4442	svint32x2_t svrshl(svint32x2_t, svint32_t);
		4443	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_s64_x2)))
		4444	svint64x2_t svrshl(svint64x2_t, svint64_t);
		4445	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_s16_x2)))
		4446	svint16x2_t svrshl(svint16x2_t, svint16_t);
		4447	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_u8_x2)))
		4448	svuint8x2_t svrshl(svuint8x2_t, svuint8_t);
		4449	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_u32_x2)))
		4450	svuint32x2_t svrshl(svuint32x2_t, svuint32_t);
		4451	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_u64_x2)))
		4452	svuint64x2_t svrshl(svuint64x2_t, svuint64_t);
		4453	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_u16_x2)))
		4454	svuint16x2_t svrshl(svuint16x2_t, svuint16_t);
		4455	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_s8_x4)))
		4456	svint8x4_t svrshl(svint8x4_t, svint8_t);
		4457	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_s32_x4)))
		4458	svint32x4_t svrshl(svint32x4_t, svint32_t);
		4459	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_s64_x4)))
		4460	svint64x4_t svrshl(svint64x4_t, svint64_t);
		4461	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_s16_x4)))
		4462	svint16x4_t svrshl(svint16x4_t, svint16_t);
		4463	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_u8_x4)))
		4464	svuint8x4_t svrshl(svuint8x4_t, svuint8_t);
		4465	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_u32_x4)))
		4466	svuint32x4_t svrshl(svuint32x4_t, svuint32_t);
		4467	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_u64_x4)))
		4468	svuint64x4_t svrshl(svuint64x4_t, svuint64_t);
		4469	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_single_u16_x4)))
		4470	svuint16x4_t svrshl(svuint16x4_t, svuint16_t);
		4471	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_s8_x2)))
		4472	svint8x2_t svrshl(svint8x2_t, svint8x2_t);
		4473	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_s32_x2)))
		4474	svint32x2_t svrshl(svint32x2_t, svint32x2_t);
		4475	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_s64_x2)))
		4476	svint64x2_t svrshl(svint64x2_t, svint64x2_t);
		4477	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_s16_x2)))
		4478	svint16x2_t svrshl(svint16x2_t, svint16x2_t);
		4479	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_u8_x2)))
		4480	svuint8x2_t svrshl(svuint8x2_t, svuint8x2_t);
		4481	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_u32_x2)))
		4482	svuint32x2_t svrshl(svuint32x2_t, svuint32x2_t);
		4483	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_u64_x2)))
		4484	svuint64x2_t svrshl(svuint64x2_t, svuint64x2_t);
		4485	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_u16_x2)))
		4486	svuint16x2_t svrshl(svuint16x2_t, svuint16x2_t);
		4487	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_s8_x4)))
		4488	svint8x4_t svrshl(svint8x4_t, svint8x4_t);
		4489	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_s32_x4)))
		4490	svint32x4_t svrshl(svint32x4_t, svint32x4_t);
		4491	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_s64_x4)))
		4492	svint64x4_t svrshl(svint64x4_t, svint64x4_t);
		4493	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_s16_x4)))
		4494	svint16x4_t svrshl(svint16x4_t, svint16x4_t);
		4495	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_u8_x4)))
		4496	svuint8x4_t svrshl(svuint8x4_t, svuint8x4_t);
		4497	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_u32_x4)))
		4498	svuint32x4_t svrshl(svuint32x4_t, svuint32x4_t);
		4499	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_u64_x4)))
		4500	svuint64x4_t svrshl(svuint64x4_t, svuint64x4_t);
		4501	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrshl_u16_x4)))
		4502	svuint16x4_t svrshl(svuint16x4_t, svuint16x4_t);
		4503	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_u8_x2)))
		4504	svuint8x2_t svsel(svcount_t, svuint8x2_t, svuint8x2_t);
		4505	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_u32_x2)))
		4506	svuint32x2_t svsel(svcount_t, svuint32x2_t, svuint32x2_t);
		4507	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_u64_x2)))
		4508	svuint64x2_t svsel(svcount_t, svuint64x2_t, svuint64x2_t);
		4509	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_u16_x2)))
		4510	svuint16x2_t svsel(svcount_t, svuint16x2_t, svuint16x2_t);
		4511	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_bf16_x2)))
		4512	svbfloat16x2_t svsel(svcount_t, svbfloat16x2_t, svbfloat16x2_t);
		4513	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_s8_x2)))
		4514	svint8x2_t svsel(svcount_t, svint8x2_t, svint8x2_t);
		4515	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_f64_x2)))
		4516	svfloat64x2_t svsel(svcount_t, svfloat64x2_t, svfloat64x2_t);
		4517	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_f32_x2)))
		4518	svfloat32x2_t svsel(svcount_t, svfloat32x2_t, svfloat32x2_t);
		4519	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_f16_x2)))
		4520	svfloat16x2_t svsel(svcount_t, svfloat16x2_t, svfloat16x2_t);
		4521	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_s32_x2)))
		4522	svint32x2_t svsel(svcount_t, svint32x2_t, svint32x2_t);
		4523	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_s64_x2)))
		4524	svint64x2_t svsel(svcount_t, svint64x2_t, svint64x2_t);
		4525	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_s16_x2)))
		4526	svint16x2_t svsel(svcount_t, svint16x2_t, svint16x2_t);
		4527	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_u8_x4)))
		4528	svuint8x4_t svsel(svcount_t, svuint8x4_t, svuint8x4_t);
		4529	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_u32_x4)))
		4530	svuint32x4_t svsel(svcount_t, svuint32x4_t, svuint32x4_t);
		4531	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_u64_x4)))
		4532	svuint64x4_t svsel(svcount_t, svuint64x4_t, svuint64x4_t);
		4533	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_u16_x4)))
		4534	svuint16x4_t svsel(svcount_t, svuint16x4_t, svuint16x4_t);
		4535	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_bf16_x4)))
		4536	svbfloat16x4_t svsel(svcount_t, svbfloat16x4_t, svbfloat16x4_t);
		4537	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_s8_x4)))
		4538	svint8x4_t svsel(svcount_t, svint8x4_t, svint8x4_t);
		4539	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_f64_x4)))
		4540	svfloat64x4_t svsel(svcount_t, svfloat64x4_t, svfloat64x4_t);
		4541	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_f32_x4)))
		4542	svfloat32x4_t svsel(svcount_t, svfloat32x4_t, svfloat32x4_t);
		4543	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_f16_x4)))
		4544	svfloat16x4_t svsel(svcount_t, svfloat16x4_t, svfloat16x4_t);
		4545	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_s32_x4)))
		4546	svint32x4_t svsel(svcount_t, svint32x4_t, svint32x4_t);
		4547	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_s64_x4)))
		4548	svint64x4_t svsel(svcount_t, svint64x4_t, svint64x4_t);
		4549	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsel_s16_x4)))
		4550	svint16x4_t svsel(svcount_t, svint16x4_t, svint16x4_t);
		4551	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_s32_s16_x2)))
		4552	svint32x2_t svunpk_s32(svint16_t);
		4553	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_s64_s32_x2)))
		4554	svint64x2_t svunpk_s64(svint32_t);
		4555	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_s16_s8_x2)))
		4556	svint16x2_t svunpk_s16(svint8_t);
		4557	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_u32_u16_x2)))
		4558	svuint32x2_t svunpk_u32(svuint16_t);
		4559	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_u64_u32_x2)))
		4560	svuint64x2_t svunpk_u64(svuint32_t);
		4561	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_u16_u8_x2)))
		4562	svuint16x2_t svunpk_u16(svuint8_t);
		4563	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_s32_s16_x4)))
		4564	svint32x4_t svunpk_s32(svint16x2_t);
		4565	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_s64_s32_x4)))
		4566	svint64x4_t svunpk_s64(svint32x2_t);
		4567	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_s16_s8_x4)))
		4568	svint16x4_t svunpk_s16(svint8x2_t);
		4569	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_u32_u16_x4)))
		4570	svuint32x4_t svunpk_u32(svuint16x2_t);
		4571	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_u64_u32_x4)))
		4572	svuint64x4_t svunpk_u64(svuint32x2_t);
		4573	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svunpk_u16_u8_x4)))
		4574	svuint16x4_t svunpk_u16(svuint8x2_t);
		4575	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_u8_x2)))
		4576	svuint8x2_t svuzp(svuint8x2_t);
		4577	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_u32_x2)))
		4578	svuint32x2_t svuzp(svuint32x2_t);
		4579	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_u64_x2)))
		4580	svuint64x2_t svuzp(svuint64x2_t);
		4581	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_u16_x2)))
		4582	svuint16x2_t svuzp(svuint16x2_t);
		4583	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_bf16_x2)))
		4584	svbfloat16x2_t svuzp(svbfloat16x2_t);
		4585	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_s8_x2)))
		4586	svint8x2_t svuzp(svint8x2_t);
		4587	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_f64_x2)))
		4588	svfloat64x2_t svuzp(svfloat64x2_t);
		4589	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_f32_x2)))
		4590	svfloat32x2_t svuzp(svfloat32x2_t);
		4591	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_f16_x2)))
		4592	svfloat16x2_t svuzp(svfloat16x2_t);
		4593	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_s32_x2)))
		4594	svint32x2_t svuzp(svint32x2_t);
		4595	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_s64_x2)))
		4596	svint64x2_t svuzp(svint64x2_t);
		4597	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_s16_x2)))
		4598	svint16x2_t svuzp(svint16x2_t);
		4599	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_u8_x4)))
		4600	svuint8x4_t svuzp(svuint8x4_t);
		4601	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_u32_x4)))
		4602	svuint32x4_t svuzp(svuint32x4_t);
		4603	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_u64_x4)))
		4604	svuint64x4_t svuzp(svuint64x4_t);
		4605	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_u16_x4)))
		4606	svuint16x4_t svuzp(svuint16x4_t);
		4607	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_bf16_x4)))
		4608	svbfloat16x4_t svuzp(svbfloat16x4_t);
		4609	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_s8_x4)))
		4610	svint8x4_t svuzp(svint8x4_t);
		4611	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_f64_x4)))
		4612	svfloat64x4_t svuzp(svfloat64x4_t);
		4613	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_f32_x4)))
		4614	svfloat32x4_t svuzp(svfloat32x4_t);
		4615	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_f16_x4)))
		4616	svfloat16x4_t svuzp(svfloat16x4_t);
		4617	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_s32_x4)))
		4618	svint32x4_t svuzp(svint32x4_t);
		4619	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_s64_x4)))
		4620	svint64x4_t svuzp(svint64x4_t);
		4621	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzp_s16_x4)))
		4622	svint16x4_t svuzp(svint16x4_t);
		4623	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_u8_x2)))
		4624	svuint8x2_t svuzpq(svuint8x2_t);
		4625	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_u32_x2)))
		4626	svuint32x2_t svuzpq(svuint32x2_t);
		4627	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_u64_x2)))
		4628	svuint64x2_t svuzpq(svuint64x2_t);
		4629	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_u16_x2)))
		4630	svuint16x2_t svuzpq(svuint16x2_t);
		4631	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_bf16_x2)))
		4632	svbfloat16x2_t svuzpq(svbfloat16x2_t);
		4633	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_s8_x2)))
		4634	svint8x2_t svuzpq(svint8x2_t);
		4635	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_f64_x2)))
		4636	svfloat64x2_t svuzpq(svfloat64x2_t);
		4637	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_f32_x2)))
		4638	svfloat32x2_t svuzpq(svfloat32x2_t);
		4639	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_f16_x2)))
		4640	svfloat16x2_t svuzpq(svfloat16x2_t);
		4641	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_s32_x2)))
		4642	svint32x2_t svuzpq(svint32x2_t);
		4643	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_s64_x2)))
		4644	svint64x2_t svuzpq(svint64x2_t);
		4645	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_s16_x2)))
		4646	svint16x2_t svuzpq(svint16x2_t);
		4647	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_u8_x4)))
		4648	svuint8x4_t svuzpq(svuint8x4_t);
		4649	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_u32_x4)))
		4650	svuint32x4_t svuzpq(svuint32x4_t);
		4651	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_u64_x4)))
		4652	svuint64x4_t svuzpq(svuint64x4_t);
		4653	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_u16_x4)))
		4654	svuint16x4_t svuzpq(svuint16x4_t);
		4655	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_bf16_x4)))
		4656	svbfloat16x4_t svuzpq(svbfloat16x4_t);
		4657	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_s8_x4)))
		4658	svint8x4_t svuzpq(svint8x4_t);
		4659	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_f64_x4)))
		4660	svfloat64x4_t svuzpq(svfloat64x4_t);
		4661	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_f32_x4)))
		4662	svfloat32x4_t svuzpq(svfloat32x4_t);
		4663	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_f16_x4)))
		4664	svfloat16x4_t svuzpq(svfloat16x4_t);
		4665	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_s32_x4)))
		4666	svint32x4_t svuzpq(svint32x4_t);
		4667	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_s64_x4)))
		4668	svint64x4_t svuzpq(svint64x4_t);
		4669	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq_s16_x4)))
		4670	svint16x4_t svuzpq(svint16x4_t);
		4671	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_u8_x2)))
		4672	svuint8x2_t svzip(svuint8x2_t);
		4673	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_u32_x2)))
		4674	svuint32x2_t svzip(svuint32x2_t);
		4675	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_u64_x2)))
		4676	svuint64x2_t svzip(svuint64x2_t);
		4677	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_u16_x2)))
		4678	svuint16x2_t svzip(svuint16x2_t);
		4679	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_bf16_x2)))
		4680	svbfloat16x2_t svzip(svbfloat16x2_t);
		4681	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_s8_x2)))
		4682	svint8x2_t svzip(svint8x2_t);
		4683	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_f64_x2)))
		4684	svfloat64x2_t svzip(svfloat64x2_t);
		4685	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_f32_x2)))
		4686	svfloat32x2_t svzip(svfloat32x2_t);
		4687	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_f16_x2)))
		4688	svfloat16x2_t svzip(svfloat16x2_t);
		4689	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_s32_x2)))
		4690	svint32x2_t svzip(svint32x2_t);
		4691	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_s64_x2)))
		4692	svint64x2_t svzip(svint64x2_t);
		4693	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_s16_x2)))
		4694	svint16x2_t svzip(svint16x2_t);
		4695	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_u8_x4)))
		4696	svuint8x4_t svzip(svuint8x4_t);
		4697	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_u32_x4)))
		4698	svuint32x4_t svzip(svuint32x4_t);
		4699	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_u64_x4)))
		4700	svuint64x4_t svzip(svuint64x4_t);
		4701	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_u16_x4)))
		4702	svuint16x4_t svzip(svuint16x4_t);
		4703	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_bf16_x4)))
		4704	svbfloat16x4_t svzip(svbfloat16x4_t);
		4705	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_s8_x4)))
		4706	svint8x4_t svzip(svint8x4_t);
		4707	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_f64_x4)))
		4708	svfloat64x4_t svzip(svfloat64x4_t);
		4709	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_f32_x4)))
		4710	svfloat32x4_t svzip(svfloat32x4_t);
		4711	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_f16_x4)))
		4712	svfloat16x4_t svzip(svfloat16x4_t);
		4713	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_s32_x4)))
		4714	svint32x4_t svzip(svint32x4_t);
		4715	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_s64_x4)))
		4716	svint64x4_t svzip(svint64x4_t);
		4717	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzip_s16_x4)))
		4718	svint16x4_t svzip(svint16x4_t);
		4719	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_u8_x2)))
		4720	svuint8x2_t svzipq(svuint8x2_t);
		4721	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_u32_x2)))
		4722	svuint32x2_t svzipq(svuint32x2_t);
		4723	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_u64_x2)))
		4724	svuint64x2_t svzipq(svuint64x2_t);
		4725	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_u16_x2)))
		4726	svuint16x2_t svzipq(svuint16x2_t);
		4727	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_bf16_x2)))
		4728	svbfloat16x2_t svzipq(svbfloat16x2_t);
		4729	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_s8_x2)))
		4730	svint8x2_t svzipq(svint8x2_t);
		4731	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_f64_x2)))
		4732	svfloat64x2_t svzipq(svfloat64x2_t);
		4733	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_f32_x2)))
		4734	svfloat32x2_t svzipq(svfloat32x2_t);
		4735	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_f16_x2)))
		4736	svfloat16x2_t svzipq(svfloat16x2_t);
		4737	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_s32_x2)))
		4738	svint32x2_t svzipq(svint32x2_t);
		4739	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_s64_x2)))
		4740	svint64x2_t svzipq(svint64x2_t);
		4741	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_s16_x2)))
		4742	svint16x2_t svzipq(svint16x2_t);
		4743	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_u8_x4)))
		4744	svuint8x4_t svzipq(svuint8x4_t);
		4745	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_u32_x4)))
		4746	svuint32x4_t svzipq(svuint32x4_t);
		4747	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_u64_x4)))
		4748	svuint64x4_t svzipq(svuint64x4_t);
		4749	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_u16_x4)))
		4750	svuint16x4_t svzipq(svuint16x4_t);
		4751	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_bf16_x4)))
		4752	svbfloat16x4_t svzipq(svbfloat16x4_t);
		4753	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_s8_x4)))
		4754	svint8x4_t svzipq(svint8x4_t);
		4755	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_f64_x4)))
		4756	svfloat64x4_t svzipq(svfloat64x4_t);
		4757	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_f32_x4)))
		4758	svfloat32x4_t svzipq(svfloat32x4_t);
		4759	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_f16_x4)))
		4760	svfloat16x4_t svzipq(svfloat16x4_t);
		4761	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_s32_x4)))
		4762	svint32x4_t svzipq(svint32x4_t);
		4763	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_s64_x4)))
		4764	svint64x4_t svzipq(svint64x4_t);
		4765	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq_s16_x4)))
		4766	svint16x4_t svzipq(svint16x4_t);
		4767	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_s16_s32_x2)))
		4768	svint16_t svqcvtn_s16_s32_x2(svint32x2_t);
		4769	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_u16_s32_x2)))
		4770	svuint16_t svqcvtn_u16_s32_x2(svint32x2_t);
		4771	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_u16_u32_x2)))
		4772	svuint16_t svqcvtn_u16_u32_x2(svuint32x2_t);
		4773	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_s16_s32_x2)))
		4774	svint16_t svqcvtn_s16(svint32x2_t);
		4775	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_u16_s32_x2)))
		4776	svuint16_t svqcvtn_u16(svint32x2_t);
		4777	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqcvtn_u16_u32_x2)))
		4778	svuint16_t svqcvtn_u16(svuint32x2_t);
846	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svabd_n_f64_m)))	4779	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svabd_n_f64_m)))
847	svfloat64_t svabd_n_f64_m(svbool_t, svfloat64_t, float64_t);	4780	svfloat64_t svabd_n_f64_m(svbool_t, svfloat64_t, float64_t);
848	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svabd_n_f32_m)))	4781	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svabd_n_f32_m)))
...	@@ -23877,6 +27810,1834 @@ __aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsm4e_u32)))	...	@@ -23877,6 +27810,1834 @@ __aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsm4e_u32)))
23877	svuint32_t svsm4e(svuint32_t, svuint32_t);	27810	svuint32_t svsm4e(svuint32_t, svuint32_t);
23878	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsm4ekey_u32)))	27811	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svsm4ekey_u32)))
23879	svuint32_t svsm4ekey(svuint32_t, svuint32_t);	27812	svuint32_t svsm4ekey(svuint32_t, svuint32_t);
		27813	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_u8)))
		27814	uint8x16_t svaddqv_u8(svbool_t, svuint8_t);
		27815	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_u32)))
		27816	uint32x4_t svaddqv_u32(svbool_t, svuint32_t);
		27817	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_u64)))
		27818	uint64x2_t svaddqv_u64(svbool_t, svuint64_t);
		27819	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_u16)))
		27820	uint16x8_t svaddqv_u16(svbool_t, svuint16_t);
		27821	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_s8)))
		27822	int8x16_t svaddqv_s8(svbool_t, svint8_t);
		27823	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_f64)))
		27824	float64x2_t svaddqv_f64(svbool_t, svfloat64_t);
		27825	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_f32)))
		27826	float32x4_t svaddqv_f32(svbool_t, svfloat32_t);
		27827	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_f16)))
		27828	float16x8_t svaddqv_f16(svbool_t, svfloat16_t);
		27829	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_s32)))
		27830	int32x4_t svaddqv_s32(svbool_t, svint32_t);
		27831	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_s64)))
		27832	int64x2_t svaddqv_s64(svbool_t, svint64_t);
		27833	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_s16)))
		27834	int16x8_t svaddqv_s16(svbool_t, svint16_t);
		27835	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svandqv_u8)))
		27836	uint8x16_t svandqv_u8(svbool_t, svuint8_t);
		27837	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svandqv_u32)))
		27838	uint32x4_t svandqv_u32(svbool_t, svuint32_t);
		27839	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svandqv_u64)))
		27840	uint64x2_t svandqv_u64(svbool_t, svuint64_t);
		27841	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svandqv_u16)))
		27842	uint16x8_t svandqv_u16(svbool_t, svuint16_t);
		27843	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svandqv_s8)))
		27844	int8x16_t svandqv_s8(svbool_t, svint8_t);
		27845	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svandqv_s32)))
		27846	int32x4_t svandqv_s32(svbool_t, svint32_t);
		27847	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svandqv_s64)))
		27848	int64x2_t svandqv_s64(svbool_t, svint64_t);
		27849	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svandqv_s16)))
		27850	int16x8_t svandqv_s16(svbool_t, svint16_t);
		27851	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_sveorqv_u8)))
		27852	uint8x16_t sveorqv_u8(svbool_t, svuint8_t);
		27853	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_sveorqv_u32)))
		27854	uint32x4_t sveorqv_u32(svbool_t, svuint32_t);
		27855	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_sveorqv_u64)))
		27856	uint64x2_t sveorqv_u64(svbool_t, svuint64_t);
		27857	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_sveorqv_u16)))
		27858	uint16x8_t sveorqv_u16(svbool_t, svuint16_t);
		27859	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_sveorqv_s8)))
		27860	int8x16_t sveorqv_s8(svbool_t, svint8_t);
		27861	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_sveorqv_s32)))
		27862	int32x4_t sveorqv_s32(svbool_t, svint32_t);
		27863	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_sveorqv_s64)))
		27864	int64x2_t sveorqv_s64(svbool_t, svint64_t);
		27865	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_sveorqv_s16)))
		27866	int16x8_t sveorqv_s16(svbool_t, svint16_t);
		27867	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_u8)))
		27868	svuint8_t svextq_u8(svuint8_t, svuint8_t, int32_t);
		27869	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_u32)))
		27870	svuint32_t svextq_u32(svuint32_t, svuint32_t, int32_t);
		27871	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_u64)))
		27872	svuint64_t svextq_u64(svuint64_t, svuint64_t, int32_t);
		27873	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_u16)))
		27874	svuint16_t svextq_u16(svuint16_t, svuint16_t, int32_t);
		27875	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_bf16)))
		27876	svbfloat16_t svextq_bf16(svbfloat16_t, svbfloat16_t, int32_t);
		27877	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_s8)))
		27878	svint8_t svextq_s8(svint8_t, svint8_t, int32_t);
		27879	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_f64)))
		27880	svfloat64_t svextq_f64(svfloat64_t, svfloat64_t, int32_t);
		27881	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_f32)))
		27882	svfloat32_t svextq_f32(svfloat32_t, svfloat32_t, int32_t);
		27883	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_f16)))
		27884	svfloat16_t svextq_f16(svfloat16_t, svfloat16_t, int32_t);
		27885	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_s32)))
		27886	svint32_t svextq_s32(svint32_t, svint32_t, int32_t);
		27887	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_s64)))
		27888	svint64_t svextq_s64(svint64_t, svint64_t, int32_t);
		27889	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_s16)))
		27890	svint16_t svextq_s16(svint16_t, svint16_t, int32_t);
		27891	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_u32)))
		27892	svuint32_t svld1q_gather_u64base_index_u32(svbool_t, svuint64_t, int64_t);
		27893	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_u64)))
		27894	svuint64_t svld1q_gather_u64base_index_u64(svbool_t, svuint64_t, int64_t);
		27895	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_u16)))
		27896	svuint16_t svld1q_gather_u64base_index_u16(svbool_t, svuint64_t, int64_t);
		27897	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_bf16)))
		27898	svbfloat16_t svld1q_gather_u64base_index_bf16(svbool_t, svuint64_t, int64_t);
		27899	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_f64)))
		27900	svfloat64_t svld1q_gather_u64base_index_f64(svbool_t, svuint64_t, int64_t);
		27901	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_f32)))
		27902	svfloat32_t svld1q_gather_u64base_index_f32(svbool_t, svuint64_t, int64_t);
		27903	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_f16)))
		27904	svfloat16_t svld1q_gather_u64base_index_f16(svbool_t, svuint64_t, int64_t);
		27905	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_s32)))
		27906	svint32_t svld1q_gather_u64base_index_s32(svbool_t, svuint64_t, int64_t);
		27907	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_s64)))
		27908	svint64_t svld1q_gather_u64base_index_s64(svbool_t, svuint64_t, int64_t);
		27909	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_s16)))
		27910	svint16_t svld1q_gather_u64base_index_s16(svbool_t, svuint64_t, int64_t);
		27911	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_u8)))
		27912	svuint8_t svld1q_gather_u64base_offset_u8(svbool_t, svuint64_t, int64_t);
		27913	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_u32)))
		27914	svuint32_t svld1q_gather_u64base_offset_u32(svbool_t, svuint64_t, int64_t);
		27915	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_u64)))
		27916	svuint64_t svld1q_gather_u64base_offset_u64(svbool_t, svuint64_t, int64_t);
		27917	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_u16)))
		27918	svuint16_t svld1q_gather_u64base_offset_u16(svbool_t, svuint64_t, int64_t);
		27919	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_bf16)))
		27920	svbfloat16_t svld1q_gather_u64base_offset_bf16(svbool_t, svuint64_t, int64_t);
		27921	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_s8)))
		27922	svint8_t svld1q_gather_u64base_offset_s8(svbool_t, svuint64_t, int64_t);
		27923	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_f64)))
		27924	svfloat64_t svld1q_gather_u64base_offset_f64(svbool_t, svuint64_t, int64_t);
		27925	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_f32)))
		27926	svfloat32_t svld1q_gather_u64base_offset_f32(svbool_t, svuint64_t, int64_t);
		27927	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_f16)))
		27928	svfloat16_t svld1q_gather_u64base_offset_f16(svbool_t, svuint64_t, int64_t);
		27929	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_s32)))
		27930	svint32_t svld1q_gather_u64base_offset_s32(svbool_t, svuint64_t, int64_t);
		27931	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_s64)))
		27932	svint64_t svld1q_gather_u64base_offset_s64(svbool_t, svuint64_t, int64_t);
		27933	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_s16)))
		27934	svint16_t svld1q_gather_u64base_offset_s16(svbool_t, svuint64_t, int64_t);
		27935	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_u8)))
		27936	svuint8_t svld1q_gather_u64base_u8(svbool_t, svuint64_t);
		27937	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_u32)))
		27938	svuint32_t svld1q_gather_u64base_u32(svbool_t, svuint64_t);
		27939	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_u64)))
		27940	svuint64_t svld1q_gather_u64base_u64(svbool_t, svuint64_t);
		27941	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_u16)))
		27942	svuint16_t svld1q_gather_u64base_u16(svbool_t, svuint64_t);
		27943	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_bf16)))
		27944	svbfloat16_t svld1q_gather_u64base_bf16(svbool_t, svuint64_t);
		27945	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_s8)))
		27946	svint8_t svld1q_gather_u64base_s8(svbool_t, svuint64_t);
		27947	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_f64)))
		27948	svfloat64_t svld1q_gather_u64base_f64(svbool_t, svuint64_t);
		27949	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_f32)))
		27950	svfloat32_t svld1q_gather_u64base_f32(svbool_t, svuint64_t);
		27951	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_f16)))
		27952	svfloat16_t svld1q_gather_u64base_f16(svbool_t, svuint64_t);
		27953	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_s32)))
		27954	svint32_t svld1q_gather_u64base_s32(svbool_t, svuint64_t);
		27955	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_s64)))
		27956	svint64_t svld1q_gather_u64base_s64(svbool_t, svuint64_t);
		27957	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_s16)))
		27958	svint16_t svld1q_gather_u64base_s16(svbool_t, svuint64_t);
		27959	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_u32)))
		27960	svuint32_t svld1q_gather_u64index_u32(svbool_t, uint32_t const *, svuint64_t);
		27961	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_u64)))
		27962	svuint64_t svld1q_gather_u64index_u64(svbool_t, uint64_t const *, svuint64_t);
		27963	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_u16)))
		27964	svuint16_t svld1q_gather_u64index_u16(svbool_t, uint16_t const *, svuint64_t);
		27965	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_bf16)))
		27966	svbfloat16_t svld1q_gather_u64index_bf16(svbool_t, bfloat16_t const *, svuint64_t);
		27967	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_f64)))
		27968	svfloat64_t svld1q_gather_u64index_f64(svbool_t, float64_t const *, svuint64_t);
		27969	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_f32)))
		27970	svfloat32_t svld1q_gather_u64index_f32(svbool_t, float32_t const *, svuint64_t);
		27971	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_f16)))
		27972	svfloat16_t svld1q_gather_u64index_f16(svbool_t, float16_t const *, svuint64_t);
		27973	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_s32)))
		27974	svint32_t svld1q_gather_u64index_s32(svbool_t, int32_t const *, svuint64_t);
		27975	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_s64)))
		27976	svint64_t svld1q_gather_u64index_s64(svbool_t, int64_t const *, svuint64_t);
		27977	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_s16)))
		27978	svint16_t svld1q_gather_u64index_s16(svbool_t, int16_t const *, svuint64_t);
		27979	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_u8)))
		27980	svuint8_t svld1q_gather_u64offset_u8(svbool_t, uint8_t const *, svuint64_t);
		27981	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_u32)))
		27982	svuint32_t svld1q_gather_u64offset_u32(svbool_t, uint32_t const *, svuint64_t);
		27983	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_u64)))
		27984	svuint64_t svld1q_gather_u64offset_u64(svbool_t, uint64_t const *, svuint64_t);
		27985	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_u16)))
		27986	svuint16_t svld1q_gather_u64offset_u16(svbool_t, uint16_t const *, svuint64_t);
		27987	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_bf16)))
		27988	svbfloat16_t svld1q_gather_u64offset_bf16(svbool_t, bfloat16_t const *, svuint64_t);
		27989	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_s8)))
		27990	svint8_t svld1q_gather_u64offset_s8(svbool_t, int8_t const *, svuint64_t);
		27991	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_f64)))
		27992	svfloat64_t svld1q_gather_u64offset_f64(svbool_t, float64_t const *, svuint64_t);
		27993	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_f32)))
		27994	svfloat32_t svld1q_gather_u64offset_f32(svbool_t, float32_t const *, svuint64_t);
		27995	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_f16)))
		27996	svfloat16_t svld1q_gather_u64offset_f16(svbool_t, float16_t const *, svuint64_t);
		27997	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_s32)))
		27998	svint32_t svld1q_gather_u64offset_s32(svbool_t, int32_t const *, svuint64_t);
		27999	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_s64)))
		28000	svint64_t svld1q_gather_u64offset_s64(svbool_t, int64_t const *, svuint64_t);
		28001	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_s16)))
		28002	svint16_t svld1q_gather_u64offset_s16(svbool_t, int16_t const *, svuint64_t);
		28003	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1udq_u64)))
		28004	svuint64_t svld1udq_u64(svbool_t, uint64_t const *);
		28005	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1udq_f64)))
		28006	svfloat64_t svld1udq_f64(svbool_t, float64_t const *);
		28007	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1udq_s64)))
		28008	svint64_t svld1udq_s64(svbool_t, int64_t const *);
		28009	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1udq_vnum_u64)))
		28010	svuint64_t svld1udq_vnum_u64(svbool_t, uint64_t const *, int64_t);
		28011	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1udq_vnum_f64)))
		28012	svfloat64_t svld1udq_vnum_f64(svbool_t, float64_t const *, int64_t);
		28013	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1udq_vnum_s64)))
		28014	svint64_t svld1udq_vnum_s64(svbool_t, int64_t const *, int64_t);
		28015	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1uwq_u32)))
		28016	svuint32_t svld1uwq_u32(svbool_t, uint32_t const *);
		28017	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1uwq_f32)))
		28018	svfloat32_t svld1uwq_f32(svbool_t, float32_t const *);
		28019	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1uwq_s32)))
		28020	svint32_t svld1uwq_s32(svbool_t, int32_t const *);
		28021	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1uwq_vnum_u32)))
		28022	svuint32_t svld1uwq_vnum_u32(svbool_t, uint32_t const *, int64_t);
		28023	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1uwq_vnum_f32)))
		28024	svfloat32_t svld1uwq_vnum_f32(svbool_t, float32_t const *, int64_t);
		28025	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1uwq_vnum_s32)))
		28026	svint32_t svld1uwq_vnum_s32(svbool_t, int32_t const *, int64_t);
		28027	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_u8)))
		28028	svuint8x2_t svld2q_u8(svbool_t, uint8_t const *);
		28029	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_u32)))
		28030	svuint32x2_t svld2q_u32(svbool_t, uint32_t const *);
		28031	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_u64)))
		28032	svuint64x2_t svld2q_u64(svbool_t, uint64_t const *);
		28033	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_u16)))
		28034	svuint16x2_t svld2q_u16(svbool_t, uint16_t const *);
		28035	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_s8)))
		28036	svint8x2_t svld2q_s8(svbool_t, int8_t const *);
		28037	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_f64)))
		28038	svfloat64x2_t svld2q_f64(svbool_t, float64_t const *);
		28039	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_f32)))
		28040	svfloat32x2_t svld2q_f32(svbool_t, float32_t const *);
		28041	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_f16)))
		28042	svfloat16x2_t svld2q_f16(svbool_t, float16_t const *);
		28043	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_s32)))
		28044	svint32x2_t svld2q_s32(svbool_t, int32_t const *);
		28045	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_s64)))
		28046	svint64x2_t svld2q_s64(svbool_t, int64_t const *);
		28047	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_s16)))
		28048	svint16x2_t svld2q_s16(svbool_t, int16_t const *);
		28049	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_bf16)))
		28050	svbfloat16x2_t svld2q_bf16(svbool_t, bfloat16_t const *);
		28051	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_u8)))
		28052	svuint8x2_t svld2q_vnum_u8(svbool_t, uint8_t const *, int64_t);
		28053	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_u32)))
		28054	svuint32x2_t svld2q_vnum_u32(svbool_t, uint32_t const *, int64_t);
		28055	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_u64)))
		28056	svuint64x2_t svld2q_vnum_u64(svbool_t, uint64_t const *, int64_t);
		28057	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_u16)))
		28058	svuint16x2_t svld2q_vnum_u16(svbool_t, uint16_t const *, int64_t);
		28059	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_s8)))
		28060	svint8x2_t svld2q_vnum_s8(svbool_t, int8_t const *, int64_t);
		28061	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_f64)))
		28062	svfloat64x2_t svld2q_vnum_f64(svbool_t, float64_t const *, int64_t);
		28063	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_f32)))
		28064	svfloat32x2_t svld2q_vnum_f32(svbool_t, float32_t const *, int64_t);
		28065	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_f16)))
		28066	svfloat16x2_t svld2q_vnum_f16(svbool_t, float16_t const *, int64_t);
		28067	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_s32)))
		28068	svint32x2_t svld2q_vnum_s32(svbool_t, int32_t const *, int64_t);
		28069	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_s64)))
		28070	svint64x2_t svld2q_vnum_s64(svbool_t, int64_t const *, int64_t);
		28071	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_s16)))
		28072	svint16x2_t svld2q_vnum_s16(svbool_t, int16_t const *, int64_t);
		28073	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_bf16)))
		28074	svbfloat16x2_t svld2q_vnum_bf16(svbool_t, bfloat16_t const *, int64_t);
		28075	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_u8)))
		28076	svuint8x3_t svld3q_u8(svbool_t, uint8_t const *);
		28077	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_u32)))
		28078	svuint32x3_t svld3q_u32(svbool_t, uint32_t const *);
		28079	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_u64)))
		28080	svuint64x3_t svld3q_u64(svbool_t, uint64_t const *);
		28081	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_u16)))
		28082	svuint16x3_t svld3q_u16(svbool_t, uint16_t const *);
		28083	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_s8)))
		28084	svint8x3_t svld3q_s8(svbool_t, int8_t const *);
		28085	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_f64)))
		28086	svfloat64x3_t svld3q_f64(svbool_t, float64_t const *);
		28087	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_f32)))
		28088	svfloat32x3_t svld3q_f32(svbool_t, float32_t const *);
		28089	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_f16)))
		28090	svfloat16x3_t svld3q_f16(svbool_t, float16_t const *);
		28091	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_s32)))
		28092	svint32x3_t svld3q_s32(svbool_t, int32_t const *);
		28093	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_s64)))
		28094	svint64x3_t svld3q_s64(svbool_t, int64_t const *);
		28095	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_s16)))
		28096	svint16x3_t svld3q_s16(svbool_t, int16_t const *);
		28097	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_bf16)))
		28098	svbfloat16x3_t svld3q_bf16(svbool_t, bfloat16_t const *);
		28099	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_u8)))
		28100	svuint8x3_t svld3q_vnum_u8(svbool_t, uint8_t const *, int64_t);
		28101	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_u32)))
		28102	svuint32x3_t svld3q_vnum_u32(svbool_t, uint32_t const *, int64_t);
		28103	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_u64)))
		28104	svuint64x3_t svld3q_vnum_u64(svbool_t, uint64_t const *, int64_t);
		28105	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_u16)))
		28106	svuint16x3_t svld3q_vnum_u16(svbool_t, uint16_t const *, int64_t);
		28107	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_s8)))
		28108	svint8x3_t svld3q_vnum_s8(svbool_t, int8_t const *, int64_t);
		28109	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_f64)))
		28110	svfloat64x3_t svld3q_vnum_f64(svbool_t, float64_t const *, int64_t);
		28111	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_f32)))
		28112	svfloat32x3_t svld3q_vnum_f32(svbool_t, float32_t const *, int64_t);
		28113	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_f16)))
		28114	svfloat16x3_t svld3q_vnum_f16(svbool_t, float16_t const *, int64_t);
		28115	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_s32)))
		28116	svint32x3_t svld3q_vnum_s32(svbool_t, int32_t const *, int64_t);
		28117	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_s64)))
		28118	svint64x3_t svld3q_vnum_s64(svbool_t, int64_t const *, int64_t);
		28119	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_s16)))
		28120	svint16x3_t svld3q_vnum_s16(svbool_t, int16_t const *, int64_t);
		28121	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_bf16)))
		28122	svbfloat16x3_t svld3q_vnum_bf16(svbool_t, bfloat16_t const *, int64_t);
		28123	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_u8)))
		28124	svuint8x4_t svld4q_u8(svbool_t, uint8_t const *);
		28125	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_u32)))
		28126	svuint32x4_t svld4q_u32(svbool_t, uint32_t const *);
		28127	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_u64)))
		28128	svuint64x4_t svld4q_u64(svbool_t, uint64_t const *);
		28129	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_u16)))
		28130	svuint16x4_t svld4q_u16(svbool_t, uint16_t const *);
		28131	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_s8)))
		28132	svint8x4_t svld4q_s8(svbool_t, int8_t const *);
		28133	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_f64)))
		28134	svfloat64x4_t svld4q_f64(svbool_t, float64_t const *);
		28135	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_f32)))
		28136	svfloat32x4_t svld4q_f32(svbool_t, float32_t const *);
		28137	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_f16)))
		28138	svfloat16x4_t svld4q_f16(svbool_t, float16_t const *);
		28139	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_s32)))
		28140	svint32x4_t svld4q_s32(svbool_t, int32_t const *);
		28141	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_s64)))
		28142	svint64x4_t svld4q_s64(svbool_t, int64_t const *);
		28143	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_s16)))
		28144	svint16x4_t svld4q_s16(svbool_t, int16_t const *);
		28145	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_bf16)))
		28146	svbfloat16x4_t svld4q_bf16(svbool_t, bfloat16_t const *);
		28147	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_u8)))
		28148	svuint8x4_t svld4q_vnum_u8(svbool_t, uint8_t const *, int64_t);
		28149	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_u32)))
		28150	svuint32x4_t svld4q_vnum_u32(svbool_t, uint32_t const *, int64_t);
		28151	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_u64)))
		28152	svuint64x4_t svld4q_vnum_u64(svbool_t, uint64_t const *, int64_t);
		28153	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_u16)))
		28154	svuint16x4_t svld4q_vnum_u16(svbool_t, uint16_t const *, int64_t);
		28155	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_s8)))
		28156	svint8x4_t svld4q_vnum_s8(svbool_t, int8_t const *, int64_t);
		28157	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_f64)))
		28158	svfloat64x4_t svld4q_vnum_f64(svbool_t, float64_t const *, int64_t);
		28159	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_f32)))
		28160	svfloat32x4_t svld4q_vnum_f32(svbool_t, float32_t const *, int64_t);
		28161	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_f16)))
		28162	svfloat16x4_t svld4q_vnum_f16(svbool_t, float16_t const *, int64_t);
		28163	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_s32)))
		28164	svint32x4_t svld4q_vnum_s32(svbool_t, int32_t const *, int64_t);
		28165	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_s64)))
		28166	svint64x4_t svld4q_vnum_s64(svbool_t, int64_t const *, int64_t);
		28167	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_s16)))
		28168	svint16x4_t svld4q_vnum_s16(svbool_t, int16_t const *, int64_t);
		28169	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_bf16)))
		28170	svbfloat16x4_t svld4q_vnum_bf16(svbool_t, bfloat16_t const *, int64_t);
		28171	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnmqv_f64)))
		28172	float64x2_t svmaxnmqv_f64(svbool_t, svfloat64_t);
		28173	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnmqv_f32)))
		28174	float32x4_t svmaxnmqv_f32(svbool_t, svfloat32_t);
		28175	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnmqv_f16)))
		28176	float16x8_t svmaxnmqv_f16(svbool_t, svfloat16_t);
		28177	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_f64)))
		28178	float64x2_t svmaxqv_f64(svbool_t, svfloat64_t);
		28179	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_f32)))
		28180	float32x4_t svmaxqv_f32(svbool_t, svfloat32_t);
		28181	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_f16)))
		28182	float16x8_t svmaxqv_f16(svbool_t, svfloat16_t);
		28183	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_s8)))
		28184	int8x16_t svmaxqv_s8(svbool_t, svint8_t);
		28185	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_s32)))
		28186	int32x4_t svmaxqv_s32(svbool_t, svint32_t);
		28187	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_s64)))
		28188	int64x2_t svmaxqv_s64(svbool_t, svint64_t);
		28189	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_s16)))
		28190	int16x8_t svmaxqv_s16(svbool_t, svint16_t);
		28191	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_u8)))
		28192	uint8x16_t svmaxqv_u8(svbool_t, svuint8_t);
		28193	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_u32)))
		28194	uint32x4_t svmaxqv_u32(svbool_t, svuint32_t);
		28195	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_u64)))
		28196	uint64x2_t svmaxqv_u64(svbool_t, svuint64_t);
		28197	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_u16)))
		28198	uint16x8_t svmaxqv_u16(svbool_t, svuint16_t);
		28199	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnmqv_f64)))
		28200	float64x2_t svminnmqv_f64(svbool_t, svfloat64_t);
		28201	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnmqv_f32)))
		28202	float32x4_t svminnmqv_f32(svbool_t, svfloat32_t);
		28203	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnmqv_f16)))
		28204	float16x8_t svminnmqv_f16(svbool_t, svfloat16_t);
		28205	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_f64)))
		28206	float64x2_t svminqv_f64(svbool_t, svfloat64_t);
		28207	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_f32)))
		28208	float32x4_t svminqv_f32(svbool_t, svfloat32_t);
		28209	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_f16)))
		28210	float16x8_t svminqv_f16(svbool_t, svfloat16_t);
		28211	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_s8)))
		28212	int8x16_t svminqv_s8(svbool_t, svint8_t);
		28213	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_s32)))
		28214	int32x4_t svminqv_s32(svbool_t, svint32_t);
		28215	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_s64)))
		28216	int64x2_t svminqv_s64(svbool_t, svint64_t);
		28217	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_s16)))
		28218	int16x8_t svminqv_s16(svbool_t, svint16_t);
		28219	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_u8)))
		28220	uint8x16_t svminqv_u8(svbool_t, svuint8_t);
		28221	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_u32)))
		28222	uint32x4_t svminqv_u32(svbool_t, svuint32_t);
		28223	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_u64)))
		28224	uint64x2_t svminqv_u64(svbool_t, svuint64_t);
		28225	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_u16)))
		28226	uint16x8_t svminqv_u16(svbool_t, svuint16_t);
		28227	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svorqv_u8)))
		28228	uint8x16_t svorqv_u8(svbool_t, svuint8_t);
		28229	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svorqv_u32)))
		28230	uint32x4_t svorqv_u32(svbool_t, svuint32_t);
		28231	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svorqv_u64)))
		28232	uint64x2_t svorqv_u64(svbool_t, svuint64_t);
		28233	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svorqv_u16)))
		28234	uint16x8_t svorqv_u16(svbool_t, svuint16_t);
		28235	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svorqv_s8)))
		28236	int8x16_t svorqv_s8(svbool_t, svint8_t);
		28237	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svorqv_s32)))
		28238	int32x4_t svorqv_s32(svbool_t, svint32_t);
		28239	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svorqv_s64)))
		28240	int64x2_t svorqv_s64(svbool_t, svint64_t);
		28241	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svorqv_s16)))
		28242	int16x8_t svorqv_s16(svbool_t, svint16_t);
		28243	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_u8)))
		28244	svbool_t svpmov_u8(svuint8_t);
		28245	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_s8)))
		28246	svbool_t svpmov_s8(svint8_t);
		28247	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_u64)))
		28248	svbool_t svpmov_u64(svuint64_t);
		28249	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_s64)))
		28250	svbool_t svpmov_s64(svint64_t);
		28251	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_u16)))
		28252	svbool_t svpmov_u16(svuint16_t);
		28253	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_s16)))
		28254	svbool_t svpmov_s16(svint16_t);
		28255	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_u32)))
		28256	svbool_t svpmov_u32(svuint32_t);
		28257	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_s32)))
		28258	svbool_t svpmov_s32(svint32_t);
		28259	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_u8)))
		28260	svbool_t svpmov_lane_u8(svuint8_t, uint64_t);
		28261	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_s8)))
		28262	svbool_t svpmov_lane_s8(svint8_t, uint64_t);
		28263	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_u64)))
		28264	svbool_t svpmov_lane_u64(svuint64_t, uint64_t);
		28265	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_s64)))
		28266	svbool_t svpmov_lane_s64(svint64_t, uint64_t);
		28267	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_u16)))
		28268	svbool_t svpmov_lane_u16(svuint16_t, uint64_t);
		28269	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_s16)))
		28270	svbool_t svpmov_lane_s16(svint16_t, uint64_t);
		28271	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_u32)))
		28272	svbool_t svpmov_lane_u32(svuint32_t, uint64_t);
		28273	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_s32)))
		28274	svbool_t svpmov_lane_s32(svint32_t, uint64_t);
		28275	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_u64_m)))
		28276	svuint64_t svpmov_lane_u64_m(svuint64_t, svbool_t, uint64_t);
		28277	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_s64_m)))
		28278	svint64_t svpmov_lane_s64_m(svint64_t, svbool_t, uint64_t);
		28279	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_u16_m)))
		28280	svuint16_t svpmov_lane_u16_m(svuint16_t, svbool_t, uint64_t);
		28281	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_s16_m)))
		28282	svint16_t svpmov_lane_s16_m(svint16_t, svbool_t, uint64_t);
		28283	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_u32_m)))
		28284	svuint32_t svpmov_lane_u32_m(svuint32_t, svbool_t, uint64_t);
		28285	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_s32_m)))
		28286	svint32_t svpmov_lane_s32_m(svint32_t, svbool_t, uint64_t);
		28287	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_u8_z)))
		28288	svuint8_t svpmov_u8_z(svbool_t);
		28289	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_s8_z)))
		28290	svint8_t svpmov_s8_z(svbool_t);
		28291	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_u64_z)))
		28292	svuint64_t svpmov_u64_z(svbool_t);
		28293	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_s64_z)))
		28294	svint64_t svpmov_s64_z(svbool_t);
		28295	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_u16_z)))
		28296	svuint16_t svpmov_u16_z(svbool_t);
		28297	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_s16_z)))
		28298	svint16_t svpmov_s16_z(svbool_t);
		28299	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_u32_z)))
		28300	svuint32_t svpmov_u32_z(svbool_t);
		28301	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_s32_z)))
		28302	svint32_t svpmov_s32_z(svbool_t);
		28303	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1dq_u64)))
		28304	void svst1dq_u64(svbool_t, uint64_t const *, svuint64_t);
		28305	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1dq_f64)))
		28306	void svst1dq_f64(svbool_t, float64_t const *, svfloat64_t);
		28307	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1dq_s64)))
		28308	void svst1dq_s64(svbool_t, int64_t const *, svint64_t);
		28309	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1dq_vnum_u64)))
		28310	void svst1dq_vnum_u64(svbool_t, uint64_t const *, int64_t, svuint64_t);
		28311	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1dq_vnum_f64)))
		28312	void svst1dq_vnum_f64(svbool_t, float64_t const *, int64_t, svfloat64_t);
		28313	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1dq_vnum_s64)))
		28314	void svst1dq_vnum_s64(svbool_t, int64_t const *, int64_t, svint64_t);
		28315	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_u8)))
		28316	void svst1q_scatter_u64base_u8(svbool_t, svuint64_t, svuint8_t);
		28317	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_u32)))
		28318	void svst1q_scatter_u64base_u32(svbool_t, svuint64_t, svuint32_t);
		28319	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_u64)))
		28320	void svst1q_scatter_u64base_u64(svbool_t, svuint64_t, svuint64_t);
		28321	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_u16)))
		28322	void svst1q_scatter_u64base_u16(svbool_t, svuint64_t, svuint16_t);
		28323	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_bf16)))
		28324	void svst1q_scatter_u64base_bf16(svbool_t, svuint64_t, svbfloat16_t);
		28325	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_s8)))
		28326	void svst1q_scatter_u64base_s8(svbool_t, svuint64_t, svint8_t);
		28327	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_f64)))
		28328	void svst1q_scatter_u64base_f64(svbool_t, svuint64_t, svfloat64_t);
		28329	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_f32)))
		28330	void svst1q_scatter_u64base_f32(svbool_t, svuint64_t, svfloat32_t);
		28331	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_f16)))
		28332	void svst1q_scatter_u64base_f16(svbool_t, svuint64_t, svfloat16_t);
		28333	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_s32)))
		28334	void svst1q_scatter_u64base_s32(svbool_t, svuint64_t, svint32_t);
		28335	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_s64)))
		28336	void svst1q_scatter_u64base_s64(svbool_t, svuint64_t, svint64_t);
		28337	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_s16)))
		28338	void svst1q_scatter_u64base_s16(svbool_t, svuint64_t, svint16_t);
		28339	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_u32)))
		28340	void svst1q_scatter_u64base_index_u32(svbool_t, svuint64_t, int64_t, svuint32_t);
		28341	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_u64)))
		28342	void svst1q_scatter_u64base_index_u64(svbool_t, svuint64_t, int64_t, svuint64_t);
		28343	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_u16)))
		28344	void svst1q_scatter_u64base_index_u16(svbool_t, svuint64_t, int64_t, svuint16_t);
		28345	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_bf16)))
		28346	void svst1q_scatter_u64base_index_bf16(svbool_t, svuint64_t, int64_t, svbfloat16_t);
		28347	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_f64)))
		28348	void svst1q_scatter_u64base_index_f64(svbool_t, svuint64_t, int64_t, svfloat64_t);
		28349	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_f32)))
		28350	void svst1q_scatter_u64base_index_f32(svbool_t, svuint64_t, int64_t, svfloat32_t);
		28351	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_f16)))
		28352	void svst1q_scatter_u64base_index_f16(svbool_t, svuint64_t, int64_t, svfloat16_t);
		28353	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_s32)))
		28354	void svst1q_scatter_u64base_index_s32(svbool_t, svuint64_t, int64_t, svint32_t);
		28355	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_s64)))
		28356	void svst1q_scatter_u64base_index_s64(svbool_t, svuint64_t, int64_t, svint64_t);
		28357	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_s16)))
		28358	void svst1q_scatter_u64base_index_s16(svbool_t, svuint64_t, int64_t, svint16_t);
		28359	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_u8)))
		28360	void svst1q_scatter_u64base_offset_u8(svbool_t, svuint64_t, int64_t, svuint8_t);
		28361	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_u32)))
		28362	void svst1q_scatter_u64base_offset_u32(svbool_t, svuint64_t, int64_t, svuint32_t);
		28363	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_u64)))
		28364	void svst1q_scatter_u64base_offset_u64(svbool_t, svuint64_t, int64_t, svuint64_t);
		28365	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_u16)))
		28366	void svst1q_scatter_u64base_offset_u16(svbool_t, svuint64_t, int64_t, svuint16_t);
		28367	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_bf16)))
		28368	void svst1q_scatter_u64base_offset_bf16(svbool_t, svuint64_t, int64_t, svbfloat16_t);
		28369	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_s8)))
		28370	void svst1q_scatter_u64base_offset_s8(svbool_t, svuint64_t, int64_t, svint8_t);
		28371	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_f64)))
		28372	void svst1q_scatter_u64base_offset_f64(svbool_t, svuint64_t, int64_t, svfloat64_t);
		28373	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_f32)))
		28374	void svst1q_scatter_u64base_offset_f32(svbool_t, svuint64_t, int64_t, svfloat32_t);
		28375	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_f16)))
		28376	void svst1q_scatter_u64base_offset_f16(svbool_t, svuint64_t, int64_t, svfloat16_t);
		28377	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_s32)))
		28378	void svst1q_scatter_u64base_offset_s32(svbool_t, svuint64_t, int64_t, svint32_t);
		28379	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_s64)))
		28380	void svst1q_scatter_u64base_offset_s64(svbool_t, svuint64_t, int64_t, svint64_t);
		28381	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_s16)))
		28382	void svst1q_scatter_u64base_offset_s16(svbool_t, svuint64_t, int64_t, svint16_t);
		28383	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_u32)))
		28384	void svst1q_scatter_u64index_u32(svbool_t, uint32_t *, svuint64_t, svuint32_t);
		28385	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_u64)))
		28386	void svst1q_scatter_u64index_u64(svbool_t, uint64_t *, svuint64_t, svuint64_t);
		28387	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_u16)))
		28388	void svst1q_scatter_u64index_u16(svbool_t, uint16_t *, svuint64_t, svuint16_t);
		28389	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_bf16)))
		28390	void svst1q_scatter_u64index_bf16(svbool_t, bfloat16_t *, svuint64_t, svbfloat16_t);
		28391	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_f64)))
		28392	void svst1q_scatter_u64index_f64(svbool_t, float64_t *, svuint64_t, svfloat64_t);
		28393	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_f32)))
		28394	void svst1q_scatter_u64index_f32(svbool_t, float32_t *, svuint64_t, svfloat32_t);
		28395	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_f16)))
		28396	void svst1q_scatter_u64index_f16(svbool_t, float16_t *, svuint64_t, svfloat16_t);
		28397	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_s32)))
		28398	void svst1q_scatter_u64index_s32(svbool_t, int32_t *, svuint64_t, svint32_t);
		28399	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_s64)))
		28400	void svst1q_scatter_u64index_s64(svbool_t, int64_t *, svuint64_t, svint64_t);
		28401	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_s16)))
		28402	void svst1q_scatter_u64index_s16(svbool_t, int16_t *, svuint64_t, svint16_t);
		28403	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_u8)))
		28404	void svst1q_scatter_u64offset_u8(svbool_t, uint8_t *, svuint64_t, svuint8_t);
		28405	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_u32)))
		28406	void svst1q_scatter_u64offset_u32(svbool_t, uint32_t *, svuint64_t, svuint32_t);
		28407	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_u64)))
		28408	void svst1q_scatter_u64offset_u64(svbool_t, uint64_t *, svuint64_t, svuint64_t);
		28409	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_u16)))
		28410	void svst1q_scatter_u64offset_u16(svbool_t, uint16_t *, svuint64_t, svuint16_t);
		28411	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_bf16)))
		28412	void svst1q_scatter_u64offset_bf16(svbool_t, bfloat16_t *, svuint64_t, svbfloat16_t);
		28413	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_s8)))
		28414	void svst1q_scatter_u64offset_s8(svbool_t, int8_t *, svuint64_t, svint8_t);
		28415	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_f64)))
		28416	void svst1q_scatter_u64offset_f64(svbool_t, float64_t *, svuint64_t, svfloat64_t);
		28417	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_f32)))
		28418	void svst1q_scatter_u64offset_f32(svbool_t, float32_t *, svuint64_t, svfloat32_t);
		28419	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_f16)))
		28420	void svst1q_scatter_u64offset_f16(svbool_t, float16_t *, svuint64_t, svfloat16_t);
		28421	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_s32)))
		28422	void svst1q_scatter_u64offset_s32(svbool_t, int32_t *, svuint64_t, svint32_t);
		28423	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_s64)))
		28424	void svst1q_scatter_u64offset_s64(svbool_t, int64_t *, svuint64_t, svint64_t);
		28425	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_s16)))
		28426	void svst1q_scatter_u64offset_s16(svbool_t, int16_t *, svuint64_t, svint16_t);
		28427	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1wq_u32)))
		28428	void svst1wq_u32(svbool_t, uint32_t const *, svuint32_t);
		28429	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1wq_f32)))
		28430	void svst1wq_f32(svbool_t, float32_t const *, svfloat32_t);
		28431	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1wq_s32)))
		28432	void svst1wq_s32(svbool_t, int32_t const *, svint32_t);
		28433	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1wq_vnum_u32)))
		28434	void svst1wq_vnum_u32(svbool_t, uint32_t const *, int64_t, svuint32_t);
		28435	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1wq_vnum_f32)))
		28436	void svst1wq_vnum_f32(svbool_t, float32_t const *, int64_t, svfloat32_t);
		28437	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1wq_vnum_s32)))
		28438	void svst1wq_vnum_s32(svbool_t, int32_t const *, int64_t, svint32_t);
		28439	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_u8)))
		28440	void svst2q_u8(svbool_t, uint8_t const *, svuint8x2_t);
		28441	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_u32)))
		28442	void svst2q_u32(svbool_t, uint32_t const *, svuint32x2_t);
		28443	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_u64)))
		28444	void svst2q_u64(svbool_t, uint64_t const *, svuint64x2_t);
		28445	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_u16)))
		28446	void svst2q_u16(svbool_t, uint16_t const *, svuint16x2_t);
		28447	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_s8)))
		28448	void svst2q_s8(svbool_t, int8_t const *, svint8x2_t);
		28449	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_f64)))
		28450	void svst2q_f64(svbool_t, float64_t const *, svfloat64x2_t);
		28451	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_f32)))
		28452	void svst2q_f32(svbool_t, float32_t const *, svfloat32x2_t);
		28453	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_f16)))
		28454	void svst2q_f16(svbool_t, float16_t const *, svfloat16x2_t);
		28455	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_s32)))
		28456	void svst2q_s32(svbool_t, int32_t const *, svint32x2_t);
		28457	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_s64)))
		28458	void svst2q_s64(svbool_t, int64_t const *, svint64x2_t);
		28459	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_s16)))
		28460	void svst2q_s16(svbool_t, int16_t const *, svint16x2_t);
		28461	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_bf16)))
		28462	void svst2q_bf16(svbool_t, bfloat16_t const *, svbfloat16x2_t);
		28463	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_u8)))
		28464	void svst2q_vnum_u8(svbool_t, uint8_t const *, int64_t, svuint8x2_t);
		28465	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_u32)))
		28466	void svst2q_vnum_u32(svbool_t, uint32_t const *, int64_t, svuint32x2_t);
		28467	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_u64)))
		28468	void svst2q_vnum_u64(svbool_t, uint64_t const *, int64_t, svuint64x2_t);
		28469	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_u16)))
		28470	void svst2q_vnum_u16(svbool_t, uint16_t const *, int64_t, svuint16x2_t);
		28471	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_s8)))
		28472	void svst2q_vnum_s8(svbool_t, int8_t const *, int64_t, svint8x2_t);
		28473	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_f64)))
		28474	void svst2q_vnum_f64(svbool_t, float64_t const *, int64_t, svfloat64x2_t);
		28475	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_f32)))
		28476	void svst2q_vnum_f32(svbool_t, float32_t const *, int64_t, svfloat32x2_t);
		28477	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_f16)))
		28478	void svst2q_vnum_f16(svbool_t, float16_t const *, int64_t, svfloat16x2_t);
		28479	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_s32)))
		28480	void svst2q_vnum_s32(svbool_t, int32_t const *, int64_t, svint32x2_t);
		28481	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_s64)))
		28482	void svst2q_vnum_s64(svbool_t, int64_t const *, int64_t, svint64x2_t);
		28483	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_s16)))
		28484	void svst2q_vnum_s16(svbool_t, int16_t const *, int64_t, svint16x2_t);
		28485	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_bf16)))
		28486	void svst2q_vnum_bf16(svbool_t, bfloat16_t const *, int64_t, svbfloat16x2_t);
		28487	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_u8)))
		28488	void svst3q_u8(svbool_t, uint8_t const *, svuint8x3_t);
		28489	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_u32)))
		28490	void svst3q_u32(svbool_t, uint32_t const *, svuint32x3_t);
		28491	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_u64)))
		28492	void svst3q_u64(svbool_t, uint64_t const *, svuint64x3_t);
		28493	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_u16)))
		28494	void svst3q_u16(svbool_t, uint16_t const *, svuint16x3_t);
		28495	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_s8)))
		28496	void svst3q_s8(svbool_t, int8_t const *, svint8x3_t);
		28497	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_f64)))
		28498	void svst3q_f64(svbool_t, float64_t const *, svfloat64x3_t);
		28499	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_f32)))
		28500	void svst3q_f32(svbool_t, float32_t const *, svfloat32x3_t);
		28501	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_f16)))
		28502	void svst3q_f16(svbool_t, float16_t const *, svfloat16x3_t);
		28503	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_s32)))
		28504	void svst3q_s32(svbool_t, int32_t const *, svint32x3_t);
		28505	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_s64)))
		28506	void svst3q_s64(svbool_t, int64_t const *, svint64x3_t);
		28507	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_s16)))
		28508	void svst3q_s16(svbool_t, int16_t const *, svint16x3_t);
		28509	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_bf16)))
		28510	void svst3q_bf16(svbool_t, bfloat16_t const *, svbfloat16x3_t);
		28511	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_u8)))
		28512	void svst3q_vnum_u8(svbool_t, uint8_t const *, int64_t, svuint8x3_t);
		28513	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_u32)))
		28514	void svst3q_vnum_u32(svbool_t, uint32_t const *, int64_t, svuint32x3_t);
		28515	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_u64)))
		28516	void svst3q_vnum_u64(svbool_t, uint64_t const *, int64_t, svuint64x3_t);
		28517	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_u16)))
		28518	void svst3q_vnum_u16(svbool_t, uint16_t const *, int64_t, svuint16x3_t);
		28519	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_s8)))
		28520	void svst3q_vnum_s8(svbool_t, int8_t const *, int64_t, svint8x3_t);
		28521	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_f64)))
		28522	void svst3q_vnum_f64(svbool_t, float64_t const *, int64_t, svfloat64x3_t);
		28523	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_f32)))
		28524	void svst3q_vnum_f32(svbool_t, float32_t const *, int64_t, svfloat32x3_t);
		28525	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_f16)))
		28526	void svst3q_vnum_f16(svbool_t, float16_t const *, int64_t, svfloat16x3_t);
		28527	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_s32)))
		28528	void svst3q_vnum_s32(svbool_t, int32_t const *, int64_t, svint32x3_t);
		28529	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_s64)))
		28530	void svst3q_vnum_s64(svbool_t, int64_t const *, int64_t, svint64x3_t);
		28531	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_s16)))
		28532	void svst3q_vnum_s16(svbool_t, int16_t const *, int64_t, svint16x3_t);
		28533	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_bf16)))
		28534	void svst3q_vnum_bf16(svbool_t, bfloat16_t const *, int64_t, svbfloat16x3_t);
		28535	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_u8)))
		28536	void svst4q_u8(svbool_t, uint8_t const *, svuint8x4_t);
		28537	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_u32)))
		28538	void svst4q_u32(svbool_t, uint32_t const *, svuint32x4_t);
		28539	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_u64)))
		28540	void svst4q_u64(svbool_t, uint64_t const *, svuint64x4_t);
		28541	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_u16)))
		28542	void svst4q_u16(svbool_t, uint16_t const *, svuint16x4_t);
		28543	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_s8)))
		28544	void svst4q_s8(svbool_t, int8_t const *, svint8x4_t);
		28545	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_f64)))
		28546	void svst4q_f64(svbool_t, float64_t const *, svfloat64x4_t);
		28547	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_f32)))
		28548	void svst4q_f32(svbool_t, float32_t const *, svfloat32x4_t);
		28549	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_f16)))
		28550	void svst4q_f16(svbool_t, float16_t const *, svfloat16x4_t);
		28551	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_s32)))
		28552	void svst4q_s32(svbool_t, int32_t const *, svint32x4_t);
		28553	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_s64)))
		28554	void svst4q_s64(svbool_t, int64_t const *, svint64x4_t);
		28555	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_s16)))
		28556	void svst4q_s16(svbool_t, int16_t const *, svint16x4_t);
		28557	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_bf16)))
		28558	void svst4q_bf16(svbool_t, bfloat16_t const *, svbfloat16x4_t);
		28559	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_u8)))
		28560	void svst4q_vnum_u8(svbool_t, uint8_t const *, int64_t, svuint8x4_t);
		28561	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_u32)))
		28562	void svst4q_vnum_u32(svbool_t, uint32_t const *, int64_t, svuint32x4_t);
		28563	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_u64)))
		28564	void svst4q_vnum_u64(svbool_t, uint64_t const *, int64_t, svuint64x4_t);
		28565	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_u16)))
		28566	void svst4q_vnum_u16(svbool_t, uint16_t const *, int64_t, svuint16x4_t);
		28567	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_s8)))
		28568	void svst4q_vnum_s8(svbool_t, int8_t const *, int64_t, svint8x4_t);
		28569	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_f64)))
		28570	void svst4q_vnum_f64(svbool_t, float64_t const *, int64_t, svfloat64x4_t);
		28571	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_f32)))
		28572	void svst4q_vnum_f32(svbool_t, float32_t const *, int64_t, svfloat32x4_t);
		28573	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_f16)))
		28574	void svst4q_vnum_f16(svbool_t, float16_t const *, int64_t, svfloat16x4_t);
		28575	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_s32)))
		28576	void svst4q_vnum_s32(svbool_t, int32_t const *, int64_t, svint32x4_t);
		28577	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_s64)))
		28578	void svst4q_vnum_s64(svbool_t, int64_t const *, int64_t, svint64x4_t);
		28579	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_s16)))
		28580	void svst4q_vnum_s16(svbool_t, int16_t const *, int64_t, svint16x4_t);
		28581	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_bf16)))
		28582	void svst4q_vnum_bf16(svbool_t, bfloat16_t const *, int64_t, svbfloat16x4_t);
		28583	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_u8)))
		28584	svuint8_t svtblq_u8(svuint8_t, svuint8_t);
		28585	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_u32)))
		28586	svuint32_t svtblq_u32(svuint32_t, svuint32_t);
		28587	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_u64)))
		28588	svuint64_t svtblq_u64(svuint64_t, svuint64_t);
		28589	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_u16)))
		28590	svuint16_t svtblq_u16(svuint16_t, svuint16_t);
		28591	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_bf16)))
		28592	svbfloat16_t svtblq_bf16(svbfloat16_t, svuint16_t);
		28593	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_s8)))
		28594	svint8_t svtblq_s8(svint8_t, svuint8_t);
		28595	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_f64)))
		28596	svfloat64_t svtblq_f64(svfloat64_t, svuint64_t);
		28597	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_f32)))
		28598	svfloat32_t svtblq_f32(svfloat32_t, svuint32_t);
		28599	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_f16)))
		28600	svfloat16_t svtblq_f16(svfloat16_t, svuint16_t);
		28601	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_s32)))
		28602	svint32_t svtblq_s32(svint32_t, svuint32_t);
		28603	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_s64)))
		28604	svint64_t svtblq_s64(svint64_t, svuint64_t);
		28605	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_s16)))
		28606	svint16_t svtblq_s16(svint16_t, svuint16_t);
		28607	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_u8)))
		28608	svuint8_t svtbxq_u8(svuint8_t, svuint8_t, svuint8_t);
		28609	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_u32)))
		28610	svuint32_t svtbxq_u32(svuint32_t, svuint32_t, svuint32_t);
		28611	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_u64)))
		28612	svuint64_t svtbxq_u64(svuint64_t, svuint64_t, svuint64_t);
		28613	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_u16)))
		28614	svuint16_t svtbxq_u16(svuint16_t, svuint16_t, svuint16_t);
		28615	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_bf16)))
		28616	svbfloat16_t svtbxq_bf16(svbfloat16_t, svbfloat16_t, svuint16_t);
		28617	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_s8)))
		28618	svint8_t svtbxq_s8(svint8_t, svint8_t, svuint8_t);
		28619	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_f64)))
		28620	svfloat64_t svtbxq_f64(svfloat64_t, svfloat64_t, svuint64_t);
		28621	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_f32)))
		28622	svfloat32_t svtbxq_f32(svfloat32_t, svfloat32_t, svuint32_t);
		28623	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_f16)))
		28624	svfloat16_t svtbxq_f16(svfloat16_t, svfloat16_t, svuint16_t);
		28625	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_s32)))
		28626	svint32_t svtbxq_s32(svint32_t, svint32_t, svuint32_t);
		28627	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_s64)))
		28628	svint64_t svtbxq_s64(svint64_t, svint64_t, svuint64_t);
		28629	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_s16)))
		28630	svint16_t svtbxq_s16(svint16_t, svint16_t, svuint16_t);
		28631	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_u8)))
		28632	svuint8_t svuzpq1_u8(svuint8_t, svuint8_t);
		28633	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_u32)))
		28634	svuint32_t svuzpq1_u32(svuint32_t, svuint32_t);
		28635	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_u64)))
		28636	svuint64_t svuzpq1_u64(svuint64_t, svuint64_t);
		28637	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_u16)))
		28638	svuint16_t svuzpq1_u16(svuint16_t, svuint16_t);
		28639	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_bf16)))
		28640	svbfloat16_t svuzpq1_bf16(svbfloat16_t, svbfloat16_t);
		28641	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_s8)))
		28642	svint8_t svuzpq1_s8(svint8_t, svint8_t);
		28643	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_f64)))
		28644	svfloat64_t svuzpq1_f64(svfloat64_t, svfloat64_t);
		28645	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_f32)))
		28646	svfloat32_t svuzpq1_f32(svfloat32_t, svfloat32_t);
		28647	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_f16)))
		28648	svfloat16_t svuzpq1_f16(svfloat16_t, svfloat16_t);
		28649	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_s32)))
		28650	svint32_t svuzpq1_s32(svint32_t, svint32_t);
		28651	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_s64)))
		28652	svint64_t svuzpq1_s64(svint64_t, svint64_t);
		28653	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_s16)))
		28654	svint16_t svuzpq1_s16(svint16_t, svint16_t);
		28655	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_u8)))
		28656	svuint8_t svuzpq2_u8(svuint8_t, svuint8_t);
		28657	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_u32)))
		28658	svuint32_t svuzpq2_u32(svuint32_t, svuint32_t);
		28659	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_u64)))
		28660	svuint64_t svuzpq2_u64(svuint64_t, svuint64_t);
		28661	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_u16)))
		28662	svuint16_t svuzpq2_u16(svuint16_t, svuint16_t);
		28663	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_bf16)))
		28664	svbfloat16_t svuzpq2_bf16(svbfloat16_t, svbfloat16_t);
		28665	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_s8)))
		28666	svint8_t svuzpq2_s8(svint8_t, svint8_t);
		28667	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_f64)))
		28668	svfloat64_t svuzpq2_f64(svfloat64_t, svfloat64_t);
		28669	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_f32)))
		28670	svfloat32_t svuzpq2_f32(svfloat32_t, svfloat32_t);
		28671	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_f16)))
		28672	svfloat16_t svuzpq2_f16(svfloat16_t, svfloat16_t);
		28673	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_s32)))
		28674	svint32_t svuzpq2_s32(svint32_t, svint32_t);
		28675	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_s64)))
		28676	svint64_t svuzpq2_s64(svint64_t, svint64_t);
		28677	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_s16)))
		28678	svint16_t svuzpq2_s16(svint16_t, svint16_t);
		28679	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_u8)))
		28680	svuint8_t svzipq1_u8(svuint8_t, svuint8_t);
		28681	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_u32)))
		28682	svuint32_t svzipq1_u32(svuint32_t, svuint32_t);
		28683	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_u64)))
		28684	svuint64_t svzipq1_u64(svuint64_t, svuint64_t);
		28685	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_u16)))
		28686	svuint16_t svzipq1_u16(svuint16_t, svuint16_t);
		28687	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_bf16)))
		28688	svbfloat16_t svzipq1_bf16(svbfloat16_t, svbfloat16_t);
		28689	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_s8)))
		28690	svint8_t svzipq1_s8(svint8_t, svint8_t);
		28691	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_f64)))
		28692	svfloat64_t svzipq1_f64(svfloat64_t, svfloat64_t);
		28693	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_f32)))
		28694	svfloat32_t svzipq1_f32(svfloat32_t, svfloat32_t);
		28695	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_f16)))
		28696	svfloat16_t svzipq1_f16(svfloat16_t, svfloat16_t);
		28697	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_s32)))
		28698	svint32_t svzipq1_s32(svint32_t, svint32_t);
		28699	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_s64)))
		28700	svint64_t svzipq1_s64(svint64_t, svint64_t);
		28701	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_s16)))
		28702	svint16_t svzipq1_s16(svint16_t, svint16_t);
		28703	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_u8)))
		28704	svuint8_t svzipq2_u8(svuint8_t, svuint8_t);
		28705	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_u32)))
		28706	svuint32_t svzipq2_u32(svuint32_t, svuint32_t);
		28707	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_u64)))
		28708	svuint64_t svzipq2_u64(svuint64_t, svuint64_t);
		28709	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_u16)))
		28710	svuint16_t svzipq2_u16(svuint16_t, svuint16_t);
		28711	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_bf16)))
		28712	svbfloat16_t svzipq2_bf16(svbfloat16_t, svbfloat16_t);
		28713	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_s8)))
		28714	svint8_t svzipq2_s8(svint8_t, svint8_t);
		28715	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_f64)))
		28716	svfloat64_t svzipq2_f64(svfloat64_t, svfloat64_t);
		28717	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_f32)))
		28718	svfloat32_t svzipq2_f32(svfloat32_t, svfloat32_t);
		28719	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_f16)))
		28720	svfloat16_t svzipq2_f16(svfloat16_t, svfloat16_t);
		28721	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_s32)))
		28722	svint32_t svzipq2_s32(svint32_t, svint32_t);
		28723	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_s64)))
		28724	svint64_t svzipq2_s64(svint64_t, svint64_t);
		28725	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_s16)))
		28726	svint16_t svzipq2_s16(svint16_t, svint16_t);
		28727	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_u8)))
		28728	uint8x16_t svaddqv(svbool_t, svuint8_t);
		28729	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_u32)))
		28730	uint32x4_t svaddqv(svbool_t, svuint32_t);
		28731	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_u64)))
		28732	uint64x2_t svaddqv(svbool_t, svuint64_t);
		28733	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_u16)))
		28734	uint16x8_t svaddqv(svbool_t, svuint16_t);
		28735	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_s8)))
		28736	int8x16_t svaddqv(svbool_t, svint8_t);
		28737	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_f64)))
		28738	float64x2_t svaddqv(svbool_t, svfloat64_t);
		28739	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_f32)))
		28740	float32x4_t svaddqv(svbool_t, svfloat32_t);
		28741	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_f16)))
		28742	float16x8_t svaddqv(svbool_t, svfloat16_t);
		28743	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_s32)))
		28744	int32x4_t svaddqv(svbool_t, svint32_t);
		28745	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_s64)))
		28746	int64x2_t svaddqv(svbool_t, svint64_t);
		28747	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svaddqv_s16)))
		28748	int16x8_t svaddqv(svbool_t, svint16_t);
		28749	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svandqv_u8)))
		28750	uint8x16_t svandqv(svbool_t, svuint8_t);
		28751	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svandqv_u32)))
		28752	uint32x4_t svandqv(svbool_t, svuint32_t);
		28753	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svandqv_u64)))
		28754	uint64x2_t svandqv(svbool_t, svuint64_t);
		28755	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svandqv_u16)))
		28756	uint16x8_t svandqv(svbool_t, svuint16_t);
		28757	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svandqv_s8)))
		28758	int8x16_t svandqv(svbool_t, svint8_t);
		28759	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svandqv_s32)))
		28760	int32x4_t svandqv(svbool_t, svint32_t);
		28761	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svandqv_s64)))
		28762	int64x2_t svandqv(svbool_t, svint64_t);
		28763	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svandqv_s16)))
		28764	int16x8_t svandqv(svbool_t, svint16_t);
		28765	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_sveorqv_u8)))
		28766	uint8x16_t sveorqv(svbool_t, svuint8_t);
		28767	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_sveorqv_u32)))
		28768	uint32x4_t sveorqv(svbool_t, svuint32_t);
		28769	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_sveorqv_u64)))
		28770	uint64x2_t sveorqv(svbool_t, svuint64_t);
		28771	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_sveorqv_u16)))
		28772	uint16x8_t sveorqv(svbool_t, svuint16_t);
		28773	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_sveorqv_s8)))
		28774	int8x16_t sveorqv(svbool_t, svint8_t);
		28775	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_sveorqv_s32)))
		28776	int32x4_t sveorqv(svbool_t, svint32_t);
		28777	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_sveorqv_s64)))
		28778	int64x2_t sveorqv(svbool_t, svint64_t);
		28779	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_sveorqv_s16)))
		28780	int16x8_t sveorqv(svbool_t, svint16_t);
		28781	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_u8)))
		28782	svuint8_t svextq(svuint8_t, svuint8_t, int32_t);
		28783	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_u32)))
		28784	svuint32_t svextq(svuint32_t, svuint32_t, int32_t);
		28785	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_u64)))
		28786	svuint64_t svextq(svuint64_t, svuint64_t, int32_t);
		28787	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_u16)))
		28788	svuint16_t svextq(svuint16_t, svuint16_t, int32_t);
		28789	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_bf16)))
		28790	svbfloat16_t svextq(svbfloat16_t, svbfloat16_t, int32_t);
		28791	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_s8)))
		28792	svint8_t svextq(svint8_t, svint8_t, int32_t);
		28793	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_f64)))
		28794	svfloat64_t svextq(svfloat64_t, svfloat64_t, int32_t);
		28795	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_f32)))
		28796	svfloat32_t svextq(svfloat32_t, svfloat32_t, int32_t);
		28797	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_f16)))
		28798	svfloat16_t svextq(svfloat16_t, svfloat16_t, int32_t);
		28799	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_s32)))
		28800	svint32_t svextq(svint32_t, svint32_t, int32_t);
		28801	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_s64)))
		28802	svint64_t svextq(svint64_t, svint64_t, int32_t);
		28803	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svextq_s16)))
		28804	svint16_t svextq(svint16_t, svint16_t, int32_t);
		28805	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_u32)))
		28806	svuint32_t svld1q_gather_index_u32(svbool_t, svuint64_t, int64_t);
		28807	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_u64)))
		28808	svuint64_t svld1q_gather_index_u64(svbool_t, svuint64_t, int64_t);
		28809	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_u16)))
		28810	svuint16_t svld1q_gather_index_u16(svbool_t, svuint64_t, int64_t);
		28811	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_bf16)))
		28812	svbfloat16_t svld1q_gather_index_bf16(svbool_t, svuint64_t, int64_t);
		28813	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_f64)))
		28814	svfloat64_t svld1q_gather_index_f64(svbool_t, svuint64_t, int64_t);
		28815	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_f32)))
		28816	svfloat32_t svld1q_gather_index_f32(svbool_t, svuint64_t, int64_t);
		28817	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_f16)))
		28818	svfloat16_t svld1q_gather_index_f16(svbool_t, svuint64_t, int64_t);
		28819	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_s32)))
		28820	svint32_t svld1q_gather_index_s32(svbool_t, svuint64_t, int64_t);
		28821	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_s64)))
		28822	svint64_t svld1q_gather_index_s64(svbool_t, svuint64_t, int64_t);
		28823	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_index_s16)))
		28824	svint16_t svld1q_gather_index_s16(svbool_t, svuint64_t, int64_t);
		28825	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_u8)))
		28826	svuint8_t svld1q_gather_offset_u8(svbool_t, svuint64_t, int64_t);
		28827	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_u32)))
		28828	svuint32_t svld1q_gather_offset_u32(svbool_t, svuint64_t, int64_t);
		28829	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_u64)))
		28830	svuint64_t svld1q_gather_offset_u64(svbool_t, svuint64_t, int64_t);
		28831	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_u16)))
		28832	svuint16_t svld1q_gather_offset_u16(svbool_t, svuint64_t, int64_t);
		28833	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_bf16)))
		28834	svbfloat16_t svld1q_gather_offset_bf16(svbool_t, svuint64_t, int64_t);
		28835	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_s8)))
		28836	svint8_t svld1q_gather_offset_s8(svbool_t, svuint64_t, int64_t);
		28837	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_f64)))
		28838	svfloat64_t svld1q_gather_offset_f64(svbool_t, svuint64_t, int64_t);
		28839	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_f32)))
		28840	svfloat32_t svld1q_gather_offset_f32(svbool_t, svuint64_t, int64_t);
		28841	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_f16)))
		28842	svfloat16_t svld1q_gather_offset_f16(svbool_t, svuint64_t, int64_t);
		28843	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_s32)))
		28844	svint32_t svld1q_gather_offset_s32(svbool_t, svuint64_t, int64_t);
		28845	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_s64)))
		28846	svint64_t svld1q_gather_offset_s64(svbool_t, svuint64_t, int64_t);
		28847	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_offset_s16)))
		28848	svint16_t svld1q_gather_offset_s16(svbool_t, svuint64_t, int64_t);
		28849	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_u8)))
		28850	svuint8_t svld1q_gather_u8(svbool_t, svuint64_t);
		28851	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_u32)))
		28852	svuint32_t svld1q_gather_u32(svbool_t, svuint64_t);
		28853	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_u64)))
		28854	svuint64_t svld1q_gather_u64(svbool_t, svuint64_t);
		28855	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_u16)))
		28856	svuint16_t svld1q_gather_u16(svbool_t, svuint64_t);
		28857	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_bf16)))
		28858	svbfloat16_t svld1q_gather_bf16(svbool_t, svuint64_t);
		28859	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_s8)))
		28860	svint8_t svld1q_gather_s8(svbool_t, svuint64_t);
		28861	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_f64)))
		28862	svfloat64_t svld1q_gather_f64(svbool_t, svuint64_t);
		28863	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_f32)))
		28864	svfloat32_t svld1q_gather_f32(svbool_t, svuint64_t);
		28865	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_f16)))
		28866	svfloat16_t svld1q_gather_f16(svbool_t, svuint64_t);
		28867	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_s32)))
		28868	svint32_t svld1q_gather_s32(svbool_t, svuint64_t);
		28869	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_s64)))
		28870	svint64_t svld1q_gather_s64(svbool_t, svuint64_t);
		28871	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64base_s16)))
		28872	svint16_t svld1q_gather_s16(svbool_t, svuint64_t);
		28873	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_u32)))
		28874	svuint32_t svld1q_gather_index(svbool_t, uint32_t const *, svuint64_t);
		28875	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_u64)))
		28876	svuint64_t svld1q_gather_index(svbool_t, uint64_t const *, svuint64_t);
		28877	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_u16)))
		28878	svuint16_t svld1q_gather_index(svbool_t, uint16_t const *, svuint64_t);
		28879	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_bf16)))
		28880	svbfloat16_t svld1q_gather_index(svbool_t, bfloat16_t const *, svuint64_t);
		28881	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_f64)))
		28882	svfloat64_t svld1q_gather_index(svbool_t, float64_t const *, svuint64_t);
		28883	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_f32)))
		28884	svfloat32_t svld1q_gather_index(svbool_t, float32_t const *, svuint64_t);
		28885	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_f16)))
		28886	svfloat16_t svld1q_gather_index(svbool_t, float16_t const *, svuint64_t);
		28887	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_s32)))
		28888	svint32_t svld1q_gather_index(svbool_t, int32_t const *, svuint64_t);
		28889	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_s64)))
		28890	svint64_t svld1q_gather_index(svbool_t, int64_t const *, svuint64_t);
		28891	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64index_s16)))
		28892	svint16_t svld1q_gather_index(svbool_t, int16_t const *, svuint64_t);
		28893	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_u8)))
		28894	svuint8_t svld1q_gather_offset(svbool_t, uint8_t const *, svuint64_t);
		28895	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_u32)))
		28896	svuint32_t svld1q_gather_offset(svbool_t, uint32_t const *, svuint64_t);
		28897	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_u64)))
		28898	svuint64_t svld1q_gather_offset(svbool_t, uint64_t const *, svuint64_t);
		28899	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_u16)))
		28900	svuint16_t svld1q_gather_offset(svbool_t, uint16_t const *, svuint64_t);
		28901	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_bf16)))
		28902	svbfloat16_t svld1q_gather_offset(svbool_t, bfloat16_t const *, svuint64_t);
		28903	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_s8)))
		28904	svint8_t svld1q_gather_offset(svbool_t, int8_t const *, svuint64_t);
		28905	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_f64)))
		28906	svfloat64_t svld1q_gather_offset(svbool_t, float64_t const *, svuint64_t);
		28907	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_f32)))
		28908	svfloat32_t svld1q_gather_offset(svbool_t, float32_t const *, svuint64_t);
		28909	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_f16)))
		28910	svfloat16_t svld1q_gather_offset(svbool_t, float16_t const *, svuint64_t);
		28911	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_s32)))
		28912	svint32_t svld1q_gather_offset(svbool_t, int32_t const *, svuint64_t);
		28913	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_s64)))
		28914	svint64_t svld1q_gather_offset(svbool_t, int64_t const *, svuint64_t);
		28915	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1q_gather_u64offset_s16)))
		28916	svint16_t svld1q_gather_offset(svbool_t, int16_t const *, svuint64_t);
		28917	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1udq_u64)))
		28918	svuint64_t svld1udq(svbool_t, uint64_t const *);
		28919	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1udq_f64)))
		28920	svfloat64_t svld1udq(svbool_t, float64_t const *);
		28921	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1udq_s64)))
		28922	svint64_t svld1udq(svbool_t, int64_t const *);
		28923	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1udq_vnum_u64)))
		28924	svuint64_t svld1udq_vnum(svbool_t, uint64_t const *, int64_t);
		28925	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1udq_vnum_f64)))
		28926	svfloat64_t svld1udq_vnum(svbool_t, float64_t const *, int64_t);
		28927	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1udq_vnum_s64)))
		28928	svint64_t svld1udq_vnum(svbool_t, int64_t const *, int64_t);
		28929	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1uwq_u32)))
		28930	svuint32_t svld1uwq(svbool_t, uint32_t const *);
		28931	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1uwq_f32)))
		28932	svfloat32_t svld1uwq(svbool_t, float32_t const *);
		28933	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1uwq_s32)))
		28934	svint32_t svld1uwq(svbool_t, int32_t const *);
		28935	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1uwq_vnum_u32)))
		28936	svuint32_t svld1uwq_vnum(svbool_t, uint32_t const *, int64_t);
		28937	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1uwq_vnum_f32)))
		28938	svfloat32_t svld1uwq_vnum(svbool_t, float32_t const *, int64_t);
		28939	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1uwq_vnum_s32)))
		28940	svint32_t svld1uwq_vnum(svbool_t, int32_t const *, int64_t);
		28941	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_u8)))
		28942	svuint8x2_t svld2q(svbool_t, uint8_t const *);
		28943	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_u32)))
		28944	svuint32x2_t svld2q(svbool_t, uint32_t const *);
		28945	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_u64)))
		28946	svuint64x2_t svld2q(svbool_t, uint64_t const *);
		28947	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_u16)))
		28948	svuint16x2_t svld2q(svbool_t, uint16_t const *);
		28949	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_s8)))
		28950	svint8x2_t svld2q(svbool_t, int8_t const *);
		28951	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_f64)))
		28952	svfloat64x2_t svld2q(svbool_t, float64_t const *);
		28953	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_f32)))
		28954	svfloat32x2_t svld2q(svbool_t, float32_t const *);
		28955	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_f16)))
		28956	svfloat16x2_t svld2q(svbool_t, float16_t const *);
		28957	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_s32)))
		28958	svint32x2_t svld2q(svbool_t, int32_t const *);
		28959	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_s64)))
		28960	svint64x2_t svld2q(svbool_t, int64_t const *);
		28961	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_s16)))
		28962	svint16x2_t svld2q(svbool_t, int16_t const *);
		28963	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_bf16)))
		28964	svbfloat16x2_t svld2q(svbool_t, bfloat16_t const *);
		28965	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_u8)))
		28966	svuint8x2_t svld2q_vnum(svbool_t, uint8_t const *, int64_t);
		28967	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_u32)))
		28968	svuint32x2_t svld2q_vnum(svbool_t, uint32_t const *, int64_t);
		28969	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_u64)))
		28970	svuint64x2_t svld2q_vnum(svbool_t, uint64_t const *, int64_t);
		28971	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_u16)))
		28972	svuint16x2_t svld2q_vnum(svbool_t, uint16_t const *, int64_t);
		28973	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_s8)))
		28974	svint8x2_t svld2q_vnum(svbool_t, int8_t const *, int64_t);
		28975	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_f64)))
		28976	svfloat64x2_t svld2q_vnum(svbool_t, float64_t const *, int64_t);
		28977	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_f32)))
		28978	svfloat32x2_t svld2q_vnum(svbool_t, float32_t const *, int64_t);
		28979	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_f16)))
		28980	svfloat16x2_t svld2q_vnum(svbool_t, float16_t const *, int64_t);
		28981	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_s32)))
		28982	svint32x2_t svld2q_vnum(svbool_t, int32_t const *, int64_t);
		28983	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_s64)))
		28984	svint64x2_t svld2q_vnum(svbool_t, int64_t const *, int64_t);
		28985	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_s16)))
		28986	svint16x2_t svld2q_vnum(svbool_t, int16_t const *, int64_t);
		28987	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld2q_vnum_bf16)))
		28988	svbfloat16x2_t svld2q_vnum(svbool_t, bfloat16_t const *, int64_t);
		28989	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_u8)))
		28990	svuint8x3_t svld3q(svbool_t, uint8_t const *);
		28991	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_u32)))
		28992	svuint32x3_t svld3q(svbool_t, uint32_t const *);
		28993	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_u64)))
		28994	svuint64x3_t svld3q(svbool_t, uint64_t const *);
		28995	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_u16)))
		28996	svuint16x3_t svld3q(svbool_t, uint16_t const *);
		28997	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_s8)))
		28998	svint8x3_t svld3q(svbool_t, int8_t const *);
		28999	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_f64)))
		29000	svfloat64x3_t svld3q(svbool_t, float64_t const *);
		29001	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_f32)))
		29002	svfloat32x3_t svld3q(svbool_t, float32_t const *);
		29003	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_f16)))
		29004	svfloat16x3_t svld3q(svbool_t, float16_t const *);
		29005	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_s32)))
		29006	svint32x3_t svld3q(svbool_t, int32_t const *);
		29007	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_s64)))
		29008	svint64x3_t svld3q(svbool_t, int64_t const *);
		29009	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_s16)))
		29010	svint16x3_t svld3q(svbool_t, int16_t const *);
		29011	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_bf16)))
		29012	svbfloat16x3_t svld3q(svbool_t, bfloat16_t const *);
		29013	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_u8)))
		29014	svuint8x3_t svld3q_vnum(svbool_t, uint8_t const *, int64_t);
		29015	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_u32)))
		29016	svuint32x3_t svld3q_vnum(svbool_t, uint32_t const *, int64_t);
		29017	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_u64)))
		29018	svuint64x3_t svld3q_vnum(svbool_t, uint64_t const *, int64_t);
		29019	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_u16)))
		29020	svuint16x3_t svld3q_vnum(svbool_t, uint16_t const *, int64_t);
		29021	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_s8)))
		29022	svint8x3_t svld3q_vnum(svbool_t, int8_t const *, int64_t);
		29023	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_f64)))
		29024	svfloat64x3_t svld3q_vnum(svbool_t, float64_t const *, int64_t);
		29025	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_f32)))
		29026	svfloat32x3_t svld3q_vnum(svbool_t, float32_t const *, int64_t);
		29027	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_f16)))
		29028	svfloat16x3_t svld3q_vnum(svbool_t, float16_t const *, int64_t);
		29029	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_s32)))
		29030	svint32x3_t svld3q_vnum(svbool_t, int32_t const *, int64_t);
		29031	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_s64)))
		29032	svint64x3_t svld3q_vnum(svbool_t, int64_t const *, int64_t);
		29033	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_s16)))
		29034	svint16x3_t svld3q_vnum(svbool_t, int16_t const *, int64_t);
		29035	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld3q_vnum_bf16)))
		29036	svbfloat16x3_t svld3q_vnum(svbool_t, bfloat16_t const *, int64_t);
		29037	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_u8)))
		29038	svuint8x4_t svld4q(svbool_t, uint8_t const *);
		29039	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_u32)))
		29040	svuint32x4_t svld4q(svbool_t, uint32_t const *);
		29041	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_u64)))
		29042	svuint64x4_t svld4q(svbool_t, uint64_t const *);
		29043	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_u16)))
		29044	svuint16x4_t svld4q(svbool_t, uint16_t const *);
		29045	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_s8)))
		29046	svint8x4_t svld4q(svbool_t, int8_t const *);
		29047	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_f64)))
		29048	svfloat64x4_t svld4q(svbool_t, float64_t const *);
		29049	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_f32)))
		29050	svfloat32x4_t svld4q(svbool_t, float32_t const *);
		29051	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_f16)))
		29052	svfloat16x4_t svld4q(svbool_t, float16_t const *);
		29053	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_s32)))
		29054	svint32x4_t svld4q(svbool_t, int32_t const *);
		29055	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_s64)))
		29056	svint64x4_t svld4q(svbool_t, int64_t const *);
		29057	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_s16)))
		29058	svint16x4_t svld4q(svbool_t, int16_t const *);
		29059	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_bf16)))
		29060	svbfloat16x4_t svld4q(svbool_t, bfloat16_t const *);
		29061	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_u8)))
		29062	svuint8x4_t svld4q_vnum(svbool_t, uint8_t const *, int64_t);
		29063	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_u32)))
		29064	svuint32x4_t svld4q_vnum(svbool_t, uint32_t const *, int64_t);
		29065	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_u64)))
		29066	svuint64x4_t svld4q_vnum(svbool_t, uint64_t const *, int64_t);
		29067	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_u16)))
		29068	svuint16x4_t svld4q_vnum(svbool_t, uint16_t const *, int64_t);
		29069	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_s8)))
		29070	svint8x4_t svld4q_vnum(svbool_t, int8_t const *, int64_t);
		29071	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_f64)))
		29072	svfloat64x4_t svld4q_vnum(svbool_t, float64_t const *, int64_t);
		29073	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_f32)))
		29074	svfloat32x4_t svld4q_vnum(svbool_t, float32_t const *, int64_t);
		29075	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_f16)))
		29076	svfloat16x4_t svld4q_vnum(svbool_t, float16_t const *, int64_t);
		29077	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_s32)))
		29078	svint32x4_t svld4q_vnum(svbool_t, int32_t const *, int64_t);
		29079	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_s64)))
		29080	svint64x4_t svld4q_vnum(svbool_t, int64_t const *, int64_t);
		29081	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_s16)))
		29082	svint16x4_t svld4q_vnum(svbool_t, int16_t const *, int64_t);
		29083	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld4q_vnum_bf16)))
		29084	svbfloat16x4_t svld4q_vnum(svbool_t, bfloat16_t const *, int64_t);
		29085	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnmqv_f64)))
		29086	float64x2_t svmaxnmqv(svbool_t, svfloat64_t);
		29087	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnmqv_f32)))
		29088	float32x4_t svmaxnmqv(svbool_t, svfloat32_t);
		29089	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxnmqv_f16)))
		29090	float16x8_t svmaxnmqv(svbool_t, svfloat16_t);
		29091	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_f64)))
		29092	float64x2_t svmaxqv(svbool_t, svfloat64_t);
		29093	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_f32)))
		29094	float32x4_t svmaxqv(svbool_t, svfloat32_t);
		29095	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_f16)))
		29096	float16x8_t svmaxqv(svbool_t, svfloat16_t);
		29097	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_s8)))
		29098	int8x16_t svmaxqv(svbool_t, svint8_t);
		29099	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_s32)))
		29100	int32x4_t svmaxqv(svbool_t, svint32_t);
		29101	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_s64)))
		29102	int64x2_t svmaxqv(svbool_t, svint64_t);
		29103	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_s16)))
		29104	int16x8_t svmaxqv(svbool_t, svint16_t);
		29105	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_u8)))
		29106	uint8x16_t svmaxqv(svbool_t, svuint8_t);
		29107	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_u32)))
		29108	uint32x4_t svmaxqv(svbool_t, svuint32_t);
		29109	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_u64)))
		29110	uint64x2_t svmaxqv(svbool_t, svuint64_t);
		29111	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svmaxqv_u16)))
		29112	uint16x8_t svmaxqv(svbool_t, svuint16_t);
		29113	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnmqv_f64)))
		29114	float64x2_t svminnmqv(svbool_t, svfloat64_t);
		29115	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnmqv_f32)))
		29116	float32x4_t svminnmqv(svbool_t, svfloat32_t);
		29117	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminnmqv_f16)))
		29118	float16x8_t svminnmqv(svbool_t, svfloat16_t);
		29119	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_f64)))
		29120	float64x2_t svminqv(svbool_t, svfloat64_t);
		29121	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_f32)))
		29122	float32x4_t svminqv(svbool_t, svfloat32_t);
		29123	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_f16)))
		29124	float16x8_t svminqv(svbool_t, svfloat16_t);
		29125	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_s8)))
		29126	int8x16_t svminqv(svbool_t, svint8_t);
		29127	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_s32)))
		29128	int32x4_t svminqv(svbool_t, svint32_t);
		29129	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_s64)))
		29130	int64x2_t svminqv(svbool_t, svint64_t);
		29131	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_s16)))
		29132	int16x8_t svminqv(svbool_t, svint16_t);
		29133	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_u8)))
		29134	uint8x16_t svminqv(svbool_t, svuint8_t);
		29135	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_u32)))
		29136	uint32x4_t svminqv(svbool_t, svuint32_t);
		29137	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_u64)))
		29138	uint64x2_t svminqv(svbool_t, svuint64_t);
		29139	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svminqv_u16)))
		29140	uint16x8_t svminqv(svbool_t, svuint16_t);
		29141	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svorqv_u8)))
		29142	uint8x16_t svorqv(svbool_t, svuint8_t);
		29143	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svorqv_u32)))
		29144	uint32x4_t svorqv(svbool_t, svuint32_t);
		29145	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svorqv_u64)))
		29146	uint64x2_t svorqv(svbool_t, svuint64_t);
		29147	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svorqv_u16)))
		29148	uint16x8_t svorqv(svbool_t, svuint16_t);
		29149	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svorqv_s8)))
		29150	int8x16_t svorqv(svbool_t, svint8_t);
		29151	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svorqv_s32)))
		29152	int32x4_t svorqv(svbool_t, svint32_t);
		29153	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svorqv_s64)))
		29154	int64x2_t svorqv(svbool_t, svint64_t);
		29155	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svorqv_s16)))
		29156	int16x8_t svorqv(svbool_t, svint16_t);
		29157	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_u8)))
		29158	svbool_t svpmov(svuint8_t);
		29159	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_s8)))
		29160	svbool_t svpmov(svint8_t);
		29161	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_u64)))
		29162	svbool_t svpmov(svuint64_t);
		29163	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_s64)))
		29164	svbool_t svpmov(svint64_t);
		29165	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_u16)))
		29166	svbool_t svpmov(svuint16_t);
		29167	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_s16)))
		29168	svbool_t svpmov(svint16_t);
		29169	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_u32)))
		29170	svbool_t svpmov(svuint32_t);
		29171	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_s32)))
		29172	svbool_t svpmov(svint32_t);
		29173	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_u8)))
		29174	svbool_t svpmov_lane(svuint8_t, uint64_t);
		29175	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_s8)))
		29176	svbool_t svpmov_lane(svint8_t, uint64_t);
		29177	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_u64)))
		29178	svbool_t svpmov_lane(svuint64_t, uint64_t);
		29179	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_s64)))
		29180	svbool_t svpmov_lane(svint64_t, uint64_t);
		29181	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_u16)))
		29182	svbool_t svpmov_lane(svuint16_t, uint64_t);
		29183	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_s16)))
		29184	svbool_t svpmov_lane(svint16_t, uint64_t);
		29185	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_u32)))
		29186	svbool_t svpmov_lane(svuint32_t, uint64_t);
		29187	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_s32)))
		29188	svbool_t svpmov_lane(svint32_t, uint64_t);
		29189	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_u64_m)))
		29190	svuint64_t svpmov_lane_m(svuint64_t, svbool_t, uint64_t);
		29191	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_s64_m)))
		29192	svint64_t svpmov_lane_m(svint64_t, svbool_t, uint64_t);
		29193	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_u16_m)))
		29194	svuint16_t svpmov_lane_m(svuint16_t, svbool_t, uint64_t);
		29195	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_s16_m)))
		29196	svint16_t svpmov_lane_m(svint16_t, svbool_t, uint64_t);
		29197	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_u32_m)))
		29198	svuint32_t svpmov_lane_m(svuint32_t, svbool_t, uint64_t);
		29199	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpmov_lane_s32_m)))
		29200	svint32_t svpmov_lane_m(svint32_t, svbool_t, uint64_t);
		29201	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1dq_u64)))
		29202	void svst1dq(svbool_t, uint64_t const *, svuint64_t);
		29203	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1dq_f64)))
		29204	void svst1dq(svbool_t, float64_t const *, svfloat64_t);
		29205	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1dq_s64)))
		29206	void svst1dq(svbool_t, int64_t const *, svint64_t);
		29207	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1dq_vnum_u64)))
		29208	void svst1dq_vnum(svbool_t, uint64_t const *, int64_t, svuint64_t);
		29209	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1dq_vnum_f64)))
		29210	void svst1dq_vnum(svbool_t, float64_t const *, int64_t, svfloat64_t);
		29211	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1dq_vnum_s64)))
		29212	void svst1dq_vnum(svbool_t, int64_t const *, int64_t, svint64_t);
		29213	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_u8)))
		29214	void svst1q_scatter(svbool_t, svuint64_t, svuint8_t);
		29215	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_u32)))
		29216	void svst1q_scatter(svbool_t, svuint64_t, svuint32_t);
		29217	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_u64)))
		29218	void svst1q_scatter(svbool_t, svuint64_t, svuint64_t);
		29219	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_u16)))
		29220	void svst1q_scatter(svbool_t, svuint64_t, svuint16_t);
		29221	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_bf16)))
		29222	void svst1q_scatter(svbool_t, svuint64_t, svbfloat16_t);
		29223	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_s8)))
		29224	void svst1q_scatter(svbool_t, svuint64_t, svint8_t);
		29225	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_f64)))
		29226	void svst1q_scatter(svbool_t, svuint64_t, svfloat64_t);
		29227	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_f32)))
		29228	void svst1q_scatter(svbool_t, svuint64_t, svfloat32_t);
		29229	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_f16)))
		29230	void svst1q_scatter(svbool_t, svuint64_t, svfloat16_t);
		29231	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_s32)))
		29232	void svst1q_scatter(svbool_t, svuint64_t, svint32_t);
		29233	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_s64)))
		29234	void svst1q_scatter(svbool_t, svuint64_t, svint64_t);
		29235	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_s16)))
		29236	void svst1q_scatter(svbool_t, svuint64_t, svint16_t);
		29237	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_u32)))
		29238	void svst1q_scatter_index(svbool_t, svuint64_t, int64_t, svuint32_t);
		29239	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_u64)))
		29240	void svst1q_scatter_index(svbool_t, svuint64_t, int64_t, svuint64_t);
		29241	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_u16)))
		29242	void svst1q_scatter_index(svbool_t, svuint64_t, int64_t, svuint16_t);
		29243	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_bf16)))
		29244	void svst1q_scatter_index(svbool_t, svuint64_t, int64_t, svbfloat16_t);
		29245	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_f64)))
		29246	void svst1q_scatter_index(svbool_t, svuint64_t, int64_t, svfloat64_t);
		29247	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_f32)))
		29248	void svst1q_scatter_index(svbool_t, svuint64_t, int64_t, svfloat32_t);
		29249	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_f16)))
		29250	void svst1q_scatter_index(svbool_t, svuint64_t, int64_t, svfloat16_t);
		29251	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_s32)))
		29252	void svst1q_scatter_index(svbool_t, svuint64_t, int64_t, svint32_t);
		29253	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_s64)))
		29254	void svst1q_scatter_index(svbool_t, svuint64_t, int64_t, svint64_t);
		29255	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_index_s16)))
		29256	void svst1q_scatter_index(svbool_t, svuint64_t, int64_t, svint16_t);
		29257	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_u8)))
		29258	void svst1q_scatter_offset(svbool_t, svuint64_t, int64_t, svuint8_t);
		29259	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_u32)))
		29260	void svst1q_scatter_offset(svbool_t, svuint64_t, int64_t, svuint32_t);
		29261	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_u64)))
		29262	void svst1q_scatter_offset(svbool_t, svuint64_t, int64_t, svuint64_t);
		29263	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_u16)))
		29264	void svst1q_scatter_offset(svbool_t, svuint64_t, int64_t, svuint16_t);
		29265	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_bf16)))
		29266	void svst1q_scatter_offset(svbool_t, svuint64_t, int64_t, svbfloat16_t);
		29267	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_s8)))
		29268	void svst1q_scatter_offset(svbool_t, svuint64_t, int64_t, svint8_t);
		29269	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_f64)))
		29270	void svst1q_scatter_offset(svbool_t, svuint64_t, int64_t, svfloat64_t);
		29271	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_f32)))
		29272	void svst1q_scatter_offset(svbool_t, svuint64_t, int64_t, svfloat32_t);
		29273	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_f16)))
		29274	void svst1q_scatter_offset(svbool_t, svuint64_t, int64_t, svfloat16_t);
		29275	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_s32)))
		29276	void svst1q_scatter_offset(svbool_t, svuint64_t, int64_t, svint32_t);
		29277	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_s64)))
		29278	void svst1q_scatter_offset(svbool_t, svuint64_t, int64_t, svint64_t);
		29279	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64base_offset_s16)))
		29280	void svst1q_scatter_offset(svbool_t, svuint64_t, int64_t, svint16_t);
		29281	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_u32)))
		29282	void svst1q_scatter_index(svbool_t, uint32_t *, svuint64_t, svuint32_t);
		29283	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_u64)))
		29284	void svst1q_scatter_index(svbool_t, uint64_t *, svuint64_t, svuint64_t);
		29285	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_u16)))
		29286	void svst1q_scatter_index(svbool_t, uint16_t *, svuint64_t, svuint16_t);
		29287	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_bf16)))
		29288	void svst1q_scatter_index(svbool_t, bfloat16_t *, svuint64_t, svbfloat16_t);
		29289	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_f64)))
		29290	void svst1q_scatter_index(svbool_t, float64_t *, svuint64_t, svfloat64_t);
		29291	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_f32)))
		29292	void svst1q_scatter_index(svbool_t, float32_t *, svuint64_t, svfloat32_t);
		29293	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_f16)))
		29294	void svst1q_scatter_index(svbool_t, float16_t *, svuint64_t, svfloat16_t);
		29295	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_s32)))
		29296	void svst1q_scatter_index(svbool_t, int32_t *, svuint64_t, svint32_t);
		29297	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_s64)))
		29298	void svst1q_scatter_index(svbool_t, int64_t *, svuint64_t, svint64_t);
		29299	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64index_s16)))
		29300	void svst1q_scatter_index(svbool_t, int16_t *, svuint64_t, svint16_t);
		29301	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_u8)))
		29302	void svst1q_scatter_offset(svbool_t, uint8_t *, svuint64_t, svuint8_t);
		29303	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_u32)))
		29304	void svst1q_scatter_offset(svbool_t, uint32_t *, svuint64_t, svuint32_t);
		29305	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_u64)))
		29306	void svst1q_scatter_offset(svbool_t, uint64_t *, svuint64_t, svuint64_t);
		29307	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_u16)))
		29308	void svst1q_scatter_offset(svbool_t, uint16_t *, svuint64_t, svuint16_t);
		29309	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_bf16)))
		29310	void svst1q_scatter_offset(svbool_t, bfloat16_t *, svuint64_t, svbfloat16_t);
		29311	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_s8)))
		29312	void svst1q_scatter_offset(svbool_t, int8_t *, svuint64_t, svint8_t);
		29313	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_f64)))
		29314	void svst1q_scatter_offset(svbool_t, float64_t *, svuint64_t, svfloat64_t);
		29315	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_f32)))
		29316	void svst1q_scatter_offset(svbool_t, float32_t *, svuint64_t, svfloat32_t);
		29317	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_f16)))
		29318	void svst1q_scatter_offset(svbool_t, float16_t *, svuint64_t, svfloat16_t);
		29319	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_s32)))
		29320	void svst1q_scatter_offset(svbool_t, int32_t *, svuint64_t, svint32_t);
		29321	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_s64)))
		29322	void svst1q_scatter_offset(svbool_t, int64_t *, svuint64_t, svint64_t);
		29323	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1q_scatter_u64offset_s16)))
		29324	void svst1q_scatter_offset(svbool_t, int16_t *, svuint64_t, svint16_t);
		29325	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1wq_u32)))
		29326	void svst1wq(svbool_t, uint32_t const *, svuint32_t);
		29327	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1wq_f32)))
		29328	void svst1wq(svbool_t, float32_t const *, svfloat32_t);
		29329	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1wq_s32)))
		29330	void svst1wq(svbool_t, int32_t const *, svint32_t);
		29331	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1wq_vnum_u32)))
		29332	void svst1wq_vnum(svbool_t, uint32_t const *, int64_t, svuint32_t);
		29333	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1wq_vnum_f32)))
		29334	void svst1wq_vnum(svbool_t, float32_t const *, int64_t, svfloat32_t);
		29335	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1wq_vnum_s32)))
		29336	void svst1wq_vnum(svbool_t, int32_t const *, int64_t, svint32_t);
		29337	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_u8)))
		29338	void svst2q(svbool_t, uint8_t const *, svuint8x2_t);
		29339	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_u32)))
		29340	void svst2q(svbool_t, uint32_t const *, svuint32x2_t);
		29341	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_u64)))
		29342	void svst2q(svbool_t, uint64_t const *, svuint64x2_t);
		29343	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_u16)))
		29344	void svst2q(svbool_t, uint16_t const *, svuint16x2_t);
		29345	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_s8)))
		29346	void svst2q(svbool_t, int8_t const *, svint8x2_t);
		29347	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_f64)))
		29348	void svst2q(svbool_t, float64_t const *, svfloat64x2_t);
		29349	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_f32)))
		29350	void svst2q(svbool_t, float32_t const *, svfloat32x2_t);
		29351	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_f16)))
		29352	void svst2q(svbool_t, float16_t const *, svfloat16x2_t);
		29353	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_s32)))
		29354	void svst2q(svbool_t, int32_t const *, svint32x2_t);
		29355	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_s64)))
		29356	void svst2q(svbool_t, int64_t const *, svint64x2_t);
		29357	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_s16)))
		29358	void svst2q(svbool_t, int16_t const *, svint16x2_t);
		29359	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_bf16)))
		29360	void svst2q(svbool_t, bfloat16_t const *, svbfloat16x2_t);
		29361	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_u8)))
		29362	void svst2q_vnum(svbool_t, uint8_t const *, int64_t, svuint8x2_t);
		29363	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_u32)))
		29364	void svst2q_vnum(svbool_t, uint32_t const *, int64_t, svuint32x2_t);
		29365	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_u64)))
		29366	void svst2q_vnum(svbool_t, uint64_t const *, int64_t, svuint64x2_t);
		29367	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_u16)))
		29368	void svst2q_vnum(svbool_t, uint16_t const *, int64_t, svuint16x2_t);
		29369	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_s8)))
		29370	void svst2q_vnum(svbool_t, int8_t const *, int64_t, svint8x2_t);
		29371	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_f64)))
		29372	void svst2q_vnum(svbool_t, float64_t const *, int64_t, svfloat64x2_t);
		29373	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_f32)))
		29374	void svst2q_vnum(svbool_t, float32_t const *, int64_t, svfloat32x2_t);
		29375	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_f16)))
		29376	void svst2q_vnum(svbool_t, float16_t const *, int64_t, svfloat16x2_t);
		29377	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_s32)))
		29378	void svst2q_vnum(svbool_t, int32_t const *, int64_t, svint32x2_t);
		29379	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_s64)))
		29380	void svst2q_vnum(svbool_t, int64_t const *, int64_t, svint64x2_t);
		29381	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_s16)))
		29382	void svst2q_vnum(svbool_t, int16_t const *, int64_t, svint16x2_t);
		29383	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst2q_vnum_bf16)))
		29384	void svst2q_vnum(svbool_t, bfloat16_t const *, int64_t, svbfloat16x2_t);
		29385	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_u8)))
		29386	void svst3q(svbool_t, uint8_t const *, svuint8x3_t);
		29387	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_u32)))
		29388	void svst3q(svbool_t, uint32_t const *, svuint32x3_t);
		29389	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_u64)))
		29390	void svst3q(svbool_t, uint64_t const *, svuint64x3_t);
		29391	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_u16)))
		29392	void svst3q(svbool_t, uint16_t const *, svuint16x3_t);
		29393	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_s8)))
		29394	void svst3q(svbool_t, int8_t const *, svint8x3_t);
		29395	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_f64)))
		29396	void svst3q(svbool_t, float64_t const *, svfloat64x3_t);
		29397	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_f32)))
		29398	void svst3q(svbool_t, float32_t const *, svfloat32x3_t);
		29399	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_f16)))
		29400	void svst3q(svbool_t, float16_t const *, svfloat16x3_t);
		29401	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_s32)))
		29402	void svst3q(svbool_t, int32_t const *, svint32x3_t);
		29403	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_s64)))
		29404	void svst3q(svbool_t, int64_t const *, svint64x3_t);
		29405	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_s16)))
		29406	void svst3q(svbool_t, int16_t const *, svint16x3_t);
		29407	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_bf16)))
		29408	void svst3q(svbool_t, bfloat16_t const *, svbfloat16x3_t);
		29409	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_u8)))
		29410	void svst3q_vnum(svbool_t, uint8_t const *, int64_t, svuint8x3_t);
		29411	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_u32)))
		29412	void svst3q_vnum(svbool_t, uint32_t const *, int64_t, svuint32x3_t);
		29413	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_u64)))
		29414	void svst3q_vnum(svbool_t, uint64_t const *, int64_t, svuint64x3_t);
		29415	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_u16)))
		29416	void svst3q_vnum(svbool_t, uint16_t const *, int64_t, svuint16x3_t);
		29417	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_s8)))
		29418	void svst3q_vnum(svbool_t, int8_t const *, int64_t, svint8x3_t);
		29419	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_f64)))
		29420	void svst3q_vnum(svbool_t, float64_t const *, int64_t, svfloat64x3_t);
		29421	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_f32)))
		29422	void svst3q_vnum(svbool_t, float32_t const *, int64_t, svfloat32x3_t);
		29423	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_f16)))
		29424	void svst3q_vnum(svbool_t, float16_t const *, int64_t, svfloat16x3_t);
		29425	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_s32)))
		29426	void svst3q_vnum(svbool_t, int32_t const *, int64_t, svint32x3_t);
		29427	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_s64)))
		29428	void svst3q_vnum(svbool_t, int64_t const *, int64_t, svint64x3_t);
		29429	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_s16)))
		29430	void svst3q_vnum(svbool_t, int16_t const *, int64_t, svint16x3_t);
		29431	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst3q_vnum_bf16)))
		29432	void svst3q_vnum(svbool_t, bfloat16_t const *, int64_t, svbfloat16x3_t);
		29433	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_u8)))
		29434	void svst4q(svbool_t, uint8_t const *, svuint8x4_t);
		29435	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_u32)))
		29436	void svst4q(svbool_t, uint32_t const *, svuint32x4_t);
		29437	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_u64)))
		29438	void svst4q(svbool_t, uint64_t const *, svuint64x4_t);
		29439	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_u16)))
		29440	void svst4q(svbool_t, uint16_t const *, svuint16x4_t);
		29441	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_s8)))
		29442	void svst4q(svbool_t, int8_t const *, svint8x4_t);
		29443	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_f64)))
		29444	void svst4q(svbool_t, float64_t const *, svfloat64x4_t);
		29445	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_f32)))
		29446	void svst4q(svbool_t, float32_t const *, svfloat32x4_t);
		29447	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_f16)))
		29448	void svst4q(svbool_t, float16_t const *, svfloat16x4_t);
		29449	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_s32)))
		29450	void svst4q(svbool_t, int32_t const *, svint32x4_t);
		29451	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_s64)))
		29452	void svst4q(svbool_t, int64_t const *, svint64x4_t);
		29453	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_s16)))
		29454	void svst4q(svbool_t, int16_t const *, svint16x4_t);
		29455	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_bf16)))
		29456	void svst4q(svbool_t, bfloat16_t const *, svbfloat16x4_t);
		29457	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_u8)))
		29458	void svst4q_vnum(svbool_t, uint8_t const *, int64_t, svuint8x4_t);
		29459	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_u32)))
		29460	void svst4q_vnum(svbool_t, uint32_t const *, int64_t, svuint32x4_t);
		29461	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_u64)))
		29462	void svst4q_vnum(svbool_t, uint64_t const *, int64_t, svuint64x4_t);
		29463	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_u16)))
		29464	void svst4q_vnum(svbool_t, uint16_t const *, int64_t, svuint16x4_t);
		29465	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_s8)))
		29466	void svst4q_vnum(svbool_t, int8_t const *, int64_t, svint8x4_t);
		29467	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_f64)))
		29468	void svst4q_vnum(svbool_t, float64_t const *, int64_t, svfloat64x4_t);
		29469	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_f32)))
		29470	void svst4q_vnum(svbool_t, float32_t const *, int64_t, svfloat32x4_t);
		29471	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_f16)))
		29472	void svst4q_vnum(svbool_t, float16_t const *, int64_t, svfloat16x4_t);
		29473	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_s32)))
		29474	void svst4q_vnum(svbool_t, int32_t const *, int64_t, svint32x4_t);
		29475	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_s64)))
		29476	void svst4q_vnum(svbool_t, int64_t const *, int64_t, svint64x4_t);
		29477	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_s16)))
		29478	void svst4q_vnum(svbool_t, int16_t const *, int64_t, svint16x4_t);
		29479	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst4q_vnum_bf16)))
		29480	void svst4q_vnum(svbool_t, bfloat16_t const *, int64_t, svbfloat16x4_t);
		29481	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_u8)))
		29482	svuint8_t svtblq(svuint8_t, svuint8_t);
		29483	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_u32)))
		29484	svuint32_t svtblq(svuint32_t, svuint32_t);
		29485	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_u64)))
		29486	svuint64_t svtblq(svuint64_t, svuint64_t);
		29487	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_u16)))
		29488	svuint16_t svtblq(svuint16_t, svuint16_t);
		29489	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_bf16)))
		29490	svbfloat16_t svtblq(svbfloat16_t, svuint16_t);
		29491	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_s8)))
		29492	svint8_t svtblq(svint8_t, svuint8_t);
		29493	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_f64)))
		29494	svfloat64_t svtblq(svfloat64_t, svuint64_t);
		29495	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_f32)))
		29496	svfloat32_t svtblq(svfloat32_t, svuint32_t);
		29497	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_f16)))
		29498	svfloat16_t svtblq(svfloat16_t, svuint16_t);
		29499	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_s32)))
		29500	svint32_t svtblq(svint32_t, svuint32_t);
		29501	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_s64)))
		29502	svint64_t svtblq(svint64_t, svuint64_t);
		29503	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtblq_s16)))
		29504	svint16_t svtblq(svint16_t, svuint16_t);
		29505	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_u8)))
		29506	svuint8_t svtbxq(svuint8_t, svuint8_t, svuint8_t);
		29507	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_u32)))
		29508	svuint32_t svtbxq(svuint32_t, svuint32_t, svuint32_t);
		29509	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_u64)))
		29510	svuint64_t svtbxq(svuint64_t, svuint64_t, svuint64_t);
		29511	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_u16)))
		29512	svuint16_t svtbxq(svuint16_t, svuint16_t, svuint16_t);
		29513	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_bf16)))
		29514	svbfloat16_t svtbxq(svbfloat16_t, svbfloat16_t, svuint16_t);
		29515	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_s8)))
		29516	svint8_t svtbxq(svint8_t, svint8_t, svuint8_t);
		29517	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_f64)))
		29518	svfloat64_t svtbxq(svfloat64_t, svfloat64_t, svuint64_t);
		29519	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_f32)))
		29520	svfloat32_t svtbxq(svfloat32_t, svfloat32_t, svuint32_t);
		29521	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_f16)))
		29522	svfloat16_t svtbxq(svfloat16_t, svfloat16_t, svuint16_t);
		29523	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_s32)))
		29524	svint32_t svtbxq(svint32_t, svint32_t, svuint32_t);
		29525	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_s64)))
		29526	svint64_t svtbxq(svint64_t, svint64_t, svuint64_t);
		29527	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svtbxq_s16)))
		29528	svint16_t svtbxq(svint16_t, svint16_t, svuint16_t);
		29529	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_u8)))
		29530	svuint8_t svuzpq1(svuint8_t, svuint8_t);
		29531	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_u32)))
		29532	svuint32_t svuzpq1(svuint32_t, svuint32_t);
		29533	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_u64)))
		29534	svuint64_t svuzpq1(svuint64_t, svuint64_t);
		29535	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_u16)))
		29536	svuint16_t svuzpq1(svuint16_t, svuint16_t);
		29537	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_bf16)))
		29538	svbfloat16_t svuzpq1(svbfloat16_t, svbfloat16_t);
		29539	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_s8)))
		29540	svint8_t svuzpq1(svint8_t, svint8_t);
		29541	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_f64)))
		29542	svfloat64_t svuzpq1(svfloat64_t, svfloat64_t);
		29543	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_f32)))
		29544	svfloat32_t svuzpq1(svfloat32_t, svfloat32_t);
		29545	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_f16)))
		29546	svfloat16_t svuzpq1(svfloat16_t, svfloat16_t);
		29547	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_s32)))
		29548	svint32_t svuzpq1(svint32_t, svint32_t);
		29549	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_s64)))
		29550	svint64_t svuzpq1(svint64_t, svint64_t);
		29551	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq1_s16)))
		29552	svint16_t svuzpq1(svint16_t, svint16_t);
		29553	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_u8)))
		29554	svuint8_t svuzpq2(svuint8_t, svuint8_t);
		29555	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_u32)))
		29556	svuint32_t svuzpq2(svuint32_t, svuint32_t);
		29557	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_u64)))
		29558	svuint64_t svuzpq2(svuint64_t, svuint64_t);
		29559	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_u16)))
		29560	svuint16_t svuzpq2(svuint16_t, svuint16_t);
		29561	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_bf16)))
		29562	svbfloat16_t svuzpq2(svbfloat16_t, svbfloat16_t);
		29563	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_s8)))
		29564	svint8_t svuzpq2(svint8_t, svint8_t);
		29565	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_f64)))
		29566	svfloat64_t svuzpq2(svfloat64_t, svfloat64_t);
		29567	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_f32)))
		29568	svfloat32_t svuzpq2(svfloat32_t, svfloat32_t);
		29569	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_f16)))
		29570	svfloat16_t svuzpq2(svfloat16_t, svfloat16_t);
		29571	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_s32)))
		29572	svint32_t svuzpq2(svint32_t, svint32_t);
		29573	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_s64)))
		29574	svint64_t svuzpq2(svint64_t, svint64_t);
		29575	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svuzpq2_s16)))
		29576	svint16_t svuzpq2(svint16_t, svint16_t);
		29577	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_u8)))
		29578	svuint8_t svzipq1(svuint8_t, svuint8_t);
		29579	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_u32)))
		29580	svuint32_t svzipq1(svuint32_t, svuint32_t);
		29581	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_u64)))
		29582	svuint64_t svzipq1(svuint64_t, svuint64_t);
		29583	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_u16)))
		29584	svuint16_t svzipq1(svuint16_t, svuint16_t);
		29585	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_bf16)))
		29586	svbfloat16_t svzipq1(svbfloat16_t, svbfloat16_t);
		29587	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_s8)))
		29588	svint8_t svzipq1(svint8_t, svint8_t);
		29589	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_f64)))
		29590	svfloat64_t svzipq1(svfloat64_t, svfloat64_t);
		29591	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_f32)))
		29592	svfloat32_t svzipq1(svfloat32_t, svfloat32_t);
		29593	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_f16)))
		29594	svfloat16_t svzipq1(svfloat16_t, svfloat16_t);
		29595	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_s32)))
		29596	svint32_t svzipq1(svint32_t, svint32_t);
		29597	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_s64)))
		29598	svint64_t svzipq1(svint64_t, svint64_t);
		29599	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq1_s16)))
		29600	svint16_t svzipq1(svint16_t, svint16_t);
		29601	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_u8)))
		29602	svuint8_t svzipq2(svuint8_t, svuint8_t);
		29603	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_u32)))
		29604	svuint32_t svzipq2(svuint32_t, svuint32_t);
		29605	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_u64)))
		29606	svuint64_t svzipq2(svuint64_t, svuint64_t);
		29607	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_u16)))
		29608	svuint16_t svzipq2(svuint16_t, svuint16_t);
		29609	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_bf16)))
		29610	svbfloat16_t svzipq2(svbfloat16_t, svbfloat16_t);
		29611	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_s8)))
		29612	svint8_t svzipq2(svint8_t, svint8_t);
		29613	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_f64)))
		29614	svfloat64_t svzipq2(svfloat64_t, svfloat64_t);
		29615	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_f32)))
		29616	svfloat32_t svzipq2(svfloat32_t, svfloat32_t);
		29617	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_f16)))
		29618	svfloat16_t svzipq2(svfloat16_t, svfloat16_t);
		29619	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_s32)))
		29620	svint32_t svzipq2(svint32_t, svint32_t);
		29621	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_s64)))
		29622	svint64_t svzipq2(svint64_t, svint64_t);
		29623	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svzipq2_s16)))
		29624	svint16_t svzipq2(svint16_t, svint16_t);
		29625	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpsel_lane_b16)))
		29626	svbool_t svpsel_lane_b16(svbool_t, svbool_t, uint32_t);
		29627	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpsel_lane_b32)))
		29628	svbool_t svpsel_lane_b32(svbool_t, svbool_t, uint32_t);
		29629	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpsel_lane_b64)))
		29630	svbool_t svpsel_lane_b64(svbool_t, svbool_t, uint32_t);
		29631	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpsel_lane_b8)))
		29632	svbool_t svpsel_lane_b8(svbool_t, svbool_t, uint32_t);
		29633	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svbfmlslb_f32)))
		29634	svfloat32_t svbfmlslb_f32(svfloat32_t, svbfloat16_t, svbfloat16_t);
		29635	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svbfmlslb_lane_f32)))
		29636	svfloat32_t svbfmlslb_lane_f32(svfloat32_t, svbfloat16_t, svbfloat16_t, uint64_t);
		29637	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svbfmlslt_f32)))
		29638	svfloat32_t svbfmlslt_f32(svfloat32_t, svbfloat16_t, svbfloat16_t);
		29639	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svbfmlslt_lane_f32)))
		29640	svfloat32_t svbfmlslt_lane_f32(svfloat32_t, svbfloat16_t, svbfloat16_t, uint64_t);
23880	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_f64)))	29641	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_f64)))
23881	svfloat64_t svclamp_f64(svfloat64_t, svfloat64_t, svfloat64_t);	29642	svfloat64_t svclamp_f64(svfloat64_t, svfloat64_t, svfloat64_t);
23882	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_f32)))	29643	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_f32)))
...	@@ -23899,6 +29660,252 @@ __ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_u64)))	...	@@ -23899,6 +29660,252 @@ __ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_u64)))
23899	svuint64_t svclamp_u64(svuint64_t, svuint64_t, svuint64_t);	29660	svuint64_t svclamp_u64(svuint64_t, svuint64_t, svuint64_t);
23900	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_u16)))	29661	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_u16)))
23901	svuint16_t svclamp_u16(svuint16_t, svuint16_t, svuint16_t);	29662	svuint16_t svclamp_u16(svuint16_t, svuint16_t, svuint16_t);
		29663	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcntp_c8)))
		29664	uint64_t svcntp_c8(svcount_t, uint64_t);
		29665	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcntp_c32)))
		29666	uint64_t svcntp_c32(svcount_t, uint64_t);
		29667	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcntp_c64)))
		29668	uint64_t svcntp_c64(svcount_t, uint64_t);
		29669	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcntp_c16)))
		29670	uint64_t svcntp_c16(svcount_t, uint64_t);
		29671	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcreate2_b)))
		29672	svboolx2_t svcreate2_b(svbool_t, svbool_t);
		29673	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcreate4_b)))
		29674	svboolx4_t svcreate4_b(svbool_t, svbool_t, svbool_t, svbool_t);
		29675	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svdot_f32_f16)))
		29676	svfloat32_t svdot_f32_f16(svfloat32_t, svfloat16_t, svfloat16_t);
		29677	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svdot_s32_s16)))
		29678	svint32_t svdot_s32_s16(svint32_t, svint16_t, svint16_t);
		29679	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svdot_u32_u16)))
		29680	svuint32_t svdot_u32_u16(svuint32_t, svuint16_t, svuint16_t);
		29681	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svdot_lane_f32_f16)))
		29682	svfloat32_t svdot_lane_f32_f16(svfloat32_t, svfloat16_t, svfloat16_t, uint64_t);
		29683	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svdot_lane_s32_s16)))
		29684	svint32_t svdot_lane_s32_s16(svint32_t, svint16_t, svint16_t, uint64_t);
		29685	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svdot_lane_u32_u16)))
		29686	svuint32_t svdot_lane_u32_u16(svuint32_t, svuint16_t, svuint16_t, uint64_t);
		29687	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svget2_b)))
		29688	svbool_t svget2_b(svboolx2_t, uint64_t);
		29689	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svget4_b)))
		29690	svbool_t svget4_b(svboolx4_t, uint64_t);
		29691	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_u8_x2)))
		29692	svuint8x2_t svld1_u8_x2(svcount_t, uint8_t const *);
		29693	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_s8_x2)))
		29694	svint8x2_t svld1_s8_x2(svcount_t, int8_t const *);
		29695	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_u64_x2)))
		29696	svuint64x2_t svld1_u64_x2(svcount_t, uint64_t const *);
		29697	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_f64_x2)))
		29698	svfloat64x2_t svld1_f64_x2(svcount_t, float64_t const *);
		29699	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_s64_x2)))
		29700	svint64x2_t svld1_s64_x2(svcount_t, int64_t const *);
		29701	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_u16_x2)))
		29702	svuint16x2_t svld1_u16_x2(svcount_t, uint16_t const *);
		29703	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_bf16_x2)))
		29704	svbfloat16x2_t svld1_bf16_x2(svcount_t, bfloat16_t const *);
		29705	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_f16_x2)))
		29706	svfloat16x2_t svld1_f16_x2(svcount_t, float16_t const *);
		29707	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_s16_x2)))
		29708	svint16x2_t svld1_s16_x2(svcount_t, int16_t const *);
		29709	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_u32_x2)))
		29710	svuint32x2_t svld1_u32_x2(svcount_t, uint32_t const *);
		29711	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_f32_x2)))
		29712	svfloat32x2_t svld1_f32_x2(svcount_t, float32_t const *);
		29713	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_s32_x2)))
		29714	svint32x2_t svld1_s32_x2(svcount_t, int32_t const *);
		29715	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_u8_x4)))
		29716	svuint8x4_t svld1_u8_x4(svcount_t, uint8_t const *);
		29717	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_s8_x4)))
		29718	svint8x4_t svld1_s8_x4(svcount_t, int8_t const *);
		29719	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_u64_x4)))
		29720	svuint64x4_t svld1_u64_x4(svcount_t, uint64_t const *);
		29721	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_f64_x4)))
		29722	svfloat64x4_t svld1_f64_x4(svcount_t, float64_t const *);
		29723	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_s64_x4)))
		29724	svint64x4_t svld1_s64_x4(svcount_t, int64_t const *);
		29725	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_u16_x4)))
		29726	svuint16x4_t svld1_u16_x4(svcount_t, uint16_t const *);
		29727	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_bf16_x4)))
		29728	svbfloat16x4_t svld1_bf16_x4(svcount_t, bfloat16_t const *);
		29729	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_f16_x4)))
		29730	svfloat16x4_t svld1_f16_x4(svcount_t, float16_t const *);
		29731	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_s16_x4)))
		29732	svint16x4_t svld1_s16_x4(svcount_t, int16_t const *);
		29733	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_u32_x4)))
		29734	svuint32x4_t svld1_u32_x4(svcount_t, uint32_t const *);
		29735	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_f32_x4)))
		29736	svfloat32x4_t svld1_f32_x4(svcount_t, float32_t const *);
		29737	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_s32_x4)))
		29738	svint32x4_t svld1_s32_x4(svcount_t, int32_t const *);
		29739	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_u8_x2)))
		29740	svuint8x2_t svld1_vnum_u8_x2(svcount_t, uint8_t const *, int64_t);
		29741	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_s8_x2)))
		29742	svint8x2_t svld1_vnum_s8_x2(svcount_t, int8_t const *, int64_t);
		29743	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_u64_x2)))
		29744	svuint64x2_t svld1_vnum_u64_x2(svcount_t, uint64_t const *, int64_t);
		29745	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_f64_x2)))
		29746	svfloat64x2_t svld1_vnum_f64_x2(svcount_t, float64_t const *, int64_t);
		29747	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_s64_x2)))
		29748	svint64x2_t svld1_vnum_s64_x2(svcount_t, int64_t const *, int64_t);
		29749	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_u16_x2)))
		29750	svuint16x2_t svld1_vnum_u16_x2(svcount_t, uint16_t const *, int64_t);
		29751	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_bf16_x2)))
		29752	svbfloat16x2_t svld1_vnum_bf16_x2(svcount_t, bfloat16_t const *, int64_t);
		29753	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_f16_x2)))
		29754	svfloat16x2_t svld1_vnum_f16_x2(svcount_t, float16_t const *, int64_t);
		29755	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_s16_x2)))
		29756	svint16x2_t svld1_vnum_s16_x2(svcount_t, int16_t const *, int64_t);
		29757	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_u32_x2)))
		29758	svuint32x2_t svld1_vnum_u32_x2(svcount_t, uint32_t const *, int64_t);
		29759	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_f32_x2)))
		29760	svfloat32x2_t svld1_vnum_f32_x2(svcount_t, float32_t const *, int64_t);
		29761	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_s32_x2)))
		29762	svint32x2_t svld1_vnum_s32_x2(svcount_t, int32_t const *, int64_t);
		29763	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_u8_x4)))
		29764	svuint8x4_t svld1_vnum_u8_x4(svcount_t, uint8_t const *, int64_t);
		29765	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_s8_x4)))
		29766	svint8x4_t svld1_vnum_s8_x4(svcount_t, int8_t const *, int64_t);
		29767	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_u64_x4)))
		29768	svuint64x4_t svld1_vnum_u64_x4(svcount_t, uint64_t const *, int64_t);
		29769	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_f64_x4)))
		29770	svfloat64x4_t svld1_vnum_f64_x4(svcount_t, float64_t const *, int64_t);
		29771	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_s64_x4)))
		29772	svint64x4_t svld1_vnum_s64_x4(svcount_t, int64_t const *, int64_t);
		29773	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_u16_x4)))
		29774	svuint16x4_t svld1_vnum_u16_x4(svcount_t, uint16_t const *, int64_t);
		29775	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_bf16_x4)))
		29776	svbfloat16x4_t svld1_vnum_bf16_x4(svcount_t, bfloat16_t const *, int64_t);
		29777	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_f16_x4)))
		29778	svfloat16x4_t svld1_vnum_f16_x4(svcount_t, float16_t const *, int64_t);
		29779	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_s16_x4)))
		29780	svint16x4_t svld1_vnum_s16_x4(svcount_t, int16_t const *, int64_t);
		29781	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_u32_x4)))
		29782	svuint32x4_t svld1_vnum_u32_x4(svcount_t, uint32_t const *, int64_t);
		29783	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_f32_x4)))
		29784	svfloat32x4_t svld1_vnum_f32_x4(svcount_t, float32_t const *, int64_t);
		29785	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_s32_x4)))
		29786	svint32x4_t svld1_vnum_s32_x4(svcount_t, int32_t const *, int64_t);
		29787	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_u8_x2)))
		29788	svuint8x2_t svldnt1_u8_x2(svcount_t, uint8_t const *);
		29789	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_s8_x2)))
		29790	svint8x2_t svldnt1_s8_x2(svcount_t, int8_t const *);
		29791	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_u64_x2)))
		29792	svuint64x2_t svldnt1_u64_x2(svcount_t, uint64_t const *);
		29793	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_f64_x2)))
		29794	svfloat64x2_t svldnt1_f64_x2(svcount_t, float64_t const *);
		29795	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_s64_x2)))
		29796	svint64x2_t svldnt1_s64_x2(svcount_t, int64_t const *);
		29797	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_u16_x2)))
		29798	svuint16x2_t svldnt1_u16_x2(svcount_t, uint16_t const *);
		29799	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_bf16_x2)))
		29800	svbfloat16x2_t svldnt1_bf16_x2(svcount_t, bfloat16_t const *);
		29801	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_f16_x2)))
		29802	svfloat16x2_t svldnt1_f16_x2(svcount_t, float16_t const *);
		29803	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_s16_x2)))
		29804	svint16x2_t svldnt1_s16_x2(svcount_t, int16_t const *);
		29805	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_u32_x2)))
		29806	svuint32x2_t svldnt1_u32_x2(svcount_t, uint32_t const *);
		29807	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_f32_x2)))
		29808	svfloat32x2_t svldnt1_f32_x2(svcount_t, float32_t const *);
		29809	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_s32_x2)))
		29810	svint32x2_t svldnt1_s32_x2(svcount_t, int32_t const *);
		29811	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_u8_x4)))
		29812	svuint8x4_t svldnt1_u8_x4(svcount_t, uint8_t const *);
		29813	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_s8_x4)))
		29814	svint8x4_t svldnt1_s8_x4(svcount_t, int8_t const *);
		29815	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_u64_x4)))
		29816	svuint64x4_t svldnt1_u64_x4(svcount_t, uint64_t const *);
		29817	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_f64_x4)))
		29818	svfloat64x4_t svldnt1_f64_x4(svcount_t, float64_t const *);
		29819	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_s64_x4)))
		29820	svint64x4_t svldnt1_s64_x4(svcount_t, int64_t const *);
		29821	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_u16_x4)))
		29822	svuint16x4_t svldnt1_u16_x4(svcount_t, uint16_t const *);
		29823	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_bf16_x4)))
		29824	svbfloat16x4_t svldnt1_bf16_x4(svcount_t, bfloat16_t const *);
		29825	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_f16_x4)))
		29826	svfloat16x4_t svldnt1_f16_x4(svcount_t, float16_t const *);
		29827	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_s16_x4)))
		29828	svint16x4_t svldnt1_s16_x4(svcount_t, int16_t const *);
		29829	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_u32_x4)))
		29830	svuint32x4_t svldnt1_u32_x4(svcount_t, uint32_t const *);
		29831	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_f32_x4)))
		29832	svfloat32x4_t svldnt1_f32_x4(svcount_t, float32_t const *);
		29833	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_s32_x4)))
		29834	svint32x4_t svldnt1_s32_x4(svcount_t, int32_t const *);
		29835	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_u8_x2)))
		29836	svuint8x2_t svldnt1_vnum_u8_x2(svcount_t, uint8_t const *, int64_t);
		29837	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_s8_x2)))
		29838	svint8x2_t svldnt1_vnum_s8_x2(svcount_t, int8_t const *, int64_t);
		29839	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_u64_x2)))
		29840	svuint64x2_t svldnt1_vnum_u64_x2(svcount_t, uint64_t const *, int64_t);
		29841	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_f64_x2)))
		29842	svfloat64x2_t svldnt1_vnum_f64_x2(svcount_t, float64_t const *, int64_t);
		29843	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_s64_x2)))
		29844	svint64x2_t svldnt1_vnum_s64_x2(svcount_t, int64_t const *, int64_t);
		29845	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_u16_x2)))
		29846	svuint16x2_t svldnt1_vnum_u16_x2(svcount_t, uint16_t const *, int64_t);
		29847	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_bf16_x2)))
		29848	svbfloat16x2_t svldnt1_vnum_bf16_x2(svcount_t, bfloat16_t const *, int64_t);
		29849	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_f16_x2)))
		29850	svfloat16x2_t svldnt1_vnum_f16_x2(svcount_t, float16_t const *, int64_t);
		29851	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_s16_x2)))
		29852	svint16x2_t svldnt1_vnum_s16_x2(svcount_t, int16_t const *, int64_t);
		29853	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_u32_x2)))
		29854	svuint32x2_t svldnt1_vnum_u32_x2(svcount_t, uint32_t const *, int64_t);
		29855	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_f32_x2)))
		29856	svfloat32x2_t svldnt1_vnum_f32_x2(svcount_t, float32_t const *, int64_t);
		29857	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_s32_x2)))
		29858	svint32x2_t svldnt1_vnum_s32_x2(svcount_t, int32_t const *, int64_t);
		29859	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_u8_x4)))
		29860	svuint8x4_t svldnt1_vnum_u8_x4(svcount_t, uint8_t const *, int64_t);
		29861	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_s8_x4)))
		29862	svint8x4_t svldnt1_vnum_s8_x4(svcount_t, int8_t const *, int64_t);
		29863	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_u64_x4)))
		29864	svuint64x4_t svldnt1_vnum_u64_x4(svcount_t, uint64_t const *, int64_t);
		29865	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_f64_x4)))
		29866	svfloat64x4_t svldnt1_vnum_f64_x4(svcount_t, float64_t const *, int64_t);
		29867	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_s64_x4)))
		29868	svint64x4_t svldnt1_vnum_s64_x4(svcount_t, int64_t const *, int64_t);
		29869	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_u16_x4)))
		29870	svuint16x4_t svldnt1_vnum_u16_x4(svcount_t, uint16_t const *, int64_t);
		29871	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_bf16_x4)))
		29872	svbfloat16x4_t svldnt1_vnum_bf16_x4(svcount_t, bfloat16_t const *, int64_t);
		29873	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_f16_x4)))
		29874	svfloat16x4_t svldnt1_vnum_f16_x4(svcount_t, float16_t const *, int64_t);
		29875	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_s16_x4)))
		29876	svint16x4_t svldnt1_vnum_s16_x4(svcount_t, int16_t const *, int64_t);
		29877	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_u32_x4)))
		29878	svuint32x4_t svldnt1_vnum_u32_x4(svcount_t, uint32_t const *, int64_t);
		29879	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_f32_x4)))
		29880	svfloat32x4_t svldnt1_vnum_f32_x4(svcount_t, float32_t const *, int64_t);
		29881	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_s32_x4)))
		29882	svint32x4_t svldnt1_vnum_s32_x4(svcount_t, int32_t const *, int64_t);
		29883	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpext_lane_c8)))
		29884	svbool_t svpext_lane_c8(svcount_t, uint64_t);
		29885	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpext_lane_c32)))
		29886	svbool_t svpext_lane_c32(svcount_t, uint64_t);
		29887	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpext_lane_c64)))
		29888	svbool_t svpext_lane_c64(svcount_t, uint64_t);
		29889	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpext_lane_c16)))
		29890	svbool_t svpext_lane_c16(svcount_t, uint64_t);
		29891	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpext_lane_c8_x2)))
		29892	svboolx2_t svpext_lane_c8_x2(svcount_t, uint64_t);
		29893	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpext_lane_c32_x2)))
		29894	svboolx2_t svpext_lane_c32_x2(svcount_t, uint64_t);
		29895	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpext_lane_c64_x2)))
		29896	svboolx2_t svpext_lane_c64_x2(svcount_t, uint64_t);
		29897	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpext_lane_c16_x2)))
		29898	svboolx2_t svpext_lane_c16_x2(svcount_t, uint64_t);
		29899	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpfalse_c)))
		29900	svcount_t svpfalse_c(void);
		29901	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpsel_lane_c16)))
		29902	svcount_t svpsel_lane_c16(svcount_t, svbool_t, uint32_t);
		29903	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpsel_lane_c32)))
		29904	svcount_t svpsel_lane_c32(svcount_t, svbool_t, uint32_t);
		29905	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpsel_lane_c64)))
		29906	svcount_t svpsel_lane_c64(svcount_t, svbool_t, uint32_t);
		29907	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svpsel_lane_c8)))
		29908	svcount_t svpsel_lane_c8(svcount_t, svbool_t, uint32_t);
23902	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svptrue_c8)))	29909	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svptrue_c8)))
23903	svcount_t svptrue_c8(void);	29910	svcount_t svptrue_c8(void);
23904	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svptrue_c32)))	29911	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svptrue_c32)))
...	@@ -23907,6 +29914,420 @@ __ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svptrue_c64)))	...	@@ -23907,6 +29914,420 @@ __ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svptrue_c64)))
23907	svcount_t svptrue_c64(void);	29914	svcount_t svptrue_c64(void);
23908	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svptrue_c16)))	29915	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svptrue_c16)))
23909	svcount_t svptrue_c16(void);	29916	svcount_t svptrue_c16(void);
		29917	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrn_n_s16_s32_x2)))
		29918	svint16_t svqrshrn_n_s16_s32_x2(svint32x2_t, uint64_t);
		29919	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrn_n_u16_u32_x2)))
		29920	svuint16_t svqrshrn_n_u16_u32_x2(svuint32x2_t, uint64_t);
		29921	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrun_n_u16_s32_x2)))
		29922	svuint16_t svqrshrun_n_u16_s32_x2(svint32x2_t, uint64_t);
		29923	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u8_m)))
		29924	svuint8_t svrevd_u8_m(svuint8_t, svbool_t, svuint8_t);
		29925	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u32_m)))
		29926	svuint32_t svrevd_u32_m(svuint32_t, svbool_t, svuint32_t);
		29927	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u64_m)))
		29928	svuint64_t svrevd_u64_m(svuint64_t, svbool_t, svuint64_t);
		29929	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u16_m)))
		29930	svuint16_t svrevd_u16_m(svuint16_t, svbool_t, svuint16_t);
		29931	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_bf16_m)))
		29932	svbfloat16_t svrevd_bf16_m(svbfloat16_t, svbool_t, svbfloat16_t);
		29933	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s8_m)))
		29934	svint8_t svrevd_s8_m(svint8_t, svbool_t, svint8_t);
		29935	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f64_m)))
		29936	svfloat64_t svrevd_f64_m(svfloat64_t, svbool_t, svfloat64_t);
		29937	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f32_m)))
		29938	svfloat32_t svrevd_f32_m(svfloat32_t, svbool_t, svfloat32_t);
		29939	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f16_m)))
		29940	svfloat16_t svrevd_f16_m(svfloat16_t, svbool_t, svfloat16_t);
		29941	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s32_m)))
		29942	svint32_t svrevd_s32_m(svint32_t, svbool_t, svint32_t);
		29943	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s64_m)))
		29944	svint64_t svrevd_s64_m(svint64_t, svbool_t, svint64_t);
		29945	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s16_m)))
		29946	svint16_t svrevd_s16_m(svint16_t, svbool_t, svint16_t);
		29947	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u8_x)))
		29948	svuint8_t svrevd_u8_x(svbool_t, svuint8_t);
		29949	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u32_x)))
		29950	svuint32_t svrevd_u32_x(svbool_t, svuint32_t);
		29951	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u64_x)))
		29952	svuint64_t svrevd_u64_x(svbool_t, svuint64_t);
		29953	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u16_x)))
		29954	svuint16_t svrevd_u16_x(svbool_t, svuint16_t);
		29955	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_bf16_x)))
		29956	svbfloat16_t svrevd_bf16_x(svbool_t, svbfloat16_t);
		29957	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s8_x)))
		29958	svint8_t svrevd_s8_x(svbool_t, svint8_t);
		29959	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f64_x)))
		29960	svfloat64_t svrevd_f64_x(svbool_t, svfloat64_t);
		29961	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f32_x)))
		29962	svfloat32_t svrevd_f32_x(svbool_t, svfloat32_t);
		29963	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f16_x)))
		29964	svfloat16_t svrevd_f16_x(svbool_t, svfloat16_t);
		29965	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s32_x)))
		29966	svint32_t svrevd_s32_x(svbool_t, svint32_t);
		29967	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s64_x)))
		29968	svint64_t svrevd_s64_x(svbool_t, svint64_t);
		29969	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s16_x)))
		29970	svint16_t svrevd_s16_x(svbool_t, svint16_t);
		29971	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u8_z)))
		29972	svuint8_t svrevd_u8_z(svbool_t, svuint8_t);
		29973	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u32_z)))
		29974	svuint32_t svrevd_u32_z(svbool_t, svuint32_t);
		29975	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u64_z)))
		29976	svuint64_t svrevd_u64_z(svbool_t, svuint64_t);
		29977	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u16_z)))
		29978	svuint16_t svrevd_u16_z(svbool_t, svuint16_t);
		29979	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_bf16_z)))
		29980	svbfloat16_t svrevd_bf16_z(svbool_t, svbfloat16_t);
		29981	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s8_z)))
		29982	svint8_t svrevd_s8_z(svbool_t, svint8_t);
		29983	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f64_z)))
		29984	svfloat64_t svrevd_f64_z(svbool_t, svfloat64_t);
		29985	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f32_z)))
		29986	svfloat32_t svrevd_f32_z(svbool_t, svfloat32_t);
		29987	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f16_z)))
		29988	svfloat16_t svrevd_f16_z(svbool_t, svfloat16_t);
		29989	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s32_z)))
		29990	svint32_t svrevd_s32_z(svbool_t, svint32_t);
		29991	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s64_z)))
		29992	svint64_t svrevd_s64_z(svbool_t, svint64_t);
		29993	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s16_z)))
		29994	svint16_t svrevd_s16_z(svbool_t, svint16_t);
		29995	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svset2_b)))
		29996	svboolx2_t svset2_b(svboolx2_t, uint64_t, svbool_t);
		29997	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svset4_b)))
		29998	svboolx4_t svset4_b(svboolx4_t, uint64_t, svbool_t);
		29999	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_u8_x2)))
		30000	void svst1_u8_x2(svcount_t, uint8_t *, svuint8x2_t);
		30001	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_s8_x2)))
		30002	void svst1_s8_x2(svcount_t, int8_t *, svint8x2_t);
		30003	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_u64_x2)))
		30004	void svst1_u64_x2(svcount_t, uint64_t *, svuint64x2_t);
		30005	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_f64_x2)))
		30006	void svst1_f64_x2(svcount_t, float64_t *, svfloat64x2_t);
		30007	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_s64_x2)))
		30008	void svst1_s64_x2(svcount_t, int64_t *, svint64x2_t);
		30009	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_u16_x2)))
		30010	void svst1_u16_x2(svcount_t, uint16_t *, svuint16x2_t);
		30011	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_bf16_x2)))
		30012	void svst1_bf16_x2(svcount_t, bfloat16_t *, svbfloat16x2_t);
		30013	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_f16_x2)))
		30014	void svst1_f16_x2(svcount_t, float16_t *, svfloat16x2_t);
		30015	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_s16_x2)))
		30016	void svst1_s16_x2(svcount_t, int16_t *, svint16x2_t);
		30017	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_u32_x2)))
		30018	void svst1_u32_x2(svcount_t, uint32_t *, svuint32x2_t);
		30019	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_f32_x2)))
		30020	void svst1_f32_x2(svcount_t, float32_t *, svfloat32x2_t);
		30021	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_s32_x2)))
		30022	void svst1_s32_x2(svcount_t, int32_t *, svint32x2_t);
		30023	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_u8_x4)))
		30024	void svst1_u8_x4(svcount_t, uint8_t *, svuint8x4_t);
		30025	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_s8_x4)))
		30026	void svst1_s8_x4(svcount_t, int8_t *, svint8x4_t);
		30027	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_u64_x4)))
		30028	void svst1_u64_x4(svcount_t, uint64_t *, svuint64x4_t);
		30029	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_f64_x4)))
		30030	void svst1_f64_x4(svcount_t, float64_t *, svfloat64x4_t);
		30031	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_s64_x4)))
		30032	void svst1_s64_x4(svcount_t, int64_t *, svint64x4_t);
		30033	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_u16_x4)))
		30034	void svst1_u16_x4(svcount_t, uint16_t *, svuint16x4_t);
		30035	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_bf16_x4)))
		30036	void svst1_bf16_x4(svcount_t, bfloat16_t *, svbfloat16x4_t);
		30037	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_f16_x4)))
		30038	void svst1_f16_x4(svcount_t, float16_t *, svfloat16x4_t);
		30039	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_s16_x4)))
		30040	void svst1_s16_x4(svcount_t, int16_t *, svint16x4_t);
		30041	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_u32_x4)))
		30042	void svst1_u32_x4(svcount_t, uint32_t *, svuint32x4_t);
		30043	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_f32_x4)))
		30044	void svst1_f32_x4(svcount_t, float32_t *, svfloat32x4_t);
		30045	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_s32_x4)))
		30046	void svst1_s32_x4(svcount_t, int32_t *, svint32x4_t);
		30047	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_u8_x2)))
		30048	void svst1_vnum_u8_x2(svcount_t, uint8_t *, int64_t, svuint8x2_t);
		30049	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_s8_x2)))
		30050	void svst1_vnum_s8_x2(svcount_t, int8_t *, int64_t, svint8x2_t);
		30051	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_u64_x2)))
		30052	void svst1_vnum_u64_x2(svcount_t, uint64_t *, int64_t, svuint64x2_t);
		30053	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_f64_x2)))
		30054	void svst1_vnum_f64_x2(svcount_t, float64_t *, int64_t, svfloat64x2_t);
		30055	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_s64_x2)))
		30056	void svst1_vnum_s64_x2(svcount_t, int64_t *, int64_t, svint64x2_t);
		30057	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_u16_x2)))
		30058	void svst1_vnum_u16_x2(svcount_t, uint16_t *, int64_t, svuint16x2_t);
		30059	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_bf16_x2)))
		30060	void svst1_vnum_bf16_x2(svcount_t, bfloat16_t *, int64_t, svbfloat16x2_t);
		30061	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_f16_x2)))
		30062	void svst1_vnum_f16_x2(svcount_t, float16_t *, int64_t, svfloat16x2_t);
		30063	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_s16_x2)))
		30064	void svst1_vnum_s16_x2(svcount_t, int16_t *, int64_t, svint16x2_t);
		30065	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_u32_x2)))
		30066	void svst1_vnum_u32_x2(svcount_t, uint32_t *, int64_t, svuint32x2_t);
		30067	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_f32_x2)))
		30068	void svst1_vnum_f32_x2(svcount_t, float32_t *, int64_t, svfloat32x2_t);
		30069	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_s32_x2)))
		30070	void svst1_vnum_s32_x2(svcount_t, int32_t *, int64_t, svint32x2_t);
		30071	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_u8_x4)))
		30072	void svst1_vnum_u8_x4(svcount_t, uint8_t *, int64_t, svuint8x4_t);
		30073	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_s8_x4)))
		30074	void svst1_vnum_s8_x4(svcount_t, int8_t *, int64_t, svint8x4_t);
		30075	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_u64_x4)))
		30076	void svst1_vnum_u64_x4(svcount_t, uint64_t *, int64_t, svuint64x4_t);
		30077	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_f64_x4)))
		30078	void svst1_vnum_f64_x4(svcount_t, float64_t *, int64_t, svfloat64x4_t);
		30079	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_s64_x4)))
		30080	void svst1_vnum_s64_x4(svcount_t, int64_t *, int64_t, svint64x4_t);
		30081	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_u16_x4)))
		30082	void svst1_vnum_u16_x4(svcount_t, uint16_t *, int64_t, svuint16x4_t);
		30083	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_bf16_x4)))
		30084	void svst1_vnum_bf16_x4(svcount_t, bfloat16_t *, int64_t, svbfloat16x4_t);
		30085	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_f16_x4)))
		30086	void svst1_vnum_f16_x4(svcount_t, float16_t *, int64_t, svfloat16x4_t);
		30087	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_s16_x4)))
		30088	void svst1_vnum_s16_x4(svcount_t, int16_t *, int64_t, svint16x4_t);
		30089	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_u32_x4)))
		30090	void svst1_vnum_u32_x4(svcount_t, uint32_t *, int64_t, svuint32x4_t);
		30091	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_f32_x4)))
		30092	void svst1_vnum_f32_x4(svcount_t, float32_t *, int64_t, svfloat32x4_t);
		30093	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_s32_x4)))
		30094	void svst1_vnum_s32_x4(svcount_t, int32_t *, int64_t, svint32x4_t);
		30095	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_u8_x2)))
		30096	void svstnt1_u8_x2(svcount_t, uint8_t *, svuint8x2_t);
		30097	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_s8_x2)))
		30098	void svstnt1_s8_x2(svcount_t, int8_t *, svint8x2_t);
		30099	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_u64_x2)))
		30100	void svstnt1_u64_x2(svcount_t, uint64_t *, svuint64x2_t);
		30101	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_f64_x2)))
		30102	void svstnt1_f64_x2(svcount_t, float64_t *, svfloat64x2_t);
		30103	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_s64_x2)))
		30104	void svstnt1_s64_x2(svcount_t, int64_t *, svint64x2_t);
		30105	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_u16_x2)))
		30106	void svstnt1_u16_x2(svcount_t, uint16_t *, svuint16x2_t);
		30107	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_bf16_x2)))
		30108	void svstnt1_bf16_x2(svcount_t, bfloat16_t *, svbfloat16x2_t);
		30109	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_f16_x2)))
		30110	void svstnt1_f16_x2(svcount_t, float16_t *, svfloat16x2_t);
		30111	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_s16_x2)))
		30112	void svstnt1_s16_x2(svcount_t, int16_t *, svint16x2_t);
		30113	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_u32_x2)))
		30114	void svstnt1_u32_x2(svcount_t, uint32_t *, svuint32x2_t);
		30115	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_f32_x2)))
		30116	void svstnt1_f32_x2(svcount_t, float32_t *, svfloat32x2_t);
		30117	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_s32_x2)))
		30118	void svstnt1_s32_x2(svcount_t, int32_t *, svint32x2_t);
		30119	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_u8_x4)))
		30120	void svstnt1_u8_x4(svcount_t, uint8_t *, svuint8x4_t);
		30121	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_s8_x4)))
		30122	void svstnt1_s8_x4(svcount_t, int8_t *, svint8x4_t);
		30123	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_u64_x4)))
		30124	void svstnt1_u64_x4(svcount_t, uint64_t *, svuint64x4_t);
		30125	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_f64_x4)))
		30126	void svstnt1_f64_x4(svcount_t, float64_t *, svfloat64x4_t);
		30127	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_s64_x4)))
		30128	void svstnt1_s64_x4(svcount_t, int64_t *, svint64x4_t);
		30129	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_u16_x4)))
		30130	void svstnt1_u16_x4(svcount_t, uint16_t *, svuint16x4_t);
		30131	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_bf16_x4)))
		30132	void svstnt1_bf16_x4(svcount_t, bfloat16_t *, svbfloat16x4_t);
		30133	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_f16_x4)))
		30134	void svstnt1_f16_x4(svcount_t, float16_t *, svfloat16x4_t);
		30135	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_s16_x4)))
		30136	void svstnt1_s16_x4(svcount_t, int16_t *, svint16x4_t);
		30137	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_u32_x4)))
		30138	void svstnt1_u32_x4(svcount_t, uint32_t *, svuint32x4_t);
		30139	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_f32_x4)))
		30140	void svstnt1_f32_x4(svcount_t, float32_t *, svfloat32x4_t);
		30141	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_s32_x4)))
		30142	void svstnt1_s32_x4(svcount_t, int32_t *, svint32x4_t);
		30143	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_u8_x2)))
		30144	void svstnt1_vnum_u8_x2(svcount_t, uint8_t *, int64_t, svuint8x2_t);
		30145	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_s8_x2)))
		30146	void svstnt1_vnum_s8_x2(svcount_t, int8_t *, int64_t, svint8x2_t);
		30147	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_u64_x2)))
		30148	void svstnt1_vnum_u64_x2(svcount_t, uint64_t *, int64_t, svuint64x2_t);
		30149	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_f64_x2)))
		30150	void svstnt1_vnum_f64_x2(svcount_t, float64_t *, int64_t, svfloat64x2_t);
		30151	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_s64_x2)))
		30152	void svstnt1_vnum_s64_x2(svcount_t, int64_t *, int64_t, svint64x2_t);
		30153	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_u16_x2)))
		30154	void svstnt1_vnum_u16_x2(svcount_t, uint16_t *, int64_t, svuint16x2_t);
		30155	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_bf16_x2)))
		30156	void svstnt1_vnum_bf16_x2(svcount_t, bfloat16_t *, int64_t, svbfloat16x2_t);
		30157	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_f16_x2)))
		30158	void svstnt1_vnum_f16_x2(svcount_t, float16_t *, int64_t, svfloat16x2_t);
		30159	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_s16_x2)))
		30160	void svstnt1_vnum_s16_x2(svcount_t, int16_t *, int64_t, svint16x2_t);
		30161	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_u32_x2)))
		30162	void svstnt1_vnum_u32_x2(svcount_t, uint32_t *, int64_t, svuint32x2_t);
		30163	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_f32_x2)))
		30164	void svstnt1_vnum_f32_x2(svcount_t, float32_t *, int64_t, svfloat32x2_t);
		30165	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_s32_x2)))
		30166	void svstnt1_vnum_s32_x2(svcount_t, int32_t *, int64_t, svint32x2_t);
		30167	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_u8_x4)))
		30168	void svstnt1_vnum_u8_x4(svcount_t, uint8_t *, int64_t, svuint8x4_t);
		30169	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_s8_x4)))
		30170	void svstnt1_vnum_s8_x4(svcount_t, int8_t *, int64_t, svint8x4_t);
		30171	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_u64_x4)))
		30172	void svstnt1_vnum_u64_x4(svcount_t, uint64_t *, int64_t, svuint64x4_t);
		30173	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_f64_x4)))
		30174	void svstnt1_vnum_f64_x4(svcount_t, float64_t *, int64_t, svfloat64x4_t);
		30175	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_s64_x4)))
		30176	void svstnt1_vnum_s64_x4(svcount_t, int64_t *, int64_t, svint64x4_t);
		30177	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_u16_x4)))
		30178	void svstnt1_vnum_u16_x4(svcount_t, uint16_t *, int64_t, svuint16x4_t);
		30179	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_bf16_x4)))
		30180	void svstnt1_vnum_bf16_x4(svcount_t, bfloat16_t *, int64_t, svbfloat16x4_t);
		30181	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_f16_x4)))
		30182	void svstnt1_vnum_f16_x4(svcount_t, float16_t *, int64_t, svfloat16x4_t);
		30183	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_s16_x4)))
		30184	void svstnt1_vnum_s16_x4(svcount_t, int16_t *, int64_t, svint16x4_t);
		30185	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_u32_x4)))
		30186	void svstnt1_vnum_u32_x4(svcount_t, uint32_t *, int64_t, svuint32x4_t);
		30187	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_f32_x4)))
		30188	void svstnt1_vnum_f32_x4(svcount_t, float32_t *, int64_t, svfloat32x4_t);
		30189	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_s32_x4)))
		30190	void svstnt1_vnum_s32_x4(svcount_t, int32_t *, int64_t, svint32x4_t);
		30191	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svundef2_b)))
		30192	svboolx2_t svundef2_b();
		30193	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svundef4_b)))
		30194	svboolx4_t svundef4_b();
		30195	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_c8_s64)))
		30196	svcount_t svwhilege_c8_s64(int64_t, int64_t, uint64_t);
		30197	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_c32_s64)))
		30198	svcount_t svwhilege_c32_s64(int64_t, int64_t, uint64_t);
		30199	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_c64_s64)))
		30200	svcount_t svwhilege_c64_s64(int64_t, int64_t, uint64_t);
		30201	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_c16_s64)))
		30202	svcount_t svwhilege_c16_s64(int64_t, int64_t, uint64_t);
		30203	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_c8_u64)))
		30204	svcount_t svwhilege_c8_u64(uint64_t, uint64_t, uint64_t);
		30205	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_c32_u64)))
		30206	svcount_t svwhilege_c32_u64(uint64_t, uint64_t, uint64_t);
		30207	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_c64_u64)))
		30208	svcount_t svwhilege_c64_u64(uint64_t, uint64_t, uint64_t);
		30209	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_c16_u64)))
		30210	svcount_t svwhilege_c16_u64(uint64_t, uint64_t, uint64_t);
		30211	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_b8_s64_x2)))
		30212	svboolx2_t svwhilege_b8_s64_x2(int64_t, int64_t);
		30213	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_b32_s64_x2)))
		30214	svboolx2_t svwhilege_b32_s64_x2(int64_t, int64_t);
		30215	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_b64_s64_x2)))
		30216	svboolx2_t svwhilege_b64_s64_x2(int64_t, int64_t);
		30217	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_b16_s64_x2)))
		30218	svboolx2_t svwhilege_b16_s64_x2(int64_t, int64_t);
		30219	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_b8_u64_x2)))
		30220	svboolx2_t svwhilege_b8_u64_x2(uint64_t, uint64_t);
		30221	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_b32_u64_x2)))
		30222	svboolx2_t svwhilege_b32_u64_x2(uint64_t, uint64_t);
		30223	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_b64_u64_x2)))
		30224	svboolx2_t svwhilege_b64_u64_x2(uint64_t, uint64_t);
		30225	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_b16_u64_x2)))
		30226	svboolx2_t svwhilege_b16_u64_x2(uint64_t, uint64_t);
		30227	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_c8_s64)))
		30228	svcount_t svwhilegt_c8_s64(int64_t, int64_t, uint64_t);
		30229	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_c32_s64)))
		30230	svcount_t svwhilegt_c32_s64(int64_t, int64_t, uint64_t);
		30231	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_c64_s64)))
		30232	svcount_t svwhilegt_c64_s64(int64_t, int64_t, uint64_t);
		30233	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_c16_s64)))
		30234	svcount_t svwhilegt_c16_s64(int64_t, int64_t, uint64_t);
		30235	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_c8_u64)))
		30236	svcount_t svwhilegt_c8_u64(uint64_t, uint64_t, uint64_t);
		30237	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_c32_u64)))
		30238	svcount_t svwhilegt_c32_u64(uint64_t, uint64_t, uint64_t);
		30239	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_c64_u64)))
		30240	svcount_t svwhilegt_c64_u64(uint64_t, uint64_t, uint64_t);
		30241	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_c16_u64)))
		30242	svcount_t svwhilegt_c16_u64(uint64_t, uint64_t, uint64_t);
		30243	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_b8_s64_x2)))
		30244	svboolx2_t svwhilegt_b8_s64_x2(int64_t, int64_t);
		30245	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_b32_s64_x2)))
		30246	svboolx2_t svwhilegt_b32_s64_x2(int64_t, int64_t);
		30247	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_b64_s64_x2)))
		30248	svboolx2_t svwhilegt_b64_s64_x2(int64_t, int64_t);
		30249	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_b16_s64_x2)))
		30250	svboolx2_t svwhilegt_b16_s64_x2(int64_t, int64_t);
		30251	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_b8_u64_x2)))
		30252	svboolx2_t svwhilegt_b8_u64_x2(uint64_t, uint64_t);
		30253	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_b32_u64_x2)))
		30254	svboolx2_t svwhilegt_b32_u64_x2(uint64_t, uint64_t);
		30255	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_b64_u64_x2)))
		30256	svboolx2_t svwhilegt_b64_u64_x2(uint64_t, uint64_t);
		30257	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_b16_u64_x2)))
		30258	svboolx2_t svwhilegt_b16_u64_x2(uint64_t, uint64_t);
		30259	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_c8_s64)))
		30260	svcount_t svwhilele_c8_s64(int64_t, int64_t, uint64_t);
		30261	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_c32_s64)))
		30262	svcount_t svwhilele_c32_s64(int64_t, int64_t, uint64_t);
		30263	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_c64_s64)))
		30264	svcount_t svwhilele_c64_s64(int64_t, int64_t, uint64_t);
		30265	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_c16_s64)))
		30266	svcount_t svwhilele_c16_s64(int64_t, int64_t, uint64_t);
		30267	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_c8_u64)))
		30268	svcount_t svwhilele_c8_u64(uint64_t, uint64_t, uint64_t);
		30269	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_c32_u64)))
		30270	svcount_t svwhilele_c32_u64(uint64_t, uint64_t, uint64_t);
		30271	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_c64_u64)))
		30272	svcount_t svwhilele_c64_u64(uint64_t, uint64_t, uint64_t);
		30273	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_c16_u64)))
		30274	svcount_t svwhilele_c16_u64(uint64_t, uint64_t, uint64_t);
		30275	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_b8_s64_x2)))
		30276	svboolx2_t svwhilele_b8_s64_x2(int64_t, int64_t);
		30277	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_b32_s64_x2)))
		30278	svboolx2_t svwhilele_b32_s64_x2(int64_t, int64_t);
		30279	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_b64_s64_x2)))
		30280	svboolx2_t svwhilele_b64_s64_x2(int64_t, int64_t);
		30281	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_b16_s64_x2)))
		30282	svboolx2_t svwhilele_b16_s64_x2(int64_t, int64_t);
		30283	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_b8_u64_x2)))
		30284	svboolx2_t svwhilele_b8_u64_x2(uint64_t, uint64_t);
		30285	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_b32_u64_x2)))
		30286	svboolx2_t svwhilele_b32_u64_x2(uint64_t, uint64_t);
		30287	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_b64_u64_x2)))
		30288	svboolx2_t svwhilele_b64_u64_x2(uint64_t, uint64_t);
		30289	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_b16_u64_x2)))
		30290	svboolx2_t svwhilele_b16_u64_x2(uint64_t, uint64_t);
		30291	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_c8_u64)))
		30292	svcount_t svwhilelt_c8_u64(uint64_t, uint64_t, uint64_t);
		30293	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_c32_u64)))
		30294	svcount_t svwhilelt_c32_u64(uint64_t, uint64_t, uint64_t);
		30295	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_c64_u64)))
		30296	svcount_t svwhilelt_c64_u64(uint64_t, uint64_t, uint64_t);
		30297	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_c16_u64)))
		30298	svcount_t svwhilelt_c16_u64(uint64_t, uint64_t, uint64_t);
		30299	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_c8_s64)))
		30300	svcount_t svwhilelt_c8_s64(int64_t, int64_t, uint64_t);
		30301	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_c32_s64)))
		30302	svcount_t svwhilelt_c32_s64(int64_t, int64_t, uint64_t);
		30303	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_c64_s64)))
		30304	svcount_t svwhilelt_c64_s64(int64_t, int64_t, uint64_t);
		30305	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_c16_s64)))
		30306	svcount_t svwhilelt_c16_s64(int64_t, int64_t, uint64_t);
		30307	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_b8_u64_x2)))
		30308	svboolx2_t svwhilelt_b8_u64_x2(uint64_t, uint64_t);
		30309	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_b32_u64_x2)))
		30310	svboolx2_t svwhilelt_b32_u64_x2(uint64_t, uint64_t);
		30311	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_b64_u64_x2)))
		30312	svboolx2_t svwhilelt_b64_u64_x2(uint64_t, uint64_t);
		30313	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_b16_u64_x2)))
		30314	svboolx2_t svwhilelt_b16_u64_x2(uint64_t, uint64_t);
		30315	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_b8_s64_x2)))
		30316	svboolx2_t svwhilelt_b8_s64_x2(int64_t, int64_t);
		30317	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_b32_s64_x2)))
		30318	svboolx2_t svwhilelt_b32_s64_x2(int64_t, int64_t);
		30319	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_b64_s64_x2)))
		30320	svboolx2_t svwhilelt_b64_s64_x2(int64_t, int64_t);
		30321	__ai __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_b16_s64_x2)))
		30322	svboolx2_t svwhilelt_b16_s64_x2(int64_t, int64_t);
		30323	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svbfmlslb_f32)))
		30324	svfloat32_t svbfmlslb(svfloat32_t, svbfloat16_t, svbfloat16_t);
		30325	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svbfmlslb_lane_f32)))
		30326	svfloat32_t svbfmlslb_lane(svfloat32_t, svbfloat16_t, svbfloat16_t, uint64_t);
		30327	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svbfmlslt_f32)))
		30328	svfloat32_t svbfmlslt(svfloat32_t, svbfloat16_t, svbfloat16_t);
		30329	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svbfmlslt_lane_f32)))
		30330	svfloat32_t svbfmlslt_lane(svfloat32_t, svbfloat16_t, svbfloat16_t, uint64_t);
23910	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_f64)))	30331	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_f64)))
23911	svfloat64_t svclamp(svfloat64_t, svfloat64_t, svfloat64_t);	30332	svfloat64_t svclamp(svfloat64_t, svfloat64_t, svfloat64_t);
23912	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_f32)))	30333	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_f32)))
...	@@ -23929,6 +30350,620 @@ __aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_u64)))	...	@@ -23929,6 +30350,620 @@ __aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_u64)))
23929	svuint64_t svclamp(svuint64_t, svuint64_t, svuint64_t);	30350	svuint64_t svclamp(svuint64_t, svuint64_t, svuint64_t);
23930	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_u16)))	30351	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svclamp_u16)))
23931	svuint16_t svclamp(svuint16_t, svuint16_t, svuint16_t);	30352	svuint16_t svclamp(svuint16_t, svuint16_t, svuint16_t);
		30353	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcreate2_b)))
		30354	svboolx2_t svcreate2(svbool_t, svbool_t);
		30355	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svcreate4_b)))
		30356	svboolx4_t svcreate4(svbool_t, svbool_t, svbool_t, svbool_t);
		30357	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svdot_f32_f16)))
		30358	svfloat32_t svdot(svfloat32_t, svfloat16_t, svfloat16_t);
		30359	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svdot_s32_s16)))
		30360	svint32_t svdot(svint32_t, svint16_t, svint16_t);
		30361	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svdot_u32_u16)))
		30362	svuint32_t svdot(svuint32_t, svuint16_t, svuint16_t);
		30363	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svdot_lane_f32_f16)))
		30364	svfloat32_t svdot_lane(svfloat32_t, svfloat16_t, svfloat16_t, uint64_t);
		30365	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svdot_lane_s32_s16)))
		30366	svint32_t svdot_lane(svint32_t, svint16_t, svint16_t, uint64_t);
		30367	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svdot_lane_u32_u16)))
		30368	svuint32_t svdot_lane(svuint32_t, svuint16_t, svuint16_t, uint64_t);
		30369	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svget2_b)))
		30370	svbool_t svget2(svboolx2_t, uint64_t);
		30371	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svget4_b)))
		30372	svbool_t svget4(svboolx4_t, uint64_t);
		30373	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_u8_x2)))
		30374	svuint8x2_t svld1_x2(svcount_t, uint8_t const *);
		30375	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_s8_x2)))
		30376	svint8x2_t svld1_x2(svcount_t, int8_t const *);
		30377	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_u64_x2)))
		30378	svuint64x2_t svld1_x2(svcount_t, uint64_t const *);
		30379	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_f64_x2)))
		30380	svfloat64x2_t svld1_x2(svcount_t, float64_t const *);
		30381	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_s64_x2)))
		30382	svint64x2_t svld1_x2(svcount_t, int64_t const *);
		30383	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_u16_x2)))
		30384	svuint16x2_t svld1_x2(svcount_t, uint16_t const *);
		30385	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_bf16_x2)))
		30386	svbfloat16x2_t svld1_x2(svcount_t, bfloat16_t const *);
		30387	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_f16_x2)))
		30388	svfloat16x2_t svld1_x2(svcount_t, float16_t const *);
		30389	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_s16_x2)))
		30390	svint16x2_t svld1_x2(svcount_t, int16_t const *);
		30391	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_u32_x2)))
		30392	svuint32x2_t svld1_x2(svcount_t, uint32_t const *);
		30393	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_f32_x2)))
		30394	svfloat32x2_t svld1_x2(svcount_t, float32_t const *);
		30395	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_s32_x2)))
		30396	svint32x2_t svld1_x2(svcount_t, int32_t const *);
		30397	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_u8_x4)))
		30398	svuint8x4_t svld1_x4(svcount_t, uint8_t const *);
		30399	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_s8_x4)))
		30400	svint8x4_t svld1_x4(svcount_t, int8_t const *);
		30401	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_u64_x4)))
		30402	svuint64x4_t svld1_x4(svcount_t, uint64_t const *);
		30403	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_f64_x4)))
		30404	svfloat64x4_t svld1_x4(svcount_t, float64_t const *);
		30405	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_s64_x4)))
		30406	svint64x4_t svld1_x4(svcount_t, int64_t const *);
		30407	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_u16_x4)))
		30408	svuint16x4_t svld1_x4(svcount_t, uint16_t const *);
		30409	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_bf16_x4)))
		30410	svbfloat16x4_t svld1_x4(svcount_t, bfloat16_t const *);
		30411	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_f16_x4)))
		30412	svfloat16x4_t svld1_x4(svcount_t, float16_t const *);
		30413	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_s16_x4)))
		30414	svint16x4_t svld1_x4(svcount_t, int16_t const *);
		30415	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_u32_x4)))
		30416	svuint32x4_t svld1_x4(svcount_t, uint32_t const *);
		30417	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_f32_x4)))
		30418	svfloat32x4_t svld1_x4(svcount_t, float32_t const *);
		30419	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_s32_x4)))
		30420	svint32x4_t svld1_x4(svcount_t, int32_t const *);
		30421	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_u8_x2)))
		30422	svuint8x2_t svld1_vnum_x2(svcount_t, uint8_t const *, int64_t);
		30423	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_s8_x2)))
		30424	svint8x2_t svld1_vnum_x2(svcount_t, int8_t const *, int64_t);
		30425	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_u64_x2)))
		30426	svuint64x2_t svld1_vnum_x2(svcount_t, uint64_t const *, int64_t);
		30427	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_f64_x2)))
		30428	svfloat64x2_t svld1_vnum_x2(svcount_t, float64_t const *, int64_t);
		30429	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_s64_x2)))
		30430	svint64x2_t svld1_vnum_x2(svcount_t, int64_t const *, int64_t);
		30431	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_u16_x2)))
		30432	svuint16x2_t svld1_vnum_x2(svcount_t, uint16_t const *, int64_t);
		30433	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_bf16_x2)))
		30434	svbfloat16x2_t svld1_vnum_x2(svcount_t, bfloat16_t const *, int64_t);
		30435	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_f16_x2)))
		30436	svfloat16x2_t svld1_vnum_x2(svcount_t, float16_t const *, int64_t);
		30437	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_s16_x2)))
		30438	svint16x2_t svld1_vnum_x2(svcount_t, int16_t const *, int64_t);
		30439	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_u32_x2)))
		30440	svuint32x2_t svld1_vnum_x2(svcount_t, uint32_t const *, int64_t);
		30441	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_f32_x2)))
		30442	svfloat32x2_t svld1_vnum_x2(svcount_t, float32_t const *, int64_t);
		30443	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_s32_x2)))
		30444	svint32x2_t svld1_vnum_x2(svcount_t, int32_t const *, int64_t);
		30445	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_u8_x4)))
		30446	svuint8x4_t svld1_vnum_x4(svcount_t, uint8_t const *, int64_t);
		30447	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_s8_x4)))
		30448	svint8x4_t svld1_vnum_x4(svcount_t, int8_t const *, int64_t);
		30449	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_u64_x4)))
		30450	svuint64x4_t svld1_vnum_x4(svcount_t, uint64_t const *, int64_t);
		30451	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_f64_x4)))
		30452	svfloat64x4_t svld1_vnum_x4(svcount_t, float64_t const *, int64_t);
		30453	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_s64_x4)))
		30454	svint64x4_t svld1_vnum_x4(svcount_t, int64_t const *, int64_t);
		30455	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_u16_x4)))
		30456	svuint16x4_t svld1_vnum_x4(svcount_t, uint16_t const *, int64_t);
		30457	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_bf16_x4)))
		30458	svbfloat16x4_t svld1_vnum_x4(svcount_t, bfloat16_t const *, int64_t);
		30459	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_f16_x4)))
		30460	svfloat16x4_t svld1_vnum_x4(svcount_t, float16_t const *, int64_t);
		30461	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_s16_x4)))
		30462	svint16x4_t svld1_vnum_x4(svcount_t, int16_t const *, int64_t);
		30463	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_u32_x4)))
		30464	svuint32x4_t svld1_vnum_x4(svcount_t, uint32_t const *, int64_t);
		30465	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_f32_x4)))
		30466	svfloat32x4_t svld1_vnum_x4(svcount_t, float32_t const *, int64_t);
		30467	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svld1_vnum_s32_x4)))
		30468	svint32x4_t svld1_vnum_x4(svcount_t, int32_t const *, int64_t);
		30469	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_u8_x2)))
		30470	svuint8x2_t svldnt1_x2(svcount_t, uint8_t const *);
		30471	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_s8_x2)))
		30472	svint8x2_t svldnt1_x2(svcount_t, int8_t const *);
		30473	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_u64_x2)))
		30474	svuint64x2_t svldnt1_x2(svcount_t, uint64_t const *);
		30475	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_f64_x2)))
		30476	svfloat64x2_t svldnt1_x2(svcount_t, float64_t const *);
		30477	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_s64_x2)))
		30478	svint64x2_t svldnt1_x2(svcount_t, int64_t const *);
		30479	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_u16_x2)))
		30480	svuint16x2_t svldnt1_x2(svcount_t, uint16_t const *);
		30481	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_bf16_x2)))
		30482	svbfloat16x2_t svldnt1_x2(svcount_t, bfloat16_t const *);
		30483	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_f16_x2)))
		30484	svfloat16x2_t svldnt1_x2(svcount_t, float16_t const *);
		30485	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_s16_x2)))
		30486	svint16x2_t svldnt1_x2(svcount_t, int16_t const *);
		30487	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_u32_x2)))
		30488	svuint32x2_t svldnt1_x2(svcount_t, uint32_t const *);
		30489	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_f32_x2)))
		30490	svfloat32x2_t svldnt1_x2(svcount_t, float32_t const *);
		30491	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_s32_x2)))
		30492	svint32x2_t svldnt1_x2(svcount_t, int32_t const *);
		30493	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_u8_x4)))
		30494	svuint8x4_t svldnt1_x4(svcount_t, uint8_t const *);
		30495	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_s8_x4)))
		30496	svint8x4_t svldnt1_x4(svcount_t, int8_t const *);
		30497	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_u64_x4)))
		30498	svuint64x4_t svldnt1_x4(svcount_t, uint64_t const *);
		30499	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_f64_x4)))
		30500	svfloat64x4_t svldnt1_x4(svcount_t, float64_t const *);
		30501	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_s64_x4)))
		30502	svint64x4_t svldnt1_x4(svcount_t, int64_t const *);
		30503	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_u16_x4)))
		30504	svuint16x4_t svldnt1_x4(svcount_t, uint16_t const *);
		30505	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_bf16_x4)))
		30506	svbfloat16x4_t svldnt1_x4(svcount_t, bfloat16_t const *);
		30507	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_f16_x4)))
		30508	svfloat16x4_t svldnt1_x4(svcount_t, float16_t const *);
		30509	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_s16_x4)))
		30510	svint16x4_t svldnt1_x4(svcount_t, int16_t const *);
		30511	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_u32_x4)))
		30512	svuint32x4_t svldnt1_x4(svcount_t, uint32_t const *);
		30513	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_f32_x4)))
		30514	svfloat32x4_t svldnt1_x4(svcount_t, float32_t const *);
		30515	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_s32_x4)))
		30516	svint32x4_t svldnt1_x4(svcount_t, int32_t const *);
		30517	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_u8_x2)))
		30518	svuint8x2_t svldnt1_vnum_x2(svcount_t, uint8_t const *, int64_t);
		30519	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_s8_x2)))
		30520	svint8x2_t svldnt1_vnum_x2(svcount_t, int8_t const *, int64_t);
		30521	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_u64_x2)))
		30522	svuint64x2_t svldnt1_vnum_x2(svcount_t, uint64_t const *, int64_t);
		30523	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_f64_x2)))
		30524	svfloat64x2_t svldnt1_vnum_x2(svcount_t, float64_t const *, int64_t);
		30525	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_s64_x2)))
		30526	svint64x2_t svldnt1_vnum_x2(svcount_t, int64_t const *, int64_t);
		30527	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_u16_x2)))
		30528	svuint16x2_t svldnt1_vnum_x2(svcount_t, uint16_t const *, int64_t);
		30529	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_bf16_x2)))
		30530	svbfloat16x2_t svldnt1_vnum_x2(svcount_t, bfloat16_t const *, int64_t);
		30531	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_f16_x2)))
		30532	svfloat16x2_t svldnt1_vnum_x2(svcount_t, float16_t const *, int64_t);
		30533	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_s16_x2)))
		30534	svint16x2_t svldnt1_vnum_x2(svcount_t, int16_t const *, int64_t);
		30535	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_u32_x2)))
		30536	svuint32x2_t svldnt1_vnum_x2(svcount_t, uint32_t const *, int64_t);
		30537	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_f32_x2)))
		30538	svfloat32x2_t svldnt1_vnum_x2(svcount_t, float32_t const *, int64_t);
		30539	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_s32_x2)))
		30540	svint32x2_t svldnt1_vnum_x2(svcount_t, int32_t const *, int64_t);
		30541	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_u8_x4)))
		30542	svuint8x4_t svldnt1_vnum_x4(svcount_t, uint8_t const *, int64_t);
		30543	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_s8_x4)))
		30544	svint8x4_t svldnt1_vnum_x4(svcount_t, int8_t const *, int64_t);
		30545	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_u64_x4)))
		30546	svuint64x4_t svldnt1_vnum_x4(svcount_t, uint64_t const *, int64_t);
		30547	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_f64_x4)))
		30548	svfloat64x4_t svldnt1_vnum_x4(svcount_t, float64_t const *, int64_t);
		30549	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_s64_x4)))
		30550	svint64x4_t svldnt1_vnum_x4(svcount_t, int64_t const *, int64_t);
		30551	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_u16_x4)))
		30552	svuint16x4_t svldnt1_vnum_x4(svcount_t, uint16_t const *, int64_t);
		30553	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_bf16_x4)))
		30554	svbfloat16x4_t svldnt1_vnum_x4(svcount_t, bfloat16_t const *, int64_t);
		30555	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_f16_x4)))
		30556	svfloat16x4_t svldnt1_vnum_x4(svcount_t, float16_t const *, int64_t);
		30557	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_s16_x4)))
		30558	svint16x4_t svldnt1_vnum_x4(svcount_t, int16_t const *, int64_t);
		30559	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_u32_x4)))
		30560	svuint32x4_t svldnt1_vnum_x4(svcount_t, uint32_t const *, int64_t);
		30561	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_f32_x4)))
		30562	svfloat32x4_t svldnt1_vnum_x4(svcount_t, float32_t const *, int64_t);
		30563	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svldnt1_vnum_s32_x4)))
		30564	svint32x4_t svldnt1_vnum_x4(svcount_t, int32_t const *, int64_t);
		30565	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrn_n_s16_s32_x2)))
		30566	svint16_t svqrshrn_s16(svint32x2_t, uint64_t);
		30567	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrn_n_u16_u32_x2)))
		30568	svuint16_t svqrshrn_u16(svuint32x2_t, uint64_t);
		30569	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svqrshrun_n_u16_s32_x2)))
		30570	svuint16_t svqrshrun_u16(svint32x2_t, uint64_t);
		30571	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u8_m)))
		30572	svuint8_t svrevd_m(svuint8_t, svbool_t, svuint8_t);
		30573	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u32_m)))
		30574	svuint32_t svrevd_m(svuint32_t, svbool_t, svuint32_t);
		30575	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u64_m)))
		30576	svuint64_t svrevd_m(svuint64_t, svbool_t, svuint64_t);
		30577	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u16_m)))
		30578	svuint16_t svrevd_m(svuint16_t, svbool_t, svuint16_t);
		30579	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_bf16_m)))
		30580	svbfloat16_t svrevd_m(svbfloat16_t, svbool_t, svbfloat16_t);
		30581	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s8_m)))
		30582	svint8_t svrevd_m(svint8_t, svbool_t, svint8_t);
		30583	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f64_m)))
		30584	svfloat64_t svrevd_m(svfloat64_t, svbool_t, svfloat64_t);
		30585	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f32_m)))
		30586	svfloat32_t svrevd_m(svfloat32_t, svbool_t, svfloat32_t);
		30587	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f16_m)))
		30588	svfloat16_t svrevd_m(svfloat16_t, svbool_t, svfloat16_t);
		30589	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s32_m)))
		30590	svint32_t svrevd_m(svint32_t, svbool_t, svint32_t);
		30591	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s64_m)))
		30592	svint64_t svrevd_m(svint64_t, svbool_t, svint64_t);
		30593	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s16_m)))
		30594	svint16_t svrevd_m(svint16_t, svbool_t, svint16_t);
		30595	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u8_x)))
		30596	svuint8_t svrevd_x(svbool_t, svuint8_t);
		30597	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u32_x)))
		30598	svuint32_t svrevd_x(svbool_t, svuint32_t);
		30599	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u64_x)))
		30600	svuint64_t svrevd_x(svbool_t, svuint64_t);
		30601	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u16_x)))
		30602	svuint16_t svrevd_x(svbool_t, svuint16_t);
		30603	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_bf16_x)))
		30604	svbfloat16_t svrevd_x(svbool_t, svbfloat16_t);
		30605	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s8_x)))
		30606	svint8_t svrevd_x(svbool_t, svint8_t);
		30607	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f64_x)))
		30608	svfloat64_t svrevd_x(svbool_t, svfloat64_t);
		30609	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f32_x)))
		30610	svfloat32_t svrevd_x(svbool_t, svfloat32_t);
		30611	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f16_x)))
		30612	svfloat16_t svrevd_x(svbool_t, svfloat16_t);
		30613	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s32_x)))
		30614	svint32_t svrevd_x(svbool_t, svint32_t);
		30615	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s64_x)))
		30616	svint64_t svrevd_x(svbool_t, svint64_t);
		30617	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s16_x)))
		30618	svint16_t svrevd_x(svbool_t, svint16_t);
		30619	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u8_z)))
		30620	svuint8_t svrevd_z(svbool_t, svuint8_t);
		30621	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u32_z)))
		30622	svuint32_t svrevd_z(svbool_t, svuint32_t);
		30623	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u64_z)))
		30624	svuint64_t svrevd_z(svbool_t, svuint64_t);
		30625	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_u16_z)))
		30626	svuint16_t svrevd_z(svbool_t, svuint16_t);
		30627	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_bf16_z)))
		30628	svbfloat16_t svrevd_z(svbool_t, svbfloat16_t);
		30629	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s8_z)))
		30630	svint8_t svrevd_z(svbool_t, svint8_t);
		30631	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f64_z)))
		30632	svfloat64_t svrevd_z(svbool_t, svfloat64_t);
		30633	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f32_z)))
		30634	svfloat32_t svrevd_z(svbool_t, svfloat32_t);
		30635	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_f16_z)))
		30636	svfloat16_t svrevd_z(svbool_t, svfloat16_t);
		30637	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s32_z)))
		30638	svint32_t svrevd_z(svbool_t, svint32_t);
		30639	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s64_z)))
		30640	svint64_t svrevd_z(svbool_t, svint64_t);
		30641	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svrevd_s16_z)))
		30642	svint16_t svrevd_z(svbool_t, svint16_t);
		30643	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svset2_b)))
		30644	svboolx2_t svset2(svboolx2_t, uint64_t, svbool_t);
		30645	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svset4_b)))
		30646	svboolx4_t svset4(svboolx4_t, uint64_t, svbool_t);
		30647	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_u8_x2)))
		30648	void svst1(svcount_t, uint8_t *, svuint8x2_t);
		30649	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_s8_x2)))
		30650	void svst1(svcount_t, int8_t *, svint8x2_t);
		30651	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_u64_x2)))
		30652	void svst1(svcount_t, uint64_t *, svuint64x2_t);
		30653	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_f64_x2)))
		30654	void svst1(svcount_t, float64_t *, svfloat64x2_t);
		30655	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_s64_x2)))
		30656	void svst1(svcount_t, int64_t *, svint64x2_t);
		30657	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_u16_x2)))
		30658	void svst1(svcount_t, uint16_t *, svuint16x2_t);
		30659	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_bf16_x2)))
		30660	void svst1(svcount_t, bfloat16_t *, svbfloat16x2_t);
		30661	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_f16_x2)))
		30662	void svst1(svcount_t, float16_t *, svfloat16x2_t);
		30663	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_s16_x2)))
		30664	void svst1(svcount_t, int16_t *, svint16x2_t);
		30665	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_u32_x2)))
		30666	void svst1(svcount_t, uint32_t *, svuint32x2_t);
		30667	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_f32_x2)))
		30668	void svst1(svcount_t, float32_t *, svfloat32x2_t);
		30669	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_s32_x2)))
		30670	void svst1(svcount_t, int32_t *, svint32x2_t);
		30671	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_u8_x4)))
		30672	void svst1(svcount_t, uint8_t *, svuint8x4_t);
		30673	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_s8_x4)))
		30674	void svst1(svcount_t, int8_t *, svint8x4_t);
		30675	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_u64_x4)))
		30676	void svst1(svcount_t, uint64_t *, svuint64x4_t);
		30677	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_f64_x4)))
		30678	void svst1(svcount_t, float64_t *, svfloat64x4_t);
		30679	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_s64_x4)))
		30680	void svst1(svcount_t, int64_t *, svint64x4_t);
		30681	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_u16_x4)))
		30682	void svst1(svcount_t, uint16_t *, svuint16x4_t);
		30683	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_bf16_x4)))
		30684	void svst1(svcount_t, bfloat16_t *, svbfloat16x4_t);
		30685	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_f16_x4)))
		30686	void svst1(svcount_t, float16_t *, svfloat16x4_t);
		30687	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_s16_x4)))
		30688	void svst1(svcount_t, int16_t *, svint16x4_t);
		30689	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_u32_x4)))
		30690	void svst1(svcount_t, uint32_t *, svuint32x4_t);
		30691	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_f32_x4)))
		30692	void svst1(svcount_t, float32_t *, svfloat32x4_t);
		30693	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_s32_x4)))
		30694	void svst1(svcount_t, int32_t *, svint32x4_t);
		30695	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_u8_x2)))
		30696	void svst1_vnum(svcount_t, uint8_t *, int64_t, svuint8x2_t);
		30697	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_s8_x2)))
		30698	void svst1_vnum(svcount_t, int8_t *, int64_t, svint8x2_t);
		30699	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_u64_x2)))
		30700	void svst1_vnum(svcount_t, uint64_t *, int64_t, svuint64x2_t);
		30701	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_f64_x2)))
		30702	void svst1_vnum(svcount_t, float64_t *, int64_t, svfloat64x2_t);
		30703	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_s64_x2)))
		30704	void svst1_vnum(svcount_t, int64_t *, int64_t, svint64x2_t);
		30705	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_u16_x2)))
		30706	void svst1_vnum(svcount_t, uint16_t *, int64_t, svuint16x2_t);
		30707	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_bf16_x2)))
		30708	void svst1_vnum(svcount_t, bfloat16_t *, int64_t, svbfloat16x2_t);
		30709	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_f16_x2)))
		30710	void svst1_vnum(svcount_t, float16_t *, int64_t, svfloat16x2_t);
		30711	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_s16_x2)))
		30712	void svst1_vnum(svcount_t, int16_t *, int64_t, svint16x2_t);
		30713	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_u32_x2)))
		30714	void svst1_vnum(svcount_t, uint32_t *, int64_t, svuint32x2_t);
		30715	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_f32_x2)))
		30716	void svst1_vnum(svcount_t, float32_t *, int64_t, svfloat32x2_t);
		30717	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_s32_x2)))
		30718	void svst1_vnum(svcount_t, int32_t *, int64_t, svint32x2_t);
		30719	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_u8_x4)))
		30720	void svst1_vnum(svcount_t, uint8_t *, int64_t, svuint8x4_t);
		30721	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_s8_x4)))
		30722	void svst1_vnum(svcount_t, int8_t *, int64_t, svint8x4_t);
		30723	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_u64_x4)))
		30724	void svst1_vnum(svcount_t, uint64_t *, int64_t, svuint64x4_t);
		30725	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_f64_x4)))
		30726	void svst1_vnum(svcount_t, float64_t *, int64_t, svfloat64x4_t);
		30727	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_s64_x4)))
		30728	void svst1_vnum(svcount_t, int64_t *, int64_t, svint64x4_t);
		30729	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_u16_x4)))
		30730	void svst1_vnum(svcount_t, uint16_t *, int64_t, svuint16x4_t);
		30731	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_bf16_x4)))
		30732	void svst1_vnum(svcount_t, bfloat16_t *, int64_t, svbfloat16x4_t);
		30733	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_f16_x4)))
		30734	void svst1_vnum(svcount_t, float16_t *, int64_t, svfloat16x4_t);
		30735	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_s16_x4)))
		30736	void svst1_vnum(svcount_t, int16_t *, int64_t, svint16x4_t);
		30737	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_u32_x4)))
		30738	void svst1_vnum(svcount_t, uint32_t *, int64_t, svuint32x4_t);
		30739	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_f32_x4)))
		30740	void svst1_vnum(svcount_t, float32_t *, int64_t, svfloat32x4_t);
		30741	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svst1_vnum_s32_x4)))
		30742	void svst1_vnum(svcount_t, int32_t *, int64_t, svint32x4_t);
		30743	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_u8_x2)))
		30744	void svstnt1(svcount_t, uint8_t *, svuint8x2_t);
		30745	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_s8_x2)))
		30746	void svstnt1(svcount_t, int8_t *, svint8x2_t);
		30747	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_u64_x2)))
		30748	void svstnt1(svcount_t, uint64_t *, svuint64x2_t);
		30749	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_f64_x2)))
		30750	void svstnt1(svcount_t, float64_t *, svfloat64x2_t);
		30751	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_s64_x2)))
		30752	void svstnt1(svcount_t, int64_t *, svint64x2_t);
		30753	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_u16_x2)))
		30754	void svstnt1(svcount_t, uint16_t *, svuint16x2_t);
		30755	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_bf16_x2)))
		30756	void svstnt1(svcount_t, bfloat16_t *, svbfloat16x2_t);
		30757	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_f16_x2)))
		30758	void svstnt1(svcount_t, float16_t *, svfloat16x2_t);
		30759	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_s16_x2)))
		30760	void svstnt1(svcount_t, int16_t *, svint16x2_t);
		30761	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_u32_x2)))
		30762	void svstnt1(svcount_t, uint32_t *, svuint32x2_t);
		30763	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_f32_x2)))
		30764	void svstnt1(svcount_t, float32_t *, svfloat32x2_t);
		30765	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_s32_x2)))
		30766	void svstnt1(svcount_t, int32_t *, svint32x2_t);
		30767	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_u8_x4)))
		30768	void svstnt1(svcount_t, uint8_t *, svuint8x4_t);
		30769	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_s8_x4)))
		30770	void svstnt1(svcount_t, int8_t *, svint8x4_t);
		30771	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_u64_x4)))
		30772	void svstnt1(svcount_t, uint64_t *, svuint64x4_t);
		30773	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_f64_x4)))
		30774	void svstnt1(svcount_t, float64_t *, svfloat64x4_t);
		30775	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_s64_x4)))
		30776	void svstnt1(svcount_t, int64_t *, svint64x4_t);
		30777	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_u16_x4)))
		30778	void svstnt1(svcount_t, uint16_t *, svuint16x4_t);
		30779	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_bf16_x4)))
		30780	void svstnt1(svcount_t, bfloat16_t *, svbfloat16x4_t);
		30781	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_f16_x4)))
		30782	void svstnt1(svcount_t, float16_t *, svfloat16x4_t);
		30783	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_s16_x4)))
		30784	void svstnt1(svcount_t, int16_t *, svint16x4_t);
		30785	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_u32_x4)))
		30786	void svstnt1(svcount_t, uint32_t *, svuint32x4_t);
		30787	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_f32_x4)))
		30788	void svstnt1(svcount_t, float32_t *, svfloat32x4_t);
		30789	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_s32_x4)))
		30790	void svstnt1(svcount_t, int32_t *, svint32x4_t);
		30791	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_u8_x2)))
		30792	void svstnt1_vnum(svcount_t, uint8_t *, int64_t, svuint8x2_t);
		30793	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_s8_x2)))
		30794	void svstnt1_vnum(svcount_t, int8_t *, int64_t, svint8x2_t);
		30795	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_u64_x2)))
		30796	void svstnt1_vnum(svcount_t, uint64_t *, int64_t, svuint64x2_t);
		30797	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_f64_x2)))
		30798	void svstnt1_vnum(svcount_t, float64_t *, int64_t, svfloat64x2_t);
		30799	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_s64_x2)))
		30800	void svstnt1_vnum(svcount_t, int64_t *, int64_t, svint64x2_t);
		30801	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_u16_x2)))
		30802	void svstnt1_vnum(svcount_t, uint16_t *, int64_t, svuint16x2_t);
		30803	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_bf16_x2)))
		30804	void svstnt1_vnum(svcount_t, bfloat16_t *, int64_t, svbfloat16x2_t);
		30805	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_f16_x2)))
		30806	void svstnt1_vnum(svcount_t, float16_t *, int64_t, svfloat16x2_t);
		30807	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_s16_x2)))
		30808	void svstnt1_vnum(svcount_t, int16_t *, int64_t, svint16x2_t);
		30809	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_u32_x2)))
		30810	void svstnt1_vnum(svcount_t, uint32_t *, int64_t, svuint32x2_t);
		30811	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_f32_x2)))
		30812	void svstnt1_vnum(svcount_t, float32_t *, int64_t, svfloat32x2_t);
		30813	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_s32_x2)))
		30814	void svstnt1_vnum(svcount_t, int32_t *, int64_t, svint32x2_t);
		30815	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_u8_x4)))
		30816	void svstnt1_vnum(svcount_t, uint8_t *, int64_t, svuint8x4_t);
		30817	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_s8_x4)))
		30818	void svstnt1_vnum(svcount_t, int8_t *, int64_t, svint8x4_t);
		30819	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_u64_x4)))
		30820	void svstnt1_vnum(svcount_t, uint64_t *, int64_t, svuint64x4_t);
		30821	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_f64_x4)))
		30822	void svstnt1_vnum(svcount_t, float64_t *, int64_t, svfloat64x4_t);
		30823	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_s64_x4)))
		30824	void svstnt1_vnum(svcount_t, int64_t *, int64_t, svint64x4_t);
		30825	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_u16_x4)))
		30826	void svstnt1_vnum(svcount_t, uint16_t *, int64_t, svuint16x4_t);
		30827	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_bf16_x4)))
		30828	void svstnt1_vnum(svcount_t, bfloat16_t *, int64_t, svbfloat16x4_t);
		30829	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_f16_x4)))
		30830	void svstnt1_vnum(svcount_t, float16_t *, int64_t, svfloat16x4_t);
		30831	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_s16_x4)))
		30832	void svstnt1_vnum(svcount_t, int16_t *, int64_t, svint16x4_t);
		30833	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_u32_x4)))
		30834	void svstnt1_vnum(svcount_t, uint32_t *, int64_t, svuint32x4_t);
		30835	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_f32_x4)))
		30836	void svstnt1_vnum(svcount_t, float32_t *, int64_t, svfloat32x4_t);
		30837	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svstnt1_vnum_s32_x4)))
		30838	void svstnt1_vnum(svcount_t, int32_t *, int64_t, svint32x4_t);
		30839	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_c8_s64)))
		30840	svcount_t svwhilege_c8(int64_t, int64_t, uint64_t);
		30841	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_c32_s64)))
		30842	svcount_t svwhilege_c32(int64_t, int64_t, uint64_t);
		30843	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_c64_s64)))
		30844	svcount_t svwhilege_c64(int64_t, int64_t, uint64_t);
		30845	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_c16_s64)))
		30846	svcount_t svwhilege_c16(int64_t, int64_t, uint64_t);
		30847	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_c8_u64)))
		30848	svcount_t svwhilege_c8(uint64_t, uint64_t, uint64_t);
		30849	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_c32_u64)))
		30850	svcount_t svwhilege_c32(uint64_t, uint64_t, uint64_t);
		30851	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_c64_u64)))
		30852	svcount_t svwhilege_c64(uint64_t, uint64_t, uint64_t);
		30853	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_c16_u64)))
		30854	svcount_t svwhilege_c16(uint64_t, uint64_t, uint64_t);
		30855	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_b8_s64_x2)))
		30856	svboolx2_t svwhilege_b8_x2(int64_t, int64_t);
		30857	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_b32_s64_x2)))
		30858	svboolx2_t svwhilege_b32_x2(int64_t, int64_t);
		30859	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_b64_s64_x2)))
		30860	svboolx2_t svwhilege_b64_x2(int64_t, int64_t);
		30861	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_b16_s64_x2)))
		30862	svboolx2_t svwhilege_b16_x2(int64_t, int64_t);
		30863	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_b8_u64_x2)))
		30864	svboolx2_t svwhilege_b8_x2(uint64_t, uint64_t);
		30865	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_b32_u64_x2)))
		30866	svboolx2_t svwhilege_b32_x2(uint64_t, uint64_t);
		30867	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_b64_u64_x2)))
		30868	svboolx2_t svwhilege_b64_x2(uint64_t, uint64_t);
		30869	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilege_b16_u64_x2)))
		30870	svboolx2_t svwhilege_b16_x2(uint64_t, uint64_t);
		30871	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_c8_s64)))
		30872	svcount_t svwhilegt_c8(int64_t, int64_t, uint64_t);
		30873	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_c32_s64)))
		30874	svcount_t svwhilegt_c32(int64_t, int64_t, uint64_t);
		30875	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_c64_s64)))
		30876	svcount_t svwhilegt_c64(int64_t, int64_t, uint64_t);
		30877	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_c16_s64)))
		30878	svcount_t svwhilegt_c16(int64_t, int64_t, uint64_t);
		30879	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_c8_u64)))
		30880	svcount_t svwhilegt_c8(uint64_t, uint64_t, uint64_t);
		30881	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_c32_u64)))
		30882	svcount_t svwhilegt_c32(uint64_t, uint64_t, uint64_t);
		30883	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_c64_u64)))
		30884	svcount_t svwhilegt_c64(uint64_t, uint64_t, uint64_t);
		30885	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_c16_u64)))
		30886	svcount_t svwhilegt_c16(uint64_t, uint64_t, uint64_t);
		30887	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_b8_s64_x2)))
		30888	svboolx2_t svwhilegt_b8_x2(int64_t, int64_t);
		30889	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_b32_s64_x2)))
		30890	svboolx2_t svwhilegt_b32_x2(int64_t, int64_t);
		30891	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_b64_s64_x2)))
		30892	svboolx2_t svwhilegt_b64_x2(int64_t, int64_t);
		30893	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_b16_s64_x2)))
		30894	svboolx2_t svwhilegt_b16_x2(int64_t, int64_t);
		30895	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_b8_u64_x2)))
		30896	svboolx2_t svwhilegt_b8_x2(uint64_t, uint64_t);
		30897	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_b32_u64_x2)))
		30898	svboolx2_t svwhilegt_b32_x2(uint64_t, uint64_t);
		30899	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_b64_u64_x2)))
		30900	svboolx2_t svwhilegt_b64_x2(uint64_t, uint64_t);
		30901	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilegt_b16_u64_x2)))
		30902	svboolx2_t svwhilegt_b16_x2(uint64_t, uint64_t);
		30903	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_c8_s64)))
		30904	svcount_t svwhilele_c8(int64_t, int64_t, uint64_t);
		30905	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_c32_s64)))
		30906	svcount_t svwhilele_c32(int64_t, int64_t, uint64_t);
		30907	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_c64_s64)))
		30908	svcount_t svwhilele_c64(int64_t, int64_t, uint64_t);
		30909	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_c16_s64)))
		30910	svcount_t svwhilele_c16(int64_t, int64_t, uint64_t);
		30911	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_c8_u64)))
		30912	svcount_t svwhilele_c8(uint64_t, uint64_t, uint64_t);
		30913	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_c32_u64)))
		30914	svcount_t svwhilele_c32(uint64_t, uint64_t, uint64_t);
		30915	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_c64_u64)))
		30916	svcount_t svwhilele_c64(uint64_t, uint64_t, uint64_t);
		30917	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_c16_u64)))
		30918	svcount_t svwhilele_c16(uint64_t, uint64_t, uint64_t);
		30919	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_b8_s64_x2)))
		30920	svboolx2_t svwhilele_b8_x2(int64_t, int64_t);
		30921	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_b32_s64_x2)))
		30922	svboolx2_t svwhilele_b32_x2(int64_t, int64_t);
		30923	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_b64_s64_x2)))
		30924	svboolx2_t svwhilele_b64_x2(int64_t, int64_t);
		30925	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_b16_s64_x2)))
		30926	svboolx2_t svwhilele_b16_x2(int64_t, int64_t);
		30927	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_b8_u64_x2)))
		30928	svboolx2_t svwhilele_b8_x2(uint64_t, uint64_t);
		30929	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_b32_u64_x2)))
		30930	svboolx2_t svwhilele_b32_x2(uint64_t, uint64_t);
		30931	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_b64_u64_x2)))
		30932	svboolx2_t svwhilele_b64_x2(uint64_t, uint64_t);
		30933	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilele_b16_u64_x2)))
		30934	svboolx2_t svwhilele_b16_x2(uint64_t, uint64_t);
		30935	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_c8_u64)))
		30936	svcount_t svwhilelt_c8(uint64_t, uint64_t, uint64_t);
		30937	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_c32_u64)))
		30938	svcount_t svwhilelt_c32(uint64_t, uint64_t, uint64_t);
		30939	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_c64_u64)))
		30940	svcount_t svwhilelt_c64(uint64_t, uint64_t, uint64_t);
		30941	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_c16_u64)))
		30942	svcount_t svwhilelt_c16(uint64_t, uint64_t, uint64_t);
		30943	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_c8_s64)))
		30944	svcount_t svwhilelt_c8(int64_t, int64_t, uint64_t);
		30945	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_c32_s64)))
		30946	svcount_t svwhilelt_c32(int64_t, int64_t, uint64_t);
		30947	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_c64_s64)))
		30948	svcount_t svwhilelt_c64(int64_t, int64_t, uint64_t);
		30949	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_c16_s64)))
		30950	svcount_t svwhilelt_c16(int64_t, int64_t, uint64_t);
		30951	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_b8_u64_x2)))
		30952	svboolx2_t svwhilelt_b8_x2(uint64_t, uint64_t);
		30953	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_b32_u64_x2)))
		30954	svboolx2_t svwhilelt_b32_x2(uint64_t, uint64_t);
		30955	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_b64_u64_x2)))
		30956	svboolx2_t svwhilelt_b64_x2(uint64_t, uint64_t);
		30957	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_b16_u64_x2)))
		30958	svboolx2_t svwhilelt_b16_x2(uint64_t, uint64_t);
		30959	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_b8_s64_x2)))
		30960	svboolx2_t svwhilelt_b8_x2(int64_t, int64_t);
		30961	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_b32_s64_x2)))
		30962	svboolx2_t svwhilelt_b32_x2(int64_t, int64_t);
		30963	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_b64_s64_x2)))
		30964	svboolx2_t svwhilelt_b64_x2(int64_t, int64_t);
		30965	__aio __attribute__((__clang_arm_builtin_alias(__builtin_sve_svwhilelt_b16_s64_x2)))
		30966	svboolx2_t svwhilelt_b16_x2(int64_t, int64_t);
23932	#define svcvtnt_bf16_x svcvtnt_bf16_m	30967	#define svcvtnt_bf16_x svcvtnt_bf16_m
23933	#define svcvtnt_bf16_f32_x svcvtnt_bf16_f32_m	30968	#define svcvtnt_bf16_f32_x svcvtnt_bf16_f32_m
23934	#define svcvtnt_f16_x svcvtnt_f16_m	30969	#define svcvtnt_f16_x svcvtnt_f16_m

lib/include/arm_vector_types.h created+345

...	@@ -0,0 +1,345 @@
1	/*===---- arm_vector_types - ARM vector type ------===
2	*
3	*
4	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
5	* See https://llvm.org/LICENSE.txt for license information.
6	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7	*
8	*===-----------------------------------------------------------------------===
9	*/
10
11	#if !defined(__ARM_NEON_H) && !defined(__ARM_SVE_H)
12	#error "This file should not be used standalone. Please include arm_neon.h or arm_sve.h instead"
13
14	#endif
15	#ifndef __ARM_NEON_TYPES_H
16	#define __ARM_NEON_TYPES_H
17	typedef float float32_t;
18	typedef __fp16 float16_t;
19	#ifdef __aarch64__
20	typedef double float64_t;
21	#endif
22
23	typedef __attribute__((neon_vector_type(8))) int8_t int8x8_t;
24	typedef __attribute__((neon_vector_type(16))) int8_t int8x16_t;
25	typedef __attribute__((neon_vector_type(4))) int16_t int16x4_t;
26	typedef __attribute__((neon_vector_type(8))) int16_t int16x8_t;
27	typedef __attribute__((neon_vector_type(2))) int32_t int32x2_t;
28	typedef __attribute__((neon_vector_type(4))) int32_t int32x4_t;
29	typedef __attribute__((neon_vector_type(1))) int64_t int64x1_t;
30	typedef __attribute__((neon_vector_type(2))) int64_t int64x2_t;
31	typedef __attribute__((neon_vector_type(8))) uint8_t uint8x8_t;
32	typedef __attribute__((neon_vector_type(16))) uint8_t uint8x16_t;
33	typedef __attribute__((neon_vector_type(4))) uint16_t uint16x4_t;
34	typedef __attribute__((neon_vector_type(8))) uint16_t uint16x8_t;
35	typedef __attribute__((neon_vector_type(2))) uint32_t uint32x2_t;
36	typedef __attribute__((neon_vector_type(4))) uint32_t uint32x4_t;
37	typedef __attribute__((neon_vector_type(1))) uint64_t uint64x1_t;
38	typedef __attribute__((neon_vector_type(2))) uint64_t uint64x2_t;
39	typedef __attribute__((neon_vector_type(4))) float16_t float16x4_t;
40	typedef __attribute__((neon_vector_type(8))) float16_t float16x8_t;
41	typedef __attribute__((neon_vector_type(2))) float32_t float32x2_t;
42	typedef __attribute__((neon_vector_type(4))) float32_t float32x4_t;
43	#ifdef __aarch64__
44	typedef __attribute__((neon_vector_type(1))) float64_t float64x1_t;
45	typedef __attribute__((neon_vector_type(2))) float64_t float64x2_t;
46	#endif
47
48	typedef struct int8x8x2_t {
49	int8x8_t val[2];
50	} int8x8x2_t;
51
52	typedef struct int8x16x2_t {
53	int8x16_t val[2];
54	} int8x16x2_t;
55
56	typedef struct int16x4x2_t {
57	int16x4_t val[2];
58	} int16x4x2_t;
59
60	typedef struct int16x8x2_t {
61	int16x8_t val[2];
62	} int16x8x2_t;
63
64	typedef struct int32x2x2_t {
65	int32x2_t val[2];
66	} int32x2x2_t;
67
68	typedef struct int32x4x2_t {
69	int32x4_t val[2];
70	} int32x4x2_t;
71
72	typedef struct int64x1x2_t {
73	int64x1_t val[2];
74	} int64x1x2_t;
75
76	typedef struct int64x2x2_t {
77	int64x2_t val[2];
78	} int64x2x2_t;
79
80	typedef struct uint8x8x2_t {
81	uint8x8_t val[2];
82	} uint8x8x2_t;
83
84	typedef struct uint8x16x2_t {
85	uint8x16_t val[2];
86	} uint8x16x2_t;
87
88	typedef struct uint16x4x2_t {
89	uint16x4_t val[2];
90	} uint16x4x2_t;
91
92	typedef struct uint16x8x2_t {
93	uint16x8_t val[2];
94	} uint16x8x2_t;
95
96	typedef struct uint32x2x2_t {
97	uint32x2_t val[2];
98	} uint32x2x2_t;
99
100	typedef struct uint32x4x2_t {
101	uint32x4_t val[2];
102	} uint32x4x2_t;
103
104	typedef struct uint64x1x2_t {
105	uint64x1_t val[2];
106	} uint64x1x2_t;
107
108	typedef struct uint64x2x2_t {
109	uint64x2_t val[2];
110	} uint64x2x2_t;
111
112	typedef struct float16x4x2_t {
113	float16x4_t val[2];
114	} float16x4x2_t;
115
116	typedef struct float16x8x2_t {
117	float16x8_t val[2];
118	} float16x8x2_t;
119
120	typedef struct float32x2x2_t {
121	float32x2_t val[2];
122	} float32x2x2_t;
123
124	typedef struct float32x4x2_t {
125	float32x4_t val[2];
126	} float32x4x2_t;
127
128	#ifdef __aarch64__
129	typedef struct float64x1x2_t {
130	float64x1_t val[2];
131	} float64x1x2_t;
132
133	typedef struct float64x2x2_t {
134	float64x2_t val[2];
135	} float64x2x2_t;
136
137	#endif
138	typedef struct int8x8x3_t {
139	int8x8_t val[3];
140	} int8x8x3_t;
141
142	typedef struct int8x16x3_t {
143	int8x16_t val[3];
144	} int8x16x3_t;
145
146	typedef struct int16x4x3_t {
147	int16x4_t val[3];
148	} int16x4x3_t;
149
150	typedef struct int16x8x3_t {
151	int16x8_t val[3];
152	} int16x8x3_t;
153
154	typedef struct int32x2x3_t {
155	int32x2_t val[3];
156	} int32x2x3_t;
157
158	typedef struct int32x4x3_t {
159	int32x4_t val[3];
160	} int32x4x3_t;
161
162	typedef struct int64x1x3_t {
163	int64x1_t val[3];
164	} int64x1x3_t;
165
166	typedef struct int64x2x3_t {
167	int64x2_t val[3];
168	} int64x2x3_t;
169
170	typedef struct uint8x8x3_t {
171	uint8x8_t val[3];
172	} uint8x8x3_t;
173
174	typedef struct uint8x16x3_t {
175	uint8x16_t val[3];
176	} uint8x16x3_t;
177
178	typedef struct uint16x4x3_t {
179	uint16x4_t val[3];
180	} uint16x4x3_t;
181
182	typedef struct uint16x8x3_t {
183	uint16x8_t val[3];
184	} uint16x8x3_t;
185
186	typedef struct uint32x2x3_t {
187	uint32x2_t val[3];
188	} uint32x2x3_t;
189
190	typedef struct uint32x4x3_t {
191	uint32x4_t val[3];
192	} uint32x4x3_t;
193
194	typedef struct uint64x1x3_t {
195	uint64x1_t val[3];
196	} uint64x1x3_t;
197
198	typedef struct uint64x2x3_t {
199	uint64x2_t val[3];
200	} uint64x2x3_t;
201
202	typedef struct float16x4x3_t {
203	float16x4_t val[3];
204	} float16x4x3_t;
205
206	typedef struct float16x8x3_t {
207	float16x8_t val[3];
208	} float16x8x3_t;
209
210	typedef struct float32x2x3_t {
211	float32x2_t val[3];
212	} float32x2x3_t;
213
214	typedef struct float32x4x3_t {
215	float32x4_t val[3];
216	} float32x4x3_t;
217
218	#ifdef __aarch64__
219	typedef struct float64x1x3_t {
220	float64x1_t val[3];
221	} float64x1x3_t;
222
223	typedef struct float64x2x3_t {
224	float64x2_t val[3];
225	} float64x2x3_t;
226
227	#endif
228	typedef struct int8x8x4_t {
229	int8x8_t val[4];
230	} int8x8x4_t;
231
232	typedef struct int8x16x4_t {
233	int8x16_t val[4];
234	} int8x16x4_t;
235
236	typedef struct int16x4x4_t {
237	int16x4_t val[4];
238	} int16x4x4_t;
239
240	typedef struct int16x8x4_t {
241	int16x8_t val[4];
242	} int16x8x4_t;
243
244	typedef struct int32x2x4_t {
245	int32x2_t val[4];
246	} int32x2x4_t;
247
248	typedef struct int32x4x4_t {
249	int32x4_t val[4];
250	} int32x4x4_t;
251
252	typedef struct int64x1x4_t {
253	int64x1_t val[4];
254	} int64x1x4_t;
255
256	typedef struct int64x2x4_t {
257	int64x2_t val[4];
258	} int64x2x4_t;
259
260	typedef struct uint8x8x4_t {
261	uint8x8_t val[4];
262	} uint8x8x4_t;
263
264	typedef struct uint8x16x4_t {
265	uint8x16_t val[4];
266	} uint8x16x4_t;
267
268	typedef struct uint16x4x4_t {
269	uint16x4_t val[4];
270	} uint16x4x4_t;
271
272	typedef struct uint16x8x4_t {
273	uint16x8_t val[4];
274	} uint16x8x4_t;
275
276	typedef struct uint32x2x4_t {
277	uint32x2_t val[4];
278	} uint32x2x4_t;
279
280	typedef struct uint32x4x4_t {
281	uint32x4_t val[4];
282	} uint32x4x4_t;
283
284	typedef struct uint64x1x4_t {
285	uint64x1_t val[4];
286	} uint64x1x4_t;
287
288	typedef struct uint64x2x4_t {
289	uint64x2_t val[4];
290	} uint64x2x4_t;
291
292	typedef struct float16x4x4_t {
293	float16x4_t val[4];
294	} float16x4x4_t;
295
296	typedef struct float16x8x4_t {
297	float16x8_t val[4];
298	} float16x8x4_t;
299
300	typedef struct float32x2x4_t {
301	float32x2_t val[4];
302	} float32x2x4_t;
303
304	typedef struct float32x4x4_t {
305	float32x4_t val[4];
306	} float32x4x4_t;
307
308	#ifdef __aarch64__
309	typedef struct float64x1x4_t {
310	float64x1_t val[4];
311	} float64x1x4_t;
312
313	typedef struct float64x2x4_t {
314	float64x2_t val[4];
315	} float64x2x4_t;
316
317	#endif
318	typedef __attribute__((neon_vector_type(4))) bfloat16_t bfloat16x4_t;
319	typedef __attribute__((neon_vector_type(8))) bfloat16_t bfloat16x8_t;
320
321	typedef struct bfloat16x4x2_t {
322	bfloat16x4_t val[2];
323	} bfloat16x4x2_t;
324
325	typedef struct bfloat16x8x2_t {
326	bfloat16x8_t val[2];
327	} bfloat16x8x2_t;
328
329	typedef struct bfloat16x4x3_t {
330	bfloat16x4_t val[3];
331	} bfloat16x4x3_t;
332
333	typedef struct bfloat16x8x3_t {
334	bfloat16x8_t val[3];
335	} bfloat16x8x3_t;
336
337	typedef struct bfloat16x4x4_t {
338	bfloat16x4_t val[4];
339	} bfloat16x4x4_t;
340
341	typedef struct bfloat16x8x4_t {
342	bfloat16x8_t val[4];
343	} bfloat16x8x4_t;
344
345	#endif // __ARM_NEON_TYPES_H

lib/include/avx2intrin.h+24-3

...	@@ -15,8 +15,12 @@	...	@@ -15,8 +15,12 @@
15	#define __AVX2INTRIN_H	15	#define __AVX2INTRIN_H
16		16
17	/* Define the default attributes for the functions in this file. */	17	/* Define the default attributes for the functions in this file. */
18	#define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("avx2"), __min_vector_width__(256)))	18	#define __DEFAULT_FN_ATTRS256 \
19	#define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx2"), __min_vector_width__(128)))	19	__attribute__((__always_inline__, __nodebug__, \
		20	__target__("avx2,no-evex512"), __min_vector_width__(256)))
		21	#define __DEFAULT_FN_ATTRS128 \
		22	__attribute__((__always_inline__, __nodebug__, \
		23	__target__("avx2,no-evex512"), __min_vector_width__(128)))
20		24
21	/* SSE4 Multiple Packed Sums of Absolute Difference. */	25	/* SSE4 Multiple Packed Sums of Absolute Difference. */
22	/// Computes sixteen sum of absolute difference (SAD) operations on sets of	26	/// Computes sixteen sum of absolute difference (SAD) operations on sets of
...	@@ -1307,6 +1311,23 @@ _mm256_min_epu32(__m256i __a, __m256i __b)	...	@@ -1307,6 +1311,23 @@ _mm256_min_epu32(__m256i __a, __m256i __b)
1307	return (__m256i)__builtin_elementwise_min((__v8su)__a, (__v8su)__b);	1311	return (__m256i)__builtin_elementwise_min((__v8su)__a, (__v8su)__b);
1308	}	1312	}
1309		1313
		1314	/// Creates a 32-bit integer mask from the most significant bit of each byte
		1315	/// in the 256-bit integer vector in \a __a and returns the result.
		1316	///
		1317	/// \code{.operation}
		1318	/// FOR i := 0 TO 31
		1319	/// j := i*8
		1320	/// result[i] := __a[j+7]
		1321	/// ENDFOR
		1322	/// \endcode
		1323	///
		1324	/// \headerfile <immintrin.h>
		1325	///
		1326	/// This intrinsic corresponds to the \c VPMOVMSKB instruction.
		1327	///
		1328	/// \param __a
		1329	/// A 256-bit integer vector containing the source bytes.
		1330	/// \returns The 32-bit integer mask.
1310	static __inline__ int __DEFAULT_FN_ATTRS256	1331	static __inline__ int __DEFAULT_FN_ATTRS256
1311	_mm256_movemask_epi8(__m256i __a)	1332	_mm256_movemask_epi8(__m256i __a)
1312	{	1333	{
...	@@ -2962,7 +2983,7 @@ _mm256_xor_si256(__m256i __a, __m256i __b)	...	@@ -2962,7 +2983,7 @@ _mm256_xor_si256(__m256i __a, __m256i __b)
2962	/// A pointer to the 32-byte aligned memory containing the vector to load.	2983	/// A pointer to the 32-byte aligned memory containing the vector to load.
2963	/// \returns A 256-bit integer vector loaded from memory.	2984	/// \returns A 256-bit integer vector loaded from memory.
2964	static __inline__ __m256i __DEFAULT_FN_ATTRS256	2985	static __inline__ __m256i __DEFAULT_FN_ATTRS256
2965	_mm256_stream_load_si256(__m256i const *__V)	2986	_mm256_stream_load_si256(const void *__V)
2966	{	2987	{
2967	typedef __v4di __v4di_aligned __attribute__((aligned(32)));	2988	typedef __v4di __v4di_aligned __attribute__((aligned(32)));
2968	return (__m256i)__builtin_nontemporal_load((const __v4di_aligned *)__V);	2989	return (__m256i)__builtin_nontemporal_load((const __v4di_aligned *)__V);

lib/include/avx512bf16intrin.h+3-2

 @@ -20,10 +20,11 @@ typedef __bf16 __m512bh __attribute__((__vector_size__(64), __aligned__(64)));
 typedef __bf16 __bfloat16 __attribute__((deprecated("use __bf16 instead")));
 #define __DEFAULT_FN_ATTRS512 \
-  __attribute__((__always_inline__, __nodebug__, __target__("avx512bf16"), \
+  __attribute__((__always_inline__, __nodebug__, __target__("avx512bf16,evex512"), \
                  __min_vector_width__(512)))
 #define __DEFAULT_FN_ATTRS                                                     \
-  __attribute__((__always_inline__, __nodebug__, __target__("avx512bf16")))
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512bf16,no-evex512")))
 /// Convert One BF16 Data to One Single Float Data.
 ///

lib/include/avx512bitalgintrin.h+4-1

 @@ -15,7 +15,10 @@
 #define __AVX512BITALGINTRIN_H
 /* Define the default attributes for the functions in this file. */
-#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512bitalg"), __min_vector_width__(512)))
+#define __DEFAULT_FN_ATTRS                                                     \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512bitalg,evex512"),                           \
+                 __min_vector_width__(512)))
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_popcnt_epi16(__m512i __A)

lib/include/avx512bwintrin.h+28-38

 @@ -18,8 +18,12 @@ typedef unsigned int __mmask32;
 typedef unsigned long long __mmask64;
 /* Define the default attributes for the functions in this file. */
-#define __DEFAULT_FN_ATTRS512 __attribute__((__always_inline__, __nodebug__, __target__("avx512bw"), __min_vector_width__(512)))
+#define __DEFAULT_FN_ATTRS512                                                  \
-#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512bw")))
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512bw,evex512"), __min_vector_width__(512)))
+#define __DEFAULT_FN_ATTRS                                                     \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512bw,no-evex512")))
 static __inline __mmask32 __DEFAULT_FN_ATTRS
 _knot_mask32(__mmask32 __M)
 @@ -27,9 +31,7 @@ _knot_mask32(__mmask32 __M)
   return __builtin_ia32_knotsi(__M);
 }
-static __inline __mmask64 __DEFAULT_FN_ATTRS
+static __inline __mmask64 __DEFAULT_FN_ATTRS _knot_mask64(__mmask64 __M) {
-_knot_mask64(__mmask64 __M)
-{
   return __builtin_ia32_knotdi(__M);
 }
 @@ -39,9 +41,8 @@ _kand_mask32(__mmask32 __A, __mmask32 __B)
   return (__mmask32)__builtin_ia32_kandsi((__mmask32)__A, (__mmask32)__B);
 }
-static __inline__ __mmask64 __DEFAULT_FN_ATTRS
+static __inline__ __mmask64 __DEFAULT_FN_ATTRS _kand_mask64(__mmask64 __A,
-_kand_mask64(__mmask64 __A, __mmask64 __B)
+                                                            __mmask64 __B) {
-{
   return (__mmask64)__builtin_ia32_kanddi((__mmask64)__A, (__mmask64)__B);
 }
 @@ -51,9 +52,8 @@ _kandn_mask32(__mmask32 __A, __mmask32 __B)
   return (__mmask32)__builtin_ia32_kandnsi((__mmask32)__A, (__mmask32)__B);
 }
-static __inline__ __mmask64 __DEFAULT_FN_ATTRS
+static __inline__ __mmask64 __DEFAULT_FN_ATTRS _kandn_mask64(__mmask64 __A,
-_kandn_mask64(__mmask64 __A, __mmask64 __B)
+                                                             __mmask64 __B) {
-{
   return (__mmask64)__builtin_ia32_kandndi((__mmask64)__A, (__mmask64)__B);
 }
 @@ -63,9 +63,8 @@ _kor_mask32(__mmask32 __A, __mmask32 __B)
   return (__mmask32)__builtin_ia32_korsi((__mmask32)__A, (__mmask32)__B);
 }
-static __inline__ __mmask64 __DEFAULT_FN_ATTRS
+static __inline__ __mmask64 __DEFAULT_FN_ATTRS _kor_mask64(__mmask64 __A,
-_kor_mask64(__mmask64 __A, __mmask64 __B)
+                                                           __mmask64 __B) {
-{
   return (__mmask64)__builtin_ia32_kordi((__mmask64)__A, (__mmask64)__B);
 }
 @@ -75,9 +74,8 @@ _kxnor_mask32(__mmask32 __A, __mmask32 __B)
   return (__mmask32)__builtin_ia32_kxnorsi((__mmask32)__A, (__mmask32)__B);
 }
-static __inline__ __mmask64 __DEFAULT_FN_ATTRS
+static __inline__ __mmask64 __DEFAULT_FN_ATTRS _kxnor_mask64(__mmask64 __A,
-_kxnor_mask64(__mmask64 __A, __mmask64 __B)
+                                                             __mmask64 __B) {
-{
   return (__mmask64)__builtin_ia32_kxnordi((__mmask64)__A, (__mmask64)__B);
 }
 @@ -87,9 +85,8 @@ _kxor_mask32(__mmask32 __A, __mmask32 __B)
   return (__mmask32)__builtin_ia32_kxorsi((__mmask32)__A, (__mmask32)__B);
 }
-static __inline__ __mmask64 __DEFAULT_FN_ATTRS
+static __inline__ __mmask64 __DEFAULT_FN_ATTRS _kxor_mask64(__mmask64 __A,
-_kxor_mask64(__mmask64 __A, __mmask64 __B)
+                                                            __mmask64 __B) {
-{
   return (__mmask64)__builtin_ia32_kxordi((__mmask64)__A, (__mmask64)__B);
 }
 @@ -112,14 +109,12 @@ _kortest_mask32_u8(__mmask32 __A, __mmask32 __B, unsigned char *__C) {
 }
 static __inline__ unsigned char __DEFAULT_FN_ATTRS
-_kortestc_mask64_u8(__mmask64 __A, __mmask64 __B)
+_kortestc_mask64_u8(__mmask64 __A, __mmask64 __B) {
-{
   return (unsigned char)__builtin_ia32_kortestcdi(__A, __B);
 }
 static __inline__ unsigned char __DEFAULT_FN_ATTRS
-_kortestz_mask64_u8(__mmask64 __A, __mmask64 __B)
+_kortestz_mask64_u8(__mmask64 __A, __mmask64 __B) {
-{
   return (unsigned char)__builtin_ia32_kortestzdi(__A, __B);
 }
 @@ -148,14 +143,12 @@ _ktest_mask32_u8(__mmask32 __A, __mmask32 __B, unsigned char *__C) {
 }
 static __inline__ unsigned char __DEFAULT_FN_ATTRS
-_ktestc_mask64_u8(__mmask64 __A, __mmask64 __B)
+_ktestc_mask64_u8(__mmask64 __A, __mmask64 __B) {
-{
   return (unsigned char)__builtin_ia32_ktestcdi(__A, __B);
 }
 static __inline__ unsigned char __DEFAULT_FN_ATTRS
-_ktestz_mask64_u8(__mmask64 __A, __mmask64 __B)
+_ktestz_mask64_u8(__mmask64 __A, __mmask64 __B) {
-{
   return (unsigned char)__builtin_ia32_ktestzdi(__A, __B);
 }
 @@ -171,9 +164,8 @@ _kadd_mask32(__mmask32 __A, __mmask32 __B)
   return (__mmask32)__builtin_ia32_kaddsi((__mmask32)__A, (__mmask32)__B);
 }
-static __inline__ __mmask64 __DEFAULT_FN_ATTRS
+static __inline__ __mmask64 __DEFAULT_FN_ATTRS _kadd_mask64(__mmask64 __A,
-_kadd_mask64(__mmask64 __A, __mmask64 __B)
+                                                            __mmask64 __B) {
-{
   return (__mmask64)__builtin_ia32_kadddi((__mmask64)__A, (__mmask64)__B);
 }
 @@ -214,8 +206,7 @@ _load_mask32(__mmask32 *__A) {
   return (__mmask32)__builtin_ia32_kmovd(*(__mmask32 *)__A);
 }
-static __inline__ __mmask64 __DEFAULT_FN_ATTRS
+static __inline__ __mmask64 __DEFAULT_FN_ATTRS _load_mask64(__mmask64 *__A) {
-_load_mask64(__mmask64 *__A) {
   return (__mmask64)__builtin_ia32_kmovq(*(__mmask64 *)__A);
 }
 @@ -224,8 +215,8 @@ _store_mask32(__mmask32 *__A, __mmask32 __B) {
   *(__mmask32 *)__A = __builtin_ia32_kmovd((__mmask32)__B);
 }
-static __inline__ void __DEFAULT_FN_ATTRS
+static __inline__ void __DEFAULT_FN_ATTRS _store_mask64(__mmask64 *__A,
-_store_mask64(__mmask64 *__A, __mmask64 __B) {
+                                                        __mmask64 __B) {
   *(__mmask64 *)__A = __builtin_ia32_kmovq((__mmask64)__B);
 }
 @@ -1714,9 +1705,8 @@ _mm512_maskz_set1_epi8 (__mmask64 __M, char __A)
                                               (__v64qi) _mm512_setzero_si512());
 }
-static __inline__ __mmask64 __DEFAULT_FN_ATTRS
+static __inline__ __mmask64 __DEFAULT_FN_ATTRS _mm512_kunpackd(__mmask64 __A,
-_mm512_kunpackd (__mmask64 __A, __mmask64 __B)
+                                                               __mmask64 __B) {
-{
   return (__mmask64) __builtin_ia32_kunpckdi ((__mmask64) __A,
                 (__mmask64) __B);
 }

lib/include/avx512cdintrin.h+3-1

 @@ -15,7 +15,9 @@
 #define __AVX512CDINTRIN_H
 /* Define the default attributes for the functions in this file. */
-#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512cd"), __min_vector_width__(512)))
+#define __DEFAULT_FN_ATTRS                                                     \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512cd,evex512"), __min_vector_width__(512)))
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_conflict_epi64 (__m512i __A)

lib/include/avx512dqintrin.h+4-2

 @@ -15,8 +15,10 @@
 #define __AVX512DQINTRIN_H
 /* Define the default attributes for the functions in this file. */
-#define __DEFAULT_FN_ATTRS512 __attribute__((__always_inline__, __nodebug__, __target__("avx512dq"), __min_vector_width__(512)))
+#define __DEFAULT_FN_ATTRS512 __attribute__((__always_inline__, __nodebug__, __target__("avx512dq,evex512"), __min_vector_width__(512)))
-#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512dq")))
+#define __DEFAULT_FN_ATTRS                                                     \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512dq,no-evex512")))
 static __inline __mmask8 __DEFAULT_FN_ATTRS
 _knot_mask8(__mmask8 __M)

lib/include/avx512fintrin.h+7-3

 @@ -167,9 +167,13 @@ typedef enum
 } _MM_MANTISSA_SIGN_ENUM;
 /* Define the default attributes for the functions in this file. */
-#define __DEFAULT_FN_ATTRS512 __attribute__((__always_inline__, __nodebug__, __target__("avx512f"), __min_vector_width__(512)))
+#define __DEFAULT_FN_ATTRS512 __attribute__((__always_inline__, __nodebug__, __target__("avx512f,evex512"), __min_vector_width__(512)))
-#define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx512f"), __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS128                                                  \
-#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512f")))
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512f,no-evex512"), __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS                                                     \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512f,no-evex512")))
 /* Create vectors with repeated elements */

lib/include/avx512fp16intrin.h+6-4

 @@ -22,13 +22,15 @@ typedef _Float16 __m512h_u __attribute__((__vector_size__(64), __aligned__(1)));
 /* Define the default attributes for the functions in this file. */
 #define __DEFAULT_FN_ATTRS512                                                  \
-  __attribute__((__always_inline__, __nodebug__, __target__("avx512fp16"),     \
+  __attribute__((__always_inline__, __nodebug__,                               \
-                 __min_vector_width__(512)))
+                 __target__("avx512fp16,evex512"), __min_vector_width__(512)))
 #define __DEFAULT_FN_ATTRS256                                                  \
-  __attribute__((__always_inline__, __nodebug__, __target__("avx512fp16"),     \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512fp16,no-evex512"),                          \
                  __min_vector_width__(256)))
 #define __DEFAULT_FN_ATTRS128                                                  \
-  __attribute__((__always_inline__, __nodebug__, __target__("avx512fp16"),     \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512fp16,no-evex512"),                          \
                  __min_vector_width__(128)))
 static __inline__ _Float16 __DEFAULT_FN_ATTRS512 _mm512_cvtsh_h(__m512h __a) {

lib/include/avx512ifmaintrin.h+3-1

 @@ -15,7 +15,9 @@
 #define __IFMAINTRIN_H
 /* Define the default attributes for the functions in this file. */
-#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512ifma"), __min_vector_width__(512)))
+#define __DEFAULT_FN_ATTRS                                                     \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512ifma,evex512"), __min_vector_width__(512)))
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_madd52hi_epu64 (__m512i __X, __m512i __Y, __m512i __Z)

lib/include/avx512ifmavlintrin.h+8-2

 @@ -15,8 +15,14 @@
 #define __IFMAVLINTRIN_H
 /* Define the default attributes for the functions in this file. */
-#define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx512ifma,avx512vl"), __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS128                                                  \
-#define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("avx512ifma,avx512vl"), __min_vector_width__(256)))
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512ifma,avx512vl,no-evex512"),                 \
+                 __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS256                                                  \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512ifma,avx512vl,no-evex512"),                 \
+                 __min_vector_width__(256)))
 #define _mm_madd52hi_epu64(X, Y, Z)                                            \
   ((__m128i)__builtin_ia32_vpmadd52huq128((__v2di)(X), (__v2di)(Y),            \

lib/include/avx512pfintrin.h-5

...	@@ -14,9 +14,6 @@	...	@@ -14,9 +14,6 @@
14	#ifndef __AVX512PFINTRIN_H	14	#ifndef __AVX512PFINTRIN_H
15	#define __AVX512PFINTRIN_H	15	#define __AVX512PFINTRIN_H
16		16
17	/* Define the default attributes for the functions in this file. */
18	#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512pf")))
19
20	#define _mm512_mask_prefetch_i32gather_pd(index, mask, addr, scale, hint) \	17	#define _mm512_mask_prefetch_i32gather_pd(index, mask, addr, scale, hint) \
21	__builtin_ia32_gatherpfdpd((__mmask8)(mask), (__v8si)(__m256i)(index), \	18	__builtin_ia32_gatherpfdpd((__mmask8)(mask), (__v8si)(__m256i)(index), \
22	(void const *)(addr), (int)(scale), \	19	(void const *)(addr), (int)(scale), \
...	@@ -92,6 +89,4 @@	...	@@ -92,6 +89,4 @@
92	__builtin_ia32_scatterpfqps((__mmask8)(mask), (__v8di)(__m512i)(index), \	89	__builtin_ia32_scatterpfqps((__mmask8)(mask), (__v8di)(__m512i)(index), \
93	(void *)(addr), (int)(scale), (int)(hint))	90	(void *)(addr), (int)(scale), (int)(hint))
94		91
95	#undef __DEFAULT_FN_ATTRS
96
97	#endif	92	#endif

lib/include/avx512vbmi2intrin.h+1-1

...	@@ -15,7 +15,7 @@	...	@@ -15,7 +15,7 @@
15	#define __AVX512VBMI2INTRIN_H	15	#define __AVX512VBMI2INTRIN_H
16		16
17	/* Define the default attributes for the functions in this file. */	17	/* Define the default attributes for the functions in this file. */
18	#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512vbmi2"), __min_vector_width__(512)))	18	#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512vbmi2,evex512"), __min_vector_width__(512)))
19		19
20		20
21	static __inline__ __m512i __DEFAULT_FN_ATTRS	21	static __inline__ __m512i __DEFAULT_FN_ATTRS

lib/include/avx512vbmiintrin.h+3-2

...	@@ -15,8 +15,9 @@	...	@@ -15,8 +15,9 @@
15	#define __VBMIINTRIN_H	15	#define __VBMIINTRIN_H
16		16
17	/* Define the default attributes for the functions in this file. */	17	/* Define the default attributes for the functions in this file. */
18	#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512vbmi"), __min_vector_width__(512)))	18	#define __DEFAULT_FN_ATTRS \
19		19	__attribute__((__always_inline__, __nodebug__, \
		20	__target__("avx512vbmi,evex512"), __min_vector_width__(512)))
20		21
21	static __inline__ __m512i __DEFAULT_FN_ATTRS	22	static __inline__ __m512i __DEFAULT_FN_ATTRS
22	_mm512_permutex2var_epi8(__m512i __A, __m512i __I, __m512i __B)	23	_mm512_permutex2var_epi8(__m512i __A, __m512i __I, __m512i __B)

lib/include/avx512vbmivlintrin.h+8-3

...	@@ -15,9 +15,14 @@	...	@@ -15,9 +15,14 @@
15	#define __VBMIVLINTRIN_H	15	#define __VBMIVLINTRIN_H
16		16
17	/* Define the default attributes for the functions in this file. */	17	/* Define the default attributes for the functions in this file. */
18	#define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx512vbmi,avx512vl"), __min_vector_width__(128)))	18	#define __DEFAULT_FN_ATTRS128 \
19	#define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("avx512vbmi,avx512vl"), __min_vector_width__(256)))	19	__attribute__((__always_inline__, __nodebug__, \
20		20	__target__("avx512vbmi,avx512vl,no-evex512"), \
		21	__min_vector_width__(128)))
		22	#define __DEFAULT_FN_ATTRS256 \
		23	__attribute__((__always_inline__, __nodebug__, \
		24	__target__("avx512vbmi,avx512vl,no-evex512"), \
		25	__min_vector_width__(256)))
21		26
22	static __inline__ __m128i __DEFAULT_FN_ATTRS128	27	static __inline__ __m128i __DEFAULT_FN_ATTRS128
23	_mm_permutex2var_epi8(__m128i __A, __m128i __I, __m128i __B)	28	_mm_permutex2var_epi8(__m128i __A, __m128i __I, __m128i __B)

lib/include/avx512vlbf16intrin.h+8-6

...	@@ -15,12 +15,14 @@	...	@@ -15,12 +15,14 @@
15	#ifndef __AVX512VLBF16INTRIN_H	15	#ifndef __AVX512VLBF16INTRIN_H
16	#define __AVX512VLBF16INTRIN_H	16	#define __AVX512VLBF16INTRIN_H
17		17
18	#define __DEFAULT_FN_ATTRS128 \	18	#define __DEFAULT_FN_ATTRS128 \
19	__attribute__((__always_inline__, __nodebug__, \	19	__attribute__((__always_inline__, __nodebug__, \
20	__target__("avx512vl, avx512bf16"), __min_vector_width__(128)))	20	__target__("avx512vl,avx512bf16,no-evex512"), \
21	#define __DEFAULT_FN_ATTRS256 \	21	__min_vector_width__(128)))
22	__attribute__((__always_inline__, __nodebug__, \	22	#define __DEFAULT_FN_ATTRS256 \
23	__target__("avx512vl, avx512bf16"), __min_vector_width__(256)))	23	__attribute__((__always_inline__, __nodebug__, \
		24	__target__("avx512vl,avx512bf16,no-evex512"), \
		25	__min_vector_width__(256)))
24		26
25	/// Convert Two Packed Single Data to One Packed BF16 Data.	27	/// Convert Two Packed Single Data to One Packed BF16 Data.
26	///	28	///

lib/include/avx512vlbitalgintrin.h+8-2

 @@ -15,8 +15,14 @@
 #define __AVX512VLBITALGINTRIN_H
 /* Define the default attributes for the functions in this file. */
-#define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512bitalg"), __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS128                                                  \
-#define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512bitalg"), __min_vector_width__(256)))
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vl,avx512bitalg,no-evex512"),               \
+                 __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS256                                                  \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vl,avx512bitalg,no-evex512"),               \
+                 __min_vector_width__(256)))
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_popcnt_epi16(__m256i __A)

lib/include/avx512vlbwintrin.h+8-2

 @@ -15,8 +15,14 @@
 #define __AVX512VLBWINTRIN_H
 /* Define the default attributes for the functions in this file. */
-#define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512bw"), __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS128                                                  \
-#define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512bw"), __min_vector_width__(256)))
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vl,avx512bw,no-evex512"),                   \
+                 __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS256                                                  \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vl,avx512bw,no-evex512"),                   \
+                 __min_vector_width__(256)))
 /* Integer compare */

lib/include/avx512vlcdintrin.h+8-3

...	@@ -14,9 +14,14 @@	...	@@ -14,9 +14,14 @@
14	#define __AVX512VLCDINTRIN_H	14	#define __AVX512VLCDINTRIN_H
15		15
16	/* Define the default attributes for the functions in this file. */	16	/* Define the default attributes for the functions in this file. */
17	#define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512cd"), __min_vector_width__(128)))	17	#define __DEFAULT_FN_ATTRS128 \
18	#define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512cd"), __min_vector_width__(256)))	18	__attribute__((__always_inline__, __nodebug__, \
19		19	__target__("avx512vl,avx512cd,no-evex512"), \
		20	__min_vector_width__(128)))
		21	#define __DEFAULT_FN_ATTRS256 \
		22	__attribute__((__always_inline__, __nodebug__, \
		23	__target__("avx512vl,avx512cd,no-evex512"), \
		24	__min_vector_width__(256)))
20		25
21	static __inline__ __m128i __DEFAULT_FN_ATTRS128	26	static __inline__ __m128i __DEFAULT_FN_ATTRS128
22	_mm_broadcastmb_epi64 (__mmask8 __A)	27	_mm_broadcastmb_epi64 (__mmask8 __A)

lib/include/avx512vldqintrin.h+8-2

 @@ -15,8 +15,14 @@
 #define __AVX512VLDQINTRIN_H
 /* Define the default attributes for the functions in this file. */
-#define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512dq"), __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS128                                                  \
-#define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512dq"), __min_vector_width__(256)))
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vl,avx512dq,no-evex512"),                   \
+                 __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS256                                                  \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vl,avx512dq,no-evex512"),                   \
+                 __min_vector_width__(256)))
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_mullo_epi64 (__m256i __A, __m256i __B) {

lib/include/avx512vlfp16intrin.h+2-2

...	@@ -19,11 +19,11 @@	...	@@ -19,11 +19,11 @@
19	/* Define the default attributes for the functions in this file. */	19	/* Define the default attributes for the functions in this file. */
20	#define __DEFAULT_FN_ATTRS256 \	20	#define __DEFAULT_FN_ATTRS256 \
21	__attribute__((__always_inline__, __nodebug__, \	21	__attribute__((__always_inline__, __nodebug__, \
22	__target__("avx512fp16, avx512vl"), \	22	__target__("avx512fp16,avx512vl,no-evex512"), \
23	__min_vector_width__(256)))	23	__min_vector_width__(256)))
24	#define __DEFAULT_FN_ATTRS128 \	24	#define __DEFAULT_FN_ATTRS128 \
25	__attribute__((__always_inline__, __nodebug__, \	25	__attribute__((__always_inline__, __nodebug__, \
26	__target__("avx512fp16, avx512vl"), \	26	__target__("avx512fp16,avx512vl,no-evex512"), \
27	__min_vector_width__(128)))	27	__min_vector_width__(128)))
28		28
29	static __inline__ _Float16 __DEFAULT_FN_ATTRS128 _mm_cvtsh_h(__m128h __a) {	29	static __inline__ _Float16 __DEFAULT_FN_ATTRS128 _mm_cvtsh_h(__m128h __a) {

lib/include/avx512vlintrin.h+8-2

 @@ -14,8 +14,14 @@
 #ifndef __AVX512VLINTRIN_H
 #define __AVX512VLINTRIN_H
-#define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl"), __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS128                                                  \
-#define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl"), __min_vector_width__(256)))
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vl,no-evex512"),                            \
+                 __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS256                                                  \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vl,no-evex512"),                            \
+                 __min_vector_width__(256)))
 typedef short __v2hi __attribute__((__vector_size__(4)));
 typedef char __v4qi __attribute__((__vector_size__(4)));

lib/include/avx512vlvbmi2intrin.h+8-2

 @@ -15,8 +15,14 @@
 #define __AVX512VLVBMI2INTRIN_H
 /* Define the default attributes for the functions in this file. */
-#define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512vbmi2"), __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS128                                                  \
-#define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512vbmi2"), __min_vector_width__(256)))
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vl,avx512vbmi2,no-evex512"),                \
+                 __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS256                                                  \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vl,avx512vbmi2,no-evex512"),                \
+                 __min_vector_width__(256)))
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_compress_epi16(__m128i __S, __mmask8 __U, __m128i __D)

lib/include/avx512vlvnniintrin.h+8-2

 @@ -15,8 +15,14 @@
 #define __AVX512VLVNNIINTRIN_H
 /* Define the default attributes for the functions in this file. */
-#define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512vnni"), __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS128                                                  \
-#define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512vnni"), __min_vector_width__(256)))
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vl,avx512vnni,no-evex512"),                 \
+                 __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS256                                                  \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vl,avx512vnni,no-evex512"),                 \
+                 __min_vector_width__(256)))
 /// Multiply groups of 4 adjacent pairs of unsigned 8-bit integers in \a A with
 /// corresponding signed 8-bit integers in \a B, producing 4 intermediate signed

lib/include/avx512vlvp2intersectintrin.h+6-4

 @@ -28,12 +28,14 @@
 #ifndef _AVX512VLVP2INTERSECT_H
 #define _AVX512VLVP2INTERSECT_H
-#define __DEFAULT_FN_ATTRS128 \
+#define __DEFAULT_FN_ATTRS128                                                  \
-  __attribute__((__always_inline__, __nodebug__,  __target__("avx512vl,avx512vp2intersect"), \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vl,avx512vp2intersect,no-evex512"),         \
                  __min_vector_width__(128)))
-#define __DEFAULT_FN_ATTRS256 \
+#define __DEFAULT_FN_ATTRS256                                                  \
-  __attribute__((__always_inline__, __nodebug__,  __target__("avx512vl,avx512vp2intersect"), \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vl,avx512vp2intersect,no-evex512"),         \
                  __min_vector_width__(256)))
 /// Store, in an even/odd pair of mask registers, the indicators of the
 /// locations of value matches between dwords in operands __a and __b.

lib/include/avx512vnniintrin.h+3-2

...	@@ -15,8 +15,9 @@	...	@@ -15,8 +15,9 @@
15	#define __AVX512VNNIINTRIN_H	15	#define __AVX512VNNIINTRIN_H
16		16
17	/* Define the default attributes for the functions in this file. */	17	/* Define the default attributes for the functions in this file. */
18	#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512vnni"), __min_vector_width__(512)))	18	#define __DEFAULT_FN_ATTRS \
19		19	__attribute__((__always_inline__, __nodebug__, \
		20	__target__("avx512vnni,evex512"), __min_vector_width__(512)))
20		21
21	static __inline__ __m512i __DEFAULT_FN_ATTRS	22	static __inline__ __m512i __DEFAULT_FN_ATTRS
22	_mm512_dpbusd_epi32(__m512i __S, __m512i __A, __m512i __B)	23	_mm512_dpbusd_epi32(__m512i __S, __m512i __A, __m512i __B)

lib/include/avx512vp2intersectintrin.h+3-2

 @@ -28,8 +28,9 @@
 #ifndef _AVX512VP2INTERSECT_H
 #define _AVX512VP2INTERSECT_H
-#define __DEFAULT_FN_ATTRS \
+#define __DEFAULT_FN_ATTRS                                                     \
-  __attribute__((__always_inline__, __nodebug__,  __target__("avx512vp2intersect"), \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vp2intersect,evex512"),                     \
                  __min_vector_width__(512)))
 /// Store, in an even/odd pair of mask registers, the indicators of the

lib/include/avx512vpopcntdqintrin.h+3-1

 @@ -17,7 +17,9 @@
 /* Define the default attributes for the functions in this file. */
 #define __DEFAULT_FN_ATTRS                                                     \
-  __attribute__((__always_inline__, __nodebug__, __target__("avx512vpopcntdq"), __min_vector_width__(512)))
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vpopcntdq,evex512"),                        \
+                 __min_vector_width__(512)))
 static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_popcnt_epi64(__m512i __A) {
   return (__m512i)__builtin_ia32_vpopcntq_512((__v8di)__A);

lib/include/avx512vpopcntdqvlintrin.h+6-2

 @@ -17,9 +17,13 @@
 /* Define the default attributes for the functions in this file. */
 #define __DEFAULT_FN_ATTRS128                                                  \
-  __attribute__((__always_inline__, __nodebug__, __target__("avx512vpopcntdq,avx512vl"), __min_vector_width__(128)))
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vpopcntdq,avx512vl,no-evex512"),            \
+                 __min_vector_width__(128)))
 #define __DEFAULT_FN_ATTRS256                                                  \
-  __attribute__((__always_inline__, __nodebug__, __target__("avx512vpopcntdq,avx512vl"), __min_vector_width__(256)))
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512vpopcntdq,avx512vl,no-evex512"),            \
+                 __min_vector_width__(256)))
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_popcnt_epi64(__m128i __A) {

lib/include/avxintrin.h+9-5

...	@@ -50,8 +50,12 @@ typedef __bf16 __m256bh __attribute__((__vector_size__(32), __aligned__(32)));	...	@@ -50,8 +50,12 @@ typedef __bf16 __m256bh __attribute__((__vector_size__(32), __aligned__(32)));
50	#endif	50	#endif
51		51
52	/* Define the default attributes for the functions in this file. */	52	/* Define the default attributes for the functions in this file. */
53	#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx"), __min_vector_width__(256)))	53	#define __DEFAULT_FN_ATTRS \
54	#define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx"), __min_vector_width__(128)))	54	__attribute__((__always_inline__, __nodebug__, __target__("avx,no-evex512"), \
		55	__min_vector_width__(256)))
		56	#define __DEFAULT_FN_ATTRS128 \
		57	__attribute__((__always_inline__, __nodebug__, __target__("avx,no-evex512"), \
		58	__min_vector_width__(128)))
55		59
56	/* Arithmetic */	60	/* Arithmetic */
57	/// Adds two 256-bit vectors of [4 x double].	61	/// Adds two 256-bit vectors of [4 x double].
...	@@ -3563,7 +3567,7 @@ _mm_maskstore_ps(float *__p, __m128i __m, __m128 __a)	...	@@ -3563,7 +3567,7 @@ _mm_maskstore_ps(float *__p, __m128i __m, __m128 __a)
3563	/// \param __b	3567	/// \param __b
3564	/// A 256-bit integer vector containing the values to be moved.	3568	/// A 256-bit integer vector containing the values to be moved.
3565	static __inline void __DEFAULT_FN_ATTRS	3569	static __inline void __DEFAULT_FN_ATTRS
3566	_mm256_stream_si256(__m256i *__a, __m256i __b)	3570	_mm256_stream_si256(void *__a, __m256i __b)
3567	{	3571	{
3568	typedef __v4di __v4di_aligned __attribute__((aligned(32)));	3572	typedef __v4di __v4di_aligned __attribute__((aligned(32)));
3569	__builtin_nontemporal_store((__v4di_aligned)__b, (__v4di_aligned*)__a);	3573	__builtin_nontemporal_store((__v4di_aligned)__b, (__v4di_aligned*)__a);
...	@@ -3583,7 +3587,7 @@ _mm256_stream_si256(__m256i *__a, __m256i __b)	...	@@ -3583,7 +3587,7 @@ _mm256_stream_si256(__m256i *__a, __m256i __b)
3583	/// \param __b	3587	/// \param __b
3584	/// A 256-bit vector of [4 x double] containing the values to be moved.	3588	/// A 256-bit vector of [4 x double] containing the values to be moved.
3585	static __inline void __DEFAULT_FN_ATTRS	3589	static __inline void __DEFAULT_FN_ATTRS
3586	_mm256_stream_pd(double *__a, __m256d __b)	3590	_mm256_stream_pd(void *__a, __m256d __b)
3587	{	3591	{
3588	typedef __v4df __v4df_aligned __attribute__((aligned(32)));	3592	typedef __v4df __v4df_aligned __attribute__((aligned(32)));
3589	__builtin_nontemporal_store((__v4df_aligned)__b, (__v4df_aligned*)__a);	3593	__builtin_nontemporal_store((__v4df_aligned)__b, (__v4df_aligned*)__a);
...	@@ -3604,7 +3608,7 @@ _mm256_stream_pd(double *__a, __m256d __b)	...	@@ -3604,7 +3608,7 @@ _mm256_stream_pd(double *__a, __m256d __b)
3604	/// \param __a	3608	/// \param __a
3605	/// A 256-bit vector of [8 x float] containing the values to be moved.	3609	/// A 256-bit vector of [8 x float] containing the values to be moved.
3606	static __inline void __DEFAULT_FN_ATTRS	3610	static __inline void __DEFAULT_FN_ATTRS
3607	_mm256_stream_ps(float *__p, __m256 __a)	3611	_mm256_stream_ps(void *__p, __m256 __a)
3608	{	3612	{
3609	typedef __v8sf __v8sf_aligned __attribute__((aligned(32)));	3613	typedef __v8sf __v8sf_aligned __attribute__((aligned(32)));
3610	__builtin_nontemporal_store((__v8sf_aligned)__a, (__v8sf_aligned*)__p);	3614	__builtin_nontemporal_store((__v8sf_aligned)__a, (__v8sf_aligned*)__p);

lib/include/bmiintrin.h+231-46

...	@@ -19,18 +19,17 @@	...	@@ -19,18 +19,17 @@
19	to use it as a potentially faster version of BSF. */	19	to use it as a potentially faster version of BSF. */
20	#define __RELAXED_FN_ATTRS __attribute__((__always_inline__, __nodebug__))	20	#define __RELAXED_FN_ATTRS __attribute__((__always_inline__, __nodebug__))
21		21
22	#define _tzcnt_u16(a) (__tzcnt_u16((a)))
23
24	/// Counts the number of trailing zero bits in the operand.	22	/// Counts the number of trailing zero bits in the operand.
25	///	23	///
26	/// \headerfile <x86intrin.h>	24	/// \headerfile <x86intrin.h>
27	///	25	///
28	/// This intrinsic corresponds to the <c> TZCNT </c> instruction.	26	/// This intrinsic corresponds to the \c TZCNT instruction.
29	///	27	///
30	/// \param __X	28	/// \param __X
31	/// An unsigned 16-bit integer whose trailing zeros are to be counted.	29	/// An unsigned 16-bit integer whose trailing zeros are to be counted.
32	/// \returns An unsigned 16-bit integer containing the number of trailing zero	30	/// \returns An unsigned 16-bit integer containing the number of trailing zero
33	/// bits in the operand.	31	/// bits in the operand.
		32	/// \see _tzcnt_u16
34	static __inline__ unsigned short __RELAXED_FN_ATTRS	33	static __inline__ unsigned short __RELAXED_FN_ATTRS
35	__tzcnt_u16(unsigned short __X)	34	__tzcnt_u16(unsigned short __X)
36	{	35	{
...	@@ -41,13 +40,30 @@ __tzcnt_u16(unsigned short __X)	...	@@ -41,13 +40,30 @@ __tzcnt_u16(unsigned short __X)
41	///	40	///
42	/// \headerfile <x86intrin.h>	41	/// \headerfile <x86intrin.h>
43	///	42	///
44	/// This intrinsic corresponds to the <c> TZCNT </c> instruction.	43	/// \code
		44	/// unsigned short _tzcnt_u16(unsigned short __X);
		45	/// \endcode
		46	///
		47	/// This intrinsic corresponds to the \c TZCNT instruction.
		48	///
		49	/// \param __X
		50	/// An unsigned 16-bit integer whose trailing zeros are to be counted.
		51	/// \returns An unsigned 16-bit integer containing the number of trailing zero
		52	/// bits in the operand.
		53	/// \see __tzcnt_u16
		54	#define _tzcnt_u16 __tzcnt_u16
		55
		56	/// Counts the number of trailing zero bits in the operand.
		57	///
		58	/// \headerfile <x86intrin.h>
		59	///
		60	/// This intrinsic corresponds to the \c TZCNT instruction.
45	///	61	///
46	/// \param __X	62	/// \param __X
47	/// An unsigned 32-bit integer whose trailing zeros are to be counted.	63	/// An unsigned 32-bit integer whose trailing zeros are to be counted.
48	/// \returns An unsigned 32-bit integer containing the number of trailing zero	64	/// \returns An unsigned 32-bit integer containing the number of trailing zero
49	/// bits in the operand.	65	/// bits in the operand.
50	/// \see _mm_tzcnt_32	66	/// \see { _mm_tzcnt_32 _tzcnt_u32 }
51	static __inline__ unsigned int __RELAXED_FN_ATTRS	67	static __inline__ unsigned int __RELAXED_FN_ATTRS
52	__tzcnt_u32(unsigned int __X)	68	__tzcnt_u32(unsigned int __X)
53	{	69	{
...	@@ -58,20 +74,35 @@ __tzcnt_u32(unsigned int __X)	...	@@ -58,20 +74,35 @@ __tzcnt_u32(unsigned int __X)
58	///	74	///
59	/// \headerfile <x86intrin.h>	75	/// \headerfile <x86intrin.h>
60	///	76	///
61	/// This intrinsic corresponds to the <c> TZCNT </c> instruction.	77	/// This intrinsic corresponds to the \c TZCNT instruction.
62	///	78	///
63	/// \param __X	79	/// \param __X
64	/// An unsigned 32-bit integer whose trailing zeros are to be counted.	80	/// An unsigned 32-bit integer whose trailing zeros are to be counted.
65	/// \returns An 32-bit integer containing the number of trailing zero bits in	81	/// \returns A 32-bit integer containing the number of trailing zero bits in
66	/// the operand.	82	/// the operand.
67	/// \see __tzcnt_u32	83	/// \see { __tzcnt_u32 _tzcnt_u32 }
68	static __inline__ int __RELAXED_FN_ATTRS	84	static __inline__ int __RELAXED_FN_ATTRS
69	_mm_tzcnt_32(unsigned int __X)	85	_mm_tzcnt_32(unsigned int __X)
70	{	86	{
71	return (int)__builtin_ia32_tzcnt_u32(__X);	87	return (int)__builtin_ia32_tzcnt_u32(__X);
72	}	88	}
73		89
74	#define _tzcnt_u32(a) (__tzcnt_u32((a)))	90	/// Counts the number of trailing zero bits in the operand.
		91	///
		92	/// \headerfile <x86intrin.h>
		93	///
		94	/// \code
		95	/// unsigned int _tzcnt_u32(unsigned int __X);
		96	/// \endcode
		97	///
		98	/// This intrinsic corresponds to the \c TZCNT instruction.
		99	///
		100	/// \param __X
		101	/// An unsigned 32-bit integer whose trailing zeros are to be counted.
		102	/// \returns An unsigned 32-bit integer containing the number of trailing zero
		103	/// bits in the operand.
		104	/// \see { _mm_tzcnt_32 __tzcnt_u32 }
		105	#define _tzcnt_u32 __tzcnt_u32
75		106
76	#ifdef __x86_64__	107	#ifdef __x86_64__
77		108
...	@@ -79,13 +110,13 @@ _mm_tzcnt_32(unsigned int __X)	...	@@ -79,13 +110,13 @@ _mm_tzcnt_32(unsigned int __X)
79	///	110	///
80	/// \headerfile <x86intrin.h>	111	/// \headerfile <x86intrin.h>
81	///	112	///
82	/// This intrinsic corresponds to the <c> TZCNT </c> instruction.	113	/// This intrinsic corresponds to the \c TZCNT instruction.
83	///	114	///
84	/// \param __X	115	/// \param __X
85	/// An unsigned 64-bit integer whose trailing zeros are to be counted.	116	/// An unsigned 64-bit integer whose trailing zeros are to be counted.
86	/// \returns An unsigned 64-bit integer containing the number of trailing zero	117	/// \returns An unsigned 64-bit integer containing the number of trailing zero
87	/// bits in the operand.	118	/// bits in the operand.
88	/// \see _mm_tzcnt_64	119	/// \see { _mm_tzcnt_64 _tzcnt_u64 }
89	static __inline__ unsigned long long __RELAXED_FN_ATTRS	120	static __inline__ unsigned long long __RELAXED_FN_ATTRS
90	__tzcnt_u64(unsigned long long __X)	121	__tzcnt_u64(unsigned long long __X)
91	{	122	{
...	@@ -96,20 +127,35 @@ __tzcnt_u64(unsigned long long __X)	...	@@ -96,20 +127,35 @@ __tzcnt_u64(unsigned long long __X)
96	///	127	///
97	/// \headerfile <x86intrin.h>	128	/// \headerfile <x86intrin.h>
98	///	129	///
99	/// This intrinsic corresponds to the <c> TZCNT </c> instruction.	130	/// This intrinsic corresponds to the \c TZCNT instruction.
100	///	131	///
101	/// \param __X	132	/// \param __X
102	/// An unsigned 64-bit integer whose trailing zeros are to be counted.	133	/// An unsigned 64-bit integer whose trailing zeros are to be counted.
103	/// \returns An 64-bit integer containing the number of trailing zero bits in	134	/// \returns An 64-bit integer containing the number of trailing zero bits in
104	/// the operand.	135	/// the operand.
105	/// \see __tzcnt_u64	136	/// \see { __tzcnt_u64 _tzcnt_u64 }
106	static __inline__ long long __RELAXED_FN_ATTRS	137	static __inline__ long long __RELAXED_FN_ATTRS
107	_mm_tzcnt_64(unsigned long long __X)	138	_mm_tzcnt_64(unsigned long long __X)
108	{	139	{
109	return (long long)__builtin_ia32_tzcnt_u64(__X);	140	return (long long)__builtin_ia32_tzcnt_u64(__X);
110	}	141	}
111		142
112	#define _tzcnt_u64(a) (__tzcnt_u64((a)))	143	/// Counts the number of trailing zero bits in the operand.
		144	///
		145	/// \headerfile <x86intrin.h>
		146	///
		147	/// \code
		148	/// unsigned long long _tzcnt_u64(unsigned long long __X);
		149	/// \endcode
		150	///
		151	/// This intrinsic corresponds to the \c TZCNT instruction.
		152	///
		153	/// \param __X
		154	/// An unsigned 64-bit integer whose trailing zeros are to be counted.
		155	/// \returns An unsigned 64-bit integer containing the number of trailing zero
		156	/// bits in the operand.
		157	/// \see { _mm_tzcnt_64 __tzcnt_u64
		158	#define _tzcnt_u64 __tzcnt_u64
113		159
114	#endif /* __x86_64__ */	160	#endif /* __x86_64__ */
115		161
...	@@ -121,21 +167,12 @@ _mm_tzcnt_64(unsigned long long __X)	...	@@ -121,21 +167,12 @@ _mm_tzcnt_64(unsigned long long __X)
121	/* Define the default attributes for the functions in this file. */	167	/* Define the default attributes for the functions in this file. */
122	#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("bmi")))	168	#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("bmi")))
123		169
124	#define _andn_u32(a, b) (__andn_u32((a), (b)))
125
126	/* _bextr_u32 != __bextr_u32 */
127	#define _blsi_u32(a) (__blsi_u32((a)))
128
129	#define _blsmsk_u32(a) (__blsmsk_u32((a)))
130
131	#define _blsr_u32(a) (__blsr_u32((a)))
132
133	/// Performs a bitwise AND of the second operand with the one's	170	/// Performs a bitwise AND of the second operand with the one's
134	/// complement of the first operand.	171	/// complement of the first operand.
135	///	172	///
136	/// \headerfile <x86intrin.h>	173	/// \headerfile <x86intrin.h>
137	///	174	///
138	/// This intrinsic corresponds to the <c> ANDN </c> instruction.	175	/// This intrinsic corresponds to the \c ANDN instruction.
139	///	176	///
140	/// \param __X	177	/// \param __X
141	/// An unsigned integer containing one of the operands.	178	/// An unsigned integer containing one of the operands.
...	@@ -143,19 +180,40 @@ _mm_tzcnt_64(unsigned long long __X)	...	@@ -143,19 +180,40 @@ _mm_tzcnt_64(unsigned long long __X)
143	/// An unsigned integer containing one of the operands.	180	/// An unsigned integer containing one of the operands.
144	/// \returns An unsigned integer containing the bitwise AND of the second	181	/// \returns An unsigned integer containing the bitwise AND of the second
145	/// operand with the one's complement of the first operand.	182	/// operand with the one's complement of the first operand.
		183	/// \see _andn_u32
146	static __inline__ unsigned int __DEFAULT_FN_ATTRS	184	static __inline__ unsigned int __DEFAULT_FN_ATTRS
147	__andn_u32(unsigned int __X, unsigned int __Y)	185	__andn_u32(unsigned int __X, unsigned int __Y)
148	{	186	{
149	return ~__X & __Y;	187	return ~__X & __Y;
150	}	188	}
151		189
		190	/// Performs a bitwise AND of the second operand with the one's
		191	/// complement of the first operand.
		192	///
		193	/// \headerfile <x86intrin.h>
		194	///
		195	/// \code
		196	/// unsigned int _andn_u32(unsigned int __X, unsigned int __Y);
		197	/// \endcode
		198	///
		199	/// This intrinsic corresponds to the \c ANDN instruction.
		200	///
		201	/// \param __X
		202	/// An unsigned integer containing one of the operands.
		203	/// \param __Y
		204	/// An unsigned integer containing one of the operands.
		205	/// \returns An unsigned integer containing the bitwise AND of the second
		206	/// operand with the one's complement of the first operand.
		207	/// \see __andn_u32
		208	#define _andn_u32 __andn_u32
		209
152	/* AMD-specified, double-leading-underscore version of BEXTR */	210	/* AMD-specified, double-leading-underscore version of BEXTR */
153	/// Extracts the specified bits from the first operand and returns them	211	/// Extracts the specified bits from the first operand and returns them
154	/// in the least significant bits of the result.	212	/// in the least significant bits of the result.
155	///	213	///
156	/// \headerfile <x86intrin.h>	214	/// \headerfile <x86intrin.h>
157	///	215	///
158	/// This intrinsic corresponds to the <c> BEXTR </c> instruction.	216	/// This intrinsic corresponds to the \c BEXTR instruction.
159	///	217	///
160	/// \param __X	218	/// \param __X
161	/// An unsigned integer whose bits are to be extracted.	219	/// An unsigned integer whose bits are to be extracted.
...	@@ -178,7 +236,7 @@ __bextr_u32(unsigned int __X, unsigned int __Y)	...	@@ -178,7 +236,7 @@ __bextr_u32(unsigned int __X, unsigned int __Y)
178	///	236	///
179	/// \headerfile <x86intrin.h>	237	/// \headerfile <x86intrin.h>
180	///	238	///
181	/// This intrinsic corresponds to the <c> BEXTR </c> instruction.	239	/// This intrinsic corresponds to the \c BEXTR instruction.
182	///	240	///
183	/// \param __X	241	/// \param __X
184	/// An unsigned integer whose bits are to be extracted.	242	/// An unsigned integer whose bits are to be extracted.
...	@@ -203,7 +261,7 @@ _bextr_u32(unsigned int __X, unsigned int __Y, unsigned int __Z)	...	@@ -203,7 +261,7 @@ _bextr_u32(unsigned int __X, unsigned int __Y, unsigned int __Z)
203	///	261	///
204	/// \headerfile <x86intrin.h>	262	/// \headerfile <x86intrin.h>
205	///	263	///
206	/// This intrinsic corresponds to the <c> BEXTR </c> instruction.	264	/// This intrinsic corresponds to the \c BEXTR instruction.
207	///	265	///
208	/// \param __X	266	/// \param __X
209	/// An unsigned integer whose bits are to be extracted.	267	/// An unsigned integer whose bits are to be extracted.
...	@@ -224,69 +282,117 @@ _bextr2_u32(unsigned int __X, unsigned int __Y) {	...	@@ -224,69 +282,117 @@ _bextr2_u32(unsigned int __X, unsigned int __Y) {
224	///	282	///
225	/// \headerfile <x86intrin.h>	283	/// \headerfile <x86intrin.h>
226	///	284	///
227	/// This intrinsic corresponds to the <c> BLSI </c> instruction.	285	/// This intrinsic corresponds to the \c BLSI instruction.
228	///	286	///
229	/// \param __X	287	/// \param __X
230	/// An unsigned integer whose bits are to be cleared.	288	/// An unsigned integer whose bits are to be cleared.
231	/// \returns An unsigned integer containing the result of clearing the bits from	289	/// \returns An unsigned integer containing the result of clearing the bits from
232	/// the source operand.	290	/// the source operand.
		291	/// \see _blsi_u32
233	static __inline__ unsigned int __DEFAULT_FN_ATTRS	292	static __inline__ unsigned int __DEFAULT_FN_ATTRS
234	__blsi_u32(unsigned int __X)	293	__blsi_u32(unsigned int __X)
235	{	294	{
236	return __X & -__X;	295	return __X & -__X;
237	}	296	}
238		297
		298	/// Clears all bits in the source except for the least significant bit
		299	/// containing a value of 1 and returns the result.
		300	///
		301	/// \headerfile <x86intrin.h>
		302	///
		303	/// \code
		304	/// unsigned int _blsi_u32(unsigned int __X);
		305	/// \endcode
		306	///
		307	/// This intrinsic corresponds to the \c BLSI instruction.
		308	///
		309	/// \param __X
		310	/// An unsigned integer whose bits are to be cleared.
		311	/// \returns An unsigned integer containing the result of clearing the bits from
		312	/// the source operand.
		313	/// \see __blsi_u32
		314	#define _blsi_u32 __blsi_u32
		315
239	/// Creates a mask whose bits are set to 1, using bit 0 up to and	316	/// Creates a mask whose bits are set to 1, using bit 0 up to and
240	/// including the least significant bit that is set to 1 in the source	317	/// including the least significant bit that is set to 1 in the source
241	/// operand and returns the result.	318	/// operand and returns the result.
242	///	319	///
243	/// \headerfile <x86intrin.h>	320	/// \headerfile <x86intrin.h>
244	///	321	///
245	/// This intrinsic corresponds to the <c> BLSMSK </c> instruction.	322	/// This intrinsic corresponds to the \c BLSMSK instruction.
246	///	323	///
247	/// \param __X	324	/// \param __X
248	/// An unsigned integer used to create the mask.	325	/// An unsigned integer used to create the mask.
249	/// \returns An unsigned integer containing the newly created mask.	326	/// \returns An unsigned integer containing the newly created mask.
		327	/// \see _blsmsk_u32
250	static __inline__ unsigned int __DEFAULT_FN_ATTRS	328	static __inline__ unsigned int __DEFAULT_FN_ATTRS
251	__blsmsk_u32(unsigned int __X)	329	__blsmsk_u32(unsigned int __X)
252	{	330	{
253	return __X ^ (__X - 1);	331	return __X ^ (__X - 1);
254	}	332	}
255		333
		334	/// Creates a mask whose bits are set to 1, using bit 0 up to and
		335	/// including the least significant bit that is set to 1 in the source
		336	/// operand and returns the result.
		337	///
		338	/// \headerfile <x86intrin.h>
		339	///
		340	/// \code
		341	/// unsigned int _blsmsk_u32(unsigned int __X);
		342	/// \endcode
		343	///
		344	/// This intrinsic corresponds to the \c BLSMSK instruction.
		345	///
		346	/// \param __X
		347	/// An unsigned integer used to create the mask.
		348	/// \returns An unsigned integer containing the newly created mask.
		349	/// \see __blsmsk_u32
		350	#define _blsmsk_u32 __blsmsk_u32
		351
256	/// Clears the least significant bit that is set to 1 in the source	352	/// Clears the least significant bit that is set to 1 in the source
257	/// operand and returns the result.	353	/// operand and returns the result.
258	///	354	///
259	/// \headerfile <x86intrin.h>	355	/// \headerfile <x86intrin.h>
260	///	356	///
261	/// This intrinsic corresponds to the <c> BLSR </c> instruction.	357	/// This intrinsic corresponds to the \c BLSR instruction.
262	///	358	///
263	/// \param __X	359	/// \param __X
264	/// An unsigned integer containing the operand to be cleared.	360	/// An unsigned integer containing the operand to be cleared.
265	/// \returns An unsigned integer containing the result of clearing the source	361	/// \returns An unsigned integer containing the result of clearing the source
266	/// operand.	362	/// operand.
		363	/// \see _blsr_u32
267	static __inline__ unsigned int __DEFAULT_FN_ATTRS	364	static __inline__ unsigned int __DEFAULT_FN_ATTRS
268	__blsr_u32(unsigned int __X)	365	__blsr_u32(unsigned int __X)
269	{	366	{
270	return __X & (__X - 1);	367	return __X & (__X - 1);
271	}	368	}
272		369
273	#ifdef __x86_64__	370	/// Clears the least significant bit that is set to 1 in the source
274		371	/// operand and returns the result.
275	#define _andn_u64(a, b) (__andn_u64((a), (b)))	372	///
276		373	/// \headerfile <x86intrin.h>
277	/* _bextr_u64 != __bextr_u64 */	374	///
278	#define _blsi_u64(a) (__blsi_u64((a)))	375	/// \code
279		376	/// unsigned int _bls4_u32(unsigned int __X);
280	#define _blsmsk_u64(a) (__blsmsk_u64((a)))	377	/// \endcode
		378	///
		379	/// This intrinsic corresponds to the \c BLSR instruction.
		380	///
		381	/// \param __X
		382	/// An unsigned integer containing the operand to be cleared.
		383	/// \returns An unsigned integer containing the result of clearing the source
		384	/// operand.
		385	/// \see __blsr_u32
		386	#define _blsr_u32 __blsr_u32
281		387
282	#define _blsr_u64(a) (__blsr_u64((a)))	388	#ifdef __x86_64__
283		389
284	/// Performs a bitwise AND of the second operand with the one's	390	/// Performs a bitwise AND of the second operand with the one's
285	/// complement of the first operand.	391	/// complement of the first operand.
286	///	392	///
287	/// \headerfile <x86intrin.h>	393	/// \headerfile <x86intrin.h>
288	///	394	///
289	/// This intrinsic corresponds to the <c> ANDN </c> instruction.	395	/// This intrinsic corresponds to the \c ANDN instruction.
290	///	396	///
291	/// \param __X	397	/// \param __X
292	/// An unsigned 64-bit integer containing one of the operands.	398	/// An unsigned 64-bit integer containing one of the operands.
...	@@ -294,19 +400,41 @@ __blsr_u32(unsigned int __X)	...	@@ -294,19 +400,41 @@ __blsr_u32(unsigned int __X)
294	/// An unsigned 64-bit integer containing one of the operands.	400	/// An unsigned 64-bit integer containing one of the operands.
295	/// \returns An unsigned 64-bit integer containing the bitwise AND of the second	401	/// \returns An unsigned 64-bit integer containing the bitwise AND of the second
296	/// operand with the one's complement of the first operand.	402	/// operand with the one's complement of the first operand.
		403	/// \see _andn_u64
297	static __inline__ unsigned long long __DEFAULT_FN_ATTRS	404	static __inline__ unsigned long long __DEFAULT_FN_ATTRS
298	__andn_u64 (unsigned long long __X, unsigned long long __Y)	405	__andn_u64 (unsigned long long __X, unsigned long long __Y)
299	{	406	{
300	return ~__X & __Y;	407	return ~__X & __Y;
301	}	408	}
302		409
		410	/// Performs a bitwise AND of the second operand with the one's
		411	/// complement of the first operand.
		412	///
		413	/// \headerfile <x86intrin.h>
		414	///
		415	/// \code
		416	/// unsigned long long _andn_u64(unsigned long long __X,
		417	/// unsigned long long __Y);
		418	/// \endcode
		419	///
		420	/// This intrinsic corresponds to the \c ANDN instruction.
		421	///
		422	/// \param __X
		423	/// An unsigned 64-bit integer containing one of the operands.
		424	/// \param __Y
		425	/// An unsigned 64-bit integer containing one of the operands.
		426	/// \returns An unsigned 64-bit integer containing the bitwise AND of the second
		427	/// operand with the one's complement of the first operand.
		428	/// \see __andn_u64
		429	#define _andn_u64 __andn_u64
		430
303	/* AMD-specified, double-leading-underscore version of BEXTR */	431	/* AMD-specified, double-leading-underscore version of BEXTR */
304	/// Extracts the specified bits from the first operand and returns them	432	/// Extracts the specified bits from the first operand and returns them
305	/// in the least significant bits of the result.	433	/// in the least significant bits of the result.
306	///	434	///
307	/// \headerfile <x86intrin.h>	435	/// \headerfile <x86intrin.h>
308	///	436	///
309	/// This intrinsic corresponds to the <c> BEXTR </c> instruction.	437	/// This intrinsic corresponds to the \c BEXTR instruction.
310	///	438	///
311	/// \param __X	439	/// \param __X
312	/// An unsigned 64-bit integer whose bits are to be extracted.	440	/// An unsigned 64-bit integer whose bits are to be extracted.
...	@@ -329,7 +457,7 @@ __bextr_u64(unsigned long long __X, unsigned long long __Y)	...	@@ -329,7 +457,7 @@ __bextr_u64(unsigned long long __X, unsigned long long __Y)
329	///	457	///
330	/// \headerfile <x86intrin.h>	458	/// \headerfile <x86intrin.h>
331	///	459	///
332	/// This intrinsic corresponds to the <c> BEXTR </c> instruction.	460	/// This intrinsic corresponds to the \c BEXTR instruction.
333	///	461	///
334	/// \param __X	462	/// \param __X
335	/// An unsigned 64-bit integer whose bits are to be extracted.	463	/// An unsigned 64-bit integer whose bits are to be extracted.
...	@@ -354,7 +482,7 @@ _bextr_u64(unsigned long long __X, unsigned int __Y, unsigned int __Z)	...	@@ -354,7 +482,7 @@ _bextr_u64(unsigned long long __X, unsigned int __Y, unsigned int __Z)
354	///	482	///
355	/// \headerfile <x86intrin.h>	483	/// \headerfile <x86intrin.h>
356	///	484	///
357	/// This intrinsic corresponds to the <c> BEXTR </c> instruction.	485	/// This intrinsic corresponds to the \c BEXTR instruction.
358	///	486	///
359	/// \param __X	487	/// \param __X
360	/// An unsigned 64-bit integer whose bits are to be extracted.	488	/// An unsigned 64-bit integer whose bits are to be extracted.
...	@@ -375,52 +503,109 @@ _bextr2_u64(unsigned long long __X, unsigned long long __Y) {	...	@@ -375,52 +503,109 @@ _bextr2_u64(unsigned long long __X, unsigned long long __Y) {
375	///	503	///
376	/// \headerfile <x86intrin.h>	504	/// \headerfile <x86intrin.h>
377	///	505	///
378	/// This intrinsic corresponds to the <c> BLSI </c> instruction.	506	/// This intrinsic corresponds to the \c BLSI instruction.
379	///	507	///
380	/// \param __X	508	/// \param __X
381	/// An unsigned 64-bit integer whose bits are to be cleared.	509	/// An unsigned 64-bit integer whose bits are to be cleared.
382	/// \returns An unsigned 64-bit integer containing the result of clearing the	510	/// \returns An unsigned 64-bit integer containing the result of clearing the
383	/// bits from the source operand.	511	/// bits from the source operand.
		512	/// \see _blsi_u64
384	static __inline__ unsigned long long __DEFAULT_FN_ATTRS	513	static __inline__ unsigned long long __DEFAULT_FN_ATTRS
385	__blsi_u64(unsigned long long __X)	514	__blsi_u64(unsigned long long __X)
386	{	515	{
387	return __X & -__X;	516	return __X & -__X;
388	}	517	}
389		518
		519	/// Clears all bits in the source except for the least significant bit
		520	/// containing a value of 1 and returns the result.
		521	///
		522	/// \headerfile <x86intrin.h>
		523	///
		524	/// \code
		525	/// unsigned long long _blsi_u64(unsigned long long __X);
		526	/// \endcode
		527	///
		528	/// This intrinsic corresponds to the \c BLSI instruction.
		529	///
		530	/// \param __X
		531	/// An unsigned 64-bit integer whose bits are to be cleared.
		532	/// \returns An unsigned 64-bit integer containing the result of clearing the
		533	/// bits from the source operand.
		534	/// \see __blsi_u64
		535	#define _blsi_u64 __blsi_u64
		536
390	/// Creates a mask whose bits are set to 1, using bit 0 up to and	537	/// Creates a mask whose bits are set to 1, using bit 0 up to and
391	/// including the least significant bit that is set to 1 in the source	538	/// including the least significant bit that is set to 1 in the source
392	/// operand and returns the result.	539	/// operand and returns the result.
393	///	540	///
394	/// \headerfile <x86intrin.h>	541	/// \headerfile <x86intrin.h>
395	///	542	///
396	/// This intrinsic corresponds to the <c> BLSMSK </c> instruction.	543	/// This intrinsic corresponds to the \c BLSMSK instruction.
397	///	544	///
398	/// \param __X	545	/// \param __X
399	/// An unsigned 64-bit integer used to create the mask.	546	/// An unsigned 64-bit integer used to create the mask.
400	/// \returns An unsigned 64-bit integer containing the newly created mask.	547	/// \returns An unsigned 64-bit integer containing the newly created mask.
		548	/// \see _blsmsk_u64
401	static __inline__ unsigned long long __DEFAULT_FN_ATTRS	549	static __inline__ unsigned long long __DEFAULT_FN_ATTRS
402	__blsmsk_u64(unsigned long long __X)	550	__blsmsk_u64(unsigned long long __X)
403	{	551	{
404	return __X ^ (__X - 1);	552	return __X ^ (__X - 1);
405	}	553	}
406		554
		555	/// Creates a mask whose bits are set to 1, using bit 0 up to and
		556	/// including the least significant bit that is set to 1 in the source
		557	/// operand and returns the result.
		558	///
		559	/// \headerfile <x86intrin.h>
		560	///
		561	/// \code
		562	/// unsigned long long _blsmsk_u64(unsigned long long __X);
		563	/// \endcode
		564	///
		565	/// This intrinsic corresponds to the \c BLSMSK instruction.
		566	///
		567	/// \param __X
		568	/// An unsigned 64-bit integer used to create the mask.
		569	/// \returns An unsigned 64-bit integer containing the newly created mask.
		570	/// \see __blsmsk_u64
		571	#define _blsmsk_u64 __blsmsk_u64
		572
407	/// Clears the least significant bit that is set to 1 in the source	573	/// Clears the least significant bit that is set to 1 in the source
408	/// operand and returns the result.	574	/// operand and returns the result.
409	///	575	///
410	/// \headerfile <x86intrin.h>	576	/// \headerfile <x86intrin.h>
411	///	577	///
412	/// This intrinsic corresponds to the <c> BLSR </c> instruction.	578	/// This intrinsic corresponds to the \c BLSR instruction.
413	///	579	///
414	/// \param __X	580	/// \param __X
415	/// An unsigned 64-bit integer containing the operand to be cleared.	581	/// An unsigned 64-bit integer containing the operand to be cleared.
416	/// \returns An unsigned 64-bit integer containing the result of clearing the	582	/// \returns An unsigned 64-bit integer containing the result of clearing the
417	/// source operand.	583	/// source operand.
		584	/// \see _blsr_u64
418	static __inline__ unsigned long long __DEFAULT_FN_ATTRS	585	static __inline__ unsigned long long __DEFAULT_FN_ATTRS
419	__blsr_u64(unsigned long long __X)	586	__blsr_u64(unsigned long long __X)
420	{	587	{
421	return __X & (__X - 1);	588	return __X & (__X - 1);
422	}	589	}
423		590
		591	/// Clears the least significant bit that is set to 1 in the source
		592	/// operand and returns the result.
		593	///
		594	/// \headerfile <x86intrin.h>
		595	///
		596	/// \code
		597	/// unsigned long long _blsr_u64(unsigned long long __X);
		598	/// \endcode
		599	///
		600	/// This intrinsic corresponds to the \c BLSR instruction.
		601	///
		602	/// \param __X
		603	/// An unsigned 64-bit integer containing the operand to be cleared.
		604	/// \returns An unsigned 64-bit integer containing the result of clearing the
		605	/// source operand.
		606	/// \see __blsr_u64
		607	#define _blsr_u64 __blsr_u64
		608
424	#endif /* __x86_64__ */	609	#endif /* __x86_64__ */
425		610
426	#undef __DEFAULT_FN_ATTRS	611	#undef __DEFAULT_FN_ATTRS

lib/include/cuda_wrappers/bits/basic_string.h created+9

...	@@ -0,0 +1,9 @@
1	// CUDA headers define __noinline__ which interferes with libstdc++'s use of
2	// `__attribute((__noinline__))`. In order to avoid compilation error,
3	// temporarily unset __noinline__ when we include affected libstdc++ header.
4
5	#pragma push_macro("__noinline__")
6	#undef __noinline__
7	#include_next "bits/basic_string.h"
8
9	#pragma pop_macro("__noinline__")

lib/include/cuda_wrappers/bits/basic_string.tcc created+9

...	@@ -0,0 +1,9 @@
1	// CUDA headers define __noinline__ which interferes with libstdc++'s use of
2	// `__attribute((__noinline__))`. In order to avoid compilation error,
3	// temporarily unset __noinline__ when we include affected libstdc++ header.
4
5	#pragma push_macro("__noinline__")
6	#undef __noinline__
7	#include_next "bits/basic_string.tcc"
8
9	#pragma pop_macro("__noinline__")

lib/include/emmintrin.h+10-10

lib/include/gfniintrin.h+23-6

 @@ -15,19 +15,36 @@
 #define __GFNIINTRIN_H
 /* Default attributes for simple form (no masking). */
-#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("gfni"), __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS                                                     \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("gfni,no-evex512"), __min_vector_width__(128)))
 /* Default attributes for YMM unmasked form. */
-#define __DEFAULT_FN_ATTRS_Y __attribute__((__always_inline__, __nodebug__, __target__("avx,gfni"), __min_vector_width__(256)))
+#define __DEFAULT_FN_ATTRS_Y                                                   \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx,gfni,no-evex512"),                            \
+                 __min_vector_width__(256)))
 /* Default attributes for ZMM unmasked forms. */
-#define __DEFAULT_FN_ATTRS_Z __attribute__((__always_inline__, __nodebug__, __target__("avx512f,gfni"), __min_vector_width__(512)))
+#define __DEFAULT_FN_ATTRS_Z                                                   \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512f,evex512,gfni"),                           \
+                 __min_vector_width__(512)))
 /* Default attributes for ZMM masked forms. */
-#define __DEFAULT_FN_ATTRS_Z_MASK __attribute__((__always_inline__, __nodebug__, __target__("avx512bw,gfni"), __min_vector_width__(512)))
+#define __DEFAULT_FN_ATTRS_Z_MASK                                              \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512bw,evex512,gfni"),                          \
+                 __min_vector_width__(512)))
 /* Default attributes for VLX masked forms. */
-#define __DEFAULT_FN_ATTRS_VL128 __attribute__((__always_inline__, __nodebug__, __target__("avx512bw,avx512vl,gfni"), __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS_VL128                                               \
-#define __DEFAULT_FN_ATTRS_VL256 __attribute__((__always_inline__, __nodebug__, __target__("avx512bw,avx512vl,gfni"), __min_vector_width__(256)))
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512bw,avx512vl,gfni,no-evex512"),              \
+                 __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS_VL256                                               \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("avx512bw,avx512vl,gfni,no-evex512"),              \
+                 __min_vector_width__(256)))
 #define _mm_gf2p8affineinv_epi64_epi8(A, B, I) \
   ((__m128i)__builtin_ia32_vgf2p8affineinvqb_v16qi((__v16qi)(__m128i)(A), \

lib/include/ia32intrin.h+620-198

...	@@ -26,167 +26,271 @@	...	@@ -26,167 +26,271 @@
26	#define __DEFAULT_FN_ATTRS_CONSTEXPR __DEFAULT_FN_ATTRS	26	#define __DEFAULT_FN_ATTRS_CONSTEXPR __DEFAULT_FN_ATTRS
27	#endif	27	#endif
28		28
29	/** Find the first set bit starting from the lsb. Result is undefined if	29	/// Find the first set bit starting from the lsb. Result is undefined if
30	* input is 0.	30	/// input is 0.
31	*	31	///
32	* \headerfile <x86intrin.h>	32	/// \headerfile <x86intrin.h>
33	*	33	///
34	* This intrinsic corresponds to the <c> BSF </c> instruction or the	34	/// This intrinsic corresponds to the \c BSF instruction or the
35	* <c> TZCNT </c> instruction.	35	/// \c TZCNT instruction.
36	*	36	///
37	* \param __A	37	/// \param __A
38	* A 32-bit integer operand.	38	/// A 32-bit integer operand.
39	* \returns A 32-bit integer containing the bit number.	39	/// \returns A 32-bit integer containing the bit number.
40	*/	40	/// \see _bit_scan_forward
41	static __inline__ int __DEFAULT_FN_ATTRS_CONSTEXPR	41	static __inline__ int __DEFAULT_FN_ATTRS_CONSTEXPR
42	__bsfd(int __A) {	42	__bsfd(int __A) {
43	return __builtin_ctz((unsigned int)__A);	43	return __builtin_ctz((unsigned int)__A);
44	}	44	}
45		45
46	/** Find the first set bit starting from the msb. Result is undefined if	46	/// Find the first set bit starting from the msb. Result is undefined if
47	* input is 0.	47	/// input is 0.
48	*	48	///
49	* \headerfile <x86intrin.h>	49	/// \headerfile <x86intrin.h>
50	*	50	///
51	* This intrinsic corresponds to the <c> BSR </c> instruction or the	51	/// This intrinsic corresponds to the \c BSR instruction or the
52	* <c> LZCNT </c> instruction and an <c> XOR </c>.	52	/// \c LZCNT instruction and an \c XOR.
53	*	53	///
54	* \param __A	54	/// \param __A
55	* A 32-bit integer operand.	55	/// A 32-bit integer operand.
56	* \returns A 32-bit integer containing the bit number.	56	/// \returns A 32-bit integer containing the bit number.
57	*/	57	/// \see _bit_scan_reverse
58	static __inline__ int __DEFAULT_FN_ATTRS_CONSTEXPR	58	static __inline__ int __DEFAULT_FN_ATTRS_CONSTEXPR
59	__bsrd(int __A) {	59	__bsrd(int __A) {
60	return 31 - __builtin_clz((unsigned int)__A);	60	return 31 - __builtin_clz((unsigned int)__A);
61	}	61	}
62		62
63	/** Swaps the bytes in the input. Converting little endian to big endian or	63	/// Swaps the bytes in the input, converting little endian to big endian or
64	* vice versa.	64	/// vice versa.
65	*	65	///
66	* \headerfile <x86intrin.h>	66	/// \headerfile <x86intrin.h>
67	*	67	///
68	* This intrinsic corresponds to the <c> BSWAP </c> instruction.	68	/// This intrinsic corresponds to the \c BSWAP instruction.
69	*	69	///
70	* \param __A	70	/// \param __A
71	* A 32-bit integer operand.	71	/// A 32-bit integer operand.
72	* \returns A 32-bit integer containing the swapped bytes.	72	/// \returns A 32-bit integer containing the swapped bytes.
73	*/
74	static __inline__ int __DEFAULT_FN_ATTRS_CONSTEXPR	73	static __inline__ int __DEFAULT_FN_ATTRS_CONSTEXPR
75	__bswapd(int __A) {	74	__bswapd(int __A) {
76	return (int)__builtin_bswap32((unsigned int)__A);	75	return (int)__builtin_bswap32((unsigned int)__A);
77	}	76	}
78		77
		78	/// Swaps the bytes in the input, converting little endian to big endian or
		79	/// vice versa.
		80	///
		81	/// \headerfile <x86intrin.h>
		82	///
		83	/// This intrinsic corresponds to the \c BSWAP instruction.
		84	///
		85	/// \param __A
		86	/// A 32-bit integer operand.
		87	/// \returns A 32-bit integer containing the swapped bytes.
79	static __inline__ int __DEFAULT_FN_ATTRS_CONSTEXPR	88	static __inline__ int __DEFAULT_FN_ATTRS_CONSTEXPR
80	_bswap(int __A) {	89	_bswap(int __A) {
81	return (int)__builtin_bswap32((unsigned int)__A);	90	return (int)__builtin_bswap32((unsigned int)__A);
82	}	91	}
83		92
		93	/// Find the first set bit starting from the lsb. Result is undefined if
		94	/// input is 0.
		95	///
		96	/// \headerfile <x86intrin.h>
		97	///
		98	/// \code
		99	/// int _bit_scan_forward(int A);
		100	/// \endcode
		101	///
		102	/// This intrinsic corresponds to the \c BSF instruction or the
		103	/// \c TZCNT instruction.
		104	///
		105	/// \param A
		106	/// A 32-bit integer operand.
		107	/// \returns A 32-bit integer containing the bit number.
		108	/// \see __bsfd
84	#define _bit_scan_forward(A) __bsfd((A))	109	#define _bit_scan_forward(A) __bsfd((A))
		110
		111	/// Find the first set bit starting from the msb. Result is undefined if
		112	/// input is 0.
		113	///
		114	/// \headerfile <x86intrin.h>
		115	///
		116	/// \code
		117	/// int _bit_scan_reverse(int A);
		118	/// \endcode
		119	///
		120	/// This intrinsic corresponds to the \c BSR instruction or the
		121	/// \c LZCNT instruction and an \c XOR.
		122	///
		123	/// \param A
		124	/// A 32-bit integer operand.
		125	/// \returns A 32-bit integer containing the bit number.
		126	/// \see __bsrd
85	#define _bit_scan_reverse(A) __bsrd((A))	127	#define _bit_scan_reverse(A) __bsrd((A))
86		128
87	#ifdef __x86_64__	129	#ifdef __x86_64__
88	/** Find the first set bit starting from the lsb. Result is undefined if	130	/// Find the first set bit starting from the lsb. Result is undefined if
89	* input is 0.	131	/// input is 0.
90	*	132	///
91	* \headerfile <x86intrin.h>	133	/// \headerfile <x86intrin.h>
92	*	134	///
93	* This intrinsic corresponds to the <c> BSF </c> instruction or the	135	/// This intrinsic corresponds to the \c BSF instruction or the
94	* <c> TZCNT </c> instruction.	136	/// \c TZCNT instruction.
95	*	137	///
96	* \param __A	138	/// \param __A
97	* A 64-bit integer operand.	139	/// A 64-bit integer operand.
98	* \returns A 32-bit integer containing the bit number.	140	/// \returns A 32-bit integer containing the bit number.
99	*/
100	static __inline__ int __DEFAULT_FN_ATTRS_CONSTEXPR	141	static __inline__ int __DEFAULT_FN_ATTRS_CONSTEXPR
101	__bsfq(long long __A) {	142	__bsfq(long long __A) {
102	return (long long)__builtin_ctzll((unsigned long long)__A);	143	return (long long)__builtin_ctzll((unsigned long long)__A);
103	}	144	}
104		145
105	/** Find the first set bit starting from the msb. Result is undefined if	146	/// Find the first set bit starting from the msb. Result is undefined if
106	* input is 0.	147	/// input is 0.
107	*	148	///
108	* \headerfile <x86intrin.h>	149	/// \headerfile <x86intrin.h>
109	*	150	///
110	* This intrinsic corresponds to the <c> BSR </c> instruction or the	151	/// This intrinsic corresponds to the \c BSR instruction or the
111	* <c> LZCNT </c> instruction and an <c> XOR </c>.	152	/// \c LZCNT instruction and an \c XOR.
112	*	153	///
113	* \param __A	154	/// \param __A
114	* A 64-bit integer operand.	155	/// A 64-bit integer operand.
115	* \returns A 32-bit integer containing the bit number.	156	/// \returns A 32-bit integer containing the bit number.
116	*/
117	static __inline__ int __DEFAULT_FN_ATTRS_CONSTEXPR	157	static __inline__ int __DEFAULT_FN_ATTRS_CONSTEXPR
118	__bsrq(long long __A) {	158	__bsrq(long long __A) {
119	return 63 - __builtin_clzll((unsigned long long)__A);	159	return 63 - __builtin_clzll((unsigned long long)__A);
120	}	160	}
121		161
122	/** Swaps the bytes in the input. Converting little endian to big endian or	162	/// Swaps the bytes in the input. Converting little endian to big endian or
123	* vice versa.	163	/// vice versa.
124	*	164	///
125	* \headerfile <x86intrin.h>	165	/// \headerfile <x86intrin.h>
126	*	166	///
127	* This intrinsic corresponds to the <c> BSWAP </c> instruction.	167	/// This intrinsic corresponds to the \c BSWAP instruction.
128	*	168	///
129	* \param __A	169	/// \param __A
130	* A 64-bit integer operand.	170	/// A 64-bit integer operand.
131	* \returns A 64-bit integer containing the swapped bytes.	171	/// \returns A 64-bit integer containing the swapped bytes.
132	*/	172	/// \see _bswap64
133	static __inline__ long long __DEFAULT_FN_ATTRS_CONSTEXPR	173	static __inline__ long long __DEFAULT_FN_ATTRS_CONSTEXPR
134	__bswapq(long long __A) {	174	__bswapq(long long __A) {
135	return (long long)__builtin_bswap64((unsigned long long)__A);	175	return (long long)__builtin_bswap64((unsigned long long)__A);
136	}	176	}
137		177
		178	/// Swaps the bytes in the input. Converting little endian to big endian or
		179	/// vice versa.
		180	///
		181	/// \headerfile <x86intrin.h>
		182	///
		183	/// \code
		184	/// long long _bswap64(long long A);
		185	/// \endcode
		186	///
		187	/// This intrinsic corresponds to the \c BSWAP instruction.
		188	///
		189	/// \param A
		190	/// A 64-bit integer operand.
		191	/// \returns A 64-bit integer containing the swapped bytes.
		192	/// \see __bswapq
138	#define _bswap64(A) __bswapq((A))	193	#define _bswap64(A) __bswapq((A))
139	#endif	194	#endif /* __x86_64__ */
140		195
141	/** Counts the number of bits in the source operand having a value of 1.	196	/// Counts the number of bits in the source operand having a value of 1.
142	*	197	///
143	* \headerfile <x86intrin.h>	198	/// \headerfile <x86intrin.h>
144	*	199	///
145	* This intrinsic corresponds to the <c> POPCNT </c> instruction or a	200	/// This intrinsic corresponds to the \c POPCNT instruction or a
146	* a sequence of arithmetic and logic ops to calculate it.	201	/// a sequence of arithmetic and logic ops to calculate it.
147	*	202	///
148	* \param __A	203	/// \param __A
149	* An unsigned 32-bit integer operand.	204	/// An unsigned 32-bit integer operand.
150	* \returns A 32-bit integer containing the number of bits with value 1 in the	205	/// \returns A 32-bit integer containing the number of bits with value 1 in the
151	* source operand.	206	/// source operand.
152	*/	207	/// \see _popcnt32
153	static __inline__ int __DEFAULT_FN_ATTRS_CONSTEXPR	208	static __inline__ int __DEFAULT_FN_ATTRS_CONSTEXPR
154	__popcntd(unsigned int __A)	209	__popcntd(unsigned int __A)
155	{	210	{
156	return __builtin_popcount(__A);	211	return __builtin_popcount(__A);
157	}	212	}
158		213
		214	/// Counts the number of bits in the source operand having a value of 1.
		215	///
		216	/// \headerfile <x86intrin.h>
		217	///
		218	/// \code
		219	/// int _popcnt32(int A);
		220	/// \endcode
		221	///
		222	/// This intrinsic corresponds to the \c POPCNT instruction or a
		223	/// a sequence of arithmetic and logic ops to calculate it.
		224	///
		225	/// \param A
		226	/// An unsigned 32-bit integer operand.
		227	/// \returns A 32-bit integer containing the number of bits with value 1 in the
		228	/// source operand.
		229	/// \see __popcntd
159	#define _popcnt32(A) __popcntd((A))	230	#define _popcnt32(A) __popcntd((A))
160		231
161	#ifdef __x86_64__	232	#ifdef __x86_64__
162	/** Counts the number of bits in the source operand having a value of 1.	233	/// Counts the number of bits in the source operand having a value of 1.
163	*	234	///
164	* \headerfile <x86intrin.h>	235	/// \headerfile <x86intrin.h>
165	*	236	///
166	* This intrinsic corresponds to the <c> POPCNT </c> instruction or a	237	/// This intrinsic corresponds to the \c POPCNT instruction or a
167	* a sequence of arithmetic and logic ops to calculate it.	238	/// a sequence of arithmetic and logic ops to calculate it.
168	*	239	///
169	* \param __A	240	/// \param __A
170	* An unsigned 64-bit integer operand.	241	/// An unsigned 64-bit integer operand.
171	* \returns A 64-bit integer containing the number of bits with value 1 in the	242	/// \returns A 64-bit integer containing the number of bits with value 1 in the
172	* source operand.	243	/// source operand.
173	*/	244	/// \see _popcnt64
174	static __inline__ long long __DEFAULT_FN_ATTRS_CONSTEXPR	245	static __inline__ long long __DEFAULT_FN_ATTRS_CONSTEXPR
175	__popcntq(unsigned long long __A)	246	__popcntq(unsigned long long __A)
176	{	247	{
177	return __builtin_popcountll(__A);	248	return __builtin_popcountll(__A);
178	}	249	}
179		250
		251	/// Counts the number of bits in the source operand having a value of 1.
		252	///
		253	/// \headerfile <x86intrin.h>
		254	///
		255	/// \code
		256	/// long long _popcnt64(unsigned long long A);
		257	/// \endcode
		258	///
		259	/// This intrinsic corresponds to the \c POPCNT instruction or a
		260	/// a sequence of arithmetic and logic ops to calculate it.
		261	///
		262	/// \param A
		263	/// An unsigned 64-bit integer operand.
		264	/// \returns A 64-bit integer containing the number of bits with value 1 in the
		265	/// source operand.
		266	/// \see __popcntq
180	#define _popcnt64(A) __popcntq((A))	267	#define _popcnt64(A) __popcntq((A))
181	#endif /* __x86_64__ */	268	#endif /* __x86_64__ */
182		269
183	#ifdef __x86_64__	270	#ifdef __x86_64__
		271	/// Returns the program status and control \c RFLAGS register with the \c VM
		272	/// and \c RF flags cleared.
		273	///
		274	/// \headerfile <x86intrin.h>
		275	///
		276	/// This intrinsic corresponds to the \c PUSHFQ + \c POP instruction sequence.
		277	///
		278	/// \returns The 64-bit value of the RFLAGS register.
184	static __inline__ unsigned long long __DEFAULT_FN_ATTRS	279	static __inline__ unsigned long long __DEFAULT_FN_ATTRS
185	__readeflags(void)	280	__readeflags(void)
186	{	281	{
187	return __builtin_ia32_readeflags_u64();	282	return __builtin_ia32_readeflags_u64();
188	}	283	}
189		284
		285	/// Writes the specified value to the program status and control \c RFLAGS
		286	/// register. Reserved bits are not affected.
		287	///
		288	/// \headerfile <x86intrin.h>
		289	///
		290	/// This intrinsic corresponds to the \c PUSH + \c POPFQ instruction sequence.
		291	///
		292	/// \param __f
		293	/// The 64-bit value to write to \c RFLAGS.
190	static __inline__ void __DEFAULT_FN_ATTRS	294	static __inline__ void __DEFAULT_FN_ATTRS
191	__writeeflags(unsigned long long __f)	295	__writeeflags(unsigned long long __f)
192	{	296	{
...	@@ -194,12 +298,29 @@ __writeeflags(unsigned long long __f)	...	@@ -194,12 +298,29 @@ __writeeflags(unsigned long long __f)
194	}	298	}
195		299
196	#else /* !__x86_64__ */	300	#else /* !__x86_64__ */
		301	/// Returns the program status and control \c EFLAGS register with the \c VM
		302	/// and \c RF flags cleared.
		303	///
		304	/// \headerfile <x86intrin.h>
		305	///
		306	/// This intrinsic corresponds to the \c PUSHFD + \c POP instruction sequence.
		307	///
		308	/// \returns The 32-bit value of the EFLAGS register.
197	static __inline__ unsigned int __DEFAULT_FN_ATTRS	309	static __inline__ unsigned int __DEFAULT_FN_ATTRS
198	__readeflags(void)	310	__readeflags(void)
199	{	311	{
200	return __builtin_ia32_readeflags_u32();	312	return __builtin_ia32_readeflags_u32();
201	}	313	}
202		314
		315	/// Writes the specified value to the program status and control \c EFLAGS
		316	/// register. Reserved bits are not affected.
		317	///
		318	/// \headerfile <x86intrin.h>
		319	///
		320	/// This intrinsic corresponds to the \c PUSH + \c POPFD instruction sequence.
		321	///
		322	/// \param __f
		323	/// The 32-bit value to write to \c EFLAGS.
203	static __inline__ void __DEFAULT_FN_ATTRS	324	static __inline__ void __DEFAULT_FN_ATTRS
204	__writeeflags(unsigned int __f)	325	__writeeflags(unsigned int __f)
205	{	326	{
...	@@ -207,123 +328,120 @@ __writeeflags(unsigned int __f)	...	@@ -207,123 +328,120 @@ __writeeflags(unsigned int __f)
207	}	328	}
208	#endif /* !__x86_64__ */	329	#endif /* !__x86_64__ */
209		330
210	/** Cast a 32-bit float value to a 32-bit unsigned integer value	331	/// Cast a 32-bit float value to a 32-bit unsigned integer value.
211	*	332	///
212	* \headerfile <x86intrin.h>	333	/// \headerfile <x86intrin.h>
213	* This intrinsic corresponds to the <c> VMOVD / MOVD </c> instruction in x86_64,	334	///
214	* and corresponds to the <c> VMOVL / MOVL </c> instruction in ia32.	335	/// This intrinsic corresponds to the \c VMOVD / \c MOVD instruction in x86_64,
215	*	336	/// and corresponds to the \c VMOVL / \c MOVL instruction in ia32.
216	* \param __A	337	///
217	* A 32-bit float value.	338	/// \param __A
218	* \returns a 32-bit unsigned integer containing the converted value.	339	/// A 32-bit float value.
219	*/	340	/// \returns a 32-bit unsigned integer containing the converted value.
220	static __inline__ unsigned int __DEFAULT_FN_ATTRS_CAST	341	static __inline__ unsigned int __DEFAULT_FN_ATTRS_CAST
221	_castf32_u32(float __A) {	342	_castf32_u32(float __A) {
222	return __builtin_bit_cast(unsigned int, __A);	343	return __builtin_bit_cast(unsigned int, __A);
223	}	344	}
224		345
225	/** Cast a 64-bit float value to a 64-bit unsigned integer value	346	/// Cast a 64-bit float value to a 64-bit unsigned integer value.
226	*	347	///
227	* \headerfile <x86intrin.h>	348	/// \headerfile <x86intrin.h>
228	* This intrinsic corresponds to the <c> VMOVQ / MOVQ </c> instruction in x86_64,	349	///
229	* and corresponds to the <c> VMOVL / MOVL </c> instruction in ia32.	350	/// This intrinsic corresponds to the \c VMOVQ / \c MOVQ instruction in x86_64,
230	*	351	/// and corresponds to the \c VMOVL / \c MOVL instruction in ia32.
231	* \param __A	352	///
232	* A 64-bit float value.	353	/// \param __A
233	* \returns a 64-bit unsigned integer containing the converted value.	354	/// A 64-bit float value.
234	*/	355	/// \returns a 64-bit unsigned integer containing the converted value.
235	static __inline__ unsigned long long __DEFAULT_FN_ATTRS_CAST	356	static __inline__ unsigned long long __DEFAULT_FN_ATTRS_CAST
236	_castf64_u64(double __A) {	357	_castf64_u64(double __A) {
237	return __builtin_bit_cast(unsigned long long, __A);	358	return __builtin_bit_cast(unsigned long long, __A);
238	}	359	}
239		360
240	/** Cast a 32-bit unsigned integer value to a 32-bit float value	361	/// Cast a 32-bit unsigned integer value to a 32-bit float value.
241	*	362	///
242	* \headerfile <x86intrin.h>	363	/// \headerfile <x86intrin.h>
243	* This intrinsic corresponds to the <c> VMOVQ / MOVQ </c> instruction in x86_64,	364	///
244	* and corresponds to the <c> FLDS </c> instruction in ia32.	365	/// This intrinsic corresponds to the \c VMOVQ / \c MOVQ instruction in x86_64,
245	*	366	/// and corresponds to the \c FLDS instruction in ia32.
246	* \param __A	367	///
247	* A 32-bit unsigned integer value.	368	/// \param __A
248	* \returns a 32-bit float value containing the converted value.	369	/// A 32-bit unsigned integer value.
249	*/	370	/// \returns a 32-bit float value containing the converted value.
250	static __inline__ float __DEFAULT_FN_ATTRS_CAST	371	static __inline__ float __DEFAULT_FN_ATTRS_CAST
251	_castu32_f32(unsigned int __A) {	372	_castu32_f32(unsigned int __A) {
252	return __builtin_bit_cast(float, __A);	373	return __builtin_bit_cast(float, __A);
253	}	374	}
254		375
255	/** Cast a 64-bit unsigned integer value to a 64-bit float value	376	/// Cast a 64-bit unsigned integer value to a 64-bit float value.
256	*	377	///
257	* \headerfile <x86intrin.h>	378	/// \headerfile <x86intrin.h>
258	* This intrinsic corresponds to the <c> VMOVQ / MOVQ </c> instruction in x86_64,	379	///
259	* and corresponds to the <c> FLDL </c> instruction in ia32.	380	/// This intrinsic corresponds to the \c VMOVQ / \c MOVQ instruction in x86_64,
260	*	381	/// and corresponds to the \c FLDL instruction in ia32.
261	* \param __A	382	///
262	* A 64-bit unsigned integer value.	383	/// \param __A
263	* \returns a 64-bit float value containing the converted value.	384	/// A 64-bit unsigned integer value.
264	*/	385	/// \returns a 64-bit float value containing the converted value.
265	static __inline__ double __DEFAULT_FN_ATTRS_CAST	386	static __inline__ double __DEFAULT_FN_ATTRS_CAST
266	_castu64_f64(unsigned long long __A) {	387	_castu64_f64(unsigned long long __A) {
267	return __builtin_bit_cast(double, __A);	388	return __builtin_bit_cast(double, __A);
268	}	389	}
269		390
270	/** Adds the unsigned integer operand to the CRC-32C checksum of the	391	/// Adds the unsigned integer operand to the CRC-32C checksum of the
271	* unsigned char operand.	392	/// unsigned char operand.
272	*	393	///
273	* \headerfile <x86intrin.h>	394	/// \headerfile <x86intrin.h>
274	*	395	///
275	* This intrinsic corresponds to the <c> CRC32B </c> instruction.	396	/// This intrinsic corresponds to the \c CRC32B instruction.
276	*	397	///
277	* \param __C	398	/// \param __C
278	* An unsigned integer operand to add to the CRC-32C checksum of operand	399	/// An unsigned integer operand to add to the CRC-32C checksum of operand
279	* \a __D.	400	/// \a __D.
280	* \param __D	401	/// \param __D
281	* An unsigned 8-bit integer operand used to compute the CRC-32C checksum.	402	/// An unsigned 8-bit integer operand used to compute the CRC-32C checksum.
282	* \returns The result of adding operand \a __C to the CRC-32C checksum of	403	/// \returns The result of adding operand \a __C to the CRC-32C checksum of
283	* operand \a __D.	404	/// operand \a __D.
284	*/
285	static __inline__ unsigned int __DEFAULT_FN_ATTRS_CRC32	405	static __inline__ unsigned int __DEFAULT_FN_ATTRS_CRC32
286	__crc32b(unsigned int __C, unsigned char __D)	406	__crc32b(unsigned int __C, unsigned char __D)
287	{	407	{
288	return __builtin_ia32_crc32qi(__C, __D);	408	return __builtin_ia32_crc32qi(__C, __D);
289	}	409	}
290		410
291	/** Adds the unsigned integer operand to the CRC-32C checksum of the	411	/// Adds the unsigned integer operand to the CRC-32C checksum of the
292	* unsigned short operand.	412	/// unsigned short operand.
293	*	413	///
294	* \headerfile <x86intrin.h>	414	/// \headerfile <x86intrin.h>
295	*	415	///
296	* This intrinsic corresponds to the <c> CRC32W </c> instruction.	416	/// This intrinsic corresponds to the \c CRC32W instruction.
297	*	417	///
298	* \param __C	418	/// \param __C
299	* An unsigned integer operand to add to the CRC-32C checksum of operand	419	/// An unsigned integer operand to add to the CRC-32C checksum of operand
300	* \a __D.	420	/// \a __D.
301	* \param __D	421	/// \param __D
302	* An unsigned 16-bit integer operand used to compute the CRC-32C checksum.	422	/// An unsigned 16-bit integer operand used to compute the CRC-32C checksum.
303	* \returns The result of adding operand \a __C to the CRC-32C checksum of	423	/// \returns The result of adding operand \a __C to the CRC-32C checksum of
304	* operand \a __D.	424	/// operand \a __D.
305	*/
306	static __inline__ unsigned int __DEFAULT_FN_ATTRS_CRC32	425	static __inline__ unsigned int __DEFAULT_FN_ATTRS_CRC32
307	__crc32w(unsigned int __C, unsigned short __D)	426	__crc32w(unsigned int __C, unsigned short __D)
308	{	427	{
309	return __builtin_ia32_crc32hi(__C, __D);	428	return __builtin_ia32_crc32hi(__C, __D);
310	}	429	}
311		430
312	/** Adds the unsigned integer operand to the CRC-32C checksum of the	431	/// Adds the unsigned integer operand to the CRC-32C checksum of the
313	* second unsigned integer operand.	432	/// second unsigned integer operand.
314	*	433	///
315	* \headerfile <x86intrin.h>	434	/// \headerfile <x86intrin.h>
316	*	435	///
317	* This intrinsic corresponds to the <c> CRC32D </c> instruction.	436	/// This intrinsic corresponds to the \c CRC32D instruction.
318	*	437	///
319	* \param __C	438	/// \param __C
320	* An unsigned integer operand to add to the CRC-32C checksum of operand	439	/// An unsigned integer operand to add to the CRC-32C checksum of operand
321	* \a __D.	440	/// \a __D.
322	* \param __D	441	/// \param __D
323	* An unsigned 32-bit integer operand used to compute the CRC-32C checksum.	442	/// An unsigned 32-bit integer operand used to compute the CRC-32C checksum.
324	* \returns The result of adding operand \a __C to the CRC-32C checksum of	443	/// \returns The result of adding operand \a __C to the CRC-32C checksum of
325	* operand \a __D.	444	/// operand \a __D.
326	*/
327	static __inline__ unsigned int __DEFAULT_FN_ATTRS_CRC32	445	static __inline__ unsigned int __DEFAULT_FN_ATTRS_CRC32
328	__crc32d(unsigned int __C, unsigned int __D)	446	__crc32d(unsigned int __C, unsigned int __D)
329	{	447	{
...	@@ -331,21 +449,20 @@ __crc32d(unsigned int __C, unsigned int __D)	...	@@ -331,21 +449,20 @@ __crc32d(unsigned int __C, unsigned int __D)
331	}	449	}
332		450
333	#ifdef __x86_64__	451	#ifdef __x86_64__
334	/** Adds the unsigned integer operand to the CRC-32C checksum of the	452	/// Adds the unsigned integer operand to the CRC-32C checksum of the
335	* unsigned 64-bit integer operand.	453	/// unsigned 64-bit integer operand.
336	*	454	///
337	* \headerfile <x86intrin.h>	455	/// \headerfile <x86intrin.h>
338	*	456	///
339	* This intrinsic corresponds to the <c> CRC32Q </c> instruction.	457	/// This intrinsic corresponds to the \c CRC32Q instruction.
340	*	458	///
341	* \param __C	459	/// \param __C
342	* An unsigned integer operand to add to the CRC-32C checksum of operand	460	/// An unsigned integer operand to add to the CRC-32C checksum of operand
343	* \a __D.	461	/// \a __D.
344	* \param __D	462	/// \param __D
345	* An unsigned 64-bit integer operand used to compute the CRC-32C checksum.	463	/// An unsigned 64-bit integer operand used to compute the CRC-32C checksum.
346	* \returns The result of adding operand \a __C to the CRC-32C checksum of	464	/// \returns The result of adding operand \a __C to the CRC-32C checksum of
347	* operand \a __D.	465	/// operand \a __D.
348	*/
349	static __inline__ unsigned long long __DEFAULT_FN_ATTRS_CRC32	466	static __inline__ unsigned long long __DEFAULT_FN_ATTRS_CRC32
350	__crc32q(unsigned long long __C, unsigned long long __D)	467	__crc32q(unsigned long long __C, unsigned long long __D)
351	{	468	{
...	@@ -353,19 +470,67 @@ __crc32q(unsigned long long __C, unsigned long long __D)	...	@@ -353,19 +470,67 @@ __crc32q(unsigned long long __C, unsigned long long __D)
353	}	470	}
354	#endif /* __x86_64__ */	471	#endif /* __x86_64__ */
355		472
		473	/// Reads the specified performance monitoring counter. Refer to your
		474	/// processor's documentation to determine which performance counters are
		475	/// supported.
		476	///
		477	/// \headerfile <x86intrin.h>
		478	///
		479	/// This intrinsic corresponds to the \c RDPMC instruction.
		480	///
		481	/// \param __A
		482	/// The performance counter to read.
		483	/// \returns The 64-bit value read from the performance counter.
		484	/// \see _rdpmc
356	static __inline__ unsigned long long __DEFAULT_FN_ATTRS	485	static __inline__ unsigned long long __DEFAULT_FN_ATTRS
357	__rdpmc(int __A) {	486	__rdpmc(int __A) {
358	return __builtin_ia32_rdpmc(__A);	487	return __builtin_ia32_rdpmc(__A);
359	}	488	}
360		489
361	/* __rdtscp */	490	/// Reads the processor's time stamp counter and the \c IA32_TSC_AUX MSR
		491	/// \c (0xc0000103).
		492	///
		493	/// \headerfile <x86intrin.h>
		494	///
		495	/// This intrinsic corresponds to the \c RDTSCP instruction.
		496	///
		497	/// \param __A
		498	/// Address of where to store the 32-bit \c IA32_TSC_AUX value.
		499	/// \returns The 64-bit value of the time stamp counter.
362	static __inline__ unsigned long long __DEFAULT_FN_ATTRS	500	static __inline__ unsigned long long __DEFAULT_FN_ATTRS
363	__rdtscp(unsigned int *__A) {	501	__rdtscp(unsigned int *__A) {
364	return __builtin_ia32_rdtscp(__A);	502	return __builtin_ia32_rdtscp(__A);
365	}	503	}
366		504
		505	/// Reads the processor's time stamp counter.
		506	///
		507	/// \headerfile <x86intrin.h>
		508	///
		509	/// \code
		510	/// unsigned long long _rdtsc();
		511	/// \endcode
		512	///
		513	/// This intrinsic corresponds to the \c RDTSC instruction.
		514	///
		515	/// \returns The 64-bit value of the time stamp counter.
367	#define _rdtsc() __rdtsc()	516	#define _rdtsc() __rdtsc()
368		517
		518	/// Reads the specified performance monitoring counter. Refer to your
		519	/// processor's documentation to determine which performance counters are
		520	/// supported.
		521	///
		522	/// \headerfile <x86intrin.h>
		523	///
		524	/// \code
		525	/// unsigned long long _rdpmc(int A);
		526	/// \endcode
		527	///
		528	/// This intrinsic corresponds to the \c RDPMC instruction.
		529	///
		530	/// \param A
		531	/// The performance counter to read.
		532	/// \returns The 64-bit value read from the performance counter.
		533	/// \see __rdpmc
369	#define _rdpmc(A) __rdpmc(A)	534	#define _rdpmc(A) __rdpmc(A)
370		535
371	static __inline__ void __DEFAULT_FN_ATTRS	536	static __inline__ void __DEFAULT_FN_ATTRS
...	@@ -373,42 +538,150 @@ _wbinvd(void) {	...	@@ -373,42 +538,150 @@ _wbinvd(void) {
373	__builtin_ia32_wbinvd();	538	__builtin_ia32_wbinvd();
374	}	539	}
375		540
		541	/// Rotates an 8-bit value to the left by the specified number of bits.
		542	/// This operation is undefined if the number of bits exceeds the size of
		543	/// the value.
		544	///
		545	/// \headerfile <x86intrin.h>
		546	///
		547	/// This intrinsic corresponds to the \c ROL instruction.
		548	///
		549	/// \param __X
		550	/// The unsigned 8-bit value to be rotated.
		551	/// \param __C
		552	/// The number of bits to rotate the value.
		553	/// \returns The rotated value.
376	static __inline__ unsigned char __DEFAULT_FN_ATTRS_CONSTEXPR	554	static __inline__ unsigned char __DEFAULT_FN_ATTRS_CONSTEXPR
377	__rolb(unsigned char __X, int __C) {	555	__rolb(unsigned char __X, int __C) {
378	return __builtin_rotateleft8(__X, __C);	556	return __builtin_rotateleft8(__X, __C);
379	}	557	}
380		558
		559	/// Rotates an 8-bit value to the right by the specified number of bits.
		560	/// This operation is undefined if the number of bits exceeds the size of
		561	/// the value.
		562	///
		563	/// \headerfile <x86intrin.h>
		564	///
		565	/// This intrinsic corresponds to the \c ROR instruction.
		566	///
		567	/// \param __X
		568	/// The unsigned 8-bit value to be rotated.
		569	/// \param __C
		570	/// The number of bits to rotate the value.
		571	/// \returns The rotated value.
381	static __inline__ unsigned char __DEFAULT_FN_ATTRS_CONSTEXPR	572	static __inline__ unsigned char __DEFAULT_FN_ATTRS_CONSTEXPR
382	__rorb(unsigned char __X, int __C) {	573	__rorb(unsigned char __X, int __C) {
383	return __builtin_rotateright8(__X, __C);	574	return __builtin_rotateright8(__X, __C);
384	}	575	}
385		576
		577	/// Rotates a 16-bit value to the left by the specified number of bits.
		578	/// This operation is undefined if the number of bits exceeds the size of
		579	/// the value.
		580	///
		581	/// \headerfile <x86intrin.h>
		582	///
		583	/// This intrinsic corresponds to the \c ROL instruction.
		584	///
		585	/// \param __X
		586	/// The unsigned 16-bit value to be rotated.
		587	/// \param __C
		588	/// The number of bits to rotate the value.
		589	/// \returns The rotated value.
		590	/// \see _rotwl
386	static __inline__ unsigned short __DEFAULT_FN_ATTRS_CONSTEXPR	591	static __inline__ unsigned short __DEFAULT_FN_ATTRS_CONSTEXPR
387	__rolw(unsigned short __X, int __C) {	592	__rolw(unsigned short __X, int __C) {
388	return __builtin_rotateleft16(__X, __C);	593	return __builtin_rotateleft16(__X, __C);
389	}	594	}
390		595
		596	/// Rotates a 16-bit value to the right by the specified number of bits.
		597	/// This operation is undefined if the number of bits exceeds the size of
		598	/// the value.
		599	///
		600	/// \headerfile <x86intrin.h>
		601	///
		602	/// This intrinsic corresponds to the \c ROR instruction.
		603	///
		604	/// \param __X
		605	/// The unsigned 16-bit value to be rotated.
		606	/// \param __C
		607	/// The number of bits to rotate the value.
		608	/// \returns The rotated value.
		609	/// \see _rotwr
391	static __inline__ unsigned short __DEFAULT_FN_ATTRS_CONSTEXPR	610	static __inline__ unsigned short __DEFAULT_FN_ATTRS_CONSTEXPR
392	__rorw(unsigned short __X, int __C) {	611	__rorw(unsigned short __X, int __C) {
393	return __builtin_rotateright16(__X, __C);	612	return __builtin_rotateright16(__X, __C);
394	}	613	}
395		614
		615	/// Rotates a 32-bit value to the left by the specified number of bits.
		616	/// This operation is undefined if the number of bits exceeds the size of
		617	/// the value.
		618	///
		619	/// \headerfile <x86intrin.h>
		620	///
		621	/// This intrinsic corresponds to the \c ROL instruction.
		622	///
		623	/// \param __X
		624	/// The unsigned 32-bit value to be rotated.
		625	/// \param __C
		626	/// The number of bits to rotate the value.
		627	/// \returns The rotated value.
		628	/// \see _rotl
396	static __inline__ unsigned int __DEFAULT_FN_ATTRS_CONSTEXPR	629	static __inline__ unsigned int __DEFAULT_FN_ATTRS_CONSTEXPR
397	__rold(unsigned int __X, int __C) {	630	__rold(unsigned int __X, int __C) {
398	return __builtin_rotateleft32(__X, (unsigned int)__C);	631	return __builtin_rotateleft32(__X, (unsigned int)__C);
399	}	632	}
400		633
		634	/// Rotates a 32-bit value to the right by the specified number of bits.
		635	/// This operation is undefined if the number of bits exceeds the size of
		636	/// the value.
		637	///
		638	/// \headerfile <x86intrin.h>
		639	///
		640	/// This intrinsic corresponds to the \c ROR instruction.
		641	///
		642	/// \param __X
		643	/// The unsigned 32-bit value to be rotated.
		644	/// \param __C
		645	/// The number of bits to rotate the value.
		646	/// \returns The rotated value.
		647	/// \see _rotr
401	static __inline__ unsigned int __DEFAULT_FN_ATTRS_CONSTEXPR	648	static __inline__ unsigned int __DEFAULT_FN_ATTRS_CONSTEXPR
402	__rord(unsigned int __X, int __C) {	649	__rord(unsigned int __X, int __C) {
403	return __builtin_rotateright32(__X, (unsigned int)__C);	650	return __builtin_rotateright32(__X, (unsigned int)__C);
404	}	651	}
405		652
406	#ifdef __x86_64__	653	#ifdef __x86_64__
		654	/// Rotates a 64-bit value to the left by the specified number of bits.
		655	/// This operation is undefined if the number of bits exceeds the size of
		656	/// the value.
		657	///
		658	/// \headerfile <x86intrin.h>
		659	///
		660	/// This intrinsic corresponds to the \c ROL instruction.
		661	///
		662	/// \param __X
		663	/// The unsigned 64-bit value to be rotated.
		664	/// \param __C
		665	/// The number of bits to rotate the value.
		666	/// \returns The rotated value.
407	static __inline__ unsigned long long __DEFAULT_FN_ATTRS_CONSTEXPR	667	static __inline__ unsigned long long __DEFAULT_FN_ATTRS_CONSTEXPR
408	__rolq(unsigned long long __X, int __C) {	668	__rolq(unsigned long long __X, int __C) {
409	return __builtin_rotateleft64(__X, (unsigned long long)__C);	669	return __builtin_rotateleft64(__X, (unsigned long long)__C);
410	}	670	}
411		671
		672	/// Rotates a 64-bit value to the right by the specified number of bits.
		673	/// This operation is undefined if the number of bits exceeds the size of
		674	/// the value.
		675	///
		676	/// \headerfile <x86intrin.h>
		677	///
		678	/// This intrinsic corresponds to the \c ROR instruction.
		679	///
		680	/// \param __X
		681	/// The unsigned 64-bit value to be rotated.
		682	/// \param __C
		683	/// The number of bits to rotate the value.
		684	/// \returns The rotated value.
412	static __inline__ unsigned long long __DEFAULT_FN_ATTRS_CONSTEXPR	685	static __inline__ unsigned long long __DEFAULT_FN_ATTRS_CONSTEXPR
413	__rorq(unsigned long long __X, int __C) {	686	__rorq(unsigned long long __X, int __C) {
414	return __builtin_rotateright64(__X, (unsigned long long)__C);	687	return __builtin_rotateright64(__X, (unsigned long long)__C);
...	@@ -419,18 +692,167 @@ __rorq(unsigned long long __X, int __C) {	...	@@ -419,18 +692,167 @@ __rorq(unsigned long long __X, int __C) {
419	/* These are already provided as builtins for MSVC. */	692	/* These are already provided as builtins for MSVC. */
420	/* Select the correct function based on the size of long. */	693	/* Select the correct function based on the size of long. */
421	#ifdef __LP64__	694	#ifdef __LP64__
		695	/// Rotates a 64-bit value to the left by the specified number of bits.
		696	/// This operation is undefined if the number of bits exceeds the size of
		697	/// the value.
		698	///
		699	/// \headerfile <x86intrin.h>
		700	///
		701	/// \code
		702	/// unsigned long long _lrotl(unsigned long long a, int b);
		703	/// \endcode
		704	///
		705	/// This intrinsic corresponds to the \c ROL instruction.
		706	///
		707	/// \param a
		708	/// The unsigned 64-bit value to be rotated.
		709	/// \param b
		710	/// The number of bits to rotate the value.
		711	/// \returns The rotated value.
		712	/// \see __rolq
422	#define _lrotl(a,b) __rolq((a), (b))	713	#define _lrotl(a,b) __rolq((a), (b))
		714
		715	/// Rotates a 64-bit value to the right by the specified number of bits.
		716	/// This operation is undefined if the number of bits exceeds the size of
		717	/// the value.
		718	///
		719	/// \headerfile <x86intrin.h>
		720	///
		721	/// \code
		722	/// unsigned long long _lrotr(unsigned long long a, int b);
		723	/// \endcode
		724	///
		725	/// This intrinsic corresponds to the \c ROR instruction.
		726	///
		727	/// \param a
		728	/// The unsigned 64-bit value to be rotated.
		729	/// \param b
		730	/// The number of bits to rotate the value.
		731	/// \returns The rotated value.
		732	/// \see __rorq
423	#define _lrotr(a,b) __rorq((a), (b))	733	#define _lrotr(a,b) __rorq((a), (b))
424	#else	734	#else // __LP64__
		735	/// Rotates a 32-bit value to the left by the specified number of bits.
		736	/// This operation is undefined if the number of bits exceeds the size of
		737	/// the value.
		738	///
		739	/// \headerfile <x86intrin.h>
		740	///
		741	/// \code
		742	/// unsigned int _lrotl(unsigned int a, int b);
		743	/// \endcode
		744	///
		745	/// This intrinsic corresponds to the \c ROL instruction.
		746	///
		747	/// \param a
		748	/// The unsigned 32-bit value to be rotated.
		749	/// \param b
		750	/// The number of bits to rotate the value.
		751	/// \returns The rotated value.
		752	/// \see __rold
425	#define _lrotl(a,b) __rold((a), (b))	753	#define _lrotl(a,b) __rold((a), (b))
		754
		755	/// Rotates a 32-bit value to the right by the specified number of bits.
		756	/// This operation is undefined if the number of bits exceeds the size of
		757	/// the value.
		758	///
		759	/// \headerfile <x86intrin.h>
		760	///
		761	/// \code
		762	/// unsigned int _lrotr(unsigned int a, int b);
		763	/// \endcode
		764	///
		765	/// This intrinsic corresponds to the \c ROR instruction.
		766	///
		767	/// \param a
		768	/// The unsigned 32-bit value to be rotated.
		769	/// \param b
		770	/// The number of bits to rotate the value.
		771	/// \returns The rotated value.
		772	/// \see __rord
426	#define _lrotr(a,b) __rord((a), (b))	773	#define _lrotr(a,b) __rord((a), (b))
427	#endif	774	#endif // __LP64__
		775
		776	/// Rotates a 32-bit value to the left by the specified number of bits.
		777	/// This operation is undefined if the number of bits exceeds the size of
		778	/// the value.
		779	///
		780	/// \headerfile <x86intrin.h>
		781	///
		782	/// \code
		783	/// unsigned int _rotl(unsigned int a, int b);
		784	/// \endcode
		785	///
		786	/// This intrinsic corresponds to the \c ROL instruction.
		787	///
		788	/// \param a
		789	/// The unsigned 32-bit value to be rotated.
		790	/// \param b
		791	/// The number of bits to rotate the value.
		792	/// \returns The rotated value.
		793	/// \see __rold
428	#define _rotl(a,b) __rold((a), (b))	794	#define _rotl(a,b) __rold((a), (b))
		795
		796	/// Rotates a 32-bit value to the right by the specified number of bits.
		797	/// This operation is undefined if the number of bits exceeds the size of
		798	/// the value.
		799	///
		800	/// \headerfile <x86intrin.h>
		801	///
		802	/// \code
		803	/// unsigned int _rotr(unsigned int a, int b);
		804	/// \endcode
		805	///
		806	/// This intrinsic corresponds to the \c ROR instruction.
		807	///
		808	/// \param a
		809	/// The unsigned 32-bit value to be rotated.
		810	/// \param b
		811	/// The number of bits to rotate the value.
		812	/// \returns The rotated value.
		813	/// \see __rord
429	#define _rotr(a,b) __rord((a), (b))	814	#define _rotr(a,b) __rord((a), (b))
430	#endif // _MSC_VER	815	#endif // _MSC_VER
431		816
432	/* These are not builtins so need to be provided in all modes. */	817	/* These are not builtins so need to be provided in all modes. */
		818	/// Rotates a 16-bit value to the left by the specified number of bits.
		819	/// This operation is undefined if the number of bits exceeds the size of
		820	/// the value.
		821	///
		822	/// \headerfile <x86intrin.h>
		823	///
		824	/// \code
		825	/// unsigned short _rotwl(unsigned short a, int b);
		826	/// \endcode
		827	///
		828	/// This intrinsic corresponds to the \c ROL instruction.
		829	///
		830	/// \param a
		831	/// The unsigned 16-bit value to be rotated.
		832	/// \param b
		833	/// The number of bits to rotate the value.
		834	/// \returns The rotated value.
		835	/// \see __rolw
433	#define _rotwl(a,b) __rolw((a), (b))	836	#define _rotwl(a,b) __rolw((a), (b))
		837
		838	/// Rotates a 16-bit value to the right by the specified number of bits.
		839	/// This operation is undefined if the number of bits exceeds the size of
		840	/// the value.
		841	///
		842	/// \headerfile <x86intrin.h>
		843	///
		844	/// \code
		845	/// unsigned short _rotwr(unsigned short a, int b);
		846	/// \endcode
		847	///
		848	/// This intrinsic corresponds to the \c ROR instruction.
		849	///
		850	/// \param a
		851	/// The unsigned 16-bit value to be rotated.
		852	/// \param b
		853	/// The number of bits to rotate the value.
		854	/// \returns The rotated value.
		855	/// \see __rorw
434	#define _rotwr(a,b) __rorw((a), (b))	856	#define _rotwr(a,b) __rorw((a), (b))
435		857
436	#undef __DEFAULT_FN_ATTRS	858	#undef __DEFAULT_FN_ATTRS

lib/include/immintrin.h+66-3

...	@@ -291,11 +291,13 @@	...	@@ -291,11 +291,13 @@
291		291
292	#if !(defined(_MSC_VER) \|\| defined(__SCE__)) \|\| __has_feature(modules) \|\| \	292	#if !(defined(_MSC_VER) \|\| defined(__SCE__)) \|\| __has_feature(modules) \|\| \
293	defined(__RDPID__)	293	defined(__RDPID__)
294	/// Returns the value of the IA32_TSC_AUX MSR (0xc0000103).	294	/// Reads the value of the IA32_TSC_AUX MSR (0xc0000103).
295	///	295	///
296	/// \headerfile <immintrin.h>	296	/// \headerfile <immintrin.h>
297	///	297	///
298	/// This intrinsic corresponds to the <c> RDPID </c> instruction.	298	/// This intrinsic corresponds to the <c> RDPID </c> instruction.
		299	///
		300	/// \returns The 32-bit contents of the MSR.
299	static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__, __target__("rdpid")))	301	static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__, __target__("rdpid")))
300	_rdpid_u32(void) {	302	_rdpid_u32(void) {
301	return __builtin_ia32_rdpid();	303	return __builtin_ia32_rdpid();
...	@@ -488,6 +490,15 @@ _writegsbase_u64(unsigned long long __V)	...	@@ -488,6 +490,15 @@ _writegsbase_u64(unsigned long long __V)
488	* field inside of it.	490	* field inside of it.
489	*/	491	*/
490		492
		493	/// Load a 16-bit value from memory and swap its bytes.
		494	///
		495	/// \headerfile <x86intrin.h>
		496	///
		497	/// This intrinsic corresponds to the MOVBE instruction.
		498	///
		499	/// \param __P
		500	/// A pointer to the 16-bit value to load.
		501	/// \returns The byte-swapped value.
491	static __inline__ short __attribute__((__always_inline__, __nodebug__, __target__("movbe")))	502	static __inline__ short __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
492	_loadbe_i16(void const * __P) {	503	_loadbe_i16(void const * __P) {
493	struct __loadu_i16 {	504	struct __loadu_i16 {
...	@@ -496,6 +507,16 @@ _loadbe_i16(void const * __P) {	...	@@ -496,6 +507,16 @@ _loadbe_i16(void const * __P) {
496	return (short)__builtin_bswap16(((const struct __loadu_i16*)__P)->__v);	507	return (short)__builtin_bswap16(((const struct __loadu_i16*)__P)->__v);
497	}	508	}
498		509
		510	/// Swap the bytes of a 16-bit value and store it to memory.
		511	///
		512	/// \headerfile <x86intrin.h>
		513	///
		514	/// This intrinsic corresponds to the MOVBE instruction.
		515	///
		516	/// \param __P
		517	/// A pointer to the memory for storing the swapped value.
		518	/// \param __D
		519	/// The 16-bit value to be byte-swapped.
499	static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("movbe")))	520	static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
500	_storebe_i16(void * __P, short __D) {	521	_storebe_i16(void * __P, short __D) {
501	struct __storeu_i16 {	522	struct __storeu_i16 {
...	@@ -504,6 +525,15 @@ _storebe_i16(void * __P, short __D) {	...	@@ -504,6 +525,15 @@ _storebe_i16(void * __P, short __D) {
504	((struct __storeu_i16*)__P)->__v = __builtin_bswap16((unsigned short)__D);	525	((struct __storeu_i16*)__P)->__v = __builtin_bswap16((unsigned short)__D);
505	}	526	}
506		527
		528	/// Load a 32-bit value from memory and swap its bytes.
		529	///
		530	/// \headerfile <x86intrin.h>
		531	///
		532	/// This intrinsic corresponds to the MOVBE instruction.
		533	///
		534	/// \param __P
		535	/// A pointer to the 32-bit value to load.
		536	/// \returns The byte-swapped value.
507	static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("movbe")))	537	static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
508	_loadbe_i32(void const * __P) {	538	_loadbe_i32(void const * __P) {
509	struct __loadu_i32 {	539	struct __loadu_i32 {
...	@@ -512,6 +542,16 @@ _loadbe_i32(void const * __P) {	...	@@ -512,6 +542,16 @@ _loadbe_i32(void const * __P) {
512	return (int)__builtin_bswap32(((const struct __loadu_i32*)__P)->__v);	542	return (int)__builtin_bswap32(((const struct __loadu_i32*)__P)->__v);
513	}	543	}
514		544
		545	/// Swap the bytes of a 32-bit value and store it to memory.
		546	///
		547	/// \headerfile <x86intrin.h>
		548	///
		549	/// This intrinsic corresponds to the MOVBE instruction.
		550	///
		551	/// \param __P
		552	/// A pointer to the memory for storing the swapped value.
		553	/// \param __D
		554	/// The 32-bit value to be byte-swapped.
515	static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("movbe")))	555	static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
516	_storebe_i32(void * __P, int __D) {	556	_storebe_i32(void * __P, int __D) {
517	struct __storeu_i32 {	557	struct __storeu_i32 {
...	@@ -521,6 +561,15 @@ _storebe_i32(void * __P, int __D) {	...	@@ -521,6 +561,15 @@ _storebe_i32(void * __P, int __D) {
521	}	561	}
522		562
523	#ifdef __x86_64__	563	#ifdef __x86_64__
		564	/// Load a 64-bit value from memory and swap its bytes.
		565	///
		566	/// \headerfile <x86intrin.h>
		567	///
		568	/// This intrinsic corresponds to the MOVBE instruction.
		569	///
		570	/// \param __P
		571	/// A pointer to the 64-bit value to load.
		572	/// \returns The byte-swapped value.
524	static __inline__ long long __attribute__((__always_inline__, __nodebug__, __target__("movbe")))	573	static __inline__ long long __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
525	_loadbe_i64(void const * __P) {	574	_loadbe_i64(void const * __P) {
526	struct __loadu_i64 {	575	struct __loadu_i64 {
...	@@ -529,6 +578,16 @@ _loadbe_i64(void const * __P) {	...	@@ -529,6 +578,16 @@ _loadbe_i64(void const * __P) {
529	return (long long)__builtin_bswap64(((const struct __loadu_i64*)__P)->__v);	578	return (long long)__builtin_bswap64(((const struct __loadu_i64*)__P)->__v);
530	}	579	}
531		580
		581	/// Swap the bytes of a 64-bit value and store it to memory.
		582	///
		583	/// \headerfile <x86intrin.h>
		584	///
		585	/// This intrinsic corresponds to the MOVBE instruction.
		586	///
		587	/// \param __P
		588	/// A pointer to the memory for storing the swapped value.
		589	/// \param __D
		590	/// The 64-bit value to be byte-swapped.
532	static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("movbe")))	591	static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
533	_storebe_i64(void * __P, long long __D) {	592	_storebe_i64(void * __P, long long __D) {
534	struct __storeu_i64 {	593	struct __storeu_i64 {
...	@@ -578,9 +637,13 @@ _storebe_i64(void * __P, long long __D) {	...	@@ -578,9 +637,13 @@ _storebe_i64(void * __P, long long __D) {
578	#include <cetintrin.h>	637	#include <cetintrin.h>
579	#endif	638	#endif
580		639
581	/* Some intrinsics inside adxintrin.h are available only on processors with ADX,	640	/* Intrinsics inside adcintrin.h are available at all times. */
582	* whereas others are also available at all times. */	641	#include <adcintrin.h>
		642
		643	#if !(defined(_MSC_VER) \|\| defined(__SCE__)) \|\| __has_feature(modules) \|\| \
		644	defined(__ADX__)
583	#include <adxintrin.h>	645	#include <adxintrin.h>
		646	#endif
584		647
585	#if !(defined(_MSC_VER) \|\| defined(__SCE__)) \|\| __has_feature(modules) \|\| \	648	#if !(defined(_MSC_VER) \|\| defined(__SCE__)) \|\| __has_feature(modules) \|\| \
586	defined(__RDSEED__)	649	defined(__RDSEED__)

lib/include/intrin.h+16

...	@@ -572,6 +572,22 @@ unsigned char __readx18byte(unsigned long offset);	...	@@ -572,6 +572,22 @@ unsigned char __readx18byte(unsigned long offset);
572	unsigned short __readx18word(unsigned long offset);	572	unsigned short __readx18word(unsigned long offset);
573	unsigned long __readx18dword(unsigned long offset);	573	unsigned long __readx18dword(unsigned long offset);
574	unsigned __int64 __readx18qword(unsigned long offset);	574	unsigned __int64 __readx18qword(unsigned long offset);
		575
		576	double _CopyDoubleFromInt64(__int64);
		577	float _CopyFloatFromInt32(__int32);
		578	__int32 _CopyInt32FromFloat(float);
		579	__int64 _CopyInt64FromDouble(double);
		580
		581	unsigned int _CountLeadingOnes(unsigned long);
		582	unsigned int _CountLeadingOnes64(unsigned __int64);
		583	unsigned int _CountLeadingSigns(long);
		584	unsigned int _CountLeadingSigns64(__int64);
		585	unsigned int _CountLeadingZeros(unsigned long);
		586	unsigned int _CountLeadingZeros64(unsigned _int64);
		587	unsigned int _CountOneBits(unsigned long);
		588	unsigned int _CountOneBits64(unsigned __int64);
		589
		590	void __cdecl __prefetch(void *);
575	#endif	591	#endif
576		592
577	/----------------------------------------------------------------------------\	593	/----------------------------------------------------------------------------\

lib/include/larchintrin.h+13-1

...	@@ -156,7 +156,7 @@ extern __inline unsigned char	...	@@ -156,7 +156,7 @@ extern __inline unsigned char
156	return (unsigned char)__builtin_loongarch_iocsrrd_b((unsigned int)_1);	156	return (unsigned char)__builtin_loongarch_iocsrrd_b((unsigned int)_1);
157	}	157	}
158		158
159	extern __inline unsigned char	159	extern __inline unsigned short
160	__attribute__((__gnu_inline__, __always_inline__, __artificial__))	160	__attribute__((__gnu_inline__, __always_inline__, __artificial__))
161	__iocsrrd_h(unsigned int _1) {	161	__iocsrrd_h(unsigned int _1) {
162	return (unsigned short)__builtin_loongarch_iocsrrd_h((unsigned int)_1);	162	return (unsigned short)__builtin_loongarch_iocsrrd_h((unsigned int)_1);
...	@@ -228,6 +228,18 @@ extern __inline void	...	@@ -228,6 +228,18 @@ extern __inline void
228	((void)__builtin_loongarch_ldpte_d((long int)(_1), (_2)))	228	((void)__builtin_loongarch_ldpte_d((long int)(_1), (_2)))
229	#endif	229	#endif
230		230
		231	#define __frecipe_s(/float/ _1) \
		232	(float)__builtin_loongarch_frecipe_s((float)_1)
		233
		234	#define __frecipe_d(/double/ _1) \
		235	(double)__builtin_loongarch_frecipe_d((double)_1)
		236
		237	#define __frsqrte_s(/float/ _1) \
		238	(float)__builtin_loongarch_frsqrte_s((float)_1)
		239
		240	#define __frsqrte_d(/double/ _1) \
		241	(double)__builtin_loongarch_frsqrte_d((double)_1)
		242
231	#ifdef __cplusplus	243	#ifdef __cplusplus
232	}	244	}
233	#endif	245	#endif

lib/include/lasxintrin.h created+3884

...	@@ -0,0 +1,3884 @@
1	/*===------------ lasxintrin.h - LoongArch LASX intrinsics -----------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	#ifndef _LOONGSON_ASXINTRIN_H
11	#define _LOONGSON_ASXINTRIN_H 1
12
13	#if defined(__loongarch_asx)
14
15	typedef signed char v32i8 __attribute__((vector_size(32), aligned(32)));
16	typedef signed char v32i8_b __attribute__((vector_size(32), aligned(1)));
17	typedef unsigned char v32u8 __attribute__((vector_size(32), aligned(32)));
18	typedef unsigned char v32u8_b __attribute__((vector_size(32), aligned(1)));
19	typedef short v16i16 __attribute__((vector_size(32), aligned(32)));
20	typedef short v16i16_h __attribute__((vector_size(32), aligned(2)));
21	typedef unsigned short v16u16 __attribute__((vector_size(32), aligned(32)));
22	typedef unsigned short v16u16_h __attribute__((vector_size(32), aligned(2)));
23	typedef int v8i32 __attribute__((vector_size(32), aligned(32)));
24	typedef int v8i32_w __attribute__((vector_size(32), aligned(4)));
25	typedef unsigned int v8u32 __attribute__((vector_size(32), aligned(32)));
26	typedef unsigned int v8u32_w __attribute__((vector_size(32), aligned(4)));
27	typedef long long v4i64 __attribute__((vector_size(32), aligned(32)));
28	typedef long long v4i64_d __attribute__((vector_size(32), aligned(8)));
29	typedef unsigned long long v4u64 __attribute__((vector_size(32), aligned(32)));
30	typedef unsigned long long v4u64_d __attribute__((vector_size(32), aligned(8)));
31	typedef float v8f32 __attribute__((vector_size(32), aligned(32)));
32	typedef float v8f32_w __attribute__((vector_size(32), aligned(4)));
33	typedef double v4f64 __attribute__((vector_size(32), aligned(32)));
34	typedef double v4f64_d __attribute__((vector_size(32), aligned(8)));
35
36	typedef double v4f64 __attribute__((vector_size(32), aligned(32)));
37	typedef double v4f64_d __attribute__((vector_size(32), aligned(8)));
38
39	typedef float __m256 __attribute__((__vector_size__(32), __may_alias__));
40	typedef long long __m256i __attribute__((__vector_size__(32), __may_alias__));
41	typedef double __m256d __attribute__((__vector_size__(32), __may_alias__));
42
43	extern __inline
44	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
45	__lasx_xvsll_b(__m256i _1, __m256i _2) {
46	return (__m256i)__builtin_lasx_xvsll_b((v32i8)_1, (v32i8)_2);
47	}
48
49	extern __inline
50	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
51	__lasx_xvsll_h(__m256i _1, __m256i _2) {
52	return (__m256i)__builtin_lasx_xvsll_h((v16i16)_1, (v16i16)_2);
53	}
54
55	extern __inline
56	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
57	__lasx_xvsll_w(__m256i _1, __m256i _2) {
58	return (__m256i)__builtin_lasx_xvsll_w((v8i32)_1, (v8i32)_2);
59	}
60
61	extern __inline
62	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
63	__lasx_xvsll_d(__m256i _1, __m256i _2) {
64	return (__m256i)__builtin_lasx_xvsll_d((v4i64)_1, (v4i64)_2);
65	}
66
67	#define __lasx_xvslli_b(/__m256i/ _1, /ui3/ _2) \
68	((__m256i)__builtin_lasx_xvslli_b((v32i8)(_1), (_2)))
69
70	#define __lasx_xvslli_h(/__m256i/ _1, /ui4/ _2) \
71	((__m256i)__builtin_lasx_xvslli_h((v16i16)(_1), (_2)))
72
73	#define __lasx_xvslli_w(/__m256i/ _1, /ui5/ _2) \
74	((__m256i)__builtin_lasx_xvslli_w((v8i32)(_1), (_2)))
75
76	#define __lasx_xvslli_d(/__m256i/ _1, /ui6/ _2) \
77	((__m256i)__builtin_lasx_xvslli_d((v4i64)(_1), (_2)))
78
79	extern __inline
80	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
81	__lasx_xvsra_b(__m256i _1, __m256i _2) {
82	return (__m256i)__builtin_lasx_xvsra_b((v32i8)_1, (v32i8)_2);
83	}
84
85	extern __inline
86	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
87	__lasx_xvsra_h(__m256i _1, __m256i _2) {
88	return (__m256i)__builtin_lasx_xvsra_h((v16i16)_1, (v16i16)_2);
89	}
90
91	extern __inline
92	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
93	__lasx_xvsra_w(__m256i _1, __m256i _2) {
94	return (__m256i)__builtin_lasx_xvsra_w((v8i32)_1, (v8i32)_2);
95	}
96
97	extern __inline
98	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
99	__lasx_xvsra_d(__m256i _1, __m256i _2) {
100	return (__m256i)__builtin_lasx_xvsra_d((v4i64)_1, (v4i64)_2);
101	}
102
103	#define __lasx_xvsrai_b(/__m256i/ _1, /ui3/ _2) \
104	((__m256i)__builtin_lasx_xvsrai_b((v32i8)(_1), (_2)))
105
106	#define __lasx_xvsrai_h(/__m256i/ _1, /ui4/ _2) \
107	((__m256i)__builtin_lasx_xvsrai_h((v16i16)(_1), (_2)))
108
109	#define __lasx_xvsrai_w(/__m256i/ _1, /ui5/ _2) \
110	((__m256i)__builtin_lasx_xvsrai_w((v8i32)(_1), (_2)))
111
112	#define __lasx_xvsrai_d(/__m256i/ _1, /ui6/ _2) \
113	((__m256i)__builtin_lasx_xvsrai_d((v4i64)(_1), (_2)))
114
115	extern __inline
116	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
117	__lasx_xvsrar_b(__m256i _1, __m256i _2) {
118	return (__m256i)__builtin_lasx_xvsrar_b((v32i8)_1, (v32i8)_2);
119	}
120
121	extern __inline
122	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
123	__lasx_xvsrar_h(__m256i _1, __m256i _2) {
124	return (__m256i)__builtin_lasx_xvsrar_h((v16i16)_1, (v16i16)_2);
125	}
126
127	extern __inline
128	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
129	__lasx_xvsrar_w(__m256i _1, __m256i _2) {
130	return (__m256i)__builtin_lasx_xvsrar_w((v8i32)_1, (v8i32)_2);
131	}
132
133	extern __inline
134	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
135	__lasx_xvsrar_d(__m256i _1, __m256i _2) {
136	return (__m256i)__builtin_lasx_xvsrar_d((v4i64)_1, (v4i64)_2);
137	}
138
139	#define __lasx_xvsrari_b(/__m256i/ _1, /ui3/ _2) \
140	((__m256i)__builtin_lasx_xvsrari_b((v32i8)(_1), (_2)))
141
142	#define __lasx_xvsrari_h(/__m256i/ _1, /ui4/ _2) \
143	((__m256i)__builtin_lasx_xvsrari_h((v16i16)(_1), (_2)))
144
145	#define __lasx_xvsrari_w(/__m256i/ _1, /ui5/ _2) \
146	((__m256i)__builtin_lasx_xvsrari_w((v8i32)(_1), (_2)))
147
148	#define __lasx_xvsrari_d(/__m256i/ _1, /ui6/ _2) \
149	((__m256i)__builtin_lasx_xvsrari_d((v4i64)(_1), (_2)))
150
151	extern __inline
152	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
153	__lasx_xvsrl_b(__m256i _1, __m256i _2) {
154	return (__m256i)__builtin_lasx_xvsrl_b((v32i8)_1, (v32i8)_2);
155	}
156
157	extern __inline
158	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
159	__lasx_xvsrl_h(__m256i _1, __m256i _2) {
160	return (__m256i)__builtin_lasx_xvsrl_h((v16i16)_1, (v16i16)_2);
161	}
162
163	extern __inline
164	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
165	__lasx_xvsrl_w(__m256i _1, __m256i _2) {
166	return (__m256i)__builtin_lasx_xvsrl_w((v8i32)_1, (v8i32)_2);
167	}
168
169	extern __inline
170	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
171	__lasx_xvsrl_d(__m256i _1, __m256i _2) {
172	return (__m256i)__builtin_lasx_xvsrl_d((v4i64)_1, (v4i64)_2);
173	}
174
175	#define __lasx_xvsrli_b(/__m256i/ _1, /ui3/ _2) \
176	((__m256i)__builtin_lasx_xvsrli_b((v32i8)(_1), (_2)))
177
178	#define __lasx_xvsrli_h(/__m256i/ _1, /ui4/ _2) \
179	((__m256i)__builtin_lasx_xvsrli_h((v16i16)(_1), (_2)))
180
181	#define __lasx_xvsrli_w(/__m256i/ _1, /ui5/ _2) \
182	((__m256i)__builtin_lasx_xvsrli_w((v8i32)(_1), (_2)))
183
184	#define __lasx_xvsrli_d(/__m256i/ _1, /ui6/ _2) \
185	((__m256i)__builtin_lasx_xvsrli_d((v4i64)(_1), (_2)))
186
187	extern __inline
188	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
189	__lasx_xvsrlr_b(__m256i _1, __m256i _2) {
190	return (__m256i)__builtin_lasx_xvsrlr_b((v32i8)_1, (v32i8)_2);
191	}
192
193	extern __inline
194	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
195	__lasx_xvsrlr_h(__m256i _1, __m256i _2) {
196	return (__m256i)__builtin_lasx_xvsrlr_h((v16i16)_1, (v16i16)_2);
197	}
198
199	extern __inline
200	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
201	__lasx_xvsrlr_w(__m256i _1, __m256i _2) {
202	return (__m256i)__builtin_lasx_xvsrlr_w((v8i32)_1, (v8i32)_2);
203	}
204
205	extern __inline
206	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
207	__lasx_xvsrlr_d(__m256i _1, __m256i _2) {
208	return (__m256i)__builtin_lasx_xvsrlr_d((v4i64)_1, (v4i64)_2);
209	}
210
211	#define __lasx_xvsrlri_b(/__m256i/ _1, /ui3/ _2) \
212	((__m256i)__builtin_lasx_xvsrlri_b((v32i8)(_1), (_2)))
213
214	#define __lasx_xvsrlri_h(/__m256i/ _1, /ui4/ _2) \
215	((__m256i)__builtin_lasx_xvsrlri_h((v16i16)(_1), (_2)))
216
217	#define __lasx_xvsrlri_w(/__m256i/ _1, /ui5/ _2) \
218	((__m256i)__builtin_lasx_xvsrlri_w((v8i32)(_1), (_2)))
219
220	#define __lasx_xvsrlri_d(/__m256i/ _1, /ui6/ _2) \
221	((__m256i)__builtin_lasx_xvsrlri_d((v4i64)(_1), (_2)))
222
223	extern __inline
224	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
225	__lasx_xvbitclr_b(__m256i _1, __m256i _2) {
226	return (__m256i)__builtin_lasx_xvbitclr_b((v32u8)_1, (v32u8)_2);
227	}
228
229	extern __inline
230	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
231	__lasx_xvbitclr_h(__m256i _1, __m256i _2) {
232	return (__m256i)__builtin_lasx_xvbitclr_h((v16u16)_1, (v16u16)_2);
233	}
234
235	extern __inline
236	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
237	__lasx_xvbitclr_w(__m256i _1, __m256i _2) {
238	return (__m256i)__builtin_lasx_xvbitclr_w((v8u32)_1, (v8u32)_2);
239	}
240
241	extern __inline
242	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
243	__lasx_xvbitclr_d(__m256i _1, __m256i _2) {
244	return (__m256i)__builtin_lasx_xvbitclr_d((v4u64)_1, (v4u64)_2);
245	}
246
247	#define __lasx_xvbitclri_b(/__m256i/ _1, /ui3/ _2) \
248	((__m256i)__builtin_lasx_xvbitclri_b((v32u8)(_1), (_2)))
249
250	#define __lasx_xvbitclri_h(/__m256i/ _1, /ui4/ _2) \
251	((__m256i)__builtin_lasx_xvbitclri_h((v16u16)(_1), (_2)))
252
253	#define __lasx_xvbitclri_w(/__m256i/ _1, /ui5/ _2) \
254	((__m256i)__builtin_lasx_xvbitclri_w((v8u32)(_1), (_2)))
255
256	#define __lasx_xvbitclri_d(/__m256i/ _1, /ui6/ _2) \
257	((__m256i)__builtin_lasx_xvbitclri_d((v4u64)(_1), (_2)))
258
259	extern __inline
260	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
261	__lasx_xvbitset_b(__m256i _1, __m256i _2) {
262	return (__m256i)__builtin_lasx_xvbitset_b((v32u8)_1, (v32u8)_2);
263	}
264
265	extern __inline
266	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
267	__lasx_xvbitset_h(__m256i _1, __m256i _2) {
268	return (__m256i)__builtin_lasx_xvbitset_h((v16u16)_1, (v16u16)_2);
269	}
270
271	extern __inline
272	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
273	__lasx_xvbitset_w(__m256i _1, __m256i _2) {
274	return (__m256i)__builtin_lasx_xvbitset_w((v8u32)_1, (v8u32)_2);
275	}
276
277	extern __inline
278	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
279	__lasx_xvbitset_d(__m256i _1, __m256i _2) {
280	return (__m256i)__builtin_lasx_xvbitset_d((v4u64)_1, (v4u64)_2);
281	}
282
283	#define __lasx_xvbitseti_b(/__m256i/ _1, /ui3/ _2) \
284	((__m256i)__builtin_lasx_xvbitseti_b((v32u8)(_1), (_2)))
285
286	#define __lasx_xvbitseti_h(/__m256i/ _1, /ui4/ _2) \
287	((__m256i)__builtin_lasx_xvbitseti_h((v16u16)(_1), (_2)))
288
289	#define __lasx_xvbitseti_w(/__m256i/ _1, /ui5/ _2) \
290	((__m256i)__builtin_lasx_xvbitseti_w((v8u32)(_1), (_2)))
291
292	#define __lasx_xvbitseti_d(/__m256i/ _1, /ui6/ _2) \
293	((__m256i)__builtin_lasx_xvbitseti_d((v4u64)(_1), (_2)))
294
295	extern __inline
296	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
297	__lasx_xvbitrev_b(__m256i _1, __m256i _2) {
298	return (__m256i)__builtin_lasx_xvbitrev_b((v32u8)_1, (v32u8)_2);
299	}
300
301	extern __inline
302	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
303	__lasx_xvbitrev_h(__m256i _1, __m256i _2) {
304	return (__m256i)__builtin_lasx_xvbitrev_h((v16u16)_1, (v16u16)_2);
305	}
306
307	extern __inline
308	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
309	__lasx_xvbitrev_w(__m256i _1, __m256i _2) {
310	return (__m256i)__builtin_lasx_xvbitrev_w((v8u32)_1, (v8u32)_2);
311	}
312
313	extern __inline
314	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
315	__lasx_xvbitrev_d(__m256i _1, __m256i _2) {
316	return (__m256i)__builtin_lasx_xvbitrev_d((v4u64)_1, (v4u64)_2);
317	}
318
319	#define __lasx_xvbitrevi_b(/__m256i/ _1, /ui3/ _2) \
320	((__m256i)__builtin_lasx_xvbitrevi_b((v32u8)(_1), (_2)))
321
322	#define __lasx_xvbitrevi_h(/__m256i/ _1, /ui4/ _2) \
323	((__m256i)__builtin_lasx_xvbitrevi_h((v16u16)(_1), (_2)))
324
325	#define __lasx_xvbitrevi_w(/__m256i/ _1, /ui5/ _2) \
326	((__m256i)__builtin_lasx_xvbitrevi_w((v8u32)(_1), (_2)))
327
328	#define __lasx_xvbitrevi_d(/__m256i/ _1, /ui6/ _2) \
329	((__m256i)__builtin_lasx_xvbitrevi_d((v4u64)(_1), (_2)))
330
331	extern __inline
332	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
333	__lasx_xvadd_b(__m256i _1, __m256i _2) {
334	return (__m256i)__builtin_lasx_xvadd_b((v32i8)_1, (v32i8)_2);
335	}
336
337	extern __inline
338	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
339	__lasx_xvadd_h(__m256i _1, __m256i _2) {
340	return (__m256i)__builtin_lasx_xvadd_h((v16i16)_1, (v16i16)_2);
341	}
342
343	extern __inline
344	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
345	__lasx_xvadd_w(__m256i _1, __m256i _2) {
346	return (__m256i)__builtin_lasx_xvadd_w((v8i32)_1, (v8i32)_2);
347	}
348
349	extern __inline
350	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
351	__lasx_xvadd_d(__m256i _1, __m256i _2) {
352	return (__m256i)__builtin_lasx_xvadd_d((v4i64)_1, (v4i64)_2);
353	}
354
355	#define __lasx_xvaddi_bu(/__m256i/ _1, /ui5/ _2) \
356	((__m256i)__builtin_lasx_xvaddi_bu((v32i8)(_1), (_2)))
357
358	#define __lasx_xvaddi_hu(/__m256i/ _1, /ui5/ _2) \
359	((__m256i)__builtin_lasx_xvaddi_hu((v16i16)(_1), (_2)))
360
361	#define __lasx_xvaddi_wu(/__m256i/ _1, /ui5/ _2) \
362	((__m256i)__builtin_lasx_xvaddi_wu((v8i32)(_1), (_2)))
363
364	#define __lasx_xvaddi_du(/__m256i/ _1, /ui5/ _2) \
365	((__m256i)__builtin_lasx_xvaddi_du((v4i64)(_1), (_2)))
366
367	extern __inline
368	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
369	__lasx_xvsub_b(__m256i _1, __m256i _2) {
370	return (__m256i)__builtin_lasx_xvsub_b((v32i8)_1, (v32i8)_2);
371	}
372
373	extern __inline
374	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
375	__lasx_xvsub_h(__m256i _1, __m256i _2) {
376	return (__m256i)__builtin_lasx_xvsub_h((v16i16)_1, (v16i16)_2);
377	}
378
379	extern __inline
380	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
381	__lasx_xvsub_w(__m256i _1, __m256i _2) {
382	return (__m256i)__builtin_lasx_xvsub_w((v8i32)_1, (v8i32)_2);
383	}
384
385	extern __inline
386	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
387	__lasx_xvsub_d(__m256i _1, __m256i _2) {
388	return (__m256i)__builtin_lasx_xvsub_d((v4i64)_1, (v4i64)_2);
389	}
390
391	#define __lasx_xvsubi_bu(/__m256i/ _1, /ui5/ _2) \
392	((__m256i)__builtin_lasx_xvsubi_bu((v32i8)(_1), (_2)))
393
394	#define __lasx_xvsubi_hu(/__m256i/ _1, /ui5/ _2) \
395	((__m256i)__builtin_lasx_xvsubi_hu((v16i16)(_1), (_2)))
396
397	#define __lasx_xvsubi_wu(/__m256i/ _1, /ui5/ _2) \
398	((__m256i)__builtin_lasx_xvsubi_wu((v8i32)(_1), (_2)))
399
400	#define __lasx_xvsubi_du(/__m256i/ _1, /ui5/ _2) \
401	((__m256i)__builtin_lasx_xvsubi_du((v4i64)(_1), (_2)))
402
403	extern __inline
404	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
405	__lasx_xvmax_b(__m256i _1, __m256i _2) {
406	return (__m256i)__builtin_lasx_xvmax_b((v32i8)_1, (v32i8)_2);
407	}
408
409	extern __inline
410	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
411	__lasx_xvmax_h(__m256i _1, __m256i _2) {
412	return (__m256i)__builtin_lasx_xvmax_h((v16i16)_1, (v16i16)_2);
413	}
414
415	extern __inline
416	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
417	__lasx_xvmax_w(__m256i _1, __m256i _2) {
418	return (__m256i)__builtin_lasx_xvmax_w((v8i32)_1, (v8i32)_2);
419	}
420
421	extern __inline
422	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
423	__lasx_xvmax_d(__m256i _1, __m256i _2) {
424	return (__m256i)__builtin_lasx_xvmax_d((v4i64)_1, (v4i64)_2);
425	}
426
427	#define __lasx_xvmaxi_b(/__m256i/ _1, /si5/ _2) \
428	((__m256i)__builtin_lasx_xvmaxi_b((v32i8)(_1), (_2)))
429
430	#define __lasx_xvmaxi_h(/__m256i/ _1, /si5/ _2) \
431	((__m256i)__builtin_lasx_xvmaxi_h((v16i16)(_1), (_2)))
432
433	#define __lasx_xvmaxi_w(/__m256i/ _1, /si5/ _2) \
434	((__m256i)__builtin_lasx_xvmaxi_w((v8i32)(_1), (_2)))
435
436	#define __lasx_xvmaxi_d(/__m256i/ _1, /si5/ _2) \
437	((__m256i)__builtin_lasx_xvmaxi_d((v4i64)(_1), (_2)))
438
439	extern __inline
440	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
441	__lasx_xvmax_bu(__m256i _1, __m256i _2) {
442	return (__m256i)__builtin_lasx_xvmax_bu((v32u8)_1, (v32u8)_2);
443	}
444
445	extern __inline
446	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
447	__lasx_xvmax_hu(__m256i _1, __m256i _2) {
448	return (__m256i)__builtin_lasx_xvmax_hu((v16u16)_1, (v16u16)_2);
449	}
450
451	extern __inline
452	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
453	__lasx_xvmax_wu(__m256i _1, __m256i _2) {
454	return (__m256i)__builtin_lasx_xvmax_wu((v8u32)_1, (v8u32)_2);
455	}
456
457	extern __inline
458	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
459	__lasx_xvmax_du(__m256i _1, __m256i _2) {
460	return (__m256i)__builtin_lasx_xvmax_du((v4u64)_1, (v4u64)_2);
461	}
462
463	#define __lasx_xvmaxi_bu(/__m256i/ _1, /ui5/ _2) \
464	((__m256i)__builtin_lasx_xvmaxi_bu((v32u8)(_1), (_2)))
465
466	#define __lasx_xvmaxi_hu(/__m256i/ _1, /ui5/ _2) \
467	((__m256i)__builtin_lasx_xvmaxi_hu((v16u16)(_1), (_2)))
468
469	#define __lasx_xvmaxi_wu(/__m256i/ _1, /ui5/ _2) \
470	((__m256i)__builtin_lasx_xvmaxi_wu((v8u32)(_1), (_2)))
471
472	#define __lasx_xvmaxi_du(/__m256i/ _1, /ui5/ _2) \
473	((__m256i)__builtin_lasx_xvmaxi_du((v4u64)(_1), (_2)))
474
475	extern __inline
476	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
477	__lasx_xvmin_b(__m256i _1, __m256i _2) {
478	return (__m256i)__builtin_lasx_xvmin_b((v32i8)_1, (v32i8)_2);
479	}
480
481	extern __inline
482	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
483	__lasx_xvmin_h(__m256i _1, __m256i _2) {
484	return (__m256i)__builtin_lasx_xvmin_h((v16i16)_1, (v16i16)_2);
485	}
486
487	extern __inline
488	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
489	__lasx_xvmin_w(__m256i _1, __m256i _2) {
490	return (__m256i)__builtin_lasx_xvmin_w((v8i32)_1, (v8i32)_2);
491	}
492
493	extern __inline
494	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
495	__lasx_xvmin_d(__m256i _1, __m256i _2) {
496	return (__m256i)__builtin_lasx_xvmin_d((v4i64)_1, (v4i64)_2);
497	}
498
499	#define __lasx_xvmini_b(/__m256i/ _1, /si5/ _2) \
500	((__m256i)__builtin_lasx_xvmini_b((v32i8)(_1), (_2)))
501
502	#define __lasx_xvmini_h(/__m256i/ _1, /si5/ _2) \
503	((__m256i)__builtin_lasx_xvmini_h((v16i16)(_1), (_2)))
504
505	#define __lasx_xvmini_w(/__m256i/ _1, /si5/ _2) \
506	((__m256i)__builtin_lasx_xvmini_w((v8i32)(_1), (_2)))
507
508	#define __lasx_xvmini_d(/__m256i/ _1, /si5/ _2) \
509	((__m256i)__builtin_lasx_xvmini_d((v4i64)(_1), (_2)))
510
511	extern __inline
512	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
513	__lasx_xvmin_bu(__m256i _1, __m256i _2) {
514	return (__m256i)__builtin_lasx_xvmin_bu((v32u8)_1, (v32u8)_2);
515	}
516
517	extern __inline
518	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
519	__lasx_xvmin_hu(__m256i _1, __m256i _2) {
520	return (__m256i)__builtin_lasx_xvmin_hu((v16u16)_1, (v16u16)_2);
521	}
522
523	extern __inline
524	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
525	__lasx_xvmin_wu(__m256i _1, __m256i _2) {
526	return (__m256i)__builtin_lasx_xvmin_wu((v8u32)_1, (v8u32)_2);
527	}
528
529	extern __inline
530	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
531	__lasx_xvmin_du(__m256i _1, __m256i _2) {
532	return (__m256i)__builtin_lasx_xvmin_du((v4u64)_1, (v4u64)_2);
533	}
534
535	#define __lasx_xvmini_bu(/__m256i/ _1, /ui5/ _2) \
536	((__m256i)__builtin_lasx_xvmini_bu((v32u8)(_1), (_2)))
537
538	#define __lasx_xvmini_hu(/__m256i/ _1, /ui5/ _2) \
539	((__m256i)__builtin_lasx_xvmini_hu((v16u16)(_1), (_2)))
540
541	#define __lasx_xvmini_wu(/__m256i/ _1, /ui5/ _2) \
542	((__m256i)__builtin_lasx_xvmini_wu((v8u32)(_1), (_2)))
543
544	#define __lasx_xvmini_du(/__m256i/ _1, /ui5/ _2) \
545	((__m256i)__builtin_lasx_xvmini_du((v4u64)(_1), (_2)))
546
547	extern __inline
548	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
549	__lasx_xvseq_b(__m256i _1, __m256i _2) {
550	return (__m256i)__builtin_lasx_xvseq_b((v32i8)_1, (v32i8)_2);
551	}
552
553	extern __inline
554	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
555	__lasx_xvseq_h(__m256i _1, __m256i _2) {
556	return (__m256i)__builtin_lasx_xvseq_h((v16i16)_1, (v16i16)_2);
557	}
558
559	extern __inline
560	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
561	__lasx_xvseq_w(__m256i _1, __m256i _2) {
562	return (__m256i)__builtin_lasx_xvseq_w((v8i32)_1, (v8i32)_2);
563	}
564
565	extern __inline
566	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
567	__lasx_xvseq_d(__m256i _1, __m256i _2) {
568	return (__m256i)__builtin_lasx_xvseq_d((v4i64)_1, (v4i64)_2);
569	}
570
571	#define __lasx_xvseqi_b(/__m256i/ _1, /si5/ _2) \
572	((__m256i)__builtin_lasx_xvseqi_b((v32i8)(_1), (_2)))
573
574	#define __lasx_xvseqi_h(/__m256i/ _1, /si5/ _2) \
575	((__m256i)__builtin_lasx_xvseqi_h((v16i16)(_1), (_2)))
576
577	#define __lasx_xvseqi_w(/__m256i/ _1, /si5/ _2) \
578	((__m256i)__builtin_lasx_xvseqi_w((v8i32)(_1), (_2)))
579
580	#define __lasx_xvseqi_d(/__m256i/ _1, /si5/ _2) \
581	((__m256i)__builtin_lasx_xvseqi_d((v4i64)(_1), (_2)))
582
583	extern __inline
584	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
585	__lasx_xvslt_b(__m256i _1, __m256i _2) {
586	return (__m256i)__builtin_lasx_xvslt_b((v32i8)_1, (v32i8)_2);
587	}
588
589	extern __inline
590	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
591	__lasx_xvslt_h(__m256i _1, __m256i _2) {
592	return (__m256i)__builtin_lasx_xvslt_h((v16i16)_1, (v16i16)_2);
593	}
594
595	extern __inline
596	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
597	__lasx_xvslt_w(__m256i _1, __m256i _2) {
598	return (__m256i)__builtin_lasx_xvslt_w((v8i32)_1, (v8i32)_2);
599	}
600
601	extern __inline
602	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
603	__lasx_xvslt_d(__m256i _1, __m256i _2) {
604	return (__m256i)__builtin_lasx_xvslt_d((v4i64)_1, (v4i64)_2);
605	}
606
607	#define __lasx_xvslti_b(/__m256i/ _1, /si5/ _2) \
608	((__m256i)__builtin_lasx_xvslti_b((v32i8)(_1), (_2)))
609
610	#define __lasx_xvslti_h(/__m256i/ _1, /si5/ _2) \
611	((__m256i)__builtin_lasx_xvslti_h((v16i16)(_1), (_2)))
612
613	#define __lasx_xvslti_w(/__m256i/ _1, /si5/ _2) \
614	((__m256i)__builtin_lasx_xvslti_w((v8i32)(_1), (_2)))
615
616	#define __lasx_xvslti_d(/__m256i/ _1, /si5/ _2) \
617	((__m256i)__builtin_lasx_xvslti_d((v4i64)(_1), (_2)))
618
619	extern __inline
620	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
621	__lasx_xvslt_bu(__m256i _1, __m256i _2) {
622	return (__m256i)__builtin_lasx_xvslt_bu((v32u8)_1, (v32u8)_2);
623	}
624
625	extern __inline
626	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
627	__lasx_xvslt_hu(__m256i _1, __m256i _2) {
628	return (__m256i)__builtin_lasx_xvslt_hu((v16u16)_1, (v16u16)_2);
629	}
630
631	extern __inline
632	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
633	__lasx_xvslt_wu(__m256i _1, __m256i _2) {
634	return (__m256i)__builtin_lasx_xvslt_wu((v8u32)_1, (v8u32)_2);
635	}
636
637	extern __inline
638	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
639	__lasx_xvslt_du(__m256i _1, __m256i _2) {
640	return (__m256i)__builtin_lasx_xvslt_du((v4u64)_1, (v4u64)_2);
641	}
642
643	#define __lasx_xvslti_bu(/__m256i/ _1, /ui5/ _2) \
644	((__m256i)__builtin_lasx_xvslti_bu((v32u8)(_1), (_2)))
645
646	#define __lasx_xvslti_hu(/__m256i/ _1, /ui5/ _2) \
647	((__m256i)__builtin_lasx_xvslti_hu((v16u16)(_1), (_2)))
648
649	#define __lasx_xvslti_wu(/__m256i/ _1, /ui5/ _2) \
650	((__m256i)__builtin_lasx_xvslti_wu((v8u32)(_1), (_2)))
651
652	#define __lasx_xvslti_du(/__m256i/ _1, /ui5/ _2) \
653	((__m256i)__builtin_lasx_xvslti_du((v4u64)(_1), (_2)))
654
655	extern __inline
656	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
657	__lasx_xvsle_b(__m256i _1, __m256i _2) {
658	return (__m256i)__builtin_lasx_xvsle_b((v32i8)_1, (v32i8)_2);
659	}
660
661	extern __inline
662	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
663	__lasx_xvsle_h(__m256i _1, __m256i _2) {
664	return (__m256i)__builtin_lasx_xvsle_h((v16i16)_1, (v16i16)_2);
665	}
666
667	extern __inline
668	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
669	__lasx_xvsle_w(__m256i _1, __m256i _2) {
670	return (__m256i)__builtin_lasx_xvsle_w((v8i32)_1, (v8i32)_2);
671	}
672
673	extern __inline
674	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
675	__lasx_xvsle_d(__m256i _1, __m256i _2) {
676	return (__m256i)__builtin_lasx_xvsle_d((v4i64)_1, (v4i64)_2);
677	}
678
679	#define __lasx_xvslei_b(/__m256i/ _1, /si5/ _2) \
680	((__m256i)__builtin_lasx_xvslei_b((v32i8)(_1), (_2)))
681
682	#define __lasx_xvslei_h(/__m256i/ _1, /si5/ _2) \
683	((__m256i)__builtin_lasx_xvslei_h((v16i16)(_1), (_2)))
684
685	#define __lasx_xvslei_w(/__m256i/ _1, /si5/ _2) \
686	((__m256i)__builtin_lasx_xvslei_w((v8i32)(_1), (_2)))
687
688	#define __lasx_xvslei_d(/__m256i/ _1, /si5/ _2) \
689	((__m256i)__builtin_lasx_xvslei_d((v4i64)(_1), (_2)))
690
691	extern __inline
692	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
693	__lasx_xvsle_bu(__m256i _1, __m256i _2) {
694	return (__m256i)__builtin_lasx_xvsle_bu((v32u8)_1, (v32u8)_2);
695	}
696
697	extern __inline
698	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
699	__lasx_xvsle_hu(__m256i _1, __m256i _2) {
700	return (__m256i)__builtin_lasx_xvsle_hu((v16u16)_1, (v16u16)_2);
701	}
702
703	extern __inline
704	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
705	__lasx_xvsle_wu(__m256i _1, __m256i _2) {
706	return (__m256i)__builtin_lasx_xvsle_wu((v8u32)_1, (v8u32)_2);
707	}
708
709	extern __inline
710	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
711	__lasx_xvsle_du(__m256i _1, __m256i _2) {
712	return (__m256i)__builtin_lasx_xvsle_du((v4u64)_1, (v4u64)_2);
713	}
714
715	#define __lasx_xvslei_bu(/__m256i/ _1, /ui5/ _2) \
716	((__m256i)__builtin_lasx_xvslei_bu((v32u8)(_1), (_2)))
717
718	#define __lasx_xvslei_hu(/__m256i/ _1, /ui5/ _2) \
719	((__m256i)__builtin_lasx_xvslei_hu((v16u16)(_1), (_2)))
720
721	#define __lasx_xvslei_wu(/__m256i/ _1, /ui5/ _2) \
722	((__m256i)__builtin_lasx_xvslei_wu((v8u32)(_1), (_2)))
723
724	#define __lasx_xvslei_du(/__m256i/ _1, /ui5/ _2) \
725	((__m256i)__builtin_lasx_xvslei_du((v4u64)(_1), (_2)))
726
727	#define __lasx_xvsat_b(/__m256i/ _1, /ui3/ _2) \
728	((__m256i)__builtin_lasx_xvsat_b((v32i8)(_1), (_2)))
729
730	#define __lasx_xvsat_h(/__m256i/ _1, /ui4/ _2) \
731	((__m256i)__builtin_lasx_xvsat_h((v16i16)(_1), (_2)))
732
733	#define __lasx_xvsat_w(/__m256i/ _1, /ui5/ _2) \
734	((__m256i)__builtin_lasx_xvsat_w((v8i32)(_1), (_2)))
735
736	#define __lasx_xvsat_d(/__m256i/ _1, /ui6/ _2) \
737	((__m256i)__builtin_lasx_xvsat_d((v4i64)(_1), (_2)))
738
739	#define __lasx_xvsat_bu(/__m256i/ _1, /ui3/ _2) \
740	((__m256i)__builtin_lasx_xvsat_bu((v32u8)(_1), (_2)))
741
742	#define __lasx_xvsat_hu(/__m256i/ _1, /ui4/ _2) \
743	((__m256i)__builtin_lasx_xvsat_hu((v16u16)(_1), (_2)))
744
745	#define __lasx_xvsat_wu(/__m256i/ _1, /ui5/ _2) \
746	((__m256i)__builtin_lasx_xvsat_wu((v8u32)(_1), (_2)))
747
748	#define __lasx_xvsat_du(/__m256i/ _1, /ui6/ _2) \
749	((__m256i)__builtin_lasx_xvsat_du((v4u64)(_1), (_2)))
750
751	extern __inline
752	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
753	__lasx_xvadda_b(__m256i _1, __m256i _2) {
754	return (__m256i)__builtin_lasx_xvadda_b((v32i8)_1, (v32i8)_2);
755	}
756
757	extern __inline
758	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
759	__lasx_xvadda_h(__m256i _1, __m256i _2) {
760	return (__m256i)__builtin_lasx_xvadda_h((v16i16)_1, (v16i16)_2);
761	}
762
763	extern __inline
764	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
765	__lasx_xvadda_w(__m256i _1, __m256i _2) {
766	return (__m256i)__builtin_lasx_xvadda_w((v8i32)_1, (v8i32)_2);
767	}
768
769	extern __inline
770	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
771	__lasx_xvadda_d(__m256i _1, __m256i _2) {
772	return (__m256i)__builtin_lasx_xvadda_d((v4i64)_1, (v4i64)_2);
773	}
774
775	extern __inline
776	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
777	__lasx_xvsadd_b(__m256i _1, __m256i _2) {
778	return (__m256i)__builtin_lasx_xvsadd_b((v32i8)_1, (v32i8)_2);
779	}
780
781	extern __inline
782	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
783	__lasx_xvsadd_h(__m256i _1, __m256i _2) {
784	return (__m256i)__builtin_lasx_xvsadd_h((v16i16)_1, (v16i16)_2);
785	}
786
787	extern __inline
788	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
789	__lasx_xvsadd_w(__m256i _1, __m256i _2) {
790	return (__m256i)__builtin_lasx_xvsadd_w((v8i32)_1, (v8i32)_2);
791	}
792
793	extern __inline
794	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
795	__lasx_xvsadd_d(__m256i _1, __m256i _2) {
796	return (__m256i)__builtin_lasx_xvsadd_d((v4i64)_1, (v4i64)_2);
797	}
798
799	extern __inline
800	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
801	__lasx_xvsadd_bu(__m256i _1, __m256i _2) {
802	return (__m256i)__builtin_lasx_xvsadd_bu((v32u8)_1, (v32u8)_2);
803	}
804
805	extern __inline
806	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
807	__lasx_xvsadd_hu(__m256i _1, __m256i _2) {
808	return (__m256i)__builtin_lasx_xvsadd_hu((v16u16)_1, (v16u16)_2);
809	}
810
811	extern __inline
812	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
813	__lasx_xvsadd_wu(__m256i _1, __m256i _2) {
814	return (__m256i)__builtin_lasx_xvsadd_wu((v8u32)_1, (v8u32)_2);
815	}
816
817	extern __inline
818	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
819	__lasx_xvsadd_du(__m256i _1, __m256i _2) {
820	return (__m256i)__builtin_lasx_xvsadd_du((v4u64)_1, (v4u64)_2);
821	}
822
823	extern __inline
824	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
825	__lasx_xvavg_b(__m256i _1, __m256i _2) {
826	return (__m256i)__builtin_lasx_xvavg_b((v32i8)_1, (v32i8)_2);
827	}
828
829	extern __inline
830	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
831	__lasx_xvavg_h(__m256i _1, __m256i _2) {
832	return (__m256i)__builtin_lasx_xvavg_h((v16i16)_1, (v16i16)_2);
833	}
834
835	extern __inline
836	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
837	__lasx_xvavg_w(__m256i _1, __m256i _2) {
838	return (__m256i)__builtin_lasx_xvavg_w((v8i32)_1, (v8i32)_2);
839	}
840
841	extern __inline
842	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
843	__lasx_xvavg_d(__m256i _1, __m256i _2) {
844	return (__m256i)__builtin_lasx_xvavg_d((v4i64)_1, (v4i64)_2);
845	}
846
847	extern __inline
848	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
849	__lasx_xvavg_bu(__m256i _1, __m256i _2) {
850	return (__m256i)__builtin_lasx_xvavg_bu((v32u8)_1, (v32u8)_2);
851	}
852
853	extern __inline
854	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
855	__lasx_xvavg_hu(__m256i _1, __m256i _2) {
856	return (__m256i)__builtin_lasx_xvavg_hu((v16u16)_1, (v16u16)_2);
857	}
858
859	extern __inline
860	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
861	__lasx_xvavg_wu(__m256i _1, __m256i _2) {
862	return (__m256i)__builtin_lasx_xvavg_wu((v8u32)_1, (v8u32)_2);
863	}
864
865	extern __inline
866	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
867	__lasx_xvavg_du(__m256i _1, __m256i _2) {
868	return (__m256i)__builtin_lasx_xvavg_du((v4u64)_1, (v4u64)_2);
869	}
870
871	extern __inline
872	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
873	__lasx_xvavgr_b(__m256i _1, __m256i _2) {
874	return (__m256i)__builtin_lasx_xvavgr_b((v32i8)_1, (v32i8)_2);
875	}
876
877	extern __inline
878	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
879	__lasx_xvavgr_h(__m256i _1, __m256i _2) {
880	return (__m256i)__builtin_lasx_xvavgr_h((v16i16)_1, (v16i16)_2);
881	}
882
883	extern __inline
884	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
885	__lasx_xvavgr_w(__m256i _1, __m256i _2) {
886	return (__m256i)__builtin_lasx_xvavgr_w((v8i32)_1, (v8i32)_2);
887	}
888
889	extern __inline
890	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
891	__lasx_xvavgr_d(__m256i _1, __m256i _2) {
892	return (__m256i)__builtin_lasx_xvavgr_d((v4i64)_1, (v4i64)_2);
893	}
894
895	extern __inline
896	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
897	__lasx_xvavgr_bu(__m256i _1, __m256i _2) {
898	return (__m256i)__builtin_lasx_xvavgr_bu((v32u8)_1, (v32u8)_2);
899	}
900
901	extern __inline
902	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
903	__lasx_xvavgr_hu(__m256i _1, __m256i _2) {
904	return (__m256i)__builtin_lasx_xvavgr_hu((v16u16)_1, (v16u16)_2);
905	}
906
907	extern __inline
908	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
909	__lasx_xvavgr_wu(__m256i _1, __m256i _2) {
910	return (__m256i)__builtin_lasx_xvavgr_wu((v8u32)_1, (v8u32)_2);
911	}
912
913	extern __inline
914	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
915	__lasx_xvavgr_du(__m256i _1, __m256i _2) {
916	return (__m256i)__builtin_lasx_xvavgr_du((v4u64)_1, (v4u64)_2);
917	}
918
919	extern __inline
920	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
921	__lasx_xvssub_b(__m256i _1, __m256i _2) {
922	return (__m256i)__builtin_lasx_xvssub_b((v32i8)_1, (v32i8)_2);
923	}
924
925	extern __inline
926	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
927	__lasx_xvssub_h(__m256i _1, __m256i _2) {
928	return (__m256i)__builtin_lasx_xvssub_h((v16i16)_1, (v16i16)_2);
929	}
930
931	extern __inline
932	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
933	__lasx_xvssub_w(__m256i _1, __m256i _2) {
934	return (__m256i)__builtin_lasx_xvssub_w((v8i32)_1, (v8i32)_2);
935	}
936
937	extern __inline
938	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
939	__lasx_xvssub_d(__m256i _1, __m256i _2) {
940	return (__m256i)__builtin_lasx_xvssub_d((v4i64)_1, (v4i64)_2);
941	}
942
943	extern __inline
944	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
945	__lasx_xvssub_bu(__m256i _1, __m256i _2) {
946	return (__m256i)__builtin_lasx_xvssub_bu((v32u8)_1, (v32u8)_2);
947	}
948
949	extern __inline
950	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
951	__lasx_xvssub_hu(__m256i _1, __m256i _2) {
952	return (__m256i)__builtin_lasx_xvssub_hu((v16u16)_1, (v16u16)_2);
953	}
954
955	extern __inline
956	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
957	__lasx_xvssub_wu(__m256i _1, __m256i _2) {
958	return (__m256i)__builtin_lasx_xvssub_wu((v8u32)_1, (v8u32)_2);
959	}
960
961	extern __inline
962	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
963	__lasx_xvssub_du(__m256i _1, __m256i _2) {
964	return (__m256i)__builtin_lasx_xvssub_du((v4u64)_1, (v4u64)_2);
965	}
966
967	extern __inline
968	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
969	__lasx_xvabsd_b(__m256i _1, __m256i _2) {
970	return (__m256i)__builtin_lasx_xvabsd_b((v32i8)_1, (v32i8)_2);
971	}
972
973	extern __inline
974	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
975	__lasx_xvabsd_h(__m256i _1, __m256i _2) {
976	return (__m256i)__builtin_lasx_xvabsd_h((v16i16)_1, (v16i16)_2);
977	}
978
979	extern __inline
980	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
981	__lasx_xvabsd_w(__m256i _1, __m256i _2) {
982	return (__m256i)__builtin_lasx_xvabsd_w((v8i32)_1, (v8i32)_2);
983	}
984
985	extern __inline
986	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
987	__lasx_xvabsd_d(__m256i _1, __m256i _2) {
988	return (__m256i)__builtin_lasx_xvabsd_d((v4i64)_1, (v4i64)_2);
989	}
990
991	extern __inline
992	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
993	__lasx_xvabsd_bu(__m256i _1, __m256i _2) {
994	return (__m256i)__builtin_lasx_xvabsd_bu((v32u8)_1, (v32u8)_2);
995	}
996
997	extern __inline
998	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
999	__lasx_xvabsd_hu(__m256i _1, __m256i _2) {
1000	return (__m256i)__builtin_lasx_xvabsd_hu((v16u16)_1, (v16u16)_2);
1001	}
1002
1003	extern __inline
1004	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1005	__lasx_xvabsd_wu(__m256i _1, __m256i _2) {
1006	return (__m256i)__builtin_lasx_xvabsd_wu((v8u32)_1, (v8u32)_2);
1007	}
1008
1009	extern __inline
1010	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1011	__lasx_xvabsd_du(__m256i _1, __m256i _2) {
1012	return (__m256i)__builtin_lasx_xvabsd_du((v4u64)_1, (v4u64)_2);
1013	}
1014
1015	extern __inline
1016	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1017	__lasx_xvmul_b(__m256i _1, __m256i _2) {
1018	return (__m256i)__builtin_lasx_xvmul_b((v32i8)_1, (v32i8)_2);
1019	}
1020
1021	extern __inline
1022	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1023	__lasx_xvmul_h(__m256i _1, __m256i _2) {
1024	return (__m256i)__builtin_lasx_xvmul_h((v16i16)_1, (v16i16)_2);
1025	}
1026
1027	extern __inline
1028	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1029	__lasx_xvmul_w(__m256i _1, __m256i _2) {
1030	return (__m256i)__builtin_lasx_xvmul_w((v8i32)_1, (v8i32)_2);
1031	}
1032
1033	extern __inline
1034	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1035	__lasx_xvmul_d(__m256i _1, __m256i _2) {
1036	return (__m256i)__builtin_lasx_xvmul_d((v4i64)_1, (v4i64)_2);
1037	}
1038
1039	extern __inline
1040	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1041	__lasx_xvmadd_b(__m256i _1, __m256i _2, __m256i _3) {
1042	return (__m256i)__builtin_lasx_xvmadd_b((v32i8)_1, (v32i8)_2, (v32i8)_3);
1043	}
1044
1045	extern __inline
1046	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1047	__lasx_xvmadd_h(__m256i _1, __m256i _2, __m256i _3) {
1048	return (__m256i)__builtin_lasx_xvmadd_h((v16i16)_1, (v16i16)_2, (v16i16)_3);
1049	}
1050
1051	extern __inline
1052	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1053	__lasx_xvmadd_w(__m256i _1, __m256i _2, __m256i _3) {
1054	return (__m256i)__builtin_lasx_xvmadd_w((v8i32)_1, (v8i32)_2, (v8i32)_3);
1055	}
1056
1057	extern __inline
1058	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1059	__lasx_xvmadd_d(__m256i _1, __m256i _2, __m256i _3) {
1060	return (__m256i)__builtin_lasx_xvmadd_d((v4i64)_1, (v4i64)_2, (v4i64)_3);
1061	}
1062
1063	extern __inline
1064	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1065	__lasx_xvmsub_b(__m256i _1, __m256i _2, __m256i _3) {
1066	return (__m256i)__builtin_lasx_xvmsub_b((v32i8)_1, (v32i8)_2, (v32i8)_3);
1067	}
1068
1069	extern __inline
1070	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1071	__lasx_xvmsub_h(__m256i _1, __m256i _2, __m256i _3) {
1072	return (__m256i)__builtin_lasx_xvmsub_h((v16i16)_1, (v16i16)_2, (v16i16)_3);
1073	}
1074
1075	extern __inline
1076	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1077	__lasx_xvmsub_w(__m256i _1, __m256i _2, __m256i _3) {
1078	return (__m256i)__builtin_lasx_xvmsub_w((v8i32)_1, (v8i32)_2, (v8i32)_3);
1079	}
1080
1081	extern __inline
1082	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1083	__lasx_xvmsub_d(__m256i _1, __m256i _2, __m256i _3) {
1084	return (__m256i)__builtin_lasx_xvmsub_d((v4i64)_1, (v4i64)_2, (v4i64)_3);
1085	}
1086
1087	extern __inline
1088	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1089	__lasx_xvdiv_b(__m256i _1, __m256i _2) {
1090	return (__m256i)__builtin_lasx_xvdiv_b((v32i8)_1, (v32i8)_2);
1091	}
1092
1093	extern __inline
1094	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1095	__lasx_xvdiv_h(__m256i _1, __m256i _2) {
1096	return (__m256i)__builtin_lasx_xvdiv_h((v16i16)_1, (v16i16)_2);
1097	}
1098
1099	extern __inline
1100	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1101	__lasx_xvdiv_w(__m256i _1, __m256i _2) {
1102	return (__m256i)__builtin_lasx_xvdiv_w((v8i32)_1, (v8i32)_2);
1103	}
1104
1105	extern __inline
1106	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1107	__lasx_xvdiv_d(__m256i _1, __m256i _2) {
1108	return (__m256i)__builtin_lasx_xvdiv_d((v4i64)_1, (v4i64)_2);
1109	}
1110
1111	extern __inline
1112	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1113	__lasx_xvdiv_bu(__m256i _1, __m256i _2) {
1114	return (__m256i)__builtin_lasx_xvdiv_bu((v32u8)_1, (v32u8)_2);
1115	}
1116
1117	extern __inline
1118	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1119	__lasx_xvdiv_hu(__m256i _1, __m256i _2) {
1120	return (__m256i)__builtin_lasx_xvdiv_hu((v16u16)_1, (v16u16)_2);
1121	}
1122
1123	extern __inline
1124	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1125	__lasx_xvdiv_wu(__m256i _1, __m256i _2) {
1126	return (__m256i)__builtin_lasx_xvdiv_wu((v8u32)_1, (v8u32)_2);
1127	}
1128
1129	extern __inline
1130	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1131	__lasx_xvdiv_du(__m256i _1, __m256i _2) {
1132	return (__m256i)__builtin_lasx_xvdiv_du((v4u64)_1, (v4u64)_2);
1133	}
1134
1135	extern __inline
1136	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1137	__lasx_xvhaddw_h_b(__m256i _1, __m256i _2) {
1138	return (__m256i)__builtin_lasx_xvhaddw_h_b((v32i8)_1, (v32i8)_2);
1139	}
1140
1141	extern __inline
1142	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1143	__lasx_xvhaddw_w_h(__m256i _1, __m256i _2) {
1144	return (__m256i)__builtin_lasx_xvhaddw_w_h((v16i16)_1, (v16i16)_2);
1145	}
1146
1147	extern __inline
1148	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1149	__lasx_xvhaddw_d_w(__m256i _1, __m256i _2) {
1150	return (__m256i)__builtin_lasx_xvhaddw_d_w((v8i32)_1, (v8i32)_2);
1151	}
1152
1153	extern __inline
1154	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1155	__lasx_xvhaddw_hu_bu(__m256i _1, __m256i _2) {
1156	return (__m256i)__builtin_lasx_xvhaddw_hu_bu((v32u8)_1, (v32u8)_2);
1157	}
1158
1159	extern __inline
1160	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1161	__lasx_xvhaddw_wu_hu(__m256i _1, __m256i _2) {
1162	return (__m256i)__builtin_lasx_xvhaddw_wu_hu((v16u16)_1, (v16u16)_2);
1163	}
1164
1165	extern __inline
1166	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1167	__lasx_xvhaddw_du_wu(__m256i _1, __m256i _2) {
1168	return (__m256i)__builtin_lasx_xvhaddw_du_wu((v8u32)_1, (v8u32)_2);
1169	}
1170
1171	extern __inline
1172	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1173	__lasx_xvhsubw_h_b(__m256i _1, __m256i _2) {
1174	return (__m256i)__builtin_lasx_xvhsubw_h_b((v32i8)_1, (v32i8)_2);
1175	}
1176
1177	extern __inline
1178	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1179	__lasx_xvhsubw_w_h(__m256i _1, __m256i _2) {
1180	return (__m256i)__builtin_lasx_xvhsubw_w_h((v16i16)_1, (v16i16)_2);
1181	}
1182
1183	extern __inline
1184	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1185	__lasx_xvhsubw_d_w(__m256i _1, __m256i _2) {
1186	return (__m256i)__builtin_lasx_xvhsubw_d_w((v8i32)_1, (v8i32)_2);
1187	}
1188
1189	extern __inline
1190	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1191	__lasx_xvhsubw_hu_bu(__m256i _1, __m256i _2) {
1192	return (__m256i)__builtin_lasx_xvhsubw_hu_bu((v32u8)_1, (v32u8)_2);
1193	}
1194
1195	extern __inline
1196	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1197	__lasx_xvhsubw_wu_hu(__m256i _1, __m256i _2) {
1198	return (__m256i)__builtin_lasx_xvhsubw_wu_hu((v16u16)_1, (v16u16)_2);
1199	}
1200
1201	extern __inline
1202	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1203	__lasx_xvhsubw_du_wu(__m256i _1, __m256i _2) {
1204	return (__m256i)__builtin_lasx_xvhsubw_du_wu((v8u32)_1, (v8u32)_2);
1205	}
1206
1207	extern __inline
1208	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1209	__lasx_xvmod_b(__m256i _1, __m256i _2) {
1210	return (__m256i)__builtin_lasx_xvmod_b((v32i8)_1, (v32i8)_2);
1211	}
1212
1213	extern __inline
1214	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1215	__lasx_xvmod_h(__m256i _1, __m256i _2) {
1216	return (__m256i)__builtin_lasx_xvmod_h((v16i16)_1, (v16i16)_2);
1217	}
1218
1219	extern __inline
1220	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1221	__lasx_xvmod_w(__m256i _1, __m256i _2) {
1222	return (__m256i)__builtin_lasx_xvmod_w((v8i32)_1, (v8i32)_2);
1223	}
1224
1225	extern __inline
1226	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1227	__lasx_xvmod_d(__m256i _1, __m256i _2) {
1228	return (__m256i)__builtin_lasx_xvmod_d((v4i64)_1, (v4i64)_2);
1229	}
1230
1231	extern __inline
1232	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1233	__lasx_xvmod_bu(__m256i _1, __m256i _2) {
1234	return (__m256i)__builtin_lasx_xvmod_bu((v32u8)_1, (v32u8)_2);
1235	}
1236
1237	extern __inline
1238	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1239	__lasx_xvmod_hu(__m256i _1, __m256i _2) {
1240	return (__m256i)__builtin_lasx_xvmod_hu((v16u16)_1, (v16u16)_2);
1241	}
1242
1243	extern __inline
1244	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1245	__lasx_xvmod_wu(__m256i _1, __m256i _2) {
1246	return (__m256i)__builtin_lasx_xvmod_wu((v8u32)_1, (v8u32)_2);
1247	}
1248
1249	extern __inline
1250	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1251	__lasx_xvmod_du(__m256i _1, __m256i _2) {
1252	return (__m256i)__builtin_lasx_xvmod_du((v4u64)_1, (v4u64)_2);
1253	}
1254
1255	#define __lasx_xvrepl128vei_b(/__m256i/ _1, /ui4/ _2) \
1256	((__m256i)__builtin_lasx_xvrepl128vei_b((v32i8)(_1), (_2)))
1257
1258	#define __lasx_xvrepl128vei_h(/__m256i/ _1, /ui3/ _2) \
1259	((__m256i)__builtin_lasx_xvrepl128vei_h((v16i16)(_1), (_2)))
1260
1261	#define __lasx_xvrepl128vei_w(/__m256i/ _1, /ui2/ _2) \
1262	((__m256i)__builtin_lasx_xvrepl128vei_w((v8i32)(_1), (_2)))
1263
1264	#define __lasx_xvrepl128vei_d(/__m256i/ _1, /ui1/ _2) \
1265	((__m256i)__builtin_lasx_xvrepl128vei_d((v4i64)(_1), (_2)))
1266
1267	extern __inline
1268	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1269	__lasx_xvpickev_b(__m256i _1, __m256i _2) {
1270	return (__m256i)__builtin_lasx_xvpickev_b((v32i8)_1, (v32i8)_2);
1271	}
1272
1273	extern __inline
1274	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1275	__lasx_xvpickev_h(__m256i _1, __m256i _2) {
1276	return (__m256i)__builtin_lasx_xvpickev_h((v16i16)_1, (v16i16)_2);
1277	}
1278
1279	extern __inline
1280	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1281	__lasx_xvpickev_w(__m256i _1, __m256i _2) {
1282	return (__m256i)__builtin_lasx_xvpickev_w((v8i32)_1, (v8i32)_2);
1283	}
1284
1285	extern __inline
1286	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1287	__lasx_xvpickev_d(__m256i _1, __m256i _2) {
1288	return (__m256i)__builtin_lasx_xvpickev_d((v4i64)_1, (v4i64)_2);
1289	}
1290
1291	extern __inline
1292	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1293	__lasx_xvpickod_b(__m256i _1, __m256i _2) {
1294	return (__m256i)__builtin_lasx_xvpickod_b((v32i8)_1, (v32i8)_2);
1295	}
1296
1297	extern __inline
1298	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1299	__lasx_xvpickod_h(__m256i _1, __m256i _2) {
1300	return (__m256i)__builtin_lasx_xvpickod_h((v16i16)_1, (v16i16)_2);
1301	}
1302
1303	extern __inline
1304	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1305	__lasx_xvpickod_w(__m256i _1, __m256i _2) {
1306	return (__m256i)__builtin_lasx_xvpickod_w((v8i32)_1, (v8i32)_2);
1307	}
1308
1309	extern __inline
1310	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1311	__lasx_xvpickod_d(__m256i _1, __m256i _2) {
1312	return (__m256i)__builtin_lasx_xvpickod_d((v4i64)_1, (v4i64)_2);
1313	}
1314
1315	extern __inline
1316	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1317	__lasx_xvilvh_b(__m256i _1, __m256i _2) {
1318	return (__m256i)__builtin_lasx_xvilvh_b((v32i8)_1, (v32i8)_2);
1319	}
1320
1321	extern __inline
1322	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1323	__lasx_xvilvh_h(__m256i _1, __m256i _2) {
1324	return (__m256i)__builtin_lasx_xvilvh_h((v16i16)_1, (v16i16)_2);
1325	}
1326
1327	extern __inline
1328	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1329	__lasx_xvilvh_w(__m256i _1, __m256i _2) {
1330	return (__m256i)__builtin_lasx_xvilvh_w((v8i32)_1, (v8i32)_2);
1331	}
1332
1333	extern __inline
1334	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1335	__lasx_xvilvh_d(__m256i _1, __m256i _2) {
1336	return (__m256i)__builtin_lasx_xvilvh_d((v4i64)_1, (v4i64)_2);
1337	}
1338
1339	extern __inline
1340	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1341	__lasx_xvilvl_b(__m256i _1, __m256i _2) {
1342	return (__m256i)__builtin_lasx_xvilvl_b((v32i8)_1, (v32i8)_2);
1343	}
1344
1345	extern __inline
1346	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1347	__lasx_xvilvl_h(__m256i _1, __m256i _2) {
1348	return (__m256i)__builtin_lasx_xvilvl_h((v16i16)_1, (v16i16)_2);
1349	}
1350
1351	extern __inline
1352	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1353	__lasx_xvilvl_w(__m256i _1, __m256i _2) {
1354	return (__m256i)__builtin_lasx_xvilvl_w((v8i32)_1, (v8i32)_2);
1355	}
1356
1357	extern __inline
1358	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1359	__lasx_xvilvl_d(__m256i _1, __m256i _2) {
1360	return (__m256i)__builtin_lasx_xvilvl_d((v4i64)_1, (v4i64)_2);
1361	}
1362
1363	extern __inline
1364	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1365	__lasx_xvpackev_b(__m256i _1, __m256i _2) {
1366	return (__m256i)__builtin_lasx_xvpackev_b((v32i8)_1, (v32i8)_2);
1367	}
1368
1369	extern __inline
1370	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1371	__lasx_xvpackev_h(__m256i _1, __m256i _2) {
1372	return (__m256i)__builtin_lasx_xvpackev_h((v16i16)_1, (v16i16)_2);
1373	}
1374
1375	extern __inline
1376	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1377	__lasx_xvpackev_w(__m256i _1, __m256i _2) {
1378	return (__m256i)__builtin_lasx_xvpackev_w((v8i32)_1, (v8i32)_2);
1379	}
1380
1381	extern __inline
1382	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1383	__lasx_xvpackev_d(__m256i _1, __m256i _2) {
1384	return (__m256i)__builtin_lasx_xvpackev_d((v4i64)_1, (v4i64)_2);
1385	}
1386
1387	extern __inline
1388	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1389	__lasx_xvpackod_b(__m256i _1, __m256i _2) {
1390	return (__m256i)__builtin_lasx_xvpackod_b((v32i8)_1, (v32i8)_2);
1391	}
1392
1393	extern __inline
1394	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1395	__lasx_xvpackod_h(__m256i _1, __m256i _2) {
1396	return (__m256i)__builtin_lasx_xvpackod_h((v16i16)_1, (v16i16)_2);
1397	}
1398
1399	extern __inline
1400	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1401	__lasx_xvpackod_w(__m256i _1, __m256i _2) {
1402	return (__m256i)__builtin_lasx_xvpackod_w((v8i32)_1, (v8i32)_2);
1403	}
1404
1405	extern __inline
1406	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1407	__lasx_xvpackod_d(__m256i _1, __m256i _2) {
1408	return (__m256i)__builtin_lasx_xvpackod_d((v4i64)_1, (v4i64)_2);
1409	}
1410
1411	extern __inline
1412	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1413	__lasx_xvshuf_b(__m256i _1, __m256i _2, __m256i _3) {
1414	return (__m256i)__builtin_lasx_xvshuf_b((v32i8)_1, (v32i8)_2, (v32i8)_3);
1415	}
1416
1417	extern __inline
1418	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1419	__lasx_xvshuf_h(__m256i _1, __m256i _2, __m256i _3) {
1420	return (__m256i)__builtin_lasx_xvshuf_h((v16i16)_1, (v16i16)_2, (v16i16)_3);
1421	}
1422
1423	extern __inline
1424	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1425	__lasx_xvshuf_w(__m256i _1, __m256i _2, __m256i _3) {
1426	return (__m256i)__builtin_lasx_xvshuf_w((v8i32)_1, (v8i32)_2, (v8i32)_3);
1427	}
1428
1429	extern __inline
1430	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1431	__lasx_xvshuf_d(__m256i _1, __m256i _2, __m256i _3) {
1432	return (__m256i)__builtin_lasx_xvshuf_d((v4i64)_1, (v4i64)_2, (v4i64)_3);
1433	}
1434
1435	extern __inline
1436	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1437	__lasx_xvand_v(__m256i _1, __m256i _2) {
1438	return (__m256i)__builtin_lasx_xvand_v((v32u8)_1, (v32u8)_2);
1439	}
1440
1441	#define __lasx_xvandi_b(/__m256i/ _1, /ui8/ _2) \
1442	((__m256i)__builtin_lasx_xvandi_b((v32u8)(_1), (_2)))
1443
1444	extern __inline
1445	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1446	__lasx_xvor_v(__m256i _1, __m256i _2) {
1447	return (__m256i)__builtin_lasx_xvor_v((v32u8)_1, (v32u8)_2);
1448	}
1449
1450	#define __lasx_xvori_b(/__m256i/ _1, /ui8/ _2) \
1451	((__m256i)__builtin_lasx_xvori_b((v32u8)(_1), (_2)))
1452
1453	extern __inline
1454	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1455	__lasx_xvnor_v(__m256i _1, __m256i _2) {
1456	return (__m256i)__builtin_lasx_xvnor_v((v32u8)_1, (v32u8)_2);
1457	}
1458
1459	#define __lasx_xvnori_b(/__m256i/ _1, /ui8/ _2) \
1460	((__m256i)__builtin_lasx_xvnori_b((v32u8)(_1), (_2)))
1461
1462	extern __inline
1463	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1464	__lasx_xvxor_v(__m256i _1, __m256i _2) {
1465	return (__m256i)__builtin_lasx_xvxor_v((v32u8)_1, (v32u8)_2);
1466	}
1467
1468	#define __lasx_xvxori_b(/__m256i/ _1, /ui8/ _2) \
1469	((__m256i)__builtin_lasx_xvxori_b((v32u8)(_1), (_2)))
1470
1471	extern __inline
1472	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1473	__lasx_xvbitsel_v(__m256i _1, __m256i _2, __m256i _3) {
1474	return (__m256i)__builtin_lasx_xvbitsel_v((v32u8)_1, (v32u8)_2, (v32u8)_3);
1475	}
1476
1477	#define __lasx_xvbitseli_b(/__m256i/ _1, /__m256i/ _2, /ui8/ _3) \
1478	((__m256i)__builtin_lasx_xvbitseli_b((v32u8)(_1), (v32u8)(_2), (_3)))
1479
1480	#define __lasx_xvshuf4i_b(/__m256i/ _1, /ui8/ _2) \
1481	((__m256i)__builtin_lasx_xvshuf4i_b((v32i8)(_1), (_2)))
1482
1483	#define __lasx_xvshuf4i_h(/__m256i/ _1, /ui8/ _2) \
1484	((__m256i)__builtin_lasx_xvshuf4i_h((v16i16)(_1), (_2)))
1485
1486	#define __lasx_xvshuf4i_w(/__m256i/ _1, /ui8/ _2) \
1487	((__m256i)__builtin_lasx_xvshuf4i_w((v8i32)(_1), (_2)))
1488
1489	extern __inline
1490	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1491	__lasx_xvreplgr2vr_b(int _1) {
1492	return (__m256i)__builtin_lasx_xvreplgr2vr_b((int)_1);
1493	}
1494
1495	extern __inline
1496	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1497	__lasx_xvreplgr2vr_h(int _1) {
1498	return (__m256i)__builtin_lasx_xvreplgr2vr_h((int)_1);
1499	}
1500
1501	extern __inline
1502	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1503	__lasx_xvreplgr2vr_w(int _1) {
1504	return (__m256i)__builtin_lasx_xvreplgr2vr_w((int)_1);
1505	}
1506
1507	extern __inline
1508	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1509	__lasx_xvreplgr2vr_d(long int _1) {
1510	return (__m256i)__builtin_lasx_xvreplgr2vr_d((long int)_1);
1511	}
1512
1513	extern __inline
1514	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1515	__lasx_xvpcnt_b(__m256i _1) {
1516	return (__m256i)__builtin_lasx_xvpcnt_b((v32i8)_1);
1517	}
1518
1519	extern __inline
1520	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1521	__lasx_xvpcnt_h(__m256i _1) {
1522	return (__m256i)__builtin_lasx_xvpcnt_h((v16i16)_1);
1523	}
1524
1525	extern __inline
1526	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1527	__lasx_xvpcnt_w(__m256i _1) {
1528	return (__m256i)__builtin_lasx_xvpcnt_w((v8i32)_1);
1529	}
1530
1531	extern __inline
1532	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1533	__lasx_xvpcnt_d(__m256i _1) {
1534	return (__m256i)__builtin_lasx_xvpcnt_d((v4i64)_1);
1535	}
1536
1537	extern __inline
1538	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1539	__lasx_xvclo_b(__m256i _1) {
1540	return (__m256i)__builtin_lasx_xvclo_b((v32i8)_1);
1541	}
1542
1543	extern __inline
1544	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1545	__lasx_xvclo_h(__m256i _1) {
1546	return (__m256i)__builtin_lasx_xvclo_h((v16i16)_1);
1547	}
1548
1549	extern __inline
1550	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1551	__lasx_xvclo_w(__m256i _1) {
1552	return (__m256i)__builtin_lasx_xvclo_w((v8i32)_1);
1553	}
1554
1555	extern __inline
1556	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1557	__lasx_xvclo_d(__m256i _1) {
1558	return (__m256i)__builtin_lasx_xvclo_d((v4i64)_1);
1559	}
1560
1561	extern __inline
1562	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1563	__lasx_xvclz_b(__m256i _1) {
1564	return (__m256i)__builtin_lasx_xvclz_b((v32i8)_1);
1565	}
1566
1567	extern __inline
1568	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1569	__lasx_xvclz_h(__m256i _1) {
1570	return (__m256i)__builtin_lasx_xvclz_h((v16i16)_1);
1571	}
1572
1573	extern __inline
1574	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1575	__lasx_xvclz_w(__m256i _1) {
1576	return (__m256i)__builtin_lasx_xvclz_w((v8i32)_1);
1577	}
1578
1579	extern __inline
1580	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1581	__lasx_xvclz_d(__m256i _1) {
1582	return (__m256i)__builtin_lasx_xvclz_d((v4i64)_1);
1583	}
1584
1585	extern __inline
1586	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1587	__lasx_xvfadd_s(__m256 _1, __m256 _2) {
1588	return (__m256)__builtin_lasx_xvfadd_s((v8f32)_1, (v8f32)_2);
1589	}
1590
1591	extern __inline
1592	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1593	__lasx_xvfadd_d(__m256d _1, __m256d _2) {
1594	return (__m256d)__builtin_lasx_xvfadd_d((v4f64)_1, (v4f64)_2);
1595	}
1596
1597	extern __inline
1598	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1599	__lasx_xvfsub_s(__m256 _1, __m256 _2) {
1600	return (__m256)__builtin_lasx_xvfsub_s((v8f32)_1, (v8f32)_2);
1601	}
1602
1603	extern __inline
1604	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1605	__lasx_xvfsub_d(__m256d _1, __m256d _2) {
1606	return (__m256d)__builtin_lasx_xvfsub_d((v4f64)_1, (v4f64)_2);
1607	}
1608
1609	extern __inline
1610	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1611	__lasx_xvfmul_s(__m256 _1, __m256 _2) {
1612	return (__m256)__builtin_lasx_xvfmul_s((v8f32)_1, (v8f32)_2);
1613	}
1614
1615	extern __inline
1616	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1617	__lasx_xvfmul_d(__m256d _1, __m256d _2) {
1618	return (__m256d)__builtin_lasx_xvfmul_d((v4f64)_1, (v4f64)_2);
1619	}
1620
1621	extern __inline
1622	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1623	__lasx_xvfdiv_s(__m256 _1, __m256 _2) {
1624	return (__m256)__builtin_lasx_xvfdiv_s((v8f32)_1, (v8f32)_2);
1625	}
1626
1627	extern __inline
1628	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1629	__lasx_xvfdiv_d(__m256d _1, __m256d _2) {
1630	return (__m256d)__builtin_lasx_xvfdiv_d((v4f64)_1, (v4f64)_2);
1631	}
1632
1633	extern __inline
1634	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1635	__lasx_xvfcvt_h_s(__m256 _1, __m256 _2) {
1636	return (__m256i)__builtin_lasx_xvfcvt_h_s((v8f32)_1, (v8f32)_2);
1637	}
1638
1639	extern __inline
1640	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1641	__lasx_xvfcvt_s_d(__m256d _1, __m256d _2) {
1642	return (__m256)__builtin_lasx_xvfcvt_s_d((v4f64)_1, (v4f64)_2);
1643	}
1644
1645	extern __inline
1646	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1647	__lasx_xvfmin_s(__m256 _1, __m256 _2) {
1648	return (__m256)__builtin_lasx_xvfmin_s((v8f32)_1, (v8f32)_2);
1649	}
1650
1651	extern __inline
1652	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1653	__lasx_xvfmin_d(__m256d _1, __m256d _2) {
1654	return (__m256d)__builtin_lasx_xvfmin_d((v4f64)_1, (v4f64)_2);
1655	}
1656
1657	extern __inline
1658	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1659	__lasx_xvfmina_s(__m256 _1, __m256 _2) {
1660	return (__m256)__builtin_lasx_xvfmina_s((v8f32)_1, (v8f32)_2);
1661	}
1662
1663	extern __inline
1664	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1665	__lasx_xvfmina_d(__m256d _1, __m256d _2) {
1666	return (__m256d)__builtin_lasx_xvfmina_d((v4f64)_1, (v4f64)_2);
1667	}
1668
1669	extern __inline
1670	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1671	__lasx_xvfmax_s(__m256 _1, __m256 _2) {
1672	return (__m256)__builtin_lasx_xvfmax_s((v8f32)_1, (v8f32)_2);
1673	}
1674
1675	extern __inline
1676	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1677	__lasx_xvfmax_d(__m256d _1, __m256d _2) {
1678	return (__m256d)__builtin_lasx_xvfmax_d((v4f64)_1, (v4f64)_2);
1679	}
1680
1681	extern __inline
1682	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1683	__lasx_xvfmaxa_s(__m256 _1, __m256 _2) {
1684	return (__m256)__builtin_lasx_xvfmaxa_s((v8f32)_1, (v8f32)_2);
1685	}
1686
1687	extern __inline
1688	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1689	__lasx_xvfmaxa_d(__m256d _1, __m256d _2) {
1690	return (__m256d)__builtin_lasx_xvfmaxa_d((v4f64)_1, (v4f64)_2);
1691	}
1692
1693	extern __inline
1694	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1695	__lasx_xvfclass_s(__m256 _1) {
1696	return (__m256i)__builtin_lasx_xvfclass_s((v8f32)_1);
1697	}
1698
1699	extern __inline
1700	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1701	__lasx_xvfclass_d(__m256d _1) {
1702	return (__m256i)__builtin_lasx_xvfclass_d((v4f64)_1);
1703	}
1704
1705	extern __inline
1706	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1707	__lasx_xvfsqrt_s(__m256 _1) {
1708	return (__m256)__builtin_lasx_xvfsqrt_s((v8f32)_1);
1709	}
1710
1711	extern __inline
1712	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1713	__lasx_xvfsqrt_d(__m256d _1) {
1714	return (__m256d)__builtin_lasx_xvfsqrt_d((v4f64)_1);
1715	}
1716
1717	extern __inline
1718	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1719	__lasx_xvfrecip_s(__m256 _1) {
1720	return (__m256)__builtin_lasx_xvfrecip_s((v8f32)_1);
1721	}
1722
1723	extern __inline
1724	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1725	__lasx_xvfrecip_d(__m256d _1) {
1726	return (__m256d)__builtin_lasx_xvfrecip_d((v4f64)_1);
1727	}
1728
1729	extern __inline
1730	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1731	__lasx_xvfrecipe_s(__m256 _1) {
1732	return (__m256)__builtin_lasx_xvfrecipe_s((v8f32)_1);
1733	}
1734
1735	extern __inline
1736	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1737	__lasx_xvfrecipe_d(__m256d _1) {
1738	return (__m256d)__builtin_lasx_xvfrecipe_d((v4f64)_1);
1739	}
1740
1741	extern __inline
1742	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1743	__lasx_xvfrint_s(__m256 _1) {
1744	return (__m256)__builtin_lasx_xvfrint_s((v8f32)_1);
1745	}
1746
1747	extern __inline
1748	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1749	__lasx_xvfrint_d(__m256d _1) {
1750	return (__m256d)__builtin_lasx_xvfrint_d((v4f64)_1);
1751	}
1752
1753	extern __inline
1754	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1755	__lasx_xvfrsqrt_s(__m256 _1) {
1756	return (__m256)__builtin_lasx_xvfrsqrt_s((v8f32)_1);
1757	}
1758
1759	extern __inline
1760	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1761	__lasx_xvfrsqrt_d(__m256d _1) {
1762	return (__m256d)__builtin_lasx_xvfrsqrt_d((v4f64)_1);
1763	}
1764
1765	extern __inline
1766	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1767	__lasx_xvfrsqrte_s(__m256 _1) {
1768	return (__m256)__builtin_lasx_xvfrsqrte_s((v8f32)_1);
1769	}
1770
1771	extern __inline
1772	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1773	__lasx_xvfrsqrte_d(__m256d _1) {
1774	return (__m256d)__builtin_lasx_xvfrsqrte_d((v4f64)_1);
1775	}
1776
1777	extern __inline
1778	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1779	__lasx_xvflogb_s(__m256 _1) {
1780	return (__m256)__builtin_lasx_xvflogb_s((v8f32)_1);
1781	}
1782
1783	extern __inline
1784	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1785	__lasx_xvflogb_d(__m256d _1) {
1786	return (__m256d)__builtin_lasx_xvflogb_d((v4f64)_1);
1787	}
1788
1789	extern __inline
1790	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1791	__lasx_xvfcvth_s_h(__m256i _1) {
1792	return (__m256)__builtin_lasx_xvfcvth_s_h((v16i16)_1);
1793	}
1794
1795	extern __inline
1796	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1797	__lasx_xvfcvth_d_s(__m256 _1) {
1798	return (__m256d)__builtin_lasx_xvfcvth_d_s((v8f32)_1);
1799	}
1800
1801	extern __inline
1802	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1803	__lasx_xvfcvtl_s_h(__m256i _1) {
1804	return (__m256)__builtin_lasx_xvfcvtl_s_h((v16i16)_1);
1805	}
1806
1807	extern __inline
1808	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1809	__lasx_xvfcvtl_d_s(__m256 _1) {
1810	return (__m256d)__builtin_lasx_xvfcvtl_d_s((v8f32)_1);
1811	}
1812
1813	extern __inline
1814	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1815	__lasx_xvftint_w_s(__m256 _1) {
1816	return (__m256i)__builtin_lasx_xvftint_w_s((v8f32)_1);
1817	}
1818
1819	extern __inline
1820	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1821	__lasx_xvftint_l_d(__m256d _1) {
1822	return (__m256i)__builtin_lasx_xvftint_l_d((v4f64)_1);
1823	}
1824
1825	extern __inline
1826	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1827	__lasx_xvftint_wu_s(__m256 _1) {
1828	return (__m256i)__builtin_lasx_xvftint_wu_s((v8f32)_1);
1829	}
1830
1831	extern __inline
1832	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1833	__lasx_xvftint_lu_d(__m256d _1) {
1834	return (__m256i)__builtin_lasx_xvftint_lu_d((v4f64)_1);
1835	}
1836
1837	extern __inline
1838	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1839	__lasx_xvftintrz_w_s(__m256 _1) {
1840	return (__m256i)__builtin_lasx_xvftintrz_w_s((v8f32)_1);
1841	}
1842
1843	extern __inline
1844	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1845	__lasx_xvftintrz_l_d(__m256d _1) {
1846	return (__m256i)__builtin_lasx_xvftintrz_l_d((v4f64)_1);
1847	}
1848
1849	extern __inline
1850	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1851	__lasx_xvftintrz_wu_s(__m256 _1) {
1852	return (__m256i)__builtin_lasx_xvftintrz_wu_s((v8f32)_1);
1853	}
1854
1855	extern __inline
1856	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1857	__lasx_xvftintrz_lu_d(__m256d _1) {
1858	return (__m256i)__builtin_lasx_xvftintrz_lu_d((v4f64)_1);
1859	}
1860
1861	extern __inline
1862	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1863	__lasx_xvffint_s_w(__m256i _1) {
1864	return (__m256)__builtin_lasx_xvffint_s_w((v8i32)_1);
1865	}
1866
1867	extern __inline
1868	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1869	__lasx_xvffint_d_l(__m256i _1) {
1870	return (__m256d)__builtin_lasx_xvffint_d_l((v4i64)_1);
1871	}
1872
1873	extern __inline
1874	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
1875	__lasx_xvffint_s_wu(__m256i _1) {
1876	return (__m256)__builtin_lasx_xvffint_s_wu((v8u32)_1);
1877	}
1878
1879	extern __inline
1880	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
1881	__lasx_xvffint_d_lu(__m256i _1) {
1882	return (__m256d)__builtin_lasx_xvffint_d_lu((v4u64)_1);
1883	}
1884
1885	extern __inline
1886	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1887	__lasx_xvreplve_b(__m256i _1, int _2) {
1888	return (__m256i)__builtin_lasx_xvreplve_b((v32i8)_1, (int)_2);
1889	}
1890
1891	extern __inline
1892	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1893	__lasx_xvreplve_h(__m256i _1, int _2) {
1894	return (__m256i)__builtin_lasx_xvreplve_h((v16i16)_1, (int)_2);
1895	}
1896
1897	extern __inline
1898	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1899	__lasx_xvreplve_w(__m256i _1, int _2) {
1900	return (__m256i)__builtin_lasx_xvreplve_w((v8i32)_1, (int)_2);
1901	}
1902
1903	extern __inline
1904	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1905	__lasx_xvreplve_d(__m256i _1, int _2) {
1906	return (__m256i)__builtin_lasx_xvreplve_d((v4i64)_1, (int)_2);
1907	}
1908
1909	#define __lasx_xvpermi_w(/__m256i/ _1, /__m256i/ _2, /ui8/ _3) \
1910	((__m256i)__builtin_lasx_xvpermi_w((v8i32)(_1), (v8i32)(_2), (_3)))
1911
1912	extern __inline
1913	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1914	__lasx_xvandn_v(__m256i _1, __m256i _2) {
1915	return (__m256i)__builtin_lasx_xvandn_v((v32u8)_1, (v32u8)_2);
1916	}
1917
1918	extern __inline
1919	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1920	__lasx_xvneg_b(__m256i _1) {
1921	return (__m256i)__builtin_lasx_xvneg_b((v32i8)_1);
1922	}
1923
1924	extern __inline
1925	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1926	__lasx_xvneg_h(__m256i _1) {
1927	return (__m256i)__builtin_lasx_xvneg_h((v16i16)_1);
1928	}
1929
1930	extern __inline
1931	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1932	__lasx_xvneg_w(__m256i _1) {
1933	return (__m256i)__builtin_lasx_xvneg_w((v8i32)_1);
1934	}
1935
1936	extern __inline
1937	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1938	__lasx_xvneg_d(__m256i _1) {
1939	return (__m256i)__builtin_lasx_xvneg_d((v4i64)_1);
1940	}
1941
1942	extern __inline
1943	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1944	__lasx_xvmuh_b(__m256i _1, __m256i _2) {
1945	return (__m256i)__builtin_lasx_xvmuh_b((v32i8)_1, (v32i8)_2);
1946	}
1947
1948	extern __inline
1949	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1950	__lasx_xvmuh_h(__m256i _1, __m256i _2) {
1951	return (__m256i)__builtin_lasx_xvmuh_h((v16i16)_1, (v16i16)_2);
1952	}
1953
1954	extern __inline
1955	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1956	__lasx_xvmuh_w(__m256i _1, __m256i _2) {
1957	return (__m256i)__builtin_lasx_xvmuh_w((v8i32)_1, (v8i32)_2);
1958	}
1959
1960	extern __inline
1961	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1962	__lasx_xvmuh_d(__m256i _1, __m256i _2) {
1963	return (__m256i)__builtin_lasx_xvmuh_d((v4i64)_1, (v4i64)_2);
1964	}
1965
1966	extern __inline
1967	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1968	__lasx_xvmuh_bu(__m256i _1, __m256i _2) {
1969	return (__m256i)__builtin_lasx_xvmuh_bu((v32u8)_1, (v32u8)_2);
1970	}
1971
1972	extern __inline
1973	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1974	__lasx_xvmuh_hu(__m256i _1, __m256i _2) {
1975	return (__m256i)__builtin_lasx_xvmuh_hu((v16u16)_1, (v16u16)_2);
1976	}
1977
1978	extern __inline
1979	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1980	__lasx_xvmuh_wu(__m256i _1, __m256i _2) {
1981	return (__m256i)__builtin_lasx_xvmuh_wu((v8u32)_1, (v8u32)_2);
1982	}
1983
1984	extern __inline
1985	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
1986	__lasx_xvmuh_du(__m256i _1, __m256i _2) {
1987	return (__m256i)__builtin_lasx_xvmuh_du((v4u64)_1, (v4u64)_2);
1988	}
1989
1990	#define __lasx_xvsllwil_h_b(/__m256i/ _1, /ui3/ _2) \
1991	((__m256i)__builtin_lasx_xvsllwil_h_b((v32i8)(_1), (_2)))
1992
1993	#define __lasx_xvsllwil_w_h(/__m256i/ _1, /ui4/ _2) \
1994	((__m256i)__builtin_lasx_xvsllwil_w_h((v16i16)(_1), (_2)))
1995
1996	#define __lasx_xvsllwil_d_w(/__m256i/ _1, /ui5/ _2) \
1997	((__m256i)__builtin_lasx_xvsllwil_d_w((v8i32)(_1), (_2)))
1998
1999	#define __lasx_xvsllwil_hu_bu(/__m256i/ _1, /ui3/ _2) \
2000	((__m256i)__builtin_lasx_xvsllwil_hu_bu((v32u8)(_1), (_2)))
2001
2002	#define __lasx_xvsllwil_wu_hu(/__m256i/ _1, /ui4/ _2) \
2003	((__m256i)__builtin_lasx_xvsllwil_wu_hu((v16u16)(_1), (_2)))
2004
2005	#define __lasx_xvsllwil_du_wu(/__m256i/ _1, /ui5/ _2) \
2006	((__m256i)__builtin_lasx_xvsllwil_du_wu((v8u32)(_1), (_2)))
2007
2008	extern __inline
2009	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2010	__lasx_xvsran_b_h(__m256i _1, __m256i _2) {
2011	return (__m256i)__builtin_lasx_xvsran_b_h((v16i16)_1, (v16i16)_2);
2012	}
2013
2014	extern __inline
2015	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2016	__lasx_xvsran_h_w(__m256i _1, __m256i _2) {
2017	return (__m256i)__builtin_lasx_xvsran_h_w((v8i32)_1, (v8i32)_2);
2018	}
2019
2020	extern __inline
2021	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2022	__lasx_xvsran_w_d(__m256i _1, __m256i _2) {
2023	return (__m256i)__builtin_lasx_xvsran_w_d((v4i64)_1, (v4i64)_2);
2024	}
2025
2026	extern __inline
2027	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2028	__lasx_xvssran_b_h(__m256i _1, __m256i _2) {
2029	return (__m256i)__builtin_lasx_xvssran_b_h((v16i16)_1, (v16i16)_2);
2030	}
2031
2032	extern __inline
2033	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2034	__lasx_xvssran_h_w(__m256i _1, __m256i _2) {
2035	return (__m256i)__builtin_lasx_xvssran_h_w((v8i32)_1, (v8i32)_2);
2036	}
2037
2038	extern __inline
2039	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2040	__lasx_xvssran_w_d(__m256i _1, __m256i _2) {
2041	return (__m256i)__builtin_lasx_xvssran_w_d((v4i64)_1, (v4i64)_2);
2042	}
2043
2044	extern __inline
2045	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2046	__lasx_xvssran_bu_h(__m256i _1, __m256i _2) {
2047	return (__m256i)__builtin_lasx_xvssran_bu_h((v16u16)_1, (v16u16)_2);
2048	}
2049
2050	extern __inline
2051	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2052	__lasx_xvssran_hu_w(__m256i _1, __m256i _2) {
2053	return (__m256i)__builtin_lasx_xvssran_hu_w((v8u32)_1, (v8u32)_2);
2054	}
2055
2056	extern __inline
2057	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2058	__lasx_xvssran_wu_d(__m256i _1, __m256i _2) {
2059	return (__m256i)__builtin_lasx_xvssran_wu_d((v4u64)_1, (v4u64)_2);
2060	}
2061
2062	extern __inline
2063	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2064	__lasx_xvsrarn_b_h(__m256i _1, __m256i _2) {
2065	return (__m256i)__builtin_lasx_xvsrarn_b_h((v16i16)_1, (v16i16)_2);
2066	}
2067
2068	extern __inline
2069	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2070	__lasx_xvsrarn_h_w(__m256i _1, __m256i _2) {
2071	return (__m256i)__builtin_lasx_xvsrarn_h_w((v8i32)_1, (v8i32)_2);
2072	}
2073
2074	extern __inline
2075	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2076	__lasx_xvsrarn_w_d(__m256i _1, __m256i _2) {
2077	return (__m256i)__builtin_lasx_xvsrarn_w_d((v4i64)_1, (v4i64)_2);
2078	}
2079
2080	extern __inline
2081	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2082	__lasx_xvssrarn_b_h(__m256i _1, __m256i _2) {
2083	return (__m256i)__builtin_lasx_xvssrarn_b_h((v16i16)_1, (v16i16)_2);
2084	}
2085
2086	extern __inline
2087	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2088	__lasx_xvssrarn_h_w(__m256i _1, __m256i _2) {
2089	return (__m256i)__builtin_lasx_xvssrarn_h_w((v8i32)_1, (v8i32)_2);
2090	}
2091
2092	extern __inline
2093	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2094	__lasx_xvssrarn_w_d(__m256i _1, __m256i _2) {
2095	return (__m256i)__builtin_lasx_xvssrarn_w_d((v4i64)_1, (v4i64)_2);
2096	}
2097
2098	extern __inline
2099	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2100	__lasx_xvssrarn_bu_h(__m256i _1, __m256i _2) {
2101	return (__m256i)__builtin_lasx_xvssrarn_bu_h((v16u16)_1, (v16u16)_2);
2102	}
2103
2104	extern __inline
2105	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2106	__lasx_xvssrarn_hu_w(__m256i _1, __m256i _2) {
2107	return (__m256i)__builtin_lasx_xvssrarn_hu_w((v8u32)_1, (v8u32)_2);
2108	}
2109
2110	extern __inline
2111	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2112	__lasx_xvssrarn_wu_d(__m256i _1, __m256i _2) {
2113	return (__m256i)__builtin_lasx_xvssrarn_wu_d((v4u64)_1, (v4u64)_2);
2114	}
2115
2116	extern __inline
2117	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2118	__lasx_xvsrln_b_h(__m256i _1, __m256i _2) {
2119	return (__m256i)__builtin_lasx_xvsrln_b_h((v16i16)_1, (v16i16)_2);
2120	}
2121
2122	extern __inline
2123	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2124	__lasx_xvsrln_h_w(__m256i _1, __m256i _2) {
2125	return (__m256i)__builtin_lasx_xvsrln_h_w((v8i32)_1, (v8i32)_2);
2126	}
2127
2128	extern __inline
2129	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2130	__lasx_xvsrln_w_d(__m256i _1, __m256i _2) {
2131	return (__m256i)__builtin_lasx_xvsrln_w_d((v4i64)_1, (v4i64)_2);
2132	}
2133
2134	extern __inline
2135	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2136	__lasx_xvssrln_bu_h(__m256i _1, __m256i _2) {
2137	return (__m256i)__builtin_lasx_xvssrln_bu_h((v16u16)_1, (v16u16)_2);
2138	}
2139
2140	extern __inline
2141	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2142	__lasx_xvssrln_hu_w(__m256i _1, __m256i _2) {
2143	return (__m256i)__builtin_lasx_xvssrln_hu_w((v8u32)_1, (v8u32)_2);
2144	}
2145
2146	extern __inline
2147	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2148	__lasx_xvssrln_wu_d(__m256i _1, __m256i _2) {
2149	return (__m256i)__builtin_lasx_xvssrln_wu_d((v4u64)_1, (v4u64)_2);
2150	}
2151
2152	extern __inline
2153	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2154	__lasx_xvsrlrn_b_h(__m256i _1, __m256i _2) {
2155	return (__m256i)__builtin_lasx_xvsrlrn_b_h((v16i16)_1, (v16i16)_2);
2156	}
2157
2158	extern __inline
2159	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2160	__lasx_xvsrlrn_h_w(__m256i _1, __m256i _2) {
2161	return (__m256i)__builtin_lasx_xvsrlrn_h_w((v8i32)_1, (v8i32)_2);
2162	}
2163
2164	extern __inline
2165	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2166	__lasx_xvsrlrn_w_d(__m256i _1, __m256i _2) {
2167	return (__m256i)__builtin_lasx_xvsrlrn_w_d((v4i64)_1, (v4i64)_2);
2168	}
2169
2170	extern __inline
2171	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2172	__lasx_xvssrlrn_bu_h(__m256i _1, __m256i _2) {
2173	return (__m256i)__builtin_lasx_xvssrlrn_bu_h((v16u16)_1, (v16u16)_2);
2174	}
2175
2176	extern __inline
2177	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2178	__lasx_xvssrlrn_hu_w(__m256i _1, __m256i _2) {
2179	return (__m256i)__builtin_lasx_xvssrlrn_hu_w((v8u32)_1, (v8u32)_2);
2180	}
2181
2182	extern __inline
2183	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2184	__lasx_xvssrlrn_wu_d(__m256i _1, __m256i _2) {
2185	return (__m256i)__builtin_lasx_xvssrlrn_wu_d((v4u64)_1, (v4u64)_2);
2186	}
2187
2188	#define __lasx_xvfrstpi_b(/__m256i/ _1, /__m256i/ _2, /ui5/ _3) \
2189	((__m256i)__builtin_lasx_xvfrstpi_b((v32i8)(_1), (v32i8)(_2), (_3)))
2190
2191	#define __lasx_xvfrstpi_h(/__m256i/ _1, /__m256i/ _2, /ui5/ _3) \
2192	((__m256i)__builtin_lasx_xvfrstpi_h((v16i16)(_1), (v16i16)(_2), (_3)))
2193
2194	extern __inline
2195	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2196	__lasx_xvfrstp_b(__m256i _1, __m256i _2, __m256i _3) {
2197	return (__m256i)__builtin_lasx_xvfrstp_b((v32i8)_1, (v32i8)_2, (v32i8)_3);
2198	}
2199
2200	extern __inline
2201	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2202	__lasx_xvfrstp_h(__m256i _1, __m256i _2, __m256i _3) {
2203	return (__m256i)__builtin_lasx_xvfrstp_h((v16i16)_1, (v16i16)_2, (v16i16)_3);
2204	}
2205
2206	#define __lasx_xvshuf4i_d(/__m256i/ _1, /__m256i/ _2, /ui8/ _3) \
2207	((__m256i)__builtin_lasx_xvshuf4i_d((v4i64)(_1), (v4i64)(_2), (_3)))
2208
2209	#define __lasx_xvbsrl_v(/__m256i/ _1, /ui5/ _2) \
2210	((__m256i)__builtin_lasx_xvbsrl_v((v32i8)(_1), (_2)))
2211
2212	#define __lasx_xvbsll_v(/__m256i/ _1, /ui5/ _2) \
2213	((__m256i)__builtin_lasx_xvbsll_v((v32i8)(_1), (_2)))
2214
2215	#define __lasx_xvextrins_b(/__m256i/ _1, /__m256i/ _2, /ui8/ _3) \
2216	((__m256i)__builtin_lasx_xvextrins_b((v32i8)(_1), (v32i8)(_2), (_3)))
2217
2218	#define __lasx_xvextrins_h(/__m256i/ _1, /__m256i/ _2, /ui8/ _3) \
2219	((__m256i)__builtin_lasx_xvextrins_h((v16i16)(_1), (v16i16)(_2), (_3)))
2220
2221	#define __lasx_xvextrins_w(/__m256i/ _1, /__m256i/ _2, /ui8/ _3) \
2222	((__m256i)__builtin_lasx_xvextrins_w((v8i32)(_1), (v8i32)(_2), (_3)))
2223
2224	#define __lasx_xvextrins_d(/__m256i/ _1, /__m256i/ _2, /ui8/ _3) \
2225	((__m256i)__builtin_lasx_xvextrins_d((v4i64)(_1), (v4i64)(_2), (_3)))
2226
2227	extern __inline
2228	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2229	__lasx_xvmskltz_b(__m256i _1) {
2230	return (__m256i)__builtin_lasx_xvmskltz_b((v32i8)_1);
2231	}
2232
2233	extern __inline
2234	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2235	__lasx_xvmskltz_h(__m256i _1) {
2236	return (__m256i)__builtin_lasx_xvmskltz_h((v16i16)_1);
2237	}
2238
2239	extern __inline
2240	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2241	__lasx_xvmskltz_w(__m256i _1) {
2242	return (__m256i)__builtin_lasx_xvmskltz_w((v8i32)_1);
2243	}
2244
2245	extern __inline
2246	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2247	__lasx_xvmskltz_d(__m256i _1) {
2248	return (__m256i)__builtin_lasx_xvmskltz_d((v4i64)_1);
2249	}
2250
2251	extern __inline
2252	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2253	__lasx_xvsigncov_b(__m256i _1, __m256i _2) {
2254	return (__m256i)__builtin_lasx_xvsigncov_b((v32i8)_1, (v32i8)_2);
2255	}
2256
2257	extern __inline
2258	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2259	__lasx_xvsigncov_h(__m256i _1, __m256i _2) {
2260	return (__m256i)__builtin_lasx_xvsigncov_h((v16i16)_1, (v16i16)_2);
2261	}
2262
2263	extern __inline
2264	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2265	__lasx_xvsigncov_w(__m256i _1, __m256i _2) {
2266	return (__m256i)__builtin_lasx_xvsigncov_w((v8i32)_1, (v8i32)_2);
2267	}
2268
2269	extern __inline
2270	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2271	__lasx_xvsigncov_d(__m256i _1, __m256i _2) {
2272	return (__m256i)__builtin_lasx_xvsigncov_d((v4i64)_1, (v4i64)_2);
2273	}
2274
2275	extern __inline
2276	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
2277	__lasx_xvfmadd_s(__m256 _1, __m256 _2, __m256 _3) {
2278	return (__m256)__builtin_lasx_xvfmadd_s((v8f32)_1, (v8f32)_2, (v8f32)_3);
2279	}
2280
2281	extern __inline
2282	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
2283	__lasx_xvfmadd_d(__m256d _1, __m256d _2, __m256d _3) {
2284	return (__m256d)__builtin_lasx_xvfmadd_d((v4f64)_1, (v4f64)_2, (v4f64)_3);
2285	}
2286
2287	extern __inline
2288	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
2289	__lasx_xvfmsub_s(__m256 _1, __m256 _2, __m256 _3) {
2290	return (__m256)__builtin_lasx_xvfmsub_s((v8f32)_1, (v8f32)_2, (v8f32)_3);
2291	}
2292
2293	extern __inline
2294	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
2295	__lasx_xvfmsub_d(__m256d _1, __m256d _2, __m256d _3) {
2296	return (__m256d)__builtin_lasx_xvfmsub_d((v4f64)_1, (v4f64)_2, (v4f64)_3);
2297	}
2298
2299	extern __inline
2300	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
2301	__lasx_xvfnmadd_s(__m256 _1, __m256 _2, __m256 _3) {
2302	return (__m256)__builtin_lasx_xvfnmadd_s((v8f32)_1, (v8f32)_2, (v8f32)_3);
2303	}
2304
2305	extern __inline
2306	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
2307	__lasx_xvfnmadd_d(__m256d _1, __m256d _2, __m256d _3) {
2308	return (__m256d)__builtin_lasx_xvfnmadd_d((v4f64)_1, (v4f64)_2, (v4f64)_3);
2309	}
2310
2311	extern __inline
2312	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
2313	__lasx_xvfnmsub_s(__m256 _1, __m256 _2, __m256 _3) {
2314	return (__m256)__builtin_lasx_xvfnmsub_s((v8f32)_1, (v8f32)_2, (v8f32)_3);
2315	}
2316
2317	extern __inline
2318	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
2319	__lasx_xvfnmsub_d(__m256d _1, __m256d _2, __m256d _3) {
2320	return (__m256d)__builtin_lasx_xvfnmsub_d((v4f64)_1, (v4f64)_2, (v4f64)_3);
2321	}
2322
2323	extern __inline
2324	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2325	__lasx_xvftintrne_w_s(__m256 _1) {
2326	return (__m256i)__builtin_lasx_xvftintrne_w_s((v8f32)_1);
2327	}
2328
2329	extern __inline
2330	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2331	__lasx_xvftintrne_l_d(__m256d _1) {
2332	return (__m256i)__builtin_lasx_xvftintrne_l_d((v4f64)_1);
2333	}
2334
2335	extern __inline
2336	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2337	__lasx_xvftintrp_w_s(__m256 _1) {
2338	return (__m256i)__builtin_lasx_xvftintrp_w_s((v8f32)_1);
2339	}
2340
2341	extern __inline
2342	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2343	__lasx_xvftintrp_l_d(__m256d _1) {
2344	return (__m256i)__builtin_lasx_xvftintrp_l_d((v4f64)_1);
2345	}
2346
2347	extern __inline
2348	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2349	__lasx_xvftintrm_w_s(__m256 _1) {
2350	return (__m256i)__builtin_lasx_xvftintrm_w_s((v8f32)_1);
2351	}
2352
2353	extern __inline
2354	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2355	__lasx_xvftintrm_l_d(__m256d _1) {
2356	return (__m256i)__builtin_lasx_xvftintrm_l_d((v4f64)_1);
2357	}
2358
2359	extern __inline
2360	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2361	__lasx_xvftint_w_d(__m256d _1, __m256d _2) {
2362	return (__m256i)__builtin_lasx_xvftint_w_d((v4f64)_1, (v4f64)_2);
2363	}
2364
2365	extern __inline
2366	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
2367	__lasx_xvffint_s_l(__m256i _1, __m256i _2) {
2368	return (__m256)__builtin_lasx_xvffint_s_l((v4i64)_1, (v4i64)_2);
2369	}
2370
2371	extern __inline
2372	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2373	__lasx_xvftintrz_w_d(__m256d _1, __m256d _2) {
2374	return (__m256i)__builtin_lasx_xvftintrz_w_d((v4f64)_1, (v4f64)_2);
2375	}
2376
2377	extern __inline
2378	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2379	__lasx_xvftintrp_w_d(__m256d _1, __m256d _2) {
2380	return (__m256i)__builtin_lasx_xvftintrp_w_d((v4f64)_1, (v4f64)_2);
2381	}
2382
2383	extern __inline
2384	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2385	__lasx_xvftintrm_w_d(__m256d _1, __m256d _2) {
2386	return (__m256i)__builtin_lasx_xvftintrm_w_d((v4f64)_1, (v4f64)_2);
2387	}
2388
2389	extern __inline
2390	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2391	__lasx_xvftintrne_w_d(__m256d _1, __m256d _2) {
2392	return (__m256i)__builtin_lasx_xvftintrne_w_d((v4f64)_1, (v4f64)_2);
2393	}
2394
2395	extern __inline
2396	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2397	__lasx_xvftinth_l_s(__m256 _1) {
2398	return (__m256i)__builtin_lasx_xvftinth_l_s((v8f32)_1);
2399	}
2400
2401	extern __inline
2402	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2403	__lasx_xvftintl_l_s(__m256 _1) {
2404	return (__m256i)__builtin_lasx_xvftintl_l_s((v8f32)_1);
2405	}
2406
2407	extern __inline
2408	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
2409	__lasx_xvffinth_d_w(__m256i _1) {
2410	return (__m256d)__builtin_lasx_xvffinth_d_w((v8i32)_1);
2411	}
2412
2413	extern __inline
2414	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
2415	__lasx_xvffintl_d_w(__m256i _1) {
2416	return (__m256d)__builtin_lasx_xvffintl_d_w((v8i32)_1);
2417	}
2418
2419	extern __inline
2420	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2421	__lasx_xvftintrzh_l_s(__m256 _1) {
2422	return (__m256i)__builtin_lasx_xvftintrzh_l_s((v8f32)_1);
2423	}
2424
2425	extern __inline
2426	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2427	__lasx_xvftintrzl_l_s(__m256 _1) {
2428	return (__m256i)__builtin_lasx_xvftintrzl_l_s((v8f32)_1);
2429	}
2430
2431	extern __inline
2432	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2433	__lasx_xvftintrph_l_s(__m256 _1) {
2434	return (__m256i)__builtin_lasx_xvftintrph_l_s((v8f32)_1);
2435	}
2436
2437	extern __inline
2438	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2439	__lasx_xvftintrpl_l_s(__m256 _1) {
2440	return (__m256i)__builtin_lasx_xvftintrpl_l_s((v8f32)_1);
2441	}
2442
2443	extern __inline
2444	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2445	__lasx_xvftintrmh_l_s(__m256 _1) {
2446	return (__m256i)__builtin_lasx_xvftintrmh_l_s((v8f32)_1);
2447	}
2448
2449	extern __inline
2450	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2451	__lasx_xvftintrml_l_s(__m256 _1) {
2452	return (__m256i)__builtin_lasx_xvftintrml_l_s((v8f32)_1);
2453	}
2454
2455	extern __inline
2456	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2457	__lasx_xvftintrneh_l_s(__m256 _1) {
2458	return (__m256i)__builtin_lasx_xvftintrneh_l_s((v8f32)_1);
2459	}
2460
2461	extern __inline
2462	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2463	__lasx_xvftintrnel_l_s(__m256 _1) {
2464	return (__m256i)__builtin_lasx_xvftintrnel_l_s((v8f32)_1);
2465	}
2466
2467	extern __inline
2468	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
2469	__lasx_xvfrintrne_s(__m256 _1) {
2470	return (__m256)__builtin_lasx_xvfrintrne_s((v8f32)_1);
2471	}
2472
2473	extern __inline
2474	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
2475	__lasx_xvfrintrne_d(__m256d _1) {
2476	return (__m256d)__builtin_lasx_xvfrintrne_d((v4f64)_1);
2477	}
2478
2479	extern __inline
2480	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
2481	__lasx_xvfrintrz_s(__m256 _1) {
2482	return (__m256)__builtin_lasx_xvfrintrz_s((v8f32)_1);
2483	}
2484
2485	extern __inline
2486	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
2487	__lasx_xvfrintrz_d(__m256d _1) {
2488	return (__m256d)__builtin_lasx_xvfrintrz_d((v4f64)_1);
2489	}
2490
2491	extern __inline
2492	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
2493	__lasx_xvfrintrp_s(__m256 _1) {
2494	return (__m256)__builtin_lasx_xvfrintrp_s((v8f32)_1);
2495	}
2496
2497	extern __inline
2498	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
2499	__lasx_xvfrintrp_d(__m256d _1) {
2500	return (__m256d)__builtin_lasx_xvfrintrp_d((v4f64)_1);
2501	}
2502
2503	extern __inline
2504	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
2505	__lasx_xvfrintrm_s(__m256 _1) {
2506	return (__m256)__builtin_lasx_xvfrintrm_s((v8f32)_1);
2507	}
2508
2509	extern __inline
2510	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
2511	__lasx_xvfrintrm_d(__m256d _1) {
2512	return (__m256d)__builtin_lasx_xvfrintrm_d((v4f64)_1);
2513	}
2514
2515	#define __lasx_xvld(/void / _1, /si12*/ _2) \
2516	((__m256i)__builtin_lasx_xvld((void const *)(_1), (_2)))
2517
2518	#define __lasx_xvst(/__m256i/ _1, /void / _2, /si12*/ _3) \
2519	((void)__builtin_lasx_xvst((v32i8)(_1), (void *)(_2), (_3)))
2520
2521	#define __lasx_xvstelm_b(/__m256i/ _1, /void / _2, /si8*/ _3, \
2522	/idx/ _4) \
2523	((void)__builtin_lasx_xvstelm_b((v32i8)(_1), (void *)(_2), (_3), (_4)))
2524
2525	#define __lasx_xvstelm_h(/__m256i/ _1, /void / _2, /si8*/ _3, \
2526	/idx/ _4) \
2527	((void)__builtin_lasx_xvstelm_h((v16i16)(_1), (void *)(_2), (_3), (_4)))
2528
2529	#define __lasx_xvstelm_w(/__m256i/ _1, /void / _2, /si8*/ _3, \
2530	/idx/ _4) \
2531	((void)__builtin_lasx_xvstelm_w((v8i32)(_1), (void *)(_2), (_3), (_4)))
2532
2533	#define __lasx_xvstelm_d(/__m256i/ _1, /void / _2, /si8*/ _3, \
2534	/idx/ _4) \
2535	((void)__builtin_lasx_xvstelm_d((v4i64)(_1), (void *)(_2), (_3), (_4)))
2536
2537	#define __lasx_xvinsve0_w(/__m256i/ _1, /__m256i/ _2, /ui3/ _3) \
2538	((__m256i)__builtin_lasx_xvinsve0_w((v8i32)(_1), (v8i32)(_2), (_3)))
2539
2540	#define __lasx_xvinsve0_d(/__m256i/ _1, /__m256i/ _2, /ui2/ _3) \
2541	((__m256i)__builtin_lasx_xvinsve0_d((v4i64)(_1), (v4i64)(_2), (_3)))
2542
2543	#define __lasx_xvpickve_w(/__m256i/ _1, /ui3/ _2) \
2544	((__m256i)__builtin_lasx_xvpickve_w((v8i32)(_1), (_2)))
2545
2546	#define __lasx_xvpickve_d(/__m256i/ _1, /ui2/ _2) \
2547	((__m256i)__builtin_lasx_xvpickve_d((v4i64)(_1), (_2)))
2548
2549	extern __inline
2550	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2551	__lasx_xvssrlrn_b_h(__m256i _1, __m256i _2) {
2552	return (__m256i)__builtin_lasx_xvssrlrn_b_h((v16i16)_1, (v16i16)_2);
2553	}
2554
2555	extern __inline
2556	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2557	__lasx_xvssrlrn_h_w(__m256i _1, __m256i _2) {
2558	return (__m256i)__builtin_lasx_xvssrlrn_h_w((v8i32)_1, (v8i32)_2);
2559	}
2560
2561	extern __inline
2562	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2563	__lasx_xvssrlrn_w_d(__m256i _1, __m256i _2) {
2564	return (__m256i)__builtin_lasx_xvssrlrn_w_d((v4i64)_1, (v4i64)_2);
2565	}
2566
2567	extern __inline
2568	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2569	__lasx_xvssrln_b_h(__m256i _1, __m256i _2) {
2570	return (__m256i)__builtin_lasx_xvssrln_b_h((v16i16)_1, (v16i16)_2);
2571	}
2572
2573	extern __inline
2574	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2575	__lasx_xvssrln_h_w(__m256i _1, __m256i _2) {
2576	return (__m256i)__builtin_lasx_xvssrln_h_w((v8i32)_1, (v8i32)_2);
2577	}
2578
2579	extern __inline
2580	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2581	__lasx_xvssrln_w_d(__m256i _1, __m256i _2) {
2582	return (__m256i)__builtin_lasx_xvssrln_w_d((v4i64)_1, (v4i64)_2);
2583	}
2584
2585	extern __inline
2586	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2587	__lasx_xvorn_v(__m256i _1, __m256i _2) {
2588	return (__m256i)__builtin_lasx_xvorn_v((v32i8)_1, (v32i8)_2);
2589	}
2590
2591	#define __lasx_xvldi(/i13/ _1) ((__m256i)__builtin_lasx_xvldi((_1)))
2592
2593	extern __inline
2594	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2595	__lasx_xvldx(void const *_1, long int _2) {
2596	return (__m256i)__builtin_lasx_xvldx((void const *)_1, (long int)_2);
2597	}
2598
2599	extern __inline
2600	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) void
2601	__lasx_xvstx(__m256i _1, void *_2, long int _3) {
2602	return (void)__builtin_lasx_xvstx((v32i8)_1, (void *)_2, (long int)_3);
2603	}
2604
2605	extern __inline
2606	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2607	__lasx_xvextl_qu_du(__m256i _1) {
2608	return (__m256i)__builtin_lasx_xvextl_qu_du((v4u64)_1);
2609	}
2610
2611	#define __lasx_xvinsgr2vr_w(/__m256i/ _1, /int/ _2, /ui3/ _3) \
2612	((__m256i)__builtin_lasx_xvinsgr2vr_w((v8i32)(_1), (int)(_2), (_3)))
2613
2614	#define __lasx_xvinsgr2vr_d(/__m256i/ _1, /long int/ _2, /ui2/ _3) \
2615	((__m256i)__builtin_lasx_xvinsgr2vr_d((v4i64)(_1), (long int)(_2), (_3)))
2616
2617	extern __inline
2618	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2619	__lasx_xvreplve0_b(__m256i _1) {
2620	return (__m256i)__builtin_lasx_xvreplve0_b((v32i8)_1);
2621	}
2622
2623	extern __inline
2624	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2625	__lasx_xvreplve0_h(__m256i _1) {
2626	return (__m256i)__builtin_lasx_xvreplve0_h((v16i16)_1);
2627	}
2628
2629	extern __inline
2630	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2631	__lasx_xvreplve0_w(__m256i _1) {
2632	return (__m256i)__builtin_lasx_xvreplve0_w((v8i32)_1);
2633	}
2634
2635	extern __inline
2636	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2637	__lasx_xvreplve0_d(__m256i _1) {
2638	return (__m256i)__builtin_lasx_xvreplve0_d((v4i64)_1);
2639	}
2640
2641	extern __inline
2642	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2643	__lasx_xvreplve0_q(__m256i _1) {
2644	return (__m256i)__builtin_lasx_xvreplve0_q((v32i8)_1);
2645	}
2646
2647	extern __inline
2648	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2649	__lasx_vext2xv_h_b(__m256i _1) {
2650	return (__m256i)__builtin_lasx_vext2xv_h_b((v32i8)_1);
2651	}
2652
2653	extern __inline
2654	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2655	__lasx_vext2xv_w_h(__m256i _1) {
2656	return (__m256i)__builtin_lasx_vext2xv_w_h((v16i16)_1);
2657	}
2658
2659	extern __inline
2660	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2661	__lasx_vext2xv_d_w(__m256i _1) {
2662	return (__m256i)__builtin_lasx_vext2xv_d_w((v8i32)_1);
2663	}
2664
2665	extern __inline
2666	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2667	__lasx_vext2xv_w_b(__m256i _1) {
2668	return (__m256i)__builtin_lasx_vext2xv_w_b((v32i8)_1);
2669	}
2670
2671	extern __inline
2672	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2673	__lasx_vext2xv_d_h(__m256i _1) {
2674	return (__m256i)__builtin_lasx_vext2xv_d_h((v16i16)_1);
2675	}
2676
2677	extern __inline
2678	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2679	__lasx_vext2xv_d_b(__m256i _1) {
2680	return (__m256i)__builtin_lasx_vext2xv_d_b((v32i8)_1);
2681	}
2682
2683	extern __inline
2684	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2685	__lasx_vext2xv_hu_bu(__m256i _1) {
2686	return (__m256i)__builtin_lasx_vext2xv_hu_bu((v32i8)_1);
2687	}
2688
2689	extern __inline
2690	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2691	__lasx_vext2xv_wu_hu(__m256i _1) {
2692	return (__m256i)__builtin_lasx_vext2xv_wu_hu((v16i16)_1);
2693	}
2694
2695	extern __inline
2696	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2697	__lasx_vext2xv_du_wu(__m256i _1) {
2698	return (__m256i)__builtin_lasx_vext2xv_du_wu((v8i32)_1);
2699	}
2700
2701	extern __inline
2702	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2703	__lasx_vext2xv_wu_bu(__m256i _1) {
2704	return (__m256i)__builtin_lasx_vext2xv_wu_bu((v32i8)_1);
2705	}
2706
2707	extern __inline
2708	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2709	__lasx_vext2xv_du_hu(__m256i _1) {
2710	return (__m256i)__builtin_lasx_vext2xv_du_hu((v16i16)_1);
2711	}
2712
2713	extern __inline
2714	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2715	__lasx_vext2xv_du_bu(__m256i _1) {
2716	return (__m256i)__builtin_lasx_vext2xv_du_bu((v32i8)_1);
2717	}
2718
2719	#define __lasx_xvpermi_q(/__m256i/ _1, /__m256i/ _2, /ui8/ _3) \
2720	((__m256i)__builtin_lasx_xvpermi_q((v32i8)(_1), (v32i8)(_2), (_3)))
2721
2722	#define __lasx_xvpermi_d(/__m256i/ _1, /ui8/ _2) \
2723	((__m256i)__builtin_lasx_xvpermi_d((v4i64)(_1), (_2)))
2724
2725	extern __inline
2726	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2727	__lasx_xvperm_w(__m256i _1, __m256i _2) {
2728	return (__m256i)__builtin_lasx_xvperm_w((v8i32)_1, (v8i32)_2);
2729	}
2730
2731	#define __lasx_xvldrepl_b(/void / _1, /si12*/ _2) \
2732	((__m256i)__builtin_lasx_xvldrepl_b((void const *)(_1), (_2)))
2733
2734	#define __lasx_xvldrepl_h(/void / _1, /si11*/ _2) \
2735	((__m256i)__builtin_lasx_xvldrepl_h((void const *)(_1), (_2)))
2736
2737	#define __lasx_xvldrepl_w(/void / _1, /si10*/ _2) \
2738	((__m256i)__builtin_lasx_xvldrepl_w((void const *)(_1), (_2)))
2739
2740	#define __lasx_xvldrepl_d(/void / _1, /si9*/ _2) \
2741	((__m256i)__builtin_lasx_xvldrepl_d((void const *)(_1), (_2)))
2742
2743	#define __lasx_xvpickve2gr_w(/__m256i/ _1, /ui3/ _2) \
2744	((int)__builtin_lasx_xvpickve2gr_w((v8i32)(_1), (_2)))
2745
2746	#define __lasx_xvpickve2gr_wu(/__m256i/ _1, /ui3/ _2) \
2747	((unsigned int)__builtin_lasx_xvpickve2gr_wu((v8i32)(_1), (_2)))
2748
2749	#define __lasx_xvpickve2gr_d(/__m256i/ _1, /ui2/ _2) \
2750	((long int)__builtin_lasx_xvpickve2gr_d((v4i64)(_1), (_2)))
2751
2752	#define __lasx_xvpickve2gr_du(/__m256i/ _1, /ui2/ _2) \
2753	((unsigned long int)__builtin_lasx_xvpickve2gr_du((v4i64)(_1), (_2)))
2754
2755	extern __inline
2756	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2757	__lasx_xvaddwev_q_d(__m256i _1, __m256i _2) {
2758	return (__m256i)__builtin_lasx_xvaddwev_q_d((v4i64)_1, (v4i64)_2);
2759	}
2760
2761	extern __inline
2762	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2763	__lasx_xvaddwev_d_w(__m256i _1, __m256i _2) {
2764	return (__m256i)__builtin_lasx_xvaddwev_d_w((v8i32)_1, (v8i32)_2);
2765	}
2766
2767	extern __inline
2768	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2769	__lasx_xvaddwev_w_h(__m256i _1, __m256i _2) {
2770	return (__m256i)__builtin_lasx_xvaddwev_w_h((v16i16)_1, (v16i16)_2);
2771	}
2772
2773	extern __inline
2774	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2775	__lasx_xvaddwev_h_b(__m256i _1, __m256i _2) {
2776	return (__m256i)__builtin_lasx_xvaddwev_h_b((v32i8)_1, (v32i8)_2);
2777	}
2778
2779	extern __inline
2780	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2781	__lasx_xvaddwev_q_du(__m256i _1, __m256i _2) {
2782	return (__m256i)__builtin_lasx_xvaddwev_q_du((v4u64)_1, (v4u64)_2);
2783	}
2784
2785	extern __inline
2786	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2787	__lasx_xvaddwev_d_wu(__m256i _1, __m256i _2) {
2788	return (__m256i)__builtin_lasx_xvaddwev_d_wu((v8u32)_1, (v8u32)_2);
2789	}
2790
2791	extern __inline
2792	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2793	__lasx_xvaddwev_w_hu(__m256i _1, __m256i _2) {
2794	return (__m256i)__builtin_lasx_xvaddwev_w_hu((v16u16)_1, (v16u16)_2);
2795	}
2796
2797	extern __inline
2798	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2799	__lasx_xvaddwev_h_bu(__m256i _1, __m256i _2) {
2800	return (__m256i)__builtin_lasx_xvaddwev_h_bu((v32u8)_1, (v32u8)_2);
2801	}
2802
2803	extern __inline
2804	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2805	__lasx_xvsubwev_q_d(__m256i _1, __m256i _2) {
2806	return (__m256i)__builtin_lasx_xvsubwev_q_d((v4i64)_1, (v4i64)_2);
2807	}
2808
2809	extern __inline
2810	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2811	__lasx_xvsubwev_d_w(__m256i _1, __m256i _2) {
2812	return (__m256i)__builtin_lasx_xvsubwev_d_w((v8i32)_1, (v8i32)_2);
2813	}
2814
2815	extern __inline
2816	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2817	__lasx_xvsubwev_w_h(__m256i _1, __m256i _2) {
2818	return (__m256i)__builtin_lasx_xvsubwev_w_h((v16i16)_1, (v16i16)_2);
2819	}
2820
2821	extern __inline
2822	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2823	__lasx_xvsubwev_h_b(__m256i _1, __m256i _2) {
2824	return (__m256i)__builtin_lasx_xvsubwev_h_b((v32i8)_1, (v32i8)_2);
2825	}
2826
2827	extern __inline
2828	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2829	__lasx_xvsubwev_q_du(__m256i _1, __m256i _2) {
2830	return (__m256i)__builtin_lasx_xvsubwev_q_du((v4u64)_1, (v4u64)_2);
2831	}
2832
2833	extern __inline
2834	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2835	__lasx_xvsubwev_d_wu(__m256i _1, __m256i _2) {
2836	return (__m256i)__builtin_lasx_xvsubwev_d_wu((v8u32)_1, (v8u32)_2);
2837	}
2838
2839	extern __inline
2840	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2841	__lasx_xvsubwev_w_hu(__m256i _1, __m256i _2) {
2842	return (__m256i)__builtin_lasx_xvsubwev_w_hu((v16u16)_1, (v16u16)_2);
2843	}
2844
2845	extern __inline
2846	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2847	__lasx_xvsubwev_h_bu(__m256i _1, __m256i _2) {
2848	return (__m256i)__builtin_lasx_xvsubwev_h_bu((v32u8)_1, (v32u8)_2);
2849	}
2850
2851	extern __inline
2852	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2853	__lasx_xvmulwev_q_d(__m256i _1, __m256i _2) {
2854	return (__m256i)__builtin_lasx_xvmulwev_q_d((v4i64)_1, (v4i64)_2);
2855	}
2856
2857	extern __inline
2858	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2859	__lasx_xvmulwev_d_w(__m256i _1, __m256i _2) {
2860	return (__m256i)__builtin_lasx_xvmulwev_d_w((v8i32)_1, (v8i32)_2);
2861	}
2862
2863	extern __inline
2864	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2865	__lasx_xvmulwev_w_h(__m256i _1, __m256i _2) {
2866	return (__m256i)__builtin_lasx_xvmulwev_w_h((v16i16)_1, (v16i16)_2);
2867	}
2868
2869	extern __inline
2870	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2871	__lasx_xvmulwev_h_b(__m256i _1, __m256i _2) {
2872	return (__m256i)__builtin_lasx_xvmulwev_h_b((v32i8)_1, (v32i8)_2);
2873	}
2874
2875	extern __inline
2876	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2877	__lasx_xvmulwev_q_du(__m256i _1, __m256i _2) {
2878	return (__m256i)__builtin_lasx_xvmulwev_q_du((v4u64)_1, (v4u64)_2);
2879	}
2880
2881	extern __inline
2882	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2883	__lasx_xvmulwev_d_wu(__m256i _1, __m256i _2) {
2884	return (__m256i)__builtin_lasx_xvmulwev_d_wu((v8u32)_1, (v8u32)_2);
2885	}
2886
2887	extern __inline
2888	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2889	__lasx_xvmulwev_w_hu(__m256i _1, __m256i _2) {
2890	return (__m256i)__builtin_lasx_xvmulwev_w_hu((v16u16)_1, (v16u16)_2);
2891	}
2892
2893	extern __inline
2894	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2895	__lasx_xvmulwev_h_bu(__m256i _1, __m256i _2) {
2896	return (__m256i)__builtin_lasx_xvmulwev_h_bu((v32u8)_1, (v32u8)_2);
2897	}
2898
2899	extern __inline
2900	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2901	__lasx_xvaddwod_q_d(__m256i _1, __m256i _2) {
2902	return (__m256i)__builtin_lasx_xvaddwod_q_d((v4i64)_1, (v4i64)_2);
2903	}
2904
2905	extern __inline
2906	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2907	__lasx_xvaddwod_d_w(__m256i _1, __m256i _2) {
2908	return (__m256i)__builtin_lasx_xvaddwod_d_w((v8i32)_1, (v8i32)_2);
2909	}
2910
2911	extern __inline
2912	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2913	__lasx_xvaddwod_w_h(__m256i _1, __m256i _2) {
2914	return (__m256i)__builtin_lasx_xvaddwod_w_h((v16i16)_1, (v16i16)_2);
2915	}
2916
2917	extern __inline
2918	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2919	__lasx_xvaddwod_h_b(__m256i _1, __m256i _2) {
2920	return (__m256i)__builtin_lasx_xvaddwod_h_b((v32i8)_1, (v32i8)_2);
2921	}
2922
2923	extern __inline
2924	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2925	__lasx_xvaddwod_q_du(__m256i _1, __m256i _2) {
2926	return (__m256i)__builtin_lasx_xvaddwod_q_du((v4u64)_1, (v4u64)_2);
2927	}
2928
2929	extern __inline
2930	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2931	__lasx_xvaddwod_d_wu(__m256i _1, __m256i _2) {
2932	return (__m256i)__builtin_lasx_xvaddwod_d_wu((v8u32)_1, (v8u32)_2);
2933	}
2934
2935	extern __inline
2936	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2937	__lasx_xvaddwod_w_hu(__m256i _1, __m256i _2) {
2938	return (__m256i)__builtin_lasx_xvaddwod_w_hu((v16u16)_1, (v16u16)_2);
2939	}
2940
2941	extern __inline
2942	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2943	__lasx_xvaddwod_h_bu(__m256i _1, __m256i _2) {
2944	return (__m256i)__builtin_lasx_xvaddwod_h_bu((v32u8)_1, (v32u8)_2);
2945	}
2946
2947	extern __inline
2948	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2949	__lasx_xvsubwod_q_d(__m256i _1, __m256i _2) {
2950	return (__m256i)__builtin_lasx_xvsubwod_q_d((v4i64)_1, (v4i64)_2);
2951	}
2952
2953	extern __inline
2954	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2955	__lasx_xvsubwod_d_w(__m256i _1, __m256i _2) {
2956	return (__m256i)__builtin_lasx_xvsubwod_d_w((v8i32)_1, (v8i32)_2);
2957	}
2958
2959	extern __inline
2960	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2961	__lasx_xvsubwod_w_h(__m256i _1, __m256i _2) {
2962	return (__m256i)__builtin_lasx_xvsubwod_w_h((v16i16)_1, (v16i16)_2);
2963	}
2964
2965	extern __inline
2966	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2967	__lasx_xvsubwod_h_b(__m256i _1, __m256i _2) {
2968	return (__m256i)__builtin_lasx_xvsubwod_h_b((v32i8)_1, (v32i8)_2);
2969	}
2970
2971	extern __inline
2972	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2973	__lasx_xvsubwod_q_du(__m256i _1, __m256i _2) {
2974	return (__m256i)__builtin_lasx_xvsubwod_q_du((v4u64)_1, (v4u64)_2);
2975	}
2976
2977	extern __inline
2978	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2979	__lasx_xvsubwod_d_wu(__m256i _1, __m256i _2) {
2980	return (__m256i)__builtin_lasx_xvsubwod_d_wu((v8u32)_1, (v8u32)_2);
2981	}
2982
2983	extern __inline
2984	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2985	__lasx_xvsubwod_w_hu(__m256i _1, __m256i _2) {
2986	return (__m256i)__builtin_lasx_xvsubwod_w_hu((v16u16)_1, (v16u16)_2);
2987	}
2988
2989	extern __inline
2990	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2991	__lasx_xvsubwod_h_bu(__m256i _1, __m256i _2) {
2992	return (__m256i)__builtin_lasx_xvsubwod_h_bu((v32u8)_1, (v32u8)_2);
2993	}
2994
2995	extern __inline
2996	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
2997	__lasx_xvmulwod_q_d(__m256i _1, __m256i _2) {
2998	return (__m256i)__builtin_lasx_xvmulwod_q_d((v4i64)_1, (v4i64)_2);
2999	}
3000
3001	extern __inline
3002	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3003	__lasx_xvmulwod_d_w(__m256i _1, __m256i _2) {
3004	return (__m256i)__builtin_lasx_xvmulwod_d_w((v8i32)_1, (v8i32)_2);
3005	}
3006
3007	extern __inline
3008	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3009	__lasx_xvmulwod_w_h(__m256i _1, __m256i _2) {
3010	return (__m256i)__builtin_lasx_xvmulwod_w_h((v16i16)_1, (v16i16)_2);
3011	}
3012
3013	extern __inline
3014	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3015	__lasx_xvmulwod_h_b(__m256i _1, __m256i _2) {
3016	return (__m256i)__builtin_lasx_xvmulwod_h_b((v32i8)_1, (v32i8)_2);
3017	}
3018
3019	extern __inline
3020	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3021	__lasx_xvmulwod_q_du(__m256i _1, __m256i _2) {
3022	return (__m256i)__builtin_lasx_xvmulwod_q_du((v4u64)_1, (v4u64)_2);
3023	}
3024
3025	extern __inline
3026	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3027	__lasx_xvmulwod_d_wu(__m256i _1, __m256i _2) {
3028	return (__m256i)__builtin_lasx_xvmulwod_d_wu((v8u32)_1, (v8u32)_2);
3029	}
3030
3031	extern __inline
3032	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3033	__lasx_xvmulwod_w_hu(__m256i _1, __m256i _2) {
3034	return (__m256i)__builtin_lasx_xvmulwod_w_hu((v16u16)_1, (v16u16)_2);
3035	}
3036
3037	extern __inline
3038	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3039	__lasx_xvmulwod_h_bu(__m256i _1, __m256i _2) {
3040	return (__m256i)__builtin_lasx_xvmulwod_h_bu((v32u8)_1, (v32u8)_2);
3041	}
3042
3043	extern __inline
3044	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3045	__lasx_xvaddwev_d_wu_w(__m256i _1, __m256i _2) {
3046	return (__m256i)__builtin_lasx_xvaddwev_d_wu_w((v8u32)_1, (v8i32)_2);
3047	}
3048
3049	extern __inline
3050	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3051	__lasx_xvaddwev_w_hu_h(__m256i _1, __m256i _2) {
3052	return (__m256i)__builtin_lasx_xvaddwev_w_hu_h((v16u16)_1, (v16i16)_2);
3053	}
3054
3055	extern __inline
3056	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3057	__lasx_xvaddwev_h_bu_b(__m256i _1, __m256i _2) {
3058	return (__m256i)__builtin_lasx_xvaddwev_h_bu_b((v32u8)_1, (v32i8)_2);
3059	}
3060
3061	extern __inline
3062	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3063	__lasx_xvmulwev_d_wu_w(__m256i _1, __m256i _2) {
3064	return (__m256i)__builtin_lasx_xvmulwev_d_wu_w((v8u32)_1, (v8i32)_2);
3065	}
3066
3067	extern __inline
3068	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3069	__lasx_xvmulwev_w_hu_h(__m256i _1, __m256i _2) {
3070	return (__m256i)__builtin_lasx_xvmulwev_w_hu_h((v16u16)_1, (v16i16)_2);
3071	}
3072
3073	extern __inline
3074	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3075	__lasx_xvmulwev_h_bu_b(__m256i _1, __m256i _2) {
3076	return (__m256i)__builtin_lasx_xvmulwev_h_bu_b((v32u8)_1, (v32i8)_2);
3077	}
3078
3079	extern __inline
3080	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3081	__lasx_xvaddwod_d_wu_w(__m256i _1, __m256i _2) {
3082	return (__m256i)__builtin_lasx_xvaddwod_d_wu_w((v8u32)_1, (v8i32)_2);
3083	}
3084
3085	extern __inline
3086	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3087	__lasx_xvaddwod_w_hu_h(__m256i _1, __m256i _2) {
3088	return (__m256i)__builtin_lasx_xvaddwod_w_hu_h((v16u16)_1, (v16i16)_2);
3089	}
3090
3091	extern __inline
3092	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3093	__lasx_xvaddwod_h_bu_b(__m256i _1, __m256i _2) {
3094	return (__m256i)__builtin_lasx_xvaddwod_h_bu_b((v32u8)_1, (v32i8)_2);
3095	}
3096
3097	extern __inline
3098	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3099	__lasx_xvmulwod_d_wu_w(__m256i _1, __m256i _2) {
3100	return (__m256i)__builtin_lasx_xvmulwod_d_wu_w((v8u32)_1, (v8i32)_2);
3101	}
3102
3103	extern __inline
3104	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3105	__lasx_xvmulwod_w_hu_h(__m256i _1, __m256i _2) {
3106	return (__m256i)__builtin_lasx_xvmulwod_w_hu_h((v16u16)_1, (v16i16)_2);
3107	}
3108
3109	extern __inline
3110	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3111	__lasx_xvmulwod_h_bu_b(__m256i _1, __m256i _2) {
3112	return (__m256i)__builtin_lasx_xvmulwod_h_bu_b((v32u8)_1, (v32i8)_2);
3113	}
3114
3115	extern __inline
3116	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3117	__lasx_xvhaddw_q_d(__m256i _1, __m256i _2) {
3118	return (__m256i)__builtin_lasx_xvhaddw_q_d((v4i64)_1, (v4i64)_2);
3119	}
3120
3121	extern __inline
3122	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3123	__lasx_xvhaddw_qu_du(__m256i _1, __m256i _2) {
3124	return (__m256i)__builtin_lasx_xvhaddw_qu_du((v4u64)_1, (v4u64)_2);
3125	}
3126
3127	extern __inline
3128	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3129	__lasx_xvhsubw_q_d(__m256i _1, __m256i _2) {
3130	return (__m256i)__builtin_lasx_xvhsubw_q_d((v4i64)_1, (v4i64)_2);
3131	}
3132
3133	extern __inline
3134	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3135	__lasx_xvhsubw_qu_du(__m256i _1, __m256i _2) {
3136	return (__m256i)__builtin_lasx_xvhsubw_qu_du((v4u64)_1, (v4u64)_2);
3137	}
3138
3139	extern __inline
3140	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3141	__lasx_xvmaddwev_q_d(__m256i _1, __m256i _2, __m256i _3) {
3142	return (__m256i)__builtin_lasx_xvmaddwev_q_d((v4i64)_1, (v4i64)_2, (v4i64)_3);
3143	}
3144
3145	extern __inline
3146	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3147	__lasx_xvmaddwev_d_w(__m256i _1, __m256i _2, __m256i _3) {
3148	return (__m256i)__builtin_lasx_xvmaddwev_d_w((v4i64)_1, (v8i32)_2, (v8i32)_3);
3149	}
3150
3151	extern __inline
3152	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3153	__lasx_xvmaddwev_w_h(__m256i _1, __m256i _2, __m256i _3) {
3154	return (__m256i)__builtin_lasx_xvmaddwev_w_h((v8i32)_1, (v16i16)_2,
3155	(v16i16)_3);
3156	}
3157
3158	extern __inline
3159	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3160	__lasx_xvmaddwev_h_b(__m256i _1, __m256i _2, __m256i _3) {
3161	return (__m256i)__builtin_lasx_xvmaddwev_h_b((v16i16)_1, (v32i8)_2,
3162	(v32i8)_3);
3163	}
3164
3165	extern __inline
3166	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3167	__lasx_xvmaddwev_q_du(__m256i _1, __m256i _2, __m256i _3) {
3168	return (__m256i)__builtin_lasx_xvmaddwev_q_du((v4u64)_1, (v4u64)_2,
3169	(v4u64)_3);
3170	}
3171
3172	extern __inline
3173	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3174	__lasx_xvmaddwev_d_wu(__m256i _1, __m256i _2, __m256i _3) {
3175	return (__m256i)__builtin_lasx_xvmaddwev_d_wu((v4u64)_1, (v8u32)_2,
3176	(v8u32)_3);
3177	}
3178
3179	extern __inline
3180	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3181	__lasx_xvmaddwev_w_hu(__m256i _1, __m256i _2, __m256i _3) {
3182	return (__m256i)__builtin_lasx_xvmaddwev_w_hu((v8u32)_1, (v16u16)_2,
3183	(v16u16)_3);
3184	}
3185
3186	extern __inline
3187	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3188	__lasx_xvmaddwev_h_bu(__m256i _1, __m256i _2, __m256i _3) {
3189	return (__m256i)__builtin_lasx_xvmaddwev_h_bu((v16u16)_1, (v32u8)_2,
3190	(v32u8)_3);
3191	}
3192
3193	extern __inline
3194	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3195	__lasx_xvmaddwod_q_d(__m256i _1, __m256i _2, __m256i _3) {
3196	return (__m256i)__builtin_lasx_xvmaddwod_q_d((v4i64)_1, (v4i64)_2, (v4i64)_3);
3197	}
3198
3199	extern __inline
3200	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3201	__lasx_xvmaddwod_d_w(__m256i _1, __m256i _2, __m256i _3) {
3202	return (__m256i)__builtin_lasx_xvmaddwod_d_w((v4i64)_1, (v8i32)_2, (v8i32)_3);
3203	}
3204
3205	extern __inline
3206	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3207	__lasx_xvmaddwod_w_h(__m256i _1, __m256i _2, __m256i _3) {
3208	return (__m256i)__builtin_lasx_xvmaddwod_w_h((v8i32)_1, (v16i16)_2,
3209	(v16i16)_3);
3210	}
3211
3212	extern __inline
3213	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3214	__lasx_xvmaddwod_h_b(__m256i _1, __m256i _2, __m256i _3) {
3215	return (__m256i)__builtin_lasx_xvmaddwod_h_b((v16i16)_1, (v32i8)_2,
3216	(v32i8)_3);
3217	}
3218
3219	extern __inline
3220	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3221	__lasx_xvmaddwod_q_du(__m256i _1, __m256i _2, __m256i _3) {
3222	return (__m256i)__builtin_lasx_xvmaddwod_q_du((v4u64)_1, (v4u64)_2,
3223	(v4u64)_3);
3224	}
3225
3226	extern __inline
3227	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3228	__lasx_xvmaddwod_d_wu(__m256i _1, __m256i _2, __m256i _3) {
3229	return (__m256i)__builtin_lasx_xvmaddwod_d_wu((v4u64)_1, (v8u32)_2,
3230	(v8u32)_3);
3231	}
3232
3233	extern __inline
3234	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3235	__lasx_xvmaddwod_w_hu(__m256i _1, __m256i _2, __m256i _3) {
3236	return (__m256i)__builtin_lasx_xvmaddwod_w_hu((v8u32)_1, (v16u16)_2,
3237	(v16u16)_3);
3238	}
3239
3240	extern __inline
3241	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3242	__lasx_xvmaddwod_h_bu(__m256i _1, __m256i _2, __m256i _3) {
3243	return (__m256i)__builtin_lasx_xvmaddwod_h_bu((v16u16)_1, (v32u8)_2,
3244	(v32u8)_3);
3245	}
3246
3247	extern __inline
3248	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3249	__lasx_xvmaddwev_q_du_d(__m256i _1, __m256i _2, __m256i _3) {
3250	return (__m256i)__builtin_lasx_xvmaddwev_q_du_d((v4i64)_1, (v4u64)_2,
3251	(v4i64)_3);
3252	}
3253
3254	extern __inline
3255	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3256	__lasx_xvmaddwev_d_wu_w(__m256i _1, __m256i _2, __m256i _3) {
3257	return (__m256i)__builtin_lasx_xvmaddwev_d_wu_w((v4i64)_1, (v8u32)_2,
3258	(v8i32)_3);
3259	}
3260
3261	extern __inline
3262	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3263	__lasx_xvmaddwev_w_hu_h(__m256i _1, __m256i _2, __m256i _3) {
3264	return (__m256i)__builtin_lasx_xvmaddwev_w_hu_h((v8i32)_1, (v16u16)_2,
3265	(v16i16)_3);
3266	}
3267
3268	extern __inline
3269	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3270	__lasx_xvmaddwev_h_bu_b(__m256i _1, __m256i _2, __m256i _3) {
3271	return (__m256i)__builtin_lasx_xvmaddwev_h_bu_b((v16i16)_1, (v32u8)_2,
3272	(v32i8)_3);
3273	}
3274
3275	extern __inline
3276	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3277	__lasx_xvmaddwod_q_du_d(__m256i _1, __m256i _2, __m256i _3) {
3278	return (__m256i)__builtin_lasx_xvmaddwod_q_du_d((v4i64)_1, (v4u64)_2,
3279	(v4i64)_3);
3280	}
3281
3282	extern __inline
3283	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3284	__lasx_xvmaddwod_d_wu_w(__m256i _1, __m256i _2, __m256i _3) {
3285	return (__m256i)__builtin_lasx_xvmaddwod_d_wu_w((v4i64)_1, (v8u32)_2,
3286	(v8i32)_3);
3287	}
3288
3289	extern __inline
3290	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3291	__lasx_xvmaddwod_w_hu_h(__m256i _1, __m256i _2, __m256i _3) {
3292	return (__m256i)__builtin_lasx_xvmaddwod_w_hu_h((v8i32)_1, (v16u16)_2,
3293	(v16i16)_3);
3294	}
3295
3296	extern __inline
3297	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3298	__lasx_xvmaddwod_h_bu_b(__m256i _1, __m256i _2, __m256i _3) {
3299	return (__m256i)__builtin_lasx_xvmaddwod_h_bu_b((v16i16)_1, (v32u8)_2,
3300	(v32i8)_3);
3301	}
3302
3303	extern __inline
3304	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3305	__lasx_xvrotr_b(__m256i _1, __m256i _2) {
3306	return (__m256i)__builtin_lasx_xvrotr_b((v32i8)_1, (v32i8)_2);
3307	}
3308
3309	extern __inline
3310	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3311	__lasx_xvrotr_h(__m256i _1, __m256i _2) {
3312	return (__m256i)__builtin_lasx_xvrotr_h((v16i16)_1, (v16i16)_2);
3313	}
3314
3315	extern __inline
3316	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3317	__lasx_xvrotr_w(__m256i _1, __m256i _2) {
3318	return (__m256i)__builtin_lasx_xvrotr_w((v8i32)_1, (v8i32)_2);
3319	}
3320
3321	extern __inline
3322	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3323	__lasx_xvrotr_d(__m256i _1, __m256i _2) {
3324	return (__m256i)__builtin_lasx_xvrotr_d((v4i64)_1, (v4i64)_2);
3325	}
3326
3327	extern __inline
3328	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3329	__lasx_xvadd_q(__m256i _1, __m256i _2) {
3330	return (__m256i)__builtin_lasx_xvadd_q((v4i64)_1, (v4i64)_2);
3331	}
3332
3333	extern __inline
3334	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3335	__lasx_xvsub_q(__m256i _1, __m256i _2) {
3336	return (__m256i)__builtin_lasx_xvsub_q((v4i64)_1, (v4i64)_2);
3337	}
3338
3339	extern __inline
3340	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3341	__lasx_xvaddwev_q_du_d(__m256i _1, __m256i _2) {
3342	return (__m256i)__builtin_lasx_xvaddwev_q_du_d((v4u64)_1, (v4i64)_2);
3343	}
3344
3345	extern __inline
3346	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3347	__lasx_xvaddwod_q_du_d(__m256i _1, __m256i _2) {
3348	return (__m256i)__builtin_lasx_xvaddwod_q_du_d((v4u64)_1, (v4i64)_2);
3349	}
3350
3351	extern __inline
3352	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3353	__lasx_xvmulwev_q_du_d(__m256i _1, __m256i _2) {
3354	return (__m256i)__builtin_lasx_xvmulwev_q_du_d((v4u64)_1, (v4i64)_2);
3355	}
3356
3357	extern __inline
3358	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3359	__lasx_xvmulwod_q_du_d(__m256i _1, __m256i _2) {
3360	return (__m256i)__builtin_lasx_xvmulwod_q_du_d((v4u64)_1, (v4i64)_2);
3361	}
3362
3363	extern __inline
3364	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3365	__lasx_xvmskgez_b(__m256i _1) {
3366	return (__m256i)__builtin_lasx_xvmskgez_b((v32i8)_1);
3367	}
3368
3369	extern __inline
3370	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3371	__lasx_xvmsknz_b(__m256i _1) {
3372	return (__m256i)__builtin_lasx_xvmsknz_b((v32i8)_1);
3373	}
3374
3375	extern __inline
3376	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3377	__lasx_xvexth_h_b(__m256i _1) {
3378	return (__m256i)__builtin_lasx_xvexth_h_b((v32i8)_1);
3379	}
3380
3381	extern __inline
3382	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3383	__lasx_xvexth_w_h(__m256i _1) {
3384	return (__m256i)__builtin_lasx_xvexth_w_h((v16i16)_1);
3385	}
3386
3387	extern __inline
3388	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3389	__lasx_xvexth_d_w(__m256i _1) {
3390	return (__m256i)__builtin_lasx_xvexth_d_w((v8i32)_1);
3391	}
3392
3393	extern __inline
3394	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3395	__lasx_xvexth_q_d(__m256i _1) {
3396	return (__m256i)__builtin_lasx_xvexth_q_d((v4i64)_1);
3397	}
3398
3399	extern __inline
3400	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3401	__lasx_xvexth_hu_bu(__m256i _1) {
3402	return (__m256i)__builtin_lasx_xvexth_hu_bu((v32u8)_1);
3403	}
3404
3405	extern __inline
3406	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3407	__lasx_xvexth_wu_hu(__m256i _1) {
3408	return (__m256i)__builtin_lasx_xvexth_wu_hu((v16u16)_1);
3409	}
3410
3411	extern __inline
3412	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3413	__lasx_xvexth_du_wu(__m256i _1) {
3414	return (__m256i)__builtin_lasx_xvexth_du_wu((v8u32)_1);
3415	}
3416
3417	extern __inline
3418	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3419	__lasx_xvexth_qu_du(__m256i _1) {
3420	return (__m256i)__builtin_lasx_xvexth_qu_du((v4u64)_1);
3421	}
3422
3423	#define __lasx_xvrotri_b(/__m256i/ _1, /ui3/ _2) \
3424	((__m256i)__builtin_lasx_xvrotri_b((v32i8)(_1), (_2)))
3425
3426	#define __lasx_xvrotri_h(/__m256i/ _1, /ui4/ _2) \
3427	((__m256i)__builtin_lasx_xvrotri_h((v16i16)(_1), (_2)))
3428
3429	#define __lasx_xvrotri_w(/__m256i/ _1, /ui5/ _2) \
3430	((__m256i)__builtin_lasx_xvrotri_w((v8i32)(_1), (_2)))
3431
3432	#define __lasx_xvrotri_d(/__m256i/ _1, /ui6/ _2) \
3433	((__m256i)__builtin_lasx_xvrotri_d((v4i64)(_1), (_2)))
3434
3435	extern __inline
3436	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3437	__lasx_xvextl_q_d(__m256i _1) {
3438	return (__m256i)__builtin_lasx_xvextl_q_d((v4i64)_1);
3439	}
3440
3441	#define __lasx_xvsrlni_b_h(/__m256i/ _1, /__m256i/ _2, /ui4/ _3) \
3442	((__m256i)__builtin_lasx_xvsrlni_b_h((v32i8)(_1), (v32i8)(_2), (_3)))
3443
3444	#define __lasx_xvsrlni_h_w(/__m256i/ _1, /__m256i/ _2, /ui5/ _3) \
3445	((__m256i)__builtin_lasx_xvsrlni_h_w((v16i16)(_1), (v16i16)(_2), (_3)))
3446
3447	#define __lasx_xvsrlni_w_d(/__m256i/ _1, /__m256i/ _2, /ui6/ _3) \
3448	((__m256i)__builtin_lasx_xvsrlni_w_d((v8i32)(_1), (v8i32)(_2), (_3)))
3449
3450	#define __lasx_xvsrlni_d_q(/__m256i/ _1, /__m256i/ _2, /ui7/ _3) \
3451	((__m256i)__builtin_lasx_xvsrlni_d_q((v4i64)(_1), (v4i64)(_2), (_3)))
3452
3453	#define __lasx_xvsrlrni_b_h(/__m256i/ _1, /__m256i/ _2, /ui4/ _3) \
3454	((__m256i)__builtin_lasx_xvsrlrni_b_h((v32i8)(_1), (v32i8)(_2), (_3)))
3455
3456	#define __lasx_xvsrlrni_h_w(/__m256i/ _1, /__m256i/ _2, /ui5/ _3) \
3457	((__m256i)__builtin_lasx_xvsrlrni_h_w((v16i16)(_1), (v16i16)(_2), (_3)))
3458
3459	#define __lasx_xvsrlrni_w_d(/__m256i/ _1, /__m256i/ _2, /ui6/ _3) \
3460	((__m256i)__builtin_lasx_xvsrlrni_w_d((v8i32)(_1), (v8i32)(_2), (_3)))
3461
3462	#define __lasx_xvsrlrni_d_q(/__m256i/ _1, /__m256i/ _2, /ui7/ _3) \
3463	((__m256i)__builtin_lasx_xvsrlrni_d_q((v4i64)(_1), (v4i64)(_2), (_3)))
3464
3465	#define __lasx_xvssrlni_b_h(/__m256i/ _1, /__m256i/ _2, /ui4/ _3) \
3466	((__m256i)__builtin_lasx_xvssrlni_b_h((v32i8)(_1), (v32i8)(_2), (_3)))
3467
3468	#define __lasx_xvssrlni_h_w(/__m256i/ _1, /__m256i/ _2, /ui5/ _3) \
3469	((__m256i)__builtin_lasx_xvssrlni_h_w((v16i16)(_1), (v16i16)(_2), (_3)))
3470
3471	#define __lasx_xvssrlni_w_d(/__m256i/ _1, /__m256i/ _2, /ui6/ _3) \
3472	((__m256i)__builtin_lasx_xvssrlni_w_d((v8i32)(_1), (v8i32)(_2), (_3)))
3473
3474	#define __lasx_xvssrlni_d_q(/__m256i/ _1, /__m256i/ _2, /ui7/ _3) \
3475	((__m256i)__builtin_lasx_xvssrlni_d_q((v4i64)(_1), (v4i64)(_2), (_3)))
3476
3477	#define __lasx_xvssrlni_bu_h(/__m256i/ _1, /__m256i/ _2, /ui4/ _3) \
3478	((__m256i)__builtin_lasx_xvssrlni_bu_h((v32u8)(_1), (v32i8)(_2), (_3)))
3479
3480	#define __lasx_xvssrlni_hu_w(/__m256i/ _1, /__m256i/ _2, /ui5/ _3) \
3481	((__m256i)__builtin_lasx_xvssrlni_hu_w((v16u16)(_1), (v16i16)(_2), (_3)))
3482
3483	#define __lasx_xvssrlni_wu_d(/__m256i/ _1, /__m256i/ _2, /ui6/ _3) \
3484	((__m256i)__builtin_lasx_xvssrlni_wu_d((v8u32)(_1), (v8i32)(_2), (_3)))
3485
3486	#define __lasx_xvssrlni_du_q(/__m256i/ _1, /__m256i/ _2, /ui7/ _3) \
3487	((__m256i)__builtin_lasx_xvssrlni_du_q((v4u64)(_1), (v4i64)(_2), (_3)))
3488
3489	#define __lasx_xvssrlrni_b_h(/__m256i/ _1, /__m256i/ _2, /ui4/ _3) \
3490	((__m256i)__builtin_lasx_xvssrlrni_b_h((v32i8)(_1), (v32i8)(_2), (_3)))
3491
3492	#define __lasx_xvssrlrni_h_w(/__m256i/ _1, /__m256i/ _2, /ui5/ _3) \
3493	((__m256i)__builtin_lasx_xvssrlrni_h_w((v16i16)(_1), (v16i16)(_2), (_3)))
3494
3495	#define __lasx_xvssrlrni_w_d(/__m256i/ _1, /__m256i/ _2, /ui6/ _3) \
3496	((__m256i)__builtin_lasx_xvssrlrni_w_d((v8i32)(_1), (v8i32)(_2), (_3)))
3497
3498	#define __lasx_xvssrlrni_d_q(/__m256i/ _1, /__m256i/ _2, /ui7/ _3) \
3499	((__m256i)__builtin_lasx_xvssrlrni_d_q((v4i64)(_1), (v4i64)(_2), (_3)))
3500
3501	#define __lasx_xvssrlrni_bu_h(/__m256i/ _1, /__m256i/ _2, /ui4/ _3) \
3502	((__m256i)__builtin_lasx_xvssrlrni_bu_h((v32u8)(_1), (v32i8)(_2), (_3)))
3503
3504	#define __lasx_xvssrlrni_hu_w(/__m256i/ _1, /__m256i/ _2, /ui5/ _3) \
3505	((__m256i)__builtin_lasx_xvssrlrni_hu_w((v16u16)(_1), (v16i16)(_2), (_3)))
3506
3507	#define __lasx_xvssrlrni_wu_d(/__m256i/ _1, /__m256i/ _2, /ui6/ _3) \
3508	((__m256i)__builtin_lasx_xvssrlrni_wu_d((v8u32)(_1), (v8i32)(_2), (_3)))
3509
3510	#define __lasx_xvssrlrni_du_q(/__m256i/ _1, /__m256i/ _2, /ui7/ _3) \
3511	((__m256i)__builtin_lasx_xvssrlrni_du_q((v4u64)(_1), (v4i64)(_2), (_3)))
3512
3513	#define __lasx_xvsrani_b_h(/__m256i/ _1, /__m256i/ _2, /ui4/ _3) \
3514	((__m256i)__builtin_lasx_xvsrani_b_h((v32i8)(_1), (v32i8)(_2), (_3)))
3515
3516	#define __lasx_xvsrani_h_w(/__m256i/ _1, /__m256i/ _2, /ui5/ _3) \
3517	((__m256i)__builtin_lasx_xvsrani_h_w((v16i16)(_1), (v16i16)(_2), (_3)))
3518
3519	#define __lasx_xvsrani_w_d(/__m256i/ _1, /__m256i/ _2, /ui6/ _3) \
3520	((__m256i)__builtin_lasx_xvsrani_w_d((v8i32)(_1), (v8i32)(_2), (_3)))
3521
3522	#define __lasx_xvsrani_d_q(/__m256i/ _1, /__m256i/ _2, /ui7/ _3) \
3523	((__m256i)__builtin_lasx_xvsrani_d_q((v4i64)(_1), (v4i64)(_2), (_3)))
3524
3525	#define __lasx_xvsrarni_b_h(/__m256i/ _1, /__m256i/ _2, /ui4/ _3) \
3526	((__m256i)__builtin_lasx_xvsrarni_b_h((v32i8)(_1), (v32i8)(_2), (_3)))
3527
3528	#define __lasx_xvsrarni_h_w(/__m256i/ _1, /__m256i/ _2, /ui5/ _3) \
3529	((__m256i)__builtin_lasx_xvsrarni_h_w((v16i16)(_1), (v16i16)(_2), (_3)))
3530
3531	#define __lasx_xvsrarni_w_d(/__m256i/ _1, /__m256i/ _2, /ui6/ _3) \
3532	((__m256i)__builtin_lasx_xvsrarni_w_d((v8i32)(_1), (v8i32)(_2), (_3)))
3533
3534	#define __lasx_xvsrarni_d_q(/__m256i/ _1, /__m256i/ _2, /ui7/ _3) \
3535	((__m256i)__builtin_lasx_xvsrarni_d_q((v4i64)(_1), (v4i64)(_2), (_3)))
3536
3537	#define __lasx_xvssrani_b_h(/__m256i/ _1, /__m256i/ _2, /ui4/ _3) \
3538	((__m256i)__builtin_lasx_xvssrani_b_h((v32i8)(_1), (v32i8)(_2), (_3)))
3539
3540	#define __lasx_xvssrani_h_w(/__m256i/ _1, /__m256i/ _2, /ui5/ _3) \
3541	((__m256i)__builtin_lasx_xvssrani_h_w((v16i16)(_1), (v16i16)(_2), (_3)))
3542
3543	#define __lasx_xvssrani_w_d(/__m256i/ _1, /__m256i/ _2, /ui6/ _3) \
3544	((__m256i)__builtin_lasx_xvssrani_w_d((v8i32)(_1), (v8i32)(_2), (_3)))
3545
3546	#define __lasx_xvssrani_d_q(/__m256i/ _1, /__m256i/ _2, /ui7/ _3) \
3547	((__m256i)__builtin_lasx_xvssrani_d_q((v4i64)(_1), (v4i64)(_2), (_3)))
3548
3549	#define __lasx_xvssrani_bu_h(/__m256i/ _1, /__m256i/ _2, /ui4/ _3) \
3550	((__m256i)__builtin_lasx_xvssrani_bu_h((v32u8)(_1), (v32i8)(_2), (_3)))
3551
3552	#define __lasx_xvssrani_hu_w(/__m256i/ _1, /__m256i/ _2, /ui5/ _3) \
3553	((__m256i)__builtin_lasx_xvssrani_hu_w((v16u16)(_1), (v16i16)(_2), (_3)))
3554
3555	#define __lasx_xvssrani_wu_d(/__m256i/ _1, /__m256i/ _2, /ui6/ _3) \
3556	((__m256i)__builtin_lasx_xvssrani_wu_d((v8u32)(_1), (v8i32)(_2), (_3)))
3557
3558	#define __lasx_xvssrani_du_q(/__m256i/ _1, /__m256i/ _2, /ui7/ _3) \
3559	((__m256i)__builtin_lasx_xvssrani_du_q((v4u64)(_1), (v4i64)(_2), (_3)))
3560
3561	#define __lasx_xvssrarni_b_h(/__m256i/ _1, /__m256i/ _2, /ui4/ _3) \
3562	((__m256i)__builtin_lasx_xvssrarni_b_h((v32i8)(_1), (v32i8)(_2), (_3)))
3563
3564	#define __lasx_xvssrarni_h_w(/__m256i/ _1, /__m256i/ _2, /ui5/ _3) \
3565	((__m256i)__builtin_lasx_xvssrarni_h_w((v16i16)(_1), (v16i16)(_2), (_3)))
3566
3567	#define __lasx_xvssrarni_w_d(/__m256i/ _1, /__m256i/ _2, /ui6/ _3) \
3568	((__m256i)__builtin_lasx_xvssrarni_w_d((v8i32)(_1), (v8i32)(_2), (_3)))
3569
3570	#define __lasx_xvssrarni_d_q(/__m256i/ _1, /__m256i/ _2, /ui7/ _3) \
3571	((__m256i)__builtin_lasx_xvssrarni_d_q((v4i64)(_1), (v4i64)(_2), (_3)))
3572
3573	#define __lasx_xvssrarni_bu_h(/__m256i/ _1, /__m256i/ _2, /ui4/ _3) \
3574	((__m256i)__builtin_lasx_xvssrarni_bu_h((v32u8)(_1), (v32i8)(_2), (_3)))
3575
3576	#define __lasx_xvssrarni_hu_w(/__m256i/ _1, /__m256i/ _2, /ui5/ _3) \
3577	((__m256i)__builtin_lasx_xvssrarni_hu_w((v16u16)(_1), (v16i16)(_2), (_3)))
3578
3579	#define __lasx_xvssrarni_wu_d(/__m256i/ _1, /__m256i/ _2, /ui6/ _3) \
3580	((__m256i)__builtin_lasx_xvssrarni_wu_d((v8u32)(_1), (v8i32)(_2), (_3)))
3581
3582	#define __lasx_xvssrarni_du_q(/__m256i/ _1, /__m256i/ _2, /ui7/ _3) \
3583	((__m256i)__builtin_lasx_xvssrarni_du_q((v4u64)(_1), (v4i64)(_2), (_3)))
3584
3585	#define __lasx_xbnz_b(/__m256i/ _1) ((int)__builtin_lasx_xbnz_b((v32u8)(_1)))
3586
3587	#define __lasx_xbnz_d(/__m256i/ _1) ((int)__builtin_lasx_xbnz_d((v4u64)(_1)))
3588
3589	#define __lasx_xbnz_h(/__m256i/ _1) ((int)__builtin_lasx_xbnz_h((v16u16)(_1)))
3590
3591	#define __lasx_xbnz_v(/__m256i/ _1) ((int)__builtin_lasx_xbnz_v((v32u8)(_1)))
3592
3593	#define __lasx_xbnz_w(/__m256i/ _1) ((int)__builtin_lasx_xbnz_w((v8u32)(_1)))
3594
3595	#define __lasx_xbz_b(/__m256i/ _1) ((int)__builtin_lasx_xbz_b((v32u8)(_1)))
3596
3597	#define __lasx_xbz_d(/__m256i/ _1) ((int)__builtin_lasx_xbz_d((v4u64)(_1)))
3598
3599	#define __lasx_xbz_h(/__m256i/ _1) ((int)__builtin_lasx_xbz_h((v16u16)(_1)))
3600
3601	#define __lasx_xbz_v(/__m256i/ _1) ((int)__builtin_lasx_xbz_v((v32u8)(_1)))
3602
3603	#define __lasx_xbz_w(/__m256i/ _1) ((int)__builtin_lasx_xbz_w((v8u32)(_1)))
3604
3605	extern __inline
3606	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3607	__lasx_xvfcmp_caf_d(__m256d _1, __m256d _2) {
3608	return (__m256i)__builtin_lasx_xvfcmp_caf_d((v4f64)_1, (v4f64)_2);
3609	}
3610
3611	extern __inline
3612	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3613	__lasx_xvfcmp_caf_s(__m256 _1, __m256 _2) {
3614	return (__m256i)__builtin_lasx_xvfcmp_caf_s((v8f32)_1, (v8f32)_2);
3615	}
3616
3617	extern __inline
3618	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3619	__lasx_xvfcmp_ceq_d(__m256d _1, __m256d _2) {
3620	return (__m256i)__builtin_lasx_xvfcmp_ceq_d((v4f64)_1, (v4f64)_2);
3621	}
3622
3623	extern __inline
3624	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3625	__lasx_xvfcmp_ceq_s(__m256 _1, __m256 _2) {
3626	return (__m256i)__builtin_lasx_xvfcmp_ceq_s((v8f32)_1, (v8f32)_2);
3627	}
3628
3629	extern __inline
3630	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3631	__lasx_xvfcmp_cle_d(__m256d _1, __m256d _2) {
3632	return (__m256i)__builtin_lasx_xvfcmp_cle_d((v4f64)_1, (v4f64)_2);
3633	}
3634
3635	extern __inline
3636	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3637	__lasx_xvfcmp_cle_s(__m256 _1, __m256 _2) {
3638	return (__m256i)__builtin_lasx_xvfcmp_cle_s((v8f32)_1, (v8f32)_2);
3639	}
3640
3641	extern __inline
3642	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3643	__lasx_xvfcmp_clt_d(__m256d _1, __m256d _2) {
3644	return (__m256i)__builtin_lasx_xvfcmp_clt_d((v4f64)_1, (v4f64)_2);
3645	}
3646
3647	extern __inline
3648	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3649	__lasx_xvfcmp_clt_s(__m256 _1, __m256 _2) {
3650	return (__m256i)__builtin_lasx_xvfcmp_clt_s((v8f32)_1, (v8f32)_2);
3651	}
3652
3653	extern __inline
3654	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3655	__lasx_xvfcmp_cne_d(__m256d _1, __m256d _2) {
3656	return (__m256i)__builtin_lasx_xvfcmp_cne_d((v4f64)_1, (v4f64)_2);
3657	}
3658
3659	extern __inline
3660	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3661	__lasx_xvfcmp_cne_s(__m256 _1, __m256 _2) {
3662	return (__m256i)__builtin_lasx_xvfcmp_cne_s((v8f32)_1, (v8f32)_2);
3663	}
3664
3665	extern __inline
3666	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3667	__lasx_xvfcmp_cor_d(__m256d _1, __m256d _2) {
3668	return (__m256i)__builtin_lasx_xvfcmp_cor_d((v4f64)_1, (v4f64)_2);
3669	}
3670
3671	extern __inline
3672	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3673	__lasx_xvfcmp_cor_s(__m256 _1, __m256 _2) {
3674	return (__m256i)__builtin_lasx_xvfcmp_cor_s((v8f32)_1, (v8f32)_2);
3675	}
3676
3677	extern __inline
3678	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3679	__lasx_xvfcmp_cueq_d(__m256d _1, __m256d _2) {
3680	return (__m256i)__builtin_lasx_xvfcmp_cueq_d((v4f64)_1, (v4f64)_2);
3681	}
3682
3683	extern __inline
3684	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3685	__lasx_xvfcmp_cueq_s(__m256 _1, __m256 _2) {
3686	return (__m256i)__builtin_lasx_xvfcmp_cueq_s((v8f32)_1, (v8f32)_2);
3687	}
3688
3689	extern __inline
3690	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3691	__lasx_xvfcmp_cule_d(__m256d _1, __m256d _2) {
3692	return (__m256i)__builtin_lasx_xvfcmp_cule_d((v4f64)_1, (v4f64)_2);
3693	}
3694
3695	extern __inline
3696	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3697	__lasx_xvfcmp_cule_s(__m256 _1, __m256 _2) {
3698	return (__m256i)__builtin_lasx_xvfcmp_cule_s((v8f32)_1, (v8f32)_2);
3699	}
3700
3701	extern __inline
3702	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3703	__lasx_xvfcmp_cult_d(__m256d _1, __m256d _2) {
3704	return (__m256i)__builtin_lasx_xvfcmp_cult_d((v4f64)_1, (v4f64)_2);
3705	}
3706
3707	extern __inline
3708	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3709	__lasx_xvfcmp_cult_s(__m256 _1, __m256 _2) {
3710	return (__m256i)__builtin_lasx_xvfcmp_cult_s((v8f32)_1, (v8f32)_2);
3711	}
3712
3713	extern __inline
3714	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3715	__lasx_xvfcmp_cun_d(__m256d _1, __m256d _2) {
3716	return (__m256i)__builtin_lasx_xvfcmp_cun_d((v4f64)_1, (v4f64)_2);
3717	}
3718
3719	extern __inline
3720	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3721	__lasx_xvfcmp_cune_d(__m256d _1, __m256d _2) {
3722	return (__m256i)__builtin_lasx_xvfcmp_cune_d((v4f64)_1, (v4f64)_2);
3723	}
3724
3725	extern __inline
3726	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3727	__lasx_xvfcmp_cune_s(__m256 _1, __m256 _2) {
3728	return (__m256i)__builtin_lasx_xvfcmp_cune_s((v8f32)_1, (v8f32)_2);
3729	}
3730
3731	extern __inline
3732	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3733	__lasx_xvfcmp_cun_s(__m256 _1, __m256 _2) {
3734	return (__m256i)__builtin_lasx_xvfcmp_cun_s((v8f32)_1, (v8f32)_2);
3735	}
3736
3737	extern __inline
3738	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3739	__lasx_xvfcmp_saf_d(__m256d _1, __m256d _2) {
3740	return (__m256i)__builtin_lasx_xvfcmp_saf_d((v4f64)_1, (v4f64)_2);
3741	}
3742
3743	extern __inline
3744	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3745	__lasx_xvfcmp_saf_s(__m256 _1, __m256 _2) {
3746	return (__m256i)__builtin_lasx_xvfcmp_saf_s((v8f32)_1, (v8f32)_2);
3747	}
3748
3749	extern __inline
3750	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3751	__lasx_xvfcmp_seq_d(__m256d _1, __m256d _2) {
3752	return (__m256i)__builtin_lasx_xvfcmp_seq_d((v4f64)_1, (v4f64)_2);
3753	}
3754
3755	extern __inline
3756	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3757	__lasx_xvfcmp_seq_s(__m256 _1, __m256 _2) {
3758	return (__m256i)__builtin_lasx_xvfcmp_seq_s((v8f32)_1, (v8f32)_2);
3759	}
3760
3761	extern __inline
3762	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3763	__lasx_xvfcmp_sle_d(__m256d _1, __m256d _2) {
3764	return (__m256i)__builtin_lasx_xvfcmp_sle_d((v4f64)_1, (v4f64)_2);
3765	}
3766
3767	extern __inline
3768	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3769	__lasx_xvfcmp_sle_s(__m256 _1, __m256 _2) {
3770	return (__m256i)__builtin_lasx_xvfcmp_sle_s((v8f32)_1, (v8f32)_2);
3771	}
3772
3773	extern __inline
3774	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3775	__lasx_xvfcmp_slt_d(__m256d _1, __m256d _2) {
3776	return (__m256i)__builtin_lasx_xvfcmp_slt_d((v4f64)_1, (v4f64)_2);
3777	}
3778
3779	extern __inline
3780	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3781	__lasx_xvfcmp_slt_s(__m256 _1, __m256 _2) {
3782	return (__m256i)__builtin_lasx_xvfcmp_slt_s((v8f32)_1, (v8f32)_2);
3783	}
3784
3785	extern __inline
3786	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3787	__lasx_xvfcmp_sne_d(__m256d _1, __m256d _2) {
3788	return (__m256i)__builtin_lasx_xvfcmp_sne_d((v4f64)_1, (v4f64)_2);
3789	}
3790
3791	extern __inline
3792	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3793	__lasx_xvfcmp_sne_s(__m256 _1, __m256 _2) {
3794	return (__m256i)__builtin_lasx_xvfcmp_sne_s((v8f32)_1, (v8f32)_2);
3795	}
3796
3797	extern __inline
3798	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3799	__lasx_xvfcmp_sor_d(__m256d _1, __m256d _2) {
3800	return (__m256i)__builtin_lasx_xvfcmp_sor_d((v4f64)_1, (v4f64)_2);
3801	}
3802
3803	extern __inline
3804	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3805	__lasx_xvfcmp_sor_s(__m256 _1, __m256 _2) {
3806	return (__m256i)__builtin_lasx_xvfcmp_sor_s((v8f32)_1, (v8f32)_2);
3807	}
3808
3809	extern __inline
3810	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3811	__lasx_xvfcmp_sueq_d(__m256d _1, __m256d _2) {
3812	return (__m256i)__builtin_lasx_xvfcmp_sueq_d((v4f64)_1, (v4f64)_2);
3813	}
3814
3815	extern __inline
3816	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3817	__lasx_xvfcmp_sueq_s(__m256 _1, __m256 _2) {
3818	return (__m256i)__builtin_lasx_xvfcmp_sueq_s((v8f32)_1, (v8f32)_2);
3819	}
3820
3821	extern __inline
3822	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3823	__lasx_xvfcmp_sule_d(__m256d _1, __m256d _2) {
3824	return (__m256i)__builtin_lasx_xvfcmp_sule_d((v4f64)_1, (v4f64)_2);
3825	}
3826
3827	extern __inline
3828	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3829	__lasx_xvfcmp_sule_s(__m256 _1, __m256 _2) {
3830	return (__m256i)__builtin_lasx_xvfcmp_sule_s((v8f32)_1, (v8f32)_2);
3831	}
3832
3833	extern __inline
3834	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3835	__lasx_xvfcmp_sult_d(__m256d _1, __m256d _2) {
3836	return (__m256i)__builtin_lasx_xvfcmp_sult_d((v4f64)_1, (v4f64)_2);
3837	}
3838
3839	extern __inline
3840	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3841	__lasx_xvfcmp_sult_s(__m256 _1, __m256 _2) {
3842	return (__m256i)__builtin_lasx_xvfcmp_sult_s((v8f32)_1, (v8f32)_2);
3843	}
3844
3845	extern __inline
3846	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3847	__lasx_xvfcmp_sun_d(__m256d _1, __m256d _2) {
3848	return (__m256i)__builtin_lasx_xvfcmp_sun_d((v4f64)_1, (v4f64)_2);
3849	}
3850
3851	extern __inline
3852	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3853	__lasx_xvfcmp_sune_d(__m256d _1, __m256d _2) {
3854	return (__m256i)__builtin_lasx_xvfcmp_sune_d((v4f64)_1, (v4f64)_2);
3855	}
3856
3857	extern __inline
3858	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3859	__lasx_xvfcmp_sune_s(__m256 _1, __m256 _2) {
3860	return (__m256i)__builtin_lasx_xvfcmp_sune_s((v8f32)_1, (v8f32)_2);
3861	}
3862
3863	extern __inline
3864	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
3865	__lasx_xvfcmp_sun_s(__m256 _1, __m256 _2) {
3866	return (__m256i)__builtin_lasx_xvfcmp_sun_s((v8f32)_1, (v8f32)_2);
3867	}
3868
3869	#define __lasx_xvpickve_d_f(/__m256d/ _1, /ui2/ _2) \
3870	((__m256d)__builtin_lasx_xvpickve_d_f((v4f64)(_1), (_2)))
3871
3872	#define __lasx_xvpickve_w_f(/__m256/ _1, /ui3/ _2) \
3873	((__m256)__builtin_lasx_xvpickve_w_f((v8f32)(_1), (_2)))
3874
3875	#define __lasx_xvrepli_b(/si10/ _1) ((__m256i)__builtin_lasx_xvrepli_b((_1)))
3876
3877	#define __lasx_xvrepli_d(/si10/ _1) ((__m256i)__builtin_lasx_xvrepli_d((_1)))
3878
3879	#define __lasx_xvrepli_h(/si10/ _1) ((__m256i)__builtin_lasx_xvrepli_h((_1)))
3880
3881	#define __lasx_xvrepli_w(/si10/ _1) ((__m256i)__builtin_lasx_xvrepli_w((_1)))
3882
3883	#endif /* defined(__loongarch_asx). */
3884	#endif /* _LOONGSON_ASXINTRIN_H. */

lib/include/limits.h+2-4

 @@ -66,10 +66,8 @@
 #define CHAR_BIT  __CHAR_BIT__
-/* C2x 5.2.4.2.1 */
+/* C23 5.2.4.2.1 */
-/* FIXME: This is using the placeholder dates Clang produces for these macros
+#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
-   in C2x mode; switch to the correct values once they've been published. */
-#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
 #define BOOL_WIDTH   __BOOL_WIDTH__
 #define CHAR_WIDTH   CHAR_BIT
 #define SCHAR_WIDTH  CHAR_BIT

lib/include/llvm_libc_wrappers/assert.h created+34

...	@@ -0,0 +1,34 @@
1	//===-- Wrapper for C standard assert.h declarations on the GPU ------------===//
2	//
3	// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	// See https://llvm.org/LICENSE.txt for license information.
5	// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	//
7	//===----------------------------------------------------------------------===//
8
9	#ifndef __CLANG_LLVM_LIBC_WRAPPERS_ASSERT_H__
10	#define __CLANG_LLVM_LIBC_WRAPPERS_ASSERT_H__
11
12	#if !defined(_OPENMP) && !defined(__HIP__) && !defined(__CUDA__)
13	#error "This file is for GPU offloading compilation only"
14	#endif
15
16	#include_next <assert.h>
17
18	#if __has_include(<llvm-libc-decls/assert.h>)
19
20	#if defined(__HIP__) \|\| defined(__CUDA__)
21	#define __LIBC_ATTRS __attribute__((device))
22	#endif
23
24	#pragma omp begin declare target
25
26	#include <llvm-libc-decls/assert.h>
27
28	#pragma omp end declare target
29
30	#undef __LIBC_ATTRS
31
32	#endif
33
34	#endif // __CLANG_LLVM_LIBC_WRAPPERS_ASSERT_H__

lib/include/llvm_libc_wrappers/ctype.h+17

 @@ -13,8 +13,19 @@
 #error "This file is for GPU offloading compilation only"
 #endif
+// The GNU headers like to define 'toupper' and 'tolower' redundantly. This is
+// necessary to prevent it from doing that and remapping our implementation.
+#if (defined(__NVPTX__) || defined(__AMDGPU__)) && defined(__GLIBC__)
+#pragma push_macro("__USE_EXTERN_INLINES")
+#undef __USE_EXTERN_INLINES
+#endif
 #include_next <ctype.h>
+#if (defined(__NVPTX__) || defined(__AMDGPU__)) && defined(__GLIBC__)
+#pragma pop_macro("__USE_EXTERN_INLINES")
+#endif
 #if __has_include(<llvm-libc-decls/ctype.h>)
 #if defined(__HIP__) || defined(__CUDA__)
 @@ -26,6 +37,7 @@
 #pragma push_macro("isalnum")
 #pragma push_macro("isalpha")
+#pragma push_macro("isascii")
 #pragma push_macro("isblank")
 #pragma push_macro("iscntrl")
 #pragma push_macro("isdigit")
 @@ -36,11 +48,13 @@
 #pragma push_macro("isspace")
 #pragma push_macro("isupper")
 #pragma push_macro("isxdigit")
+#pragma push_macro("toascii")
 #pragma push_macro("tolower")
 #pragma push_macro("toupper")
 #undef isalnum
 #undef isalpha
+#undef isascii
 #undef iscntrl
 #undef isdigit
 #undef islower
 @@ -51,6 +65,7 @@
 #undef isupper
 #undef isblank
 #undef isxdigit
+#undef toascii
 #undef tolower
 #undef toupper
 @@ -64,6 +79,7 @@
 #if !defined(__NVPTX__) && !defined(__AMDGPU__)
 #pragma pop_macro("isalnum")
 #pragma pop_macro("isalpha")
+#pragma pop_macro("isascii")
 #pragma pop_macro("isblank")
 #pragma pop_macro("iscntrl")
 #pragma pop_macro("isdigit")
 @@ -74,6 +90,7 @@
 #pragma pop_macro("isspace")
 #pragma pop_macro("isupper")
 #pragma pop_macro("isxdigit")
+#pragma pop_macro("toascii")
 #pragma pop_macro("tolower")
 #pragma pop_macro("toupper")
 #endif

lib/include/llvm_libc_wrappers/stdio.h+49-3

...	@@ -6,21 +6,58 @@	...	@@ -6,21 +6,58 @@
6	//	6	//
7	//===----------------------------------------------------------------------===//	7	//===----------------------------------------------------------------------===//
8		8
9	#ifndef __CLANG_LLVM_LIBC_WRAPPERS_STDIO_H__
10	#define __CLANG_LLVM_LIBC_WRAPPERS_STDIO_H__
11
12	#if !defined(_OPENMP) && !defined(__HIP__) && !defined(__CUDA__)	9	#if !defined(_OPENMP) && !defined(__HIP__) && !defined(__CUDA__)
13	#error "This file is for GPU offloading compilation only"	10	#error "This file is for GPU offloading compilation only"
14	#endif	11	#endif
15		12
16	#include_next <stdio.h>	13	#include_next <stdio.h>
17		14
		15	// In some old versions of glibc, other standard headers sometimes define
		16	// special macros (e.g., __need_FILE) before including stdio.h to cause stdio.h
		17	// to produce special definitions. Future includes of stdio.h when those
		18	// special macros are undefined are expected to produce the normal definitions
		19	// from stdio.h.
		20	//
		21	// We do not apply our include guard (__CLANG_LLVM_LIBC_WRAPPERS_STDIO_H__)
		22	// unconditionally to the above include_next. Otherwise, after an occurrence of
		23	// the first glibc stdio.h use case described above, the include_next would be
		24	// skipped for remaining includes of stdio.h, leaving required symbols
		25	// undefined.
		26	//
		27	// We make the following assumptions to handle all use cases:
		28	//
		29	// 1. If the above include_next produces special glibc definitions, then (a) it
		30	// does not produce the normal definitions that we must intercept below, (b)
		31	// the current file was included from a glibc header that already defined
		32	// __GLIBC__ (usually by including glibc's <features.h>), and (c) the above
		33	// include_next does not define _STDIO_H. In that case, we skip the rest of
		34	// the current file and don't guard against future includes.
		35	// 2. If the above include_next produces the normal stdio.h definitions, then
		36	// either (a) __GLIBC__ is not defined because C headers are from some other
		37	// libc implementation or (b) the above include_next defines _STDIO_H to
		38	// prevent the above include_next from having any effect in the future.
		39	#if !defined(__GLIBC__) \|\| defined(_STDIO_H)
		40
		41	#ifndef __CLANG_LLVM_LIBC_WRAPPERS_STDIO_H__
		42	#define __CLANG_LLVM_LIBC_WRAPPERS_STDIO_H__
		43
18	#if __has_include(<llvm-libc-decls/stdio.h>)	44	#if __has_include(<llvm-libc-decls/stdio.h>)
19		45
20	#if defined(__HIP__) \|\| defined(__CUDA__)	46	#if defined(__HIP__) \|\| defined(__CUDA__)
21	#define __LIBC_ATTRS __attribute__((device))	47	#define __LIBC_ATTRS __attribute__((device))
22	#endif	48	#endif
23		49
		50	// Some headers provide these as macros. Temporarily undefine them so they do
		51	// not conflict with any definitions for the GPU.
		52
		53	#pragma push_macro("stdout")
		54	#pragma push_macro("stdin")
		55	#pragma push_macro("stderr")
		56
		57	#undef stdout
		58	#undef stderr
		59	#undef stdin
		60
24	#pragma omp begin declare target	61	#pragma omp begin declare target
25		62
26	#include <llvm-libc-decls/stdio.h>	63	#include <llvm-libc-decls/stdio.h>
...	@@ -29,6 +66,15 @@	...	@@ -29,6 +66,15 @@
29		66
30	#undef __LIBC_ATTRS	67	#undef __LIBC_ATTRS
31		68
		69	// Restore the original macros when compiling on the host.
		70	#if !defined(__NVPTX__) && !defined(__AMDGPU__)
		71	#pragma pop_macro("stdout")
		72	#pragma pop_macro("stderr")
		73	#pragma pop_macro("stdin")
		74	#endif
		75
32	#endif	76	#endif
33		77
34	#endif // __CLANG_LLVM_LIBC_WRAPPERS_STDIO_H__	78	#endif // __CLANG_LLVM_LIBC_WRAPPERS_STDIO_H__
		79
		80	#endif

lib/include/llvm_libc_wrappers/stdlib.h+4-1

 @@ -23,8 +23,11 @@
 #pragma omp begin declare target
-// The LLVM C library uses this type so we forward declare it.
+// The LLVM C library uses these named types so we forward declare them.
 typedef void (*__atexithandler_t)(void);
+typedef int (*__bsearchcompare_t)(const void *, const void *);
+typedef int (*__qsortcompare_t)(const void *, const void *);
+typedef int (*__qsortrcompare_t)(const void *, const void *, void *);
 // Enforce ABI compatibility with the structs used by the LLVM C library.
 _Static_assert(__builtin_offsetof(div_t, quot) == 0, "ABI mismatch!");

lib/include/llvm_libc_wrappers/string.h+62-3

...	@@ -13,9 +13,6 @@	...	@@ -13,9 +13,6 @@
13	#error "This file is for GPU offloading compilation only"	13	#error "This file is for GPU offloading compilation only"
14	#endif	14	#endif
15		15
16	// FIXME: The GNU headers provide C++ standard compliant headers when in C++
17	// mode and the LLVM libc does not. We cannot enable memchr, strchr, strchrnul,
18	// strpbrk, strrchr, strstr, or strcasestr until this is addressed.
19	#include_next <string.h>	16	#include_next <string.h>
20		17
21	#if __has_include(<llvm-libc-decls/string.h>)	18	#if __has_include(<llvm-libc-decls/string.h>)
...	@@ -26,8 +23,70 @@	...	@@ -26,8 +23,70 @@
26		23
27	#pragma omp begin declare target	24	#pragma omp begin declare target
28		25
		26	// The GNU headers provide C++ standard compliant headers when in C++ mode and
		27	// the LLVM libc does not. We need to manually provide the definitions using the
		28	// same prototypes.
		29	#if defined(__cplusplus) && defined(__GLIBC__) && \
		30	defined(__CORRECT_ISO_CPP_STRING_H_PROTO)
		31
		32	#ifndef __LIBC_ATTRS
		33	#define __LIBC_ATTRS
		34	#endif
		35
		36	extern "C" {
		37	void memccpy(void __restrict, const void *__restrict, int,
		38	size_t) __LIBC_ATTRS;
		39	int memcmp(const void , const void , size_t) __LIBC_ATTRS;
		40	void memcpy(void __restrict, const void *__restrict, size_t) __LIBC_ATTRS;
		41	void memmem(const void , size_t, const void *, size_t) __LIBC_ATTRS;
		42	void memmove(void , const void *, size_t) __LIBC_ATTRS;
		43	void mempcpy(void __restrict, const void *__restrict, size_t) __LIBC_ATTRS;
		44	void memset(void , int, size_t) __LIBC_ATTRS;
		45	char stpcpy(char __restrict, const char *__restrict) __LIBC_ATTRS;
		46	char stpncpy(char __restrict, const char *__restrict, size_t) __LIBC_ATTRS;
		47	char strcat(char __restrict, const char *__restrict) __LIBC_ATTRS;
		48	int strcmp(const char , const char ) __LIBC_ATTRS;
		49	int strcoll(const char , const char ) __LIBC_ATTRS;
		50	char strcpy(char __restrict, const char *__restrict) __LIBC_ATTRS;
		51	size_t strcspn(const char , const char ) __LIBC_ATTRS;
		52	char strdup(const char ) __LIBC_ATTRS;
		53	size_t strlen(const char *) __LIBC_ATTRS;
		54	char strncat(char __restrict, const char *__restrict, size_t) __LIBC_ATTRS;
		55	int strncmp(const char , const char , size_t) __LIBC_ATTRS;
		56	char strncpy(char __restrict, const char *__restrict, size_t) __LIBC_ATTRS;
		57	char strndup(const char , size_t) __LIBC_ATTRS;
		58	size_t strnlen(const char *, size_t) __LIBC_ATTRS;
		59	size_t strspn(const char , const char ) __LIBC_ATTRS;
		60	char strtok(char __restrict, const char *__restrict) __LIBC_ATTRS;
		61	char strtok_r(char __restrict, const char *__restrict,
		62	char **__restrict) __LIBC_ATTRS;
		63	size_t strxfrm(char __restrict, const char __restrict, size_t) __LIBC_ATTRS;
		64	}
		65
		66	extern "C++" {
		67	char strstr(char , const char *) noexcept __LIBC_ATTRS;
		68	const char strstr(const char , const char *) noexcept __LIBC_ATTRS;
		69	char strpbrk(char , const char *) noexcept __LIBC_ATTRS;
		70	const char strpbrk(const char , const char *) noexcept __LIBC_ATTRS;
		71	char strrchr(char , int) noexcept __LIBC_ATTRS;
		72	const char strrchr(const char , int) noexcept __LIBC_ATTRS;
		73	char strchr(char , int) noexcept __LIBC_ATTRS;
		74	const char strchr(const char , int) noexcept __LIBC_ATTRS;
		75	char strchrnul(char , int) noexcept __LIBC_ATTRS;
		76	const char strchrnul(const char , int) noexcept __LIBC_ATTRS;
		77	char strcasestr(char , const char *) noexcept __LIBC_ATTRS;
		78	const char strcasestr(const char , const char *) noexcept __LIBC_ATTRS;
		79	void memrchr(void __s, int __c, size_t __n) noexcept __LIBC_ATTRS;
		80	const void memrchr(const void __s, int __c, size_t __n) noexcept __LIBC_ATTRS;
		81	void memchr(void __s, int __c, size_t __n) noexcept __LIBC_ATTRS;
		82	const void memchr(const void __s, int __c, size_t __n) noexcept __LIBC_ATTRS;
		83	}
		84
		85	#else
29	#include <llvm-libc-decls/string.h>	86	#include <llvm-libc-decls/string.h>
30		87
		88	#endif
		89
31	#pragma omp end declare target	90	#pragma omp end declare target
32		91
33	#undef __LIBC_ATTRS	92	#undef __LIBC_ATTRS

lib/include/llvm_libc_wrappers/time.h created+34

...	@@ -0,0 +1,34 @@
1	//===-- Wrapper for C standard time.h declarations on the GPU -------------===//
2	//
3	// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	// See https://llvm.org/LICENSE.txt for license information.
5	// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	//
7	//===----------------------------------------------------------------------===//
8
9	#ifndef __CLANG_LLVM_LIBC_WRAPPERS_TIME_H__
10	#define __CLANG_LLVM_LIBC_WRAPPERS_TIME_H__
11
12	#if !defined(_OPENMP) && !defined(__HIP__) && !defined(__CUDA__)
13	#error "This file is for GPU offloading compilation only"
14	#endif
15
16	#include_next <time.h>
17
18	#if __has_include(<llvm-libc-decls/time.h>)
19
20	#if defined(__HIP__) \|\| defined(__CUDA__)
21	#define __LIBC_ATTRS __attribute__((device))
22	#endif
23
24	#pragma omp begin declare target
25
26	_Static_assert(sizeof(clock_t) == sizeof(long), "ABI mismatch!");
27
28	#include <llvm-libc-decls/time.h>
29
30	#pragma omp end declare target
31
32	#endif
33
34	#endif // __CLANG_LLVM_LIBC_WRAPPERS_TIME_H__

lib/include/lsxintrin.h created+3750

...	@@ -0,0 +1,3750 @@
1	/*===------------- lsxintrin.h - LoongArch LSX intrinsics ------------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	#ifndef _LOONGSON_SXINTRIN_H
11	#define _LOONGSON_SXINTRIN_H 1
12
13	#if defined(__loongarch_sx)
14	typedef signed char v16i8 __attribute__((vector_size(16), aligned(16)));
15	typedef signed char v16i8_b __attribute__((vector_size(16), aligned(1)));
16	typedef unsigned char v16u8 __attribute__((vector_size(16), aligned(16)));
17	typedef unsigned char v16u8_b __attribute__((vector_size(16), aligned(1)));
18	typedef short v8i16 __attribute__((vector_size(16), aligned(16)));
19	typedef short v8i16_h __attribute__((vector_size(16), aligned(2)));
20	typedef unsigned short v8u16 __attribute__((vector_size(16), aligned(16)));
21	typedef unsigned short v8u16_h __attribute__((vector_size(16), aligned(2)));
22	typedef int v4i32 __attribute__((vector_size(16), aligned(16)));
23	typedef int v4i32_w __attribute__((vector_size(16), aligned(4)));
24	typedef unsigned int v4u32 __attribute__((vector_size(16), aligned(16)));
25	typedef unsigned int v4u32_w __attribute__((vector_size(16), aligned(4)));
26	typedef long long v2i64 __attribute__((vector_size(16), aligned(16)));
27	typedef long long v2i64_d __attribute__((vector_size(16), aligned(8)));
28	typedef unsigned long long v2u64 __attribute__((vector_size(16), aligned(16)));
29	typedef unsigned long long v2u64_d __attribute__((vector_size(16), aligned(8)));
30	typedef float v4f32 __attribute__((vector_size(16), aligned(16)));
31	typedef float v4f32_w __attribute__((vector_size(16), aligned(4)));
32	typedef double v2f64 __attribute__((vector_size(16), aligned(16)));
33	typedef double v2f64_d __attribute__((vector_size(16), aligned(8)));
34
35	typedef long long __m128i __attribute__((__vector_size__(16), __may_alias__));
36	typedef float __m128 __attribute__((__vector_size__(16), __may_alias__));
37	typedef double __m128d __attribute__((__vector_size__(16), __may_alias__));
38
39	extern __inline
40	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
41	__lsx_vsll_b(__m128i _1, __m128i _2) {
42	return (__m128i)__builtin_lsx_vsll_b((v16i8)_1, (v16i8)_2);
43	}
44
45	extern __inline
46	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
47	__lsx_vsll_h(__m128i _1, __m128i _2) {
48	return (__m128i)__builtin_lsx_vsll_h((v8i16)_1, (v8i16)_2);
49	}
50
51	extern __inline
52	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
53	__lsx_vsll_w(__m128i _1, __m128i _2) {
54	return (__m128i)__builtin_lsx_vsll_w((v4i32)_1, (v4i32)_2);
55	}
56
57	extern __inline
58	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
59	__lsx_vsll_d(__m128i _1, __m128i _2) {
60	return (__m128i)__builtin_lsx_vsll_d((v2i64)_1, (v2i64)_2);
61	}
62
63	#define __lsx_vslli_b(/__m128i/ _1, /ui3/ _2) \
64	((__m128i)__builtin_lsx_vslli_b((v16i8)(_1), (_2)))
65
66	#define __lsx_vslli_h(/__m128i/ _1, /ui4/ _2) \
67	((__m128i)__builtin_lsx_vslli_h((v8i16)(_1), (_2)))
68
69	#define __lsx_vslli_w(/__m128i/ _1, /ui5/ _2) \
70	((__m128i)__builtin_lsx_vslli_w((v4i32)(_1), (_2)))
71
72	#define __lsx_vslli_d(/__m128i/ _1, /ui6/ _2) \
73	((__m128i)__builtin_lsx_vslli_d((v2i64)(_1), (_2)))
74
75	extern __inline
76	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
77	__lsx_vsra_b(__m128i _1, __m128i _2) {
78	return (__m128i)__builtin_lsx_vsra_b((v16i8)_1, (v16i8)_2);
79	}
80
81	extern __inline
82	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
83	__lsx_vsra_h(__m128i _1, __m128i _2) {
84	return (__m128i)__builtin_lsx_vsra_h((v8i16)_1, (v8i16)_2);
85	}
86
87	extern __inline
88	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
89	__lsx_vsra_w(__m128i _1, __m128i _2) {
90	return (__m128i)__builtin_lsx_vsra_w((v4i32)_1, (v4i32)_2);
91	}
92
93	extern __inline
94	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
95	__lsx_vsra_d(__m128i _1, __m128i _2) {
96	return (__m128i)__builtin_lsx_vsra_d((v2i64)_1, (v2i64)_2);
97	}
98
99	#define __lsx_vsrai_b(/__m128i/ _1, /ui3/ _2) \
100	((__m128i)__builtin_lsx_vsrai_b((v16i8)(_1), (_2)))
101
102	#define __lsx_vsrai_h(/__m128i/ _1, /ui4/ _2) \
103	((__m128i)__builtin_lsx_vsrai_h((v8i16)(_1), (_2)))
104
105	#define __lsx_vsrai_w(/__m128i/ _1, /ui5/ _2) \
106	((__m128i)__builtin_lsx_vsrai_w((v4i32)(_1), (_2)))
107
108	#define __lsx_vsrai_d(/__m128i/ _1, /ui6/ _2) \
109	((__m128i)__builtin_lsx_vsrai_d((v2i64)(_1), (_2)))
110
111	extern __inline
112	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
113	__lsx_vsrar_b(__m128i _1, __m128i _2) {
114	return (__m128i)__builtin_lsx_vsrar_b((v16i8)_1, (v16i8)_2);
115	}
116
117	extern __inline
118	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
119	__lsx_vsrar_h(__m128i _1, __m128i _2) {
120	return (__m128i)__builtin_lsx_vsrar_h((v8i16)_1, (v8i16)_2);
121	}
122
123	extern __inline
124	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
125	__lsx_vsrar_w(__m128i _1, __m128i _2) {
126	return (__m128i)__builtin_lsx_vsrar_w((v4i32)_1, (v4i32)_2);
127	}
128
129	extern __inline
130	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
131	__lsx_vsrar_d(__m128i _1, __m128i _2) {
132	return (__m128i)__builtin_lsx_vsrar_d((v2i64)_1, (v2i64)_2);
133	}
134
135	#define __lsx_vsrari_b(/__m128i/ _1, /ui3/ _2) \
136	((__m128i)__builtin_lsx_vsrari_b((v16i8)(_1), (_2)))
137
138	#define __lsx_vsrari_h(/__m128i/ _1, /ui4/ _2) \
139	((__m128i)__builtin_lsx_vsrari_h((v8i16)(_1), (_2)))
140
141	#define __lsx_vsrari_w(/__m128i/ _1, /ui5/ _2) \
142	((__m128i)__builtin_lsx_vsrari_w((v4i32)(_1), (_2)))
143
144	#define __lsx_vsrari_d(/__m128i/ _1, /ui6/ _2) \
145	((__m128i)__builtin_lsx_vsrari_d((v2i64)(_1), (_2)))
146
147	extern __inline
148	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
149	__lsx_vsrl_b(__m128i _1, __m128i _2) {
150	return (__m128i)__builtin_lsx_vsrl_b((v16i8)_1, (v16i8)_2);
151	}
152
153	extern __inline
154	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
155	__lsx_vsrl_h(__m128i _1, __m128i _2) {
156	return (__m128i)__builtin_lsx_vsrl_h((v8i16)_1, (v8i16)_2);
157	}
158
159	extern __inline
160	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
161	__lsx_vsrl_w(__m128i _1, __m128i _2) {
162	return (__m128i)__builtin_lsx_vsrl_w((v4i32)_1, (v4i32)_2);
163	}
164
165	extern __inline
166	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
167	__lsx_vsrl_d(__m128i _1, __m128i _2) {
168	return (__m128i)__builtin_lsx_vsrl_d((v2i64)_1, (v2i64)_2);
169	}
170
171	#define __lsx_vsrli_b(/__m128i/ _1, /ui3/ _2) \
172	((__m128i)__builtin_lsx_vsrli_b((v16i8)(_1), (_2)))
173
174	#define __lsx_vsrli_h(/__m128i/ _1, /ui4/ _2) \
175	((__m128i)__builtin_lsx_vsrli_h((v8i16)(_1), (_2)))
176
177	#define __lsx_vsrli_w(/__m128i/ _1, /ui5/ _2) \
178	((__m128i)__builtin_lsx_vsrli_w((v4i32)(_1), (_2)))
179
180	#define __lsx_vsrli_d(/__m128i/ _1, /ui6/ _2) \
181	((__m128i)__builtin_lsx_vsrli_d((v2i64)(_1), (_2)))
182
183	extern __inline
184	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
185	__lsx_vsrlr_b(__m128i _1, __m128i _2) {
186	return (__m128i)__builtin_lsx_vsrlr_b((v16i8)_1, (v16i8)_2);
187	}
188
189	extern __inline
190	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
191	__lsx_vsrlr_h(__m128i _1, __m128i _2) {
192	return (__m128i)__builtin_lsx_vsrlr_h((v8i16)_1, (v8i16)_2);
193	}
194
195	extern __inline
196	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
197	__lsx_vsrlr_w(__m128i _1, __m128i _2) {
198	return (__m128i)__builtin_lsx_vsrlr_w((v4i32)_1, (v4i32)_2);
199	}
200
201	extern __inline
202	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
203	__lsx_vsrlr_d(__m128i _1, __m128i _2) {
204	return (__m128i)__builtin_lsx_vsrlr_d((v2i64)_1, (v2i64)_2);
205	}
206
207	#define __lsx_vsrlri_b(/__m128i/ _1, /ui3/ _2) \
208	((__m128i)__builtin_lsx_vsrlri_b((v16i8)(_1), (_2)))
209
210	#define __lsx_vsrlri_h(/__m128i/ _1, /ui4/ _2) \
211	((__m128i)__builtin_lsx_vsrlri_h((v8i16)(_1), (_2)))
212
213	#define __lsx_vsrlri_w(/__m128i/ _1, /ui5/ _2) \
214	((__m128i)__builtin_lsx_vsrlri_w((v4i32)(_1), (_2)))
215
216	#define __lsx_vsrlri_d(/__m128i/ _1, /ui6/ _2) \
217	((__m128i)__builtin_lsx_vsrlri_d((v2i64)(_1), (_2)))
218
219	extern __inline
220	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
221	__lsx_vbitclr_b(__m128i _1, __m128i _2) {
222	return (__m128i)__builtin_lsx_vbitclr_b((v16u8)_1, (v16u8)_2);
223	}
224
225	extern __inline
226	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
227	__lsx_vbitclr_h(__m128i _1, __m128i _2) {
228	return (__m128i)__builtin_lsx_vbitclr_h((v8u16)_1, (v8u16)_2);
229	}
230
231	extern __inline
232	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
233	__lsx_vbitclr_w(__m128i _1, __m128i _2) {
234	return (__m128i)__builtin_lsx_vbitclr_w((v4u32)_1, (v4u32)_2);
235	}
236
237	extern __inline
238	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
239	__lsx_vbitclr_d(__m128i _1, __m128i _2) {
240	return (__m128i)__builtin_lsx_vbitclr_d((v2u64)_1, (v2u64)_2);
241	}
242
243	#define __lsx_vbitclri_b(/__m128i/ _1, /ui3/ _2) \
244	((__m128i)__builtin_lsx_vbitclri_b((v16u8)(_1), (_2)))
245
246	#define __lsx_vbitclri_h(/__m128i/ _1, /ui4/ _2) \
247	((__m128i)__builtin_lsx_vbitclri_h((v8u16)(_1), (_2)))
248
249	#define __lsx_vbitclri_w(/__m128i/ _1, /ui5/ _2) \
250	((__m128i)__builtin_lsx_vbitclri_w((v4u32)(_1), (_2)))
251
252	#define __lsx_vbitclri_d(/__m128i/ _1, /ui6/ _2) \
253	((__m128i)__builtin_lsx_vbitclri_d((v2u64)(_1), (_2)))
254
255	extern __inline
256	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
257	__lsx_vbitset_b(__m128i _1, __m128i _2) {
258	return (__m128i)__builtin_lsx_vbitset_b((v16u8)_1, (v16u8)_2);
259	}
260
261	extern __inline
262	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
263	__lsx_vbitset_h(__m128i _1, __m128i _2) {
264	return (__m128i)__builtin_lsx_vbitset_h((v8u16)_1, (v8u16)_2);
265	}
266
267	extern __inline
268	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
269	__lsx_vbitset_w(__m128i _1, __m128i _2) {
270	return (__m128i)__builtin_lsx_vbitset_w((v4u32)_1, (v4u32)_2);
271	}
272
273	extern __inline
274	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
275	__lsx_vbitset_d(__m128i _1, __m128i _2) {
276	return (__m128i)__builtin_lsx_vbitset_d((v2u64)_1, (v2u64)_2);
277	}
278
279	#define __lsx_vbitseti_b(/__m128i/ _1, /ui3/ _2) \
280	((__m128i)__builtin_lsx_vbitseti_b((v16u8)(_1), (_2)))
281
282	#define __lsx_vbitseti_h(/__m128i/ _1, /ui4/ _2) \
283	((__m128i)__builtin_lsx_vbitseti_h((v8u16)(_1), (_2)))
284
285	#define __lsx_vbitseti_w(/__m128i/ _1, /ui5/ _2) \
286	((__m128i)__builtin_lsx_vbitseti_w((v4u32)(_1), (_2)))
287
288	#define __lsx_vbitseti_d(/__m128i/ _1, /ui6/ _2) \
289	((__m128i)__builtin_lsx_vbitseti_d((v2u64)(_1), (_2)))
290
291	extern __inline
292	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
293	__lsx_vbitrev_b(__m128i _1, __m128i _2) {
294	return (__m128i)__builtin_lsx_vbitrev_b((v16u8)_1, (v16u8)_2);
295	}
296
297	extern __inline
298	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
299	__lsx_vbitrev_h(__m128i _1, __m128i _2) {
300	return (__m128i)__builtin_lsx_vbitrev_h((v8u16)_1, (v8u16)_2);
301	}
302
303	extern __inline
304	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
305	__lsx_vbitrev_w(__m128i _1, __m128i _2) {
306	return (__m128i)__builtin_lsx_vbitrev_w((v4u32)_1, (v4u32)_2);
307	}
308
309	extern __inline
310	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
311	__lsx_vbitrev_d(__m128i _1, __m128i _2) {
312	return (__m128i)__builtin_lsx_vbitrev_d((v2u64)_1, (v2u64)_2);
313	}
314
315	#define __lsx_vbitrevi_b(/__m128i/ _1, /ui3/ _2) \
316	((__m128i)__builtin_lsx_vbitrevi_b((v16u8)(_1), (_2)))
317
318	#define __lsx_vbitrevi_h(/__m128i/ _1, /ui4/ _2) \
319	((__m128i)__builtin_lsx_vbitrevi_h((v8u16)(_1), (_2)))
320
321	#define __lsx_vbitrevi_w(/__m128i/ _1, /ui5/ _2) \
322	((__m128i)__builtin_lsx_vbitrevi_w((v4u32)(_1), (_2)))
323
324	#define __lsx_vbitrevi_d(/__m128i/ _1, /ui6/ _2) \
325	((__m128i)__builtin_lsx_vbitrevi_d((v2u64)(_1), (_2)))
326
327	extern __inline
328	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
329	__lsx_vadd_b(__m128i _1, __m128i _2) {
330	return (__m128i)__builtin_lsx_vadd_b((v16i8)_1, (v16i8)_2);
331	}
332
333	extern __inline
334	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
335	__lsx_vadd_h(__m128i _1, __m128i _2) {
336	return (__m128i)__builtin_lsx_vadd_h((v8i16)_1, (v8i16)_2);
337	}
338
339	extern __inline
340	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
341	__lsx_vadd_w(__m128i _1, __m128i _2) {
342	return (__m128i)__builtin_lsx_vadd_w((v4i32)_1, (v4i32)_2);
343	}
344
345	extern __inline
346	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
347	__lsx_vadd_d(__m128i _1, __m128i _2) {
348	return (__m128i)__builtin_lsx_vadd_d((v2i64)_1, (v2i64)_2);
349	}
350
351	#define __lsx_vaddi_bu(/__m128i/ _1, /ui5/ _2) \
352	((__m128i)__builtin_lsx_vaddi_bu((v16i8)(_1), (_2)))
353
354	#define __lsx_vaddi_hu(/__m128i/ _1, /ui5/ _2) \
355	((__m128i)__builtin_lsx_vaddi_hu((v8i16)(_1), (_2)))
356
357	#define __lsx_vaddi_wu(/__m128i/ _1, /ui5/ _2) \
358	((__m128i)__builtin_lsx_vaddi_wu((v4i32)(_1), (_2)))
359
360	#define __lsx_vaddi_du(/__m128i/ _1, /ui5/ _2) \
361	((__m128i)__builtin_lsx_vaddi_du((v2i64)(_1), (_2)))
362
363	extern __inline
364	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
365	__lsx_vsub_b(__m128i _1, __m128i _2) {
366	return (__m128i)__builtin_lsx_vsub_b((v16i8)_1, (v16i8)_2);
367	}
368
369	extern __inline
370	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
371	__lsx_vsub_h(__m128i _1, __m128i _2) {
372	return (__m128i)__builtin_lsx_vsub_h((v8i16)_1, (v8i16)_2);
373	}
374
375	extern __inline
376	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
377	__lsx_vsub_w(__m128i _1, __m128i _2) {
378	return (__m128i)__builtin_lsx_vsub_w((v4i32)_1, (v4i32)_2);
379	}
380
381	extern __inline
382	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
383	__lsx_vsub_d(__m128i _1, __m128i _2) {
384	return (__m128i)__builtin_lsx_vsub_d((v2i64)_1, (v2i64)_2);
385	}
386
387	#define __lsx_vsubi_bu(/__m128i/ _1, /ui5/ _2) \
388	((__m128i)__builtin_lsx_vsubi_bu((v16i8)(_1), (_2)))
389
390	#define __lsx_vsubi_hu(/__m128i/ _1, /ui5/ _2) \
391	((__m128i)__builtin_lsx_vsubi_hu((v8i16)(_1), (_2)))
392
393	#define __lsx_vsubi_wu(/__m128i/ _1, /ui5/ _2) \
394	((__m128i)__builtin_lsx_vsubi_wu((v4i32)(_1), (_2)))
395
396	#define __lsx_vsubi_du(/__m128i/ _1, /ui5/ _2) \
397	((__m128i)__builtin_lsx_vsubi_du((v2i64)(_1), (_2)))
398
399	extern __inline
400	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
401	__lsx_vmax_b(__m128i _1, __m128i _2) {
402	return (__m128i)__builtin_lsx_vmax_b((v16i8)_1, (v16i8)_2);
403	}
404
405	extern __inline
406	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
407	__lsx_vmax_h(__m128i _1, __m128i _2) {
408	return (__m128i)__builtin_lsx_vmax_h((v8i16)_1, (v8i16)_2);
409	}
410
411	extern __inline
412	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
413	__lsx_vmax_w(__m128i _1, __m128i _2) {
414	return (__m128i)__builtin_lsx_vmax_w((v4i32)_1, (v4i32)_2);
415	}
416
417	extern __inline
418	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
419	__lsx_vmax_d(__m128i _1, __m128i _2) {
420	return (__m128i)__builtin_lsx_vmax_d((v2i64)_1, (v2i64)_2);
421	}
422
423	#define __lsx_vmaxi_b(/__m128i/ _1, /si5/ _2) \
424	((__m128i)__builtin_lsx_vmaxi_b((v16i8)(_1), (_2)))
425
426	#define __lsx_vmaxi_h(/__m128i/ _1, /si5/ _2) \
427	((__m128i)__builtin_lsx_vmaxi_h((v8i16)(_1), (_2)))
428
429	#define __lsx_vmaxi_w(/__m128i/ _1, /si5/ _2) \
430	((__m128i)__builtin_lsx_vmaxi_w((v4i32)(_1), (_2)))
431
432	#define __lsx_vmaxi_d(/__m128i/ _1, /si5/ _2) \
433	((__m128i)__builtin_lsx_vmaxi_d((v2i64)(_1), (_2)))
434
435	extern __inline
436	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
437	__lsx_vmax_bu(__m128i _1, __m128i _2) {
438	return (__m128i)__builtin_lsx_vmax_bu((v16u8)_1, (v16u8)_2);
439	}
440
441	extern __inline
442	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
443	__lsx_vmax_hu(__m128i _1, __m128i _2) {
444	return (__m128i)__builtin_lsx_vmax_hu((v8u16)_1, (v8u16)_2);
445	}
446
447	extern __inline
448	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
449	__lsx_vmax_wu(__m128i _1, __m128i _2) {
450	return (__m128i)__builtin_lsx_vmax_wu((v4u32)_1, (v4u32)_2);
451	}
452
453	extern __inline
454	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
455	__lsx_vmax_du(__m128i _1, __m128i _2) {
456	return (__m128i)__builtin_lsx_vmax_du((v2u64)_1, (v2u64)_2);
457	}
458
459	#define __lsx_vmaxi_bu(/__m128i/ _1, /ui5/ _2) \
460	((__m128i)__builtin_lsx_vmaxi_bu((v16u8)(_1), (_2)))
461
462	#define __lsx_vmaxi_hu(/__m128i/ _1, /ui5/ _2) \
463	((__m128i)__builtin_lsx_vmaxi_hu((v8u16)(_1), (_2)))
464
465	#define __lsx_vmaxi_wu(/__m128i/ _1, /ui5/ _2) \
466	((__m128i)__builtin_lsx_vmaxi_wu((v4u32)(_1), (_2)))
467
468	#define __lsx_vmaxi_du(/__m128i/ _1, /ui5/ _2) \
469	((__m128i)__builtin_lsx_vmaxi_du((v2u64)(_1), (_2)))
470
471	extern __inline
472	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
473	__lsx_vmin_b(__m128i _1, __m128i _2) {
474	return (__m128i)__builtin_lsx_vmin_b((v16i8)_1, (v16i8)_2);
475	}
476
477	extern __inline
478	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
479	__lsx_vmin_h(__m128i _1, __m128i _2) {
480	return (__m128i)__builtin_lsx_vmin_h((v8i16)_1, (v8i16)_2);
481	}
482
483	extern __inline
484	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
485	__lsx_vmin_w(__m128i _1, __m128i _2) {
486	return (__m128i)__builtin_lsx_vmin_w((v4i32)_1, (v4i32)_2);
487	}
488
489	extern __inline
490	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
491	__lsx_vmin_d(__m128i _1, __m128i _2) {
492	return (__m128i)__builtin_lsx_vmin_d((v2i64)_1, (v2i64)_2);
493	}
494
495	#define __lsx_vmini_b(/__m128i/ _1, /si5/ _2) \
496	((__m128i)__builtin_lsx_vmini_b((v16i8)(_1), (_2)))
497
498	#define __lsx_vmini_h(/__m128i/ _1, /si5/ _2) \
499	((__m128i)__builtin_lsx_vmini_h((v8i16)(_1), (_2)))
500
501	#define __lsx_vmini_w(/__m128i/ _1, /si5/ _2) \
502	((__m128i)__builtin_lsx_vmini_w((v4i32)(_1), (_2)))
503
504	#define __lsx_vmini_d(/__m128i/ _1, /si5/ _2) \
505	((__m128i)__builtin_lsx_vmini_d((v2i64)(_1), (_2)))
506
507	extern __inline
508	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
509	__lsx_vmin_bu(__m128i _1, __m128i _2) {
510	return (__m128i)__builtin_lsx_vmin_bu((v16u8)_1, (v16u8)_2);
511	}
512
513	extern __inline
514	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
515	__lsx_vmin_hu(__m128i _1, __m128i _2) {
516	return (__m128i)__builtin_lsx_vmin_hu((v8u16)_1, (v8u16)_2);
517	}
518
519	extern __inline
520	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
521	__lsx_vmin_wu(__m128i _1, __m128i _2) {
522	return (__m128i)__builtin_lsx_vmin_wu((v4u32)_1, (v4u32)_2);
523	}
524
525	extern __inline
526	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
527	__lsx_vmin_du(__m128i _1, __m128i _2) {
528	return (__m128i)__builtin_lsx_vmin_du((v2u64)_1, (v2u64)_2);
529	}
530
531	#define __lsx_vmini_bu(/__m128i/ _1, /ui5/ _2) \
532	((__m128i)__builtin_lsx_vmini_bu((v16u8)(_1), (_2)))
533
534	#define __lsx_vmini_hu(/__m128i/ _1, /ui5/ _2) \
535	((__m128i)__builtin_lsx_vmini_hu((v8u16)(_1), (_2)))
536
537	#define __lsx_vmini_wu(/__m128i/ _1, /ui5/ _2) \
538	((__m128i)__builtin_lsx_vmini_wu((v4u32)(_1), (_2)))
539
540	#define __lsx_vmini_du(/__m128i/ _1, /ui5/ _2) \
541	((__m128i)__builtin_lsx_vmini_du((v2u64)(_1), (_2)))
542
543	extern __inline
544	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
545	__lsx_vseq_b(__m128i _1, __m128i _2) {
546	return (__m128i)__builtin_lsx_vseq_b((v16i8)_1, (v16i8)_2);
547	}
548
549	extern __inline
550	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
551	__lsx_vseq_h(__m128i _1, __m128i _2) {
552	return (__m128i)__builtin_lsx_vseq_h((v8i16)_1, (v8i16)_2);
553	}
554
555	extern __inline
556	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
557	__lsx_vseq_w(__m128i _1, __m128i _2) {
558	return (__m128i)__builtin_lsx_vseq_w((v4i32)_1, (v4i32)_2);
559	}
560
561	extern __inline
562	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
563	__lsx_vseq_d(__m128i _1, __m128i _2) {
564	return (__m128i)__builtin_lsx_vseq_d((v2i64)_1, (v2i64)_2);
565	}
566
567	#define __lsx_vseqi_b(/__m128i/ _1, /si5/ _2) \
568	((__m128i)__builtin_lsx_vseqi_b((v16i8)(_1), (_2)))
569
570	#define __lsx_vseqi_h(/__m128i/ _1, /si5/ _2) \
571	((__m128i)__builtin_lsx_vseqi_h((v8i16)(_1), (_2)))
572
573	#define __lsx_vseqi_w(/__m128i/ _1, /si5/ _2) \
574	((__m128i)__builtin_lsx_vseqi_w((v4i32)(_1), (_2)))
575
576	#define __lsx_vseqi_d(/__m128i/ _1, /si5/ _2) \
577	((__m128i)__builtin_lsx_vseqi_d((v2i64)(_1), (_2)))
578
579	#define __lsx_vslti_b(/__m128i/ _1, /si5/ _2) \
580	((__m128i)__builtin_lsx_vslti_b((v16i8)(_1), (_2)))
581
582	extern __inline
583	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
584	__lsx_vslt_b(__m128i _1, __m128i _2) {
585	return (__m128i)__builtin_lsx_vslt_b((v16i8)_1, (v16i8)_2);
586	}
587
588	extern __inline
589	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
590	__lsx_vslt_h(__m128i _1, __m128i _2) {
591	return (__m128i)__builtin_lsx_vslt_h((v8i16)_1, (v8i16)_2);
592	}
593
594	extern __inline
595	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
596	__lsx_vslt_w(__m128i _1, __m128i _2) {
597	return (__m128i)__builtin_lsx_vslt_w((v4i32)_1, (v4i32)_2);
598	}
599
600	extern __inline
601	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
602	__lsx_vslt_d(__m128i _1, __m128i _2) {
603	return (__m128i)__builtin_lsx_vslt_d((v2i64)_1, (v2i64)_2);
604	}
605
606	#define __lsx_vslti_h(/__m128i/ _1, /si5/ _2) \
607	((__m128i)__builtin_lsx_vslti_h((v8i16)(_1), (_2)))
608
609	#define __lsx_vslti_w(/__m128i/ _1, /si5/ _2) \
610	((__m128i)__builtin_lsx_vslti_w((v4i32)(_1), (_2)))
611
612	#define __lsx_vslti_d(/__m128i/ _1, /si5/ _2) \
613	((__m128i)__builtin_lsx_vslti_d((v2i64)(_1), (_2)))
614
615	extern __inline
616	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
617	__lsx_vslt_bu(__m128i _1, __m128i _2) {
618	return (__m128i)__builtin_lsx_vslt_bu((v16u8)_1, (v16u8)_2);
619	}
620
621	extern __inline
622	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
623	__lsx_vslt_hu(__m128i _1, __m128i _2) {
624	return (__m128i)__builtin_lsx_vslt_hu((v8u16)_1, (v8u16)_2);
625	}
626
627	extern __inline
628	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
629	__lsx_vslt_wu(__m128i _1, __m128i _2) {
630	return (__m128i)__builtin_lsx_vslt_wu((v4u32)_1, (v4u32)_2);
631	}
632
633	extern __inline
634	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
635	__lsx_vslt_du(__m128i _1, __m128i _2) {
636	return (__m128i)__builtin_lsx_vslt_du((v2u64)_1, (v2u64)_2);
637	}
638
639	#define __lsx_vslti_bu(/__m128i/ _1, /ui5/ _2) \
640	((__m128i)__builtin_lsx_vslti_bu((v16u8)(_1), (_2)))
641
642	#define __lsx_vslti_hu(/__m128i/ _1, /ui5/ _2) \
643	((__m128i)__builtin_lsx_vslti_hu((v8u16)(_1), (_2)))
644
645	#define __lsx_vslti_wu(/__m128i/ _1, /ui5/ _2) \
646	((__m128i)__builtin_lsx_vslti_wu((v4u32)(_1), (_2)))
647
648	#define __lsx_vslti_du(/__m128i/ _1, /ui5/ _2) \
649	((__m128i)__builtin_lsx_vslti_du((v2u64)(_1), (_2)))
650
651	extern __inline
652	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
653	__lsx_vsle_b(__m128i _1, __m128i _2) {
654	return (__m128i)__builtin_lsx_vsle_b((v16i8)_1, (v16i8)_2);
655	}
656
657	extern __inline
658	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
659	__lsx_vsle_h(__m128i _1, __m128i _2) {
660	return (__m128i)__builtin_lsx_vsle_h((v8i16)_1, (v8i16)_2);
661	}
662
663	extern __inline
664	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
665	__lsx_vsle_w(__m128i _1, __m128i _2) {
666	return (__m128i)__builtin_lsx_vsle_w((v4i32)_1, (v4i32)_2);
667	}
668
669	extern __inline
670	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
671	__lsx_vsle_d(__m128i _1, __m128i _2) {
672	return (__m128i)__builtin_lsx_vsle_d((v2i64)_1, (v2i64)_2);
673	}
674
675	#define __lsx_vslei_b(/__m128i/ _1, /si5/ _2) \
676	((__m128i)__builtin_lsx_vslei_b((v16i8)(_1), (_2)))
677
678	#define __lsx_vslei_h(/__m128i/ _1, /si5/ _2) \
679	((__m128i)__builtin_lsx_vslei_h((v8i16)(_1), (_2)))
680
681	#define __lsx_vslei_w(/__m128i/ _1, /si5/ _2) \
682	((__m128i)__builtin_lsx_vslei_w((v4i32)(_1), (_2)))
683
684	#define __lsx_vslei_d(/__m128i/ _1, /si5/ _2) \
685	((__m128i)__builtin_lsx_vslei_d((v2i64)(_1), (_2)))
686
687	extern __inline
688	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
689	__lsx_vsle_bu(__m128i _1, __m128i _2) {
690	return (__m128i)__builtin_lsx_vsle_bu((v16u8)_1, (v16u8)_2);
691	}
692
693	extern __inline
694	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
695	__lsx_vsle_hu(__m128i _1, __m128i _2) {
696	return (__m128i)__builtin_lsx_vsle_hu((v8u16)_1, (v8u16)_2);
697	}
698
699	extern __inline
700	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
701	__lsx_vsle_wu(__m128i _1, __m128i _2) {
702	return (__m128i)__builtin_lsx_vsle_wu((v4u32)_1, (v4u32)_2);
703	}
704
705	extern __inline
706	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
707	__lsx_vsle_du(__m128i _1, __m128i _2) {
708	return (__m128i)__builtin_lsx_vsle_du((v2u64)_1, (v2u64)_2);
709	}
710
711	#define __lsx_vslei_bu(/__m128i/ _1, /ui5/ _2) \
712	((__m128i)__builtin_lsx_vslei_bu((v16u8)(_1), (_2)))
713
714	#define __lsx_vslei_hu(/__m128i/ _1, /ui5/ _2) \
715	((__m128i)__builtin_lsx_vslei_hu((v8u16)(_1), (_2)))
716
717	#define __lsx_vslei_wu(/__m128i/ _1, /ui5/ _2) \
718	((__m128i)__builtin_lsx_vslei_wu((v4u32)(_1), (_2)))
719
720	#define __lsx_vslei_du(/__m128i/ _1, /ui5/ _2) \
721	((__m128i)__builtin_lsx_vslei_du((v2u64)(_1), (_2)))
722
723	#define __lsx_vsat_b(/__m128i/ _1, /ui3/ _2) \
724	((__m128i)__builtin_lsx_vsat_b((v16i8)(_1), (_2)))
725
726	#define __lsx_vsat_h(/__m128i/ _1, /ui4/ _2) \
727	((__m128i)__builtin_lsx_vsat_h((v8i16)(_1), (_2)))
728
729	#define __lsx_vsat_w(/__m128i/ _1, /ui5/ _2) \
730	((__m128i)__builtin_lsx_vsat_w((v4i32)(_1), (_2)))
731
732	#define __lsx_vsat_d(/__m128i/ _1, /ui6/ _2) \
733	((__m128i)__builtin_lsx_vsat_d((v2i64)(_1), (_2)))
734
735	#define __lsx_vsat_bu(/__m128i/ _1, /ui3/ _2) \
736	((__m128i)__builtin_lsx_vsat_bu((v16u8)(_1), (_2)))
737
738	#define __lsx_vsat_hu(/__m128i/ _1, /ui4/ _2) \
739	((__m128i)__builtin_lsx_vsat_hu((v8u16)(_1), (_2)))
740
741	#define __lsx_vsat_wu(/__m128i/ _1, /ui5/ _2) \
742	((__m128i)__builtin_lsx_vsat_wu((v4u32)(_1), (_2)))
743
744	#define __lsx_vsat_du(/__m128i/ _1, /ui6/ _2) \
745	((__m128i)__builtin_lsx_vsat_du((v2u64)(_1), (_2)))
746
747	extern __inline
748	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
749	__lsx_vadda_b(__m128i _1, __m128i _2) {
750	return (__m128i)__builtin_lsx_vadda_b((v16i8)_1, (v16i8)_2);
751	}
752
753	extern __inline
754	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
755	__lsx_vadda_h(__m128i _1, __m128i _2) {
756	return (__m128i)__builtin_lsx_vadda_h((v8i16)_1, (v8i16)_2);
757	}
758
759	extern __inline
760	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
761	__lsx_vadda_w(__m128i _1, __m128i _2) {
762	return (__m128i)__builtin_lsx_vadda_w((v4i32)_1, (v4i32)_2);
763	}
764
765	extern __inline
766	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
767	__lsx_vadda_d(__m128i _1, __m128i _2) {
768	return (__m128i)__builtin_lsx_vadda_d((v2i64)_1, (v2i64)_2);
769	}
770
771	extern __inline
772	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
773	__lsx_vsadd_b(__m128i _1, __m128i _2) {
774	return (__m128i)__builtin_lsx_vsadd_b((v16i8)_1, (v16i8)_2);
775	}
776
777	extern __inline
778	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
779	__lsx_vsadd_h(__m128i _1, __m128i _2) {
780	return (__m128i)__builtin_lsx_vsadd_h((v8i16)_1, (v8i16)_2);
781	}
782
783	extern __inline
784	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
785	__lsx_vsadd_w(__m128i _1, __m128i _2) {
786	return (__m128i)__builtin_lsx_vsadd_w((v4i32)_1, (v4i32)_2);
787	}
788
789	extern __inline
790	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
791	__lsx_vsadd_d(__m128i _1, __m128i _2) {
792	return (__m128i)__builtin_lsx_vsadd_d((v2i64)_1, (v2i64)_2);
793	}
794
795	extern __inline
796	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
797	__lsx_vsadd_bu(__m128i _1, __m128i _2) {
798	return (__m128i)__builtin_lsx_vsadd_bu((v16u8)_1, (v16u8)_2);
799	}
800
801	extern __inline
802	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
803	__lsx_vsadd_hu(__m128i _1, __m128i _2) {
804	return (__m128i)__builtin_lsx_vsadd_hu((v8u16)_1, (v8u16)_2);
805	}
806
807	extern __inline
808	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
809	__lsx_vsadd_wu(__m128i _1, __m128i _2) {
810	return (__m128i)__builtin_lsx_vsadd_wu((v4u32)_1, (v4u32)_2);
811	}
812
813	extern __inline
814	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
815	__lsx_vsadd_du(__m128i _1, __m128i _2) {
816	return (__m128i)__builtin_lsx_vsadd_du((v2u64)_1, (v2u64)_2);
817	}
818
819	extern __inline
820	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
821	__lsx_vavg_b(__m128i _1, __m128i _2) {
822	return (__m128i)__builtin_lsx_vavg_b((v16i8)_1, (v16i8)_2);
823	}
824
825	extern __inline
826	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
827	__lsx_vavg_h(__m128i _1, __m128i _2) {
828	return (__m128i)__builtin_lsx_vavg_h((v8i16)_1, (v8i16)_2);
829	}
830
831	extern __inline
832	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
833	__lsx_vavg_w(__m128i _1, __m128i _2) {
834	return (__m128i)__builtin_lsx_vavg_w((v4i32)_1, (v4i32)_2);
835	}
836
837	extern __inline
838	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
839	__lsx_vavg_d(__m128i _1, __m128i _2) {
840	return (__m128i)__builtin_lsx_vavg_d((v2i64)_1, (v2i64)_2);
841	}
842
843	extern __inline
844	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
845	__lsx_vavg_bu(__m128i _1, __m128i _2) {
846	return (__m128i)__builtin_lsx_vavg_bu((v16u8)_1, (v16u8)_2);
847	}
848
849	extern __inline
850	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
851	__lsx_vavg_hu(__m128i _1, __m128i _2) {
852	return (__m128i)__builtin_lsx_vavg_hu((v8u16)_1, (v8u16)_2);
853	}
854
855	extern __inline
856	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
857	__lsx_vavg_wu(__m128i _1, __m128i _2) {
858	return (__m128i)__builtin_lsx_vavg_wu((v4u32)_1, (v4u32)_2);
859	}
860
861	extern __inline
862	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
863	__lsx_vavg_du(__m128i _1, __m128i _2) {
864	return (__m128i)__builtin_lsx_vavg_du((v2u64)_1, (v2u64)_2);
865	}
866
867	extern __inline
868	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
869	__lsx_vavgr_b(__m128i _1, __m128i _2) {
870	return (__m128i)__builtin_lsx_vavgr_b((v16i8)_1, (v16i8)_2);
871	}
872
873	extern __inline
874	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
875	__lsx_vavgr_h(__m128i _1, __m128i _2) {
876	return (__m128i)__builtin_lsx_vavgr_h((v8i16)_1, (v8i16)_2);
877	}
878
879	extern __inline
880	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
881	__lsx_vavgr_w(__m128i _1, __m128i _2) {
882	return (__m128i)__builtin_lsx_vavgr_w((v4i32)_1, (v4i32)_2);
883	}
884
885	extern __inline
886	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
887	__lsx_vavgr_d(__m128i _1, __m128i _2) {
888	return (__m128i)__builtin_lsx_vavgr_d((v2i64)_1, (v2i64)_2);
889	}
890
891	extern __inline
892	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
893	__lsx_vavgr_bu(__m128i _1, __m128i _2) {
894	return (__m128i)__builtin_lsx_vavgr_bu((v16u8)_1, (v16u8)_2);
895	}
896
897	extern __inline
898	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
899	__lsx_vavgr_hu(__m128i _1, __m128i _2) {
900	return (__m128i)__builtin_lsx_vavgr_hu((v8u16)_1, (v8u16)_2);
901	}
902
903	extern __inline
904	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
905	__lsx_vavgr_wu(__m128i _1, __m128i _2) {
906	return (__m128i)__builtin_lsx_vavgr_wu((v4u32)_1, (v4u32)_2);
907	}
908
909	extern __inline
910	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
911	__lsx_vavgr_du(__m128i _1, __m128i _2) {
912	return (__m128i)__builtin_lsx_vavgr_du((v2u64)_1, (v2u64)_2);
913	}
914
915	extern __inline
916	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
917	__lsx_vssub_b(__m128i _1, __m128i _2) {
918	return (__m128i)__builtin_lsx_vssub_b((v16i8)_1, (v16i8)_2);
919	}
920
921	extern __inline
922	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
923	__lsx_vssub_h(__m128i _1, __m128i _2) {
924	return (__m128i)__builtin_lsx_vssub_h((v8i16)_1, (v8i16)_2);
925	}
926
927	extern __inline
928	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
929	__lsx_vssub_w(__m128i _1, __m128i _2) {
930	return (__m128i)__builtin_lsx_vssub_w((v4i32)_1, (v4i32)_2);
931	}
932
933	extern __inline
934	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
935	__lsx_vssub_d(__m128i _1, __m128i _2) {
936	return (__m128i)__builtin_lsx_vssub_d((v2i64)_1, (v2i64)_2);
937	}
938
939	extern __inline
940	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
941	__lsx_vssub_bu(__m128i _1, __m128i _2) {
942	return (__m128i)__builtin_lsx_vssub_bu((v16u8)_1, (v16u8)_2);
943	}
944
945	extern __inline
946	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
947	__lsx_vssub_hu(__m128i _1, __m128i _2) {
948	return (__m128i)__builtin_lsx_vssub_hu((v8u16)_1, (v8u16)_2);
949	}
950
951	extern __inline
952	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
953	__lsx_vssub_wu(__m128i _1, __m128i _2) {
954	return (__m128i)__builtin_lsx_vssub_wu((v4u32)_1, (v4u32)_2);
955	}
956
957	extern __inline
958	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
959	__lsx_vssub_du(__m128i _1, __m128i _2) {
960	return (__m128i)__builtin_lsx_vssub_du((v2u64)_1, (v2u64)_2);
961	}
962
963	extern __inline
964	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
965	__lsx_vabsd_b(__m128i _1, __m128i _2) {
966	return (__m128i)__builtin_lsx_vabsd_b((v16i8)_1, (v16i8)_2);
967	}
968
969	extern __inline
970	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
971	__lsx_vabsd_h(__m128i _1, __m128i _2) {
972	return (__m128i)__builtin_lsx_vabsd_h((v8i16)_1, (v8i16)_2);
973	}
974
975	extern __inline
976	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
977	__lsx_vabsd_w(__m128i _1, __m128i _2) {
978	return (__m128i)__builtin_lsx_vabsd_w((v4i32)_1, (v4i32)_2);
979	}
980
981	extern __inline
982	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
983	__lsx_vabsd_d(__m128i _1, __m128i _2) {
984	return (__m128i)__builtin_lsx_vabsd_d((v2i64)_1, (v2i64)_2);
985	}
986
987	extern __inline
988	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
989	__lsx_vabsd_bu(__m128i _1, __m128i _2) {
990	return (__m128i)__builtin_lsx_vabsd_bu((v16u8)_1, (v16u8)_2);
991	}
992
993	extern __inline
994	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
995	__lsx_vabsd_hu(__m128i _1, __m128i _2) {
996	return (__m128i)__builtin_lsx_vabsd_hu((v8u16)_1, (v8u16)_2);
997	}
998
999	extern __inline
1000	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1001	__lsx_vabsd_wu(__m128i _1, __m128i _2) {
1002	return (__m128i)__builtin_lsx_vabsd_wu((v4u32)_1, (v4u32)_2);
1003	}
1004
1005	extern __inline
1006	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1007	__lsx_vabsd_du(__m128i _1, __m128i _2) {
1008	return (__m128i)__builtin_lsx_vabsd_du((v2u64)_1, (v2u64)_2);
1009	}
1010
1011	extern __inline
1012	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1013	__lsx_vmul_b(__m128i _1, __m128i _2) {
1014	return (__m128i)__builtin_lsx_vmul_b((v16i8)_1, (v16i8)_2);
1015	}
1016
1017	extern __inline
1018	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1019	__lsx_vmul_h(__m128i _1, __m128i _2) {
1020	return (__m128i)__builtin_lsx_vmul_h((v8i16)_1, (v8i16)_2);
1021	}
1022
1023	extern __inline
1024	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1025	__lsx_vmul_w(__m128i _1, __m128i _2) {
1026	return (__m128i)__builtin_lsx_vmul_w((v4i32)_1, (v4i32)_2);
1027	}
1028
1029	extern __inline
1030	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1031	__lsx_vmul_d(__m128i _1, __m128i _2) {
1032	return (__m128i)__builtin_lsx_vmul_d((v2i64)_1, (v2i64)_2);
1033	}
1034
1035	extern __inline
1036	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1037	__lsx_vmadd_b(__m128i _1, __m128i _2, __m128i _3) {
1038	return (__m128i)__builtin_lsx_vmadd_b((v16i8)_1, (v16i8)_2, (v16i8)_3);
1039	}
1040
1041	extern __inline
1042	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1043	__lsx_vmadd_h(__m128i _1, __m128i _2, __m128i _3) {
1044	return (__m128i)__builtin_lsx_vmadd_h((v8i16)_1, (v8i16)_2, (v8i16)_3);
1045	}
1046
1047	extern __inline
1048	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1049	__lsx_vmadd_w(__m128i _1, __m128i _2, __m128i _3) {
1050	return (__m128i)__builtin_lsx_vmadd_w((v4i32)_1, (v4i32)_2, (v4i32)_3);
1051	}
1052
1053	extern __inline
1054	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1055	__lsx_vmadd_d(__m128i _1, __m128i _2, __m128i _3) {
1056	return (__m128i)__builtin_lsx_vmadd_d((v2i64)_1, (v2i64)_2, (v2i64)_3);
1057	}
1058
1059	extern __inline
1060	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1061	__lsx_vmsub_b(__m128i _1, __m128i _2, __m128i _3) {
1062	return (__m128i)__builtin_lsx_vmsub_b((v16i8)_1, (v16i8)_2, (v16i8)_3);
1063	}
1064
1065	extern __inline
1066	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1067	__lsx_vmsub_h(__m128i _1, __m128i _2, __m128i _3) {
1068	return (__m128i)__builtin_lsx_vmsub_h((v8i16)_1, (v8i16)_2, (v8i16)_3);
1069	}
1070
1071	extern __inline
1072	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1073	__lsx_vmsub_w(__m128i _1, __m128i _2, __m128i _3) {
1074	return (__m128i)__builtin_lsx_vmsub_w((v4i32)_1, (v4i32)_2, (v4i32)_3);
1075	}
1076
1077	extern __inline
1078	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1079	__lsx_vmsub_d(__m128i _1, __m128i _2, __m128i _3) {
1080	return (__m128i)__builtin_lsx_vmsub_d((v2i64)_1, (v2i64)_2, (v2i64)_3);
1081	}
1082
1083	extern __inline
1084	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1085	__lsx_vdiv_b(__m128i _1, __m128i _2) {
1086	return (__m128i)__builtin_lsx_vdiv_b((v16i8)_1, (v16i8)_2);
1087	}
1088
1089	extern __inline
1090	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1091	__lsx_vdiv_h(__m128i _1, __m128i _2) {
1092	return (__m128i)__builtin_lsx_vdiv_h((v8i16)_1, (v8i16)_2);
1093	}
1094
1095	extern __inline
1096	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1097	__lsx_vdiv_w(__m128i _1, __m128i _2) {
1098	return (__m128i)__builtin_lsx_vdiv_w((v4i32)_1, (v4i32)_2);
1099	}
1100
1101	extern __inline
1102	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1103	__lsx_vdiv_d(__m128i _1, __m128i _2) {
1104	return (__m128i)__builtin_lsx_vdiv_d((v2i64)_1, (v2i64)_2);
1105	}
1106
1107	extern __inline
1108	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1109	__lsx_vdiv_bu(__m128i _1, __m128i _2) {
1110	return (__m128i)__builtin_lsx_vdiv_bu((v16u8)_1, (v16u8)_2);
1111	}
1112
1113	extern __inline
1114	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1115	__lsx_vdiv_hu(__m128i _1, __m128i _2) {
1116	return (__m128i)__builtin_lsx_vdiv_hu((v8u16)_1, (v8u16)_2);
1117	}
1118
1119	extern __inline
1120	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1121	__lsx_vdiv_wu(__m128i _1, __m128i _2) {
1122	return (__m128i)__builtin_lsx_vdiv_wu((v4u32)_1, (v4u32)_2);
1123	}
1124
1125	extern __inline
1126	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1127	__lsx_vdiv_du(__m128i _1, __m128i _2) {
1128	return (__m128i)__builtin_lsx_vdiv_du((v2u64)_1, (v2u64)_2);
1129	}
1130
1131	extern __inline
1132	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1133	__lsx_vhaddw_h_b(__m128i _1, __m128i _2) {
1134	return (__m128i)__builtin_lsx_vhaddw_h_b((v16i8)_1, (v16i8)_2);
1135	}
1136
1137	extern __inline
1138	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1139	__lsx_vhaddw_w_h(__m128i _1, __m128i _2) {
1140	return (__m128i)__builtin_lsx_vhaddw_w_h((v8i16)_1, (v8i16)_2);
1141	}
1142
1143	extern __inline
1144	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1145	__lsx_vhaddw_d_w(__m128i _1, __m128i _2) {
1146	return (__m128i)__builtin_lsx_vhaddw_d_w((v4i32)_1, (v4i32)_2);
1147	}
1148
1149	extern __inline
1150	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1151	__lsx_vhaddw_hu_bu(__m128i _1, __m128i _2) {
1152	return (__m128i)__builtin_lsx_vhaddw_hu_bu((v16u8)_1, (v16u8)_2);
1153	}
1154
1155	extern __inline
1156	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1157	__lsx_vhaddw_wu_hu(__m128i _1, __m128i _2) {
1158	return (__m128i)__builtin_lsx_vhaddw_wu_hu((v8u16)_1, (v8u16)_2);
1159	}
1160
1161	extern __inline
1162	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1163	__lsx_vhaddw_du_wu(__m128i _1, __m128i _2) {
1164	return (__m128i)__builtin_lsx_vhaddw_du_wu((v4u32)_1, (v4u32)_2);
1165	}
1166
1167	extern __inline
1168	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1169	__lsx_vhsubw_h_b(__m128i _1, __m128i _2) {
1170	return (__m128i)__builtin_lsx_vhsubw_h_b((v16i8)_1, (v16i8)_2);
1171	}
1172
1173	extern __inline
1174	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1175	__lsx_vhsubw_w_h(__m128i _1, __m128i _2) {
1176	return (__m128i)__builtin_lsx_vhsubw_w_h((v8i16)_1, (v8i16)_2);
1177	}
1178
1179	extern __inline
1180	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1181	__lsx_vhsubw_d_w(__m128i _1, __m128i _2) {
1182	return (__m128i)__builtin_lsx_vhsubw_d_w((v4i32)_1, (v4i32)_2);
1183	}
1184
1185	extern __inline
1186	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1187	__lsx_vhsubw_hu_bu(__m128i _1, __m128i _2) {
1188	return (__m128i)__builtin_lsx_vhsubw_hu_bu((v16u8)_1, (v16u8)_2);
1189	}
1190
1191	extern __inline
1192	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1193	__lsx_vhsubw_wu_hu(__m128i _1, __m128i _2) {
1194	return (__m128i)__builtin_lsx_vhsubw_wu_hu((v8u16)_1, (v8u16)_2);
1195	}
1196
1197	extern __inline
1198	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1199	__lsx_vhsubw_du_wu(__m128i _1, __m128i _2) {
1200	return (__m128i)__builtin_lsx_vhsubw_du_wu((v4u32)_1, (v4u32)_2);
1201	}
1202
1203	extern __inline
1204	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1205	__lsx_vmod_b(__m128i _1, __m128i _2) {
1206	return (__m128i)__builtin_lsx_vmod_b((v16i8)_1, (v16i8)_2);
1207	}
1208
1209	extern __inline
1210	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1211	__lsx_vmod_h(__m128i _1, __m128i _2) {
1212	return (__m128i)__builtin_lsx_vmod_h((v8i16)_1, (v8i16)_2);
1213	}
1214
1215	extern __inline
1216	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1217	__lsx_vmod_w(__m128i _1, __m128i _2) {
1218	return (__m128i)__builtin_lsx_vmod_w((v4i32)_1, (v4i32)_2);
1219	}
1220
1221	extern __inline
1222	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1223	__lsx_vmod_d(__m128i _1, __m128i _2) {
1224	return (__m128i)__builtin_lsx_vmod_d((v2i64)_1, (v2i64)_2);
1225	}
1226
1227	extern __inline
1228	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1229	__lsx_vmod_bu(__m128i _1, __m128i _2) {
1230	return (__m128i)__builtin_lsx_vmod_bu((v16u8)_1, (v16u8)_2);
1231	}
1232
1233	extern __inline
1234	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1235	__lsx_vmod_hu(__m128i _1, __m128i _2) {
1236	return (__m128i)__builtin_lsx_vmod_hu((v8u16)_1, (v8u16)_2);
1237	}
1238
1239	extern __inline
1240	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1241	__lsx_vmod_wu(__m128i _1, __m128i _2) {
1242	return (__m128i)__builtin_lsx_vmod_wu((v4u32)_1, (v4u32)_2);
1243	}
1244
1245	extern __inline
1246	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1247	__lsx_vmod_du(__m128i _1, __m128i _2) {
1248	return (__m128i)__builtin_lsx_vmod_du((v2u64)_1, (v2u64)_2);
1249	}
1250
1251	extern __inline
1252	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1253	__lsx_vreplve_b(__m128i _1, int _2) {
1254	return (__m128i)__builtin_lsx_vreplve_b((v16i8)_1, (int)_2);
1255	}
1256
1257	extern __inline
1258	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1259	__lsx_vreplve_h(__m128i _1, int _2) {
1260	return (__m128i)__builtin_lsx_vreplve_h((v8i16)_1, (int)_2);
1261	}
1262
1263	extern __inline
1264	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1265	__lsx_vreplve_w(__m128i _1, int _2) {
1266	return (__m128i)__builtin_lsx_vreplve_w((v4i32)_1, (int)_2);
1267	}
1268
1269	extern __inline
1270	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1271	__lsx_vreplve_d(__m128i _1, int _2) {
1272	return (__m128i)__builtin_lsx_vreplve_d((v2i64)_1, (int)_2);
1273	}
1274
1275	#define __lsx_vreplvei_b(/__m128i/ _1, /ui4/ _2) \
1276	((__m128i)__builtin_lsx_vreplvei_b((v16i8)(_1), (_2)))
1277
1278	#define __lsx_vreplvei_h(/__m128i/ _1, /ui3/ _2) \
1279	((__m128i)__builtin_lsx_vreplvei_h((v8i16)(_1), (_2)))
1280
1281	#define __lsx_vreplvei_w(/__m128i/ _1, /ui2/ _2) \
1282	((__m128i)__builtin_lsx_vreplvei_w((v4i32)(_1), (_2)))
1283
1284	#define __lsx_vreplvei_d(/__m128i/ _1, /ui1/ _2) \
1285	((__m128i)__builtin_lsx_vreplvei_d((v2i64)(_1), (_2)))
1286
1287	extern __inline
1288	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1289	__lsx_vpickev_b(__m128i _1, __m128i _2) {
1290	return (__m128i)__builtin_lsx_vpickev_b((v16i8)_1, (v16i8)_2);
1291	}
1292
1293	extern __inline
1294	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1295	__lsx_vpickev_h(__m128i _1, __m128i _2) {
1296	return (__m128i)__builtin_lsx_vpickev_h((v8i16)_1, (v8i16)_2);
1297	}
1298
1299	extern __inline
1300	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1301	__lsx_vpickev_w(__m128i _1, __m128i _2) {
1302	return (__m128i)__builtin_lsx_vpickev_w((v4i32)_1, (v4i32)_2);
1303	}
1304
1305	extern __inline
1306	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1307	__lsx_vpickev_d(__m128i _1, __m128i _2) {
1308	return (__m128i)__builtin_lsx_vpickev_d((v2i64)_1, (v2i64)_2);
1309	}
1310
1311	extern __inline
1312	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1313	__lsx_vpickod_b(__m128i _1, __m128i _2) {
1314	return (__m128i)__builtin_lsx_vpickod_b((v16i8)_1, (v16i8)_2);
1315	}
1316
1317	extern __inline
1318	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1319	__lsx_vpickod_h(__m128i _1, __m128i _2) {
1320	return (__m128i)__builtin_lsx_vpickod_h((v8i16)_1, (v8i16)_2);
1321	}
1322
1323	extern __inline
1324	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1325	__lsx_vpickod_w(__m128i _1, __m128i _2) {
1326	return (__m128i)__builtin_lsx_vpickod_w((v4i32)_1, (v4i32)_2);
1327	}
1328
1329	extern __inline
1330	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1331	__lsx_vpickod_d(__m128i _1, __m128i _2) {
1332	return (__m128i)__builtin_lsx_vpickod_d((v2i64)_1, (v2i64)_2);
1333	}
1334
1335	extern __inline
1336	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1337	__lsx_vilvh_b(__m128i _1, __m128i _2) {
1338	return (__m128i)__builtin_lsx_vilvh_b((v16i8)_1, (v16i8)_2);
1339	}
1340
1341	extern __inline
1342	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1343	__lsx_vilvh_h(__m128i _1, __m128i _2) {
1344	return (__m128i)__builtin_lsx_vilvh_h((v8i16)_1, (v8i16)_2);
1345	}
1346
1347	extern __inline
1348	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1349	__lsx_vilvh_w(__m128i _1, __m128i _2) {
1350	return (__m128i)__builtin_lsx_vilvh_w((v4i32)_1, (v4i32)_2);
1351	}
1352
1353	extern __inline
1354	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1355	__lsx_vilvh_d(__m128i _1, __m128i _2) {
1356	return (__m128i)__builtin_lsx_vilvh_d((v2i64)_1, (v2i64)_2);
1357	}
1358
1359	extern __inline
1360	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1361	__lsx_vilvl_b(__m128i _1, __m128i _2) {
1362	return (__m128i)__builtin_lsx_vilvl_b((v16i8)_1, (v16i8)_2);
1363	}
1364
1365	extern __inline
1366	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1367	__lsx_vilvl_h(__m128i _1, __m128i _2) {
1368	return (__m128i)__builtin_lsx_vilvl_h((v8i16)_1, (v8i16)_2);
1369	}
1370
1371	extern __inline
1372	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1373	__lsx_vilvl_w(__m128i _1, __m128i _2) {
1374	return (__m128i)__builtin_lsx_vilvl_w((v4i32)_1, (v4i32)_2);
1375	}
1376
1377	extern __inline
1378	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1379	__lsx_vilvl_d(__m128i _1, __m128i _2) {
1380	return (__m128i)__builtin_lsx_vilvl_d((v2i64)_1, (v2i64)_2);
1381	}
1382
1383	extern __inline
1384	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1385	__lsx_vpackev_b(__m128i _1, __m128i _2) {
1386	return (__m128i)__builtin_lsx_vpackev_b((v16i8)_1, (v16i8)_2);
1387	}
1388
1389	extern __inline
1390	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1391	__lsx_vpackev_h(__m128i _1, __m128i _2) {
1392	return (__m128i)__builtin_lsx_vpackev_h((v8i16)_1, (v8i16)_2);
1393	}
1394
1395	extern __inline
1396	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1397	__lsx_vpackev_w(__m128i _1, __m128i _2) {
1398	return (__m128i)__builtin_lsx_vpackev_w((v4i32)_1, (v4i32)_2);
1399	}
1400
1401	extern __inline
1402	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1403	__lsx_vpackev_d(__m128i _1, __m128i _2) {
1404	return (__m128i)__builtin_lsx_vpackev_d((v2i64)_1, (v2i64)_2);
1405	}
1406
1407	extern __inline
1408	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1409	__lsx_vpackod_b(__m128i _1, __m128i _2) {
1410	return (__m128i)__builtin_lsx_vpackod_b((v16i8)_1, (v16i8)_2);
1411	}
1412
1413	extern __inline
1414	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1415	__lsx_vpackod_h(__m128i _1, __m128i _2) {
1416	return (__m128i)__builtin_lsx_vpackod_h((v8i16)_1, (v8i16)_2);
1417	}
1418
1419	extern __inline
1420	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1421	__lsx_vpackod_w(__m128i _1, __m128i _2) {
1422	return (__m128i)__builtin_lsx_vpackod_w((v4i32)_1, (v4i32)_2);
1423	}
1424
1425	extern __inline
1426	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1427	__lsx_vpackod_d(__m128i _1, __m128i _2) {
1428	return (__m128i)__builtin_lsx_vpackod_d((v2i64)_1, (v2i64)_2);
1429	}
1430
1431	extern __inline
1432	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1433	__lsx_vshuf_h(__m128i _1, __m128i _2, __m128i _3) {
1434	return (__m128i)__builtin_lsx_vshuf_h((v8i16)_1, (v8i16)_2, (v8i16)_3);
1435	}
1436
1437	extern __inline
1438	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1439	__lsx_vshuf_w(__m128i _1, __m128i _2, __m128i _3) {
1440	return (__m128i)__builtin_lsx_vshuf_w((v4i32)_1, (v4i32)_2, (v4i32)_3);
1441	}
1442
1443	extern __inline
1444	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1445	__lsx_vshuf_d(__m128i _1, __m128i _2, __m128i _3) {
1446	return (__m128i)__builtin_lsx_vshuf_d((v2i64)_1, (v2i64)_2, (v2i64)_3);
1447	}
1448
1449	extern __inline
1450	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1451	__lsx_vand_v(__m128i _1, __m128i _2) {
1452	return (__m128i)__builtin_lsx_vand_v((v16u8)_1, (v16u8)_2);
1453	}
1454
1455	#define __lsx_vandi_b(/__m128i/ _1, /ui8/ _2) \
1456	((__m128i)__builtin_lsx_vandi_b((v16u8)(_1), (_2)))
1457
1458	extern __inline
1459	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1460	__lsx_vor_v(__m128i _1, __m128i _2) {
1461	return (__m128i)__builtin_lsx_vor_v((v16u8)_1, (v16u8)_2);
1462	}
1463
1464	#define __lsx_vori_b(/__m128i/ _1, /ui8/ _2) \
1465	((__m128i)__builtin_lsx_vori_b((v16u8)(_1), (_2)))
1466
1467	extern __inline
1468	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1469	__lsx_vnor_v(__m128i _1, __m128i _2) {
1470	return (__m128i)__builtin_lsx_vnor_v((v16u8)_1, (v16u8)_2);
1471	}
1472
1473	#define __lsx_vnori_b(/__m128i/ _1, /ui8/ _2) \
1474	((__m128i)__builtin_lsx_vnori_b((v16u8)(_1), (_2)))
1475
1476	extern __inline
1477	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1478	__lsx_vxor_v(__m128i _1, __m128i _2) {
1479	return (__m128i)__builtin_lsx_vxor_v((v16u8)_1, (v16u8)_2);
1480	}
1481
1482	#define __lsx_vxori_b(/__m128i/ _1, /ui8/ _2) \
1483	((__m128i)__builtin_lsx_vxori_b((v16u8)(_1), (_2)))
1484
1485	extern __inline
1486	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1487	__lsx_vbitsel_v(__m128i _1, __m128i _2, __m128i _3) {
1488	return (__m128i)__builtin_lsx_vbitsel_v((v16u8)_1, (v16u8)_2, (v16u8)_3);
1489	}
1490
1491	#define __lsx_vbitseli_b(/__m128i/ _1, /__m128i/ _2, /ui8/ _3) \
1492	((__m128i)__builtin_lsx_vbitseli_b((v16u8)(_1), (v16u8)(_2), (_3)))
1493
1494	#define __lsx_vshuf4i_b(/__m128i/ _1, /ui8/ _2) \
1495	((__m128i)__builtin_lsx_vshuf4i_b((v16i8)(_1), (_2)))
1496
1497	#define __lsx_vshuf4i_h(/__m128i/ _1, /ui8/ _2) \
1498	((__m128i)__builtin_lsx_vshuf4i_h((v8i16)(_1), (_2)))
1499
1500	#define __lsx_vshuf4i_w(/__m128i/ _1, /ui8/ _2) \
1501	((__m128i)__builtin_lsx_vshuf4i_w((v4i32)(_1), (_2)))
1502
1503	extern __inline
1504	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1505	__lsx_vreplgr2vr_b(int _1) {
1506	return (__m128i)__builtin_lsx_vreplgr2vr_b((int)_1);
1507	}
1508
1509	extern __inline
1510	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1511	__lsx_vreplgr2vr_h(int _1) {
1512	return (__m128i)__builtin_lsx_vreplgr2vr_h((int)_1);
1513	}
1514
1515	extern __inline
1516	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1517	__lsx_vreplgr2vr_w(int _1) {
1518	return (__m128i)__builtin_lsx_vreplgr2vr_w((int)_1);
1519	}
1520
1521	extern __inline
1522	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1523	__lsx_vreplgr2vr_d(long int _1) {
1524	return (__m128i)__builtin_lsx_vreplgr2vr_d((long int)_1);
1525	}
1526
1527	extern __inline
1528	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1529	__lsx_vpcnt_b(__m128i _1) {
1530	return (__m128i)__builtin_lsx_vpcnt_b((v16i8)_1);
1531	}
1532
1533	extern __inline
1534	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1535	__lsx_vpcnt_h(__m128i _1) {
1536	return (__m128i)__builtin_lsx_vpcnt_h((v8i16)_1);
1537	}
1538
1539	extern __inline
1540	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1541	__lsx_vpcnt_w(__m128i _1) {
1542	return (__m128i)__builtin_lsx_vpcnt_w((v4i32)_1);
1543	}
1544
1545	extern __inline
1546	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1547	__lsx_vpcnt_d(__m128i _1) {
1548	return (__m128i)__builtin_lsx_vpcnt_d((v2i64)_1);
1549	}
1550
1551	extern __inline
1552	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1553	__lsx_vclo_b(__m128i _1) {
1554	return (__m128i)__builtin_lsx_vclo_b((v16i8)_1);
1555	}
1556
1557	extern __inline
1558	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1559	__lsx_vclo_h(__m128i _1) {
1560	return (__m128i)__builtin_lsx_vclo_h((v8i16)_1);
1561	}
1562
1563	extern __inline
1564	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1565	__lsx_vclo_w(__m128i _1) {
1566	return (__m128i)__builtin_lsx_vclo_w((v4i32)_1);
1567	}
1568
1569	extern __inline
1570	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1571	__lsx_vclo_d(__m128i _1) {
1572	return (__m128i)__builtin_lsx_vclo_d((v2i64)_1);
1573	}
1574
1575	extern __inline
1576	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1577	__lsx_vclz_b(__m128i _1) {
1578	return (__m128i)__builtin_lsx_vclz_b((v16i8)_1);
1579	}
1580
1581	extern __inline
1582	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1583	__lsx_vclz_h(__m128i _1) {
1584	return (__m128i)__builtin_lsx_vclz_h((v8i16)_1);
1585	}
1586
1587	extern __inline
1588	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1589	__lsx_vclz_w(__m128i _1) {
1590	return (__m128i)__builtin_lsx_vclz_w((v4i32)_1);
1591	}
1592
1593	extern __inline
1594	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1595	__lsx_vclz_d(__m128i _1) {
1596	return (__m128i)__builtin_lsx_vclz_d((v2i64)_1);
1597	}
1598
1599	#define __lsx_vpickve2gr_b(/__m128i/ _1, /ui4/ _2) \
1600	((int)__builtin_lsx_vpickve2gr_b((v16i8)(_1), (_2)))
1601
1602	#define __lsx_vpickve2gr_h(/__m128i/ _1, /ui3/ _2) \
1603	((int)__builtin_lsx_vpickve2gr_h((v8i16)(_1), (_2)))
1604
1605	#define __lsx_vpickve2gr_w(/__m128i/ _1, /ui2/ _2) \
1606	((int)__builtin_lsx_vpickve2gr_w((v4i32)(_1), (_2)))
1607
1608	#define __lsx_vpickve2gr_d(/__m128i/ _1, /ui1/ _2) \
1609	((long int)__builtin_lsx_vpickve2gr_d((v2i64)(_1), (_2)))
1610
1611	#define __lsx_vpickve2gr_bu(/__m128i/ _1, /ui4/ _2) \
1612	((unsigned int)__builtin_lsx_vpickve2gr_bu((v16i8)(_1), (_2)))
1613
1614	#define __lsx_vpickve2gr_hu(/__m128i/ _1, /ui3/ _2) \
1615	((unsigned int)__builtin_lsx_vpickve2gr_hu((v8i16)(_1), (_2)))
1616
1617	#define __lsx_vpickve2gr_wu(/__m128i/ _1, /ui2/ _2) \
1618	((unsigned int)__builtin_lsx_vpickve2gr_wu((v4i32)(_1), (_2)))
1619
1620	#define __lsx_vpickve2gr_du(/__m128i/ _1, /ui1/ _2) \
1621	((unsigned long int)__builtin_lsx_vpickve2gr_du((v2i64)(_1), (_2)))
1622
1623	#define __lsx_vinsgr2vr_b(/__m128i/ _1, /int/ _2, /ui4/ _3) \
1624	((__m128i)__builtin_lsx_vinsgr2vr_b((v16i8)(_1), (int)(_2), (_3)))
1625
1626	#define __lsx_vinsgr2vr_h(/__m128i/ _1, /int/ _2, /ui3/ _3) \
1627	((__m128i)__builtin_lsx_vinsgr2vr_h((v8i16)(_1), (int)(_2), (_3)))
1628
1629	#define __lsx_vinsgr2vr_w(/__m128i/ _1, /int/ _2, /ui2/ _3) \
1630	((__m128i)__builtin_lsx_vinsgr2vr_w((v4i32)(_1), (int)(_2), (_3)))
1631
1632	#define __lsx_vinsgr2vr_d(/__m128i/ _1, /long int/ _2, /ui1/ _3) \
1633	((__m128i)__builtin_lsx_vinsgr2vr_d((v2i64)(_1), (long int)(_2), (_3)))
1634
1635	extern __inline
1636	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1637	__lsx_vfadd_s(__m128 _1, __m128 _2) {
1638	return (__m128)__builtin_lsx_vfadd_s((v4f32)_1, (v4f32)_2);
1639	}
1640
1641	extern __inline
1642	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1643	__lsx_vfadd_d(__m128d _1, __m128d _2) {
1644	return (__m128d)__builtin_lsx_vfadd_d((v2f64)_1, (v2f64)_2);
1645	}
1646
1647	extern __inline
1648	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1649	__lsx_vfsub_s(__m128 _1, __m128 _2) {
1650	return (__m128)__builtin_lsx_vfsub_s((v4f32)_1, (v4f32)_2);
1651	}
1652
1653	extern __inline
1654	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1655	__lsx_vfsub_d(__m128d _1, __m128d _2) {
1656	return (__m128d)__builtin_lsx_vfsub_d((v2f64)_1, (v2f64)_2);
1657	}
1658
1659	extern __inline
1660	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1661	__lsx_vfmul_s(__m128 _1, __m128 _2) {
1662	return (__m128)__builtin_lsx_vfmul_s((v4f32)_1, (v4f32)_2);
1663	}
1664
1665	extern __inline
1666	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1667	__lsx_vfmul_d(__m128d _1, __m128d _2) {
1668	return (__m128d)__builtin_lsx_vfmul_d((v2f64)_1, (v2f64)_2);
1669	}
1670
1671	extern __inline
1672	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1673	__lsx_vfdiv_s(__m128 _1, __m128 _2) {
1674	return (__m128)__builtin_lsx_vfdiv_s((v4f32)_1, (v4f32)_2);
1675	}
1676
1677	extern __inline
1678	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1679	__lsx_vfdiv_d(__m128d _1, __m128d _2) {
1680	return (__m128d)__builtin_lsx_vfdiv_d((v2f64)_1, (v2f64)_2);
1681	}
1682
1683	extern __inline
1684	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1685	__lsx_vfcvt_h_s(__m128 _1, __m128 _2) {
1686	return (__m128i)__builtin_lsx_vfcvt_h_s((v4f32)_1, (v4f32)_2);
1687	}
1688
1689	extern __inline
1690	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1691	__lsx_vfcvt_s_d(__m128d _1, __m128d _2) {
1692	return (__m128)__builtin_lsx_vfcvt_s_d((v2f64)_1, (v2f64)_2);
1693	}
1694
1695	extern __inline
1696	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1697	__lsx_vfmin_s(__m128 _1, __m128 _2) {
1698	return (__m128)__builtin_lsx_vfmin_s((v4f32)_1, (v4f32)_2);
1699	}
1700
1701	extern __inline
1702	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1703	__lsx_vfmin_d(__m128d _1, __m128d _2) {
1704	return (__m128d)__builtin_lsx_vfmin_d((v2f64)_1, (v2f64)_2);
1705	}
1706
1707	extern __inline
1708	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1709	__lsx_vfmina_s(__m128 _1, __m128 _2) {
1710	return (__m128)__builtin_lsx_vfmina_s((v4f32)_1, (v4f32)_2);
1711	}
1712
1713	extern __inline
1714	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1715	__lsx_vfmina_d(__m128d _1, __m128d _2) {
1716	return (__m128d)__builtin_lsx_vfmina_d((v2f64)_1, (v2f64)_2);
1717	}
1718
1719	extern __inline
1720	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1721	__lsx_vfmax_s(__m128 _1, __m128 _2) {
1722	return (__m128)__builtin_lsx_vfmax_s((v4f32)_1, (v4f32)_2);
1723	}
1724
1725	extern __inline
1726	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1727	__lsx_vfmax_d(__m128d _1, __m128d _2) {
1728	return (__m128d)__builtin_lsx_vfmax_d((v2f64)_1, (v2f64)_2);
1729	}
1730
1731	extern __inline
1732	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1733	__lsx_vfmaxa_s(__m128 _1, __m128 _2) {
1734	return (__m128)__builtin_lsx_vfmaxa_s((v4f32)_1, (v4f32)_2);
1735	}
1736
1737	extern __inline
1738	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1739	__lsx_vfmaxa_d(__m128d _1, __m128d _2) {
1740	return (__m128d)__builtin_lsx_vfmaxa_d((v2f64)_1, (v2f64)_2);
1741	}
1742
1743	extern __inline
1744	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1745	__lsx_vfclass_s(__m128 _1) {
1746	return (__m128i)__builtin_lsx_vfclass_s((v4f32)_1);
1747	}
1748
1749	extern __inline
1750	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1751	__lsx_vfclass_d(__m128d _1) {
1752	return (__m128i)__builtin_lsx_vfclass_d((v2f64)_1);
1753	}
1754
1755	extern __inline
1756	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1757	__lsx_vfsqrt_s(__m128 _1) {
1758	return (__m128)__builtin_lsx_vfsqrt_s((v4f32)_1);
1759	}
1760
1761	extern __inline
1762	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1763	__lsx_vfsqrt_d(__m128d _1) {
1764	return (__m128d)__builtin_lsx_vfsqrt_d((v2f64)_1);
1765	}
1766
1767	extern __inline
1768	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1769	__lsx_vfrecip_s(__m128 _1) {
1770	return (__m128)__builtin_lsx_vfrecip_s((v4f32)_1);
1771	}
1772
1773	extern __inline
1774	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1775	__lsx_vfrecip_d(__m128d _1) {
1776	return (__m128d)__builtin_lsx_vfrecip_d((v2f64)_1);
1777	}
1778
1779	extern __inline
1780	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1781	__lsx_vfrecipe_s(__m128 _1) {
1782	return (__m128)__builtin_lsx_vfrecipe_s((v4f32)_1);
1783	}
1784
1785	extern __inline
1786	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1787	__lsx_vfrecipe_d(__m128d _1) {
1788	return (__m128d)__builtin_lsx_vfrecipe_d((v2f64)_1);
1789	}
1790
1791	extern __inline
1792	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1793	__lsx_vfrint_s(__m128 _1) {
1794	return (__m128)__builtin_lsx_vfrint_s((v4f32)_1);
1795	}
1796
1797	extern __inline
1798	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1799	__lsx_vfrint_d(__m128d _1) {
1800	return (__m128d)__builtin_lsx_vfrint_d((v2f64)_1);
1801	}
1802
1803	extern __inline
1804	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1805	__lsx_vfrsqrt_s(__m128 _1) {
1806	return (__m128)__builtin_lsx_vfrsqrt_s((v4f32)_1);
1807	}
1808
1809	extern __inline
1810	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1811	__lsx_vfrsqrt_d(__m128d _1) {
1812	return (__m128d)__builtin_lsx_vfrsqrt_d((v2f64)_1);
1813	}
1814
1815	extern __inline
1816	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1817	__lsx_vfrsqrte_s(__m128 _1) {
1818	return (__m128)__builtin_lsx_vfrsqrte_s((v4f32)_1);
1819	}
1820
1821	extern __inline
1822	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1823	__lsx_vfrsqrte_d(__m128d _1) {
1824	return (__m128d)__builtin_lsx_vfrsqrte_d((v2f64)_1);
1825	}
1826
1827	extern __inline
1828	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1829	__lsx_vflogb_s(__m128 _1) {
1830	return (__m128)__builtin_lsx_vflogb_s((v4f32)_1);
1831	}
1832
1833	extern __inline
1834	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1835	__lsx_vflogb_d(__m128d _1) {
1836	return (__m128d)__builtin_lsx_vflogb_d((v2f64)_1);
1837	}
1838
1839	extern __inline
1840	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1841	__lsx_vfcvth_s_h(__m128i _1) {
1842	return (__m128)__builtin_lsx_vfcvth_s_h((v8i16)_1);
1843	}
1844
1845	extern __inline
1846	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1847	__lsx_vfcvth_d_s(__m128 _1) {
1848	return (__m128d)__builtin_lsx_vfcvth_d_s((v4f32)_1);
1849	}
1850
1851	extern __inline
1852	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1853	__lsx_vfcvtl_s_h(__m128i _1) {
1854	return (__m128)__builtin_lsx_vfcvtl_s_h((v8i16)_1);
1855	}
1856
1857	extern __inline
1858	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1859	__lsx_vfcvtl_d_s(__m128 _1) {
1860	return (__m128d)__builtin_lsx_vfcvtl_d_s((v4f32)_1);
1861	}
1862
1863	extern __inline
1864	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1865	__lsx_vftint_w_s(__m128 _1) {
1866	return (__m128i)__builtin_lsx_vftint_w_s((v4f32)_1);
1867	}
1868
1869	extern __inline
1870	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1871	__lsx_vftint_l_d(__m128d _1) {
1872	return (__m128i)__builtin_lsx_vftint_l_d((v2f64)_1);
1873	}
1874
1875	extern __inline
1876	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1877	__lsx_vftint_wu_s(__m128 _1) {
1878	return (__m128i)__builtin_lsx_vftint_wu_s((v4f32)_1);
1879	}
1880
1881	extern __inline
1882	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1883	__lsx_vftint_lu_d(__m128d _1) {
1884	return (__m128i)__builtin_lsx_vftint_lu_d((v2f64)_1);
1885	}
1886
1887	extern __inline
1888	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1889	__lsx_vftintrz_w_s(__m128 _1) {
1890	return (__m128i)__builtin_lsx_vftintrz_w_s((v4f32)_1);
1891	}
1892
1893	extern __inline
1894	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1895	__lsx_vftintrz_l_d(__m128d _1) {
1896	return (__m128i)__builtin_lsx_vftintrz_l_d((v2f64)_1);
1897	}
1898
1899	extern __inline
1900	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1901	__lsx_vftintrz_wu_s(__m128 _1) {
1902	return (__m128i)__builtin_lsx_vftintrz_wu_s((v4f32)_1);
1903	}
1904
1905	extern __inline
1906	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1907	__lsx_vftintrz_lu_d(__m128d _1) {
1908	return (__m128i)__builtin_lsx_vftintrz_lu_d((v2f64)_1);
1909	}
1910
1911	extern __inline
1912	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1913	__lsx_vffint_s_w(__m128i _1) {
1914	return (__m128)__builtin_lsx_vffint_s_w((v4i32)_1);
1915	}
1916
1917	extern __inline
1918	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1919	__lsx_vffint_d_l(__m128i _1) {
1920	return (__m128d)__builtin_lsx_vffint_d_l((v2i64)_1);
1921	}
1922
1923	extern __inline
1924	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
1925	__lsx_vffint_s_wu(__m128i _1) {
1926	return (__m128)__builtin_lsx_vffint_s_wu((v4u32)_1);
1927	}
1928
1929	extern __inline
1930	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
1931	__lsx_vffint_d_lu(__m128i _1) {
1932	return (__m128d)__builtin_lsx_vffint_d_lu((v2u64)_1);
1933	}
1934
1935	extern __inline
1936	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1937	__lsx_vandn_v(__m128i _1, __m128i _2) {
1938	return (__m128i)__builtin_lsx_vandn_v((v16u8)_1, (v16u8)_2);
1939	}
1940
1941	extern __inline
1942	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1943	__lsx_vneg_b(__m128i _1) {
1944	return (__m128i)__builtin_lsx_vneg_b((v16i8)_1);
1945	}
1946
1947	extern __inline
1948	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1949	__lsx_vneg_h(__m128i _1) {
1950	return (__m128i)__builtin_lsx_vneg_h((v8i16)_1);
1951	}
1952
1953	extern __inline
1954	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1955	__lsx_vneg_w(__m128i _1) {
1956	return (__m128i)__builtin_lsx_vneg_w((v4i32)_1);
1957	}
1958
1959	extern __inline
1960	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1961	__lsx_vneg_d(__m128i _1) {
1962	return (__m128i)__builtin_lsx_vneg_d((v2i64)_1);
1963	}
1964
1965	extern __inline
1966	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1967	__lsx_vmuh_b(__m128i _1, __m128i _2) {
1968	return (__m128i)__builtin_lsx_vmuh_b((v16i8)_1, (v16i8)_2);
1969	}
1970
1971	extern __inline
1972	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1973	__lsx_vmuh_h(__m128i _1, __m128i _2) {
1974	return (__m128i)__builtin_lsx_vmuh_h((v8i16)_1, (v8i16)_2);
1975	}
1976
1977	extern __inline
1978	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1979	__lsx_vmuh_w(__m128i _1, __m128i _2) {
1980	return (__m128i)__builtin_lsx_vmuh_w((v4i32)_1, (v4i32)_2);
1981	}
1982
1983	extern __inline
1984	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1985	__lsx_vmuh_d(__m128i _1, __m128i _2) {
1986	return (__m128i)__builtin_lsx_vmuh_d((v2i64)_1, (v2i64)_2);
1987	}
1988
1989	extern __inline
1990	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1991	__lsx_vmuh_bu(__m128i _1, __m128i _2) {
1992	return (__m128i)__builtin_lsx_vmuh_bu((v16u8)_1, (v16u8)_2);
1993	}
1994
1995	extern __inline
1996	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
1997	__lsx_vmuh_hu(__m128i _1, __m128i _2) {
1998	return (__m128i)__builtin_lsx_vmuh_hu((v8u16)_1, (v8u16)_2);
1999	}
2000
2001	extern __inline
2002	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2003	__lsx_vmuh_wu(__m128i _1, __m128i _2) {
2004	return (__m128i)__builtin_lsx_vmuh_wu((v4u32)_1, (v4u32)_2);
2005	}
2006
2007	extern __inline
2008	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2009	__lsx_vmuh_du(__m128i _1, __m128i _2) {
2010	return (__m128i)__builtin_lsx_vmuh_du((v2u64)_1, (v2u64)_2);
2011	}
2012
2013	#define __lsx_vsllwil_h_b(/__m128i/ _1, /ui3/ _2) \
2014	((__m128i)__builtin_lsx_vsllwil_h_b((v16i8)(_1), (_2)))
2015
2016	#define __lsx_vsllwil_w_h(/__m128i/ _1, /ui4/ _2) \
2017	((__m128i)__builtin_lsx_vsllwil_w_h((v8i16)(_1), (_2)))
2018
2019	#define __lsx_vsllwil_d_w(/__m128i/ _1, /ui5/ _2) \
2020	((__m128i)__builtin_lsx_vsllwil_d_w((v4i32)(_1), (_2)))
2021
2022	#define __lsx_vsllwil_hu_bu(/__m128i/ _1, /ui3/ _2) \
2023	((__m128i)__builtin_lsx_vsllwil_hu_bu((v16u8)(_1), (_2)))
2024
2025	#define __lsx_vsllwil_wu_hu(/__m128i/ _1, /ui4/ _2) \
2026	((__m128i)__builtin_lsx_vsllwil_wu_hu((v8u16)(_1), (_2)))
2027
2028	#define __lsx_vsllwil_du_wu(/__m128i/ _1, /ui5/ _2) \
2029	((__m128i)__builtin_lsx_vsllwil_du_wu((v4u32)(_1), (_2)))
2030
2031	extern __inline
2032	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2033	__lsx_vsran_b_h(__m128i _1, __m128i _2) {
2034	return (__m128i)__builtin_lsx_vsran_b_h((v8i16)_1, (v8i16)_2);
2035	}
2036
2037	extern __inline
2038	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2039	__lsx_vsran_h_w(__m128i _1, __m128i _2) {
2040	return (__m128i)__builtin_lsx_vsran_h_w((v4i32)_1, (v4i32)_2);
2041	}
2042
2043	extern __inline
2044	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2045	__lsx_vsran_w_d(__m128i _1, __m128i _2) {
2046	return (__m128i)__builtin_lsx_vsran_w_d((v2i64)_1, (v2i64)_2);
2047	}
2048
2049	extern __inline
2050	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2051	__lsx_vssran_b_h(__m128i _1, __m128i _2) {
2052	return (__m128i)__builtin_lsx_vssran_b_h((v8i16)_1, (v8i16)_2);
2053	}
2054
2055	extern __inline
2056	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2057	__lsx_vssran_h_w(__m128i _1, __m128i _2) {
2058	return (__m128i)__builtin_lsx_vssran_h_w((v4i32)_1, (v4i32)_2);
2059	}
2060
2061	extern __inline
2062	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2063	__lsx_vssran_w_d(__m128i _1, __m128i _2) {
2064	return (__m128i)__builtin_lsx_vssran_w_d((v2i64)_1, (v2i64)_2);
2065	}
2066
2067	extern __inline
2068	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2069	__lsx_vssran_bu_h(__m128i _1, __m128i _2) {
2070	return (__m128i)__builtin_lsx_vssran_bu_h((v8u16)_1, (v8u16)_2);
2071	}
2072
2073	extern __inline
2074	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2075	__lsx_vssran_hu_w(__m128i _1, __m128i _2) {
2076	return (__m128i)__builtin_lsx_vssran_hu_w((v4u32)_1, (v4u32)_2);
2077	}
2078
2079	extern __inline
2080	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2081	__lsx_vssran_wu_d(__m128i _1, __m128i _2) {
2082	return (__m128i)__builtin_lsx_vssran_wu_d((v2u64)_1, (v2u64)_2);
2083	}
2084
2085	extern __inline
2086	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2087	__lsx_vsrarn_b_h(__m128i _1, __m128i _2) {
2088	return (__m128i)__builtin_lsx_vsrarn_b_h((v8i16)_1, (v8i16)_2);
2089	}
2090
2091	extern __inline
2092	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2093	__lsx_vsrarn_h_w(__m128i _1, __m128i _2) {
2094	return (__m128i)__builtin_lsx_vsrarn_h_w((v4i32)_1, (v4i32)_2);
2095	}
2096
2097	extern __inline
2098	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2099	__lsx_vsrarn_w_d(__m128i _1, __m128i _2) {
2100	return (__m128i)__builtin_lsx_vsrarn_w_d((v2i64)_1, (v2i64)_2);
2101	}
2102
2103	extern __inline
2104	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2105	__lsx_vssrarn_b_h(__m128i _1, __m128i _2) {
2106	return (__m128i)__builtin_lsx_vssrarn_b_h((v8i16)_1, (v8i16)_2);
2107	}
2108
2109	extern __inline
2110	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2111	__lsx_vssrarn_h_w(__m128i _1, __m128i _2) {
2112	return (__m128i)__builtin_lsx_vssrarn_h_w((v4i32)_1, (v4i32)_2);
2113	}
2114
2115	extern __inline
2116	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2117	__lsx_vssrarn_w_d(__m128i _1, __m128i _2) {
2118	return (__m128i)__builtin_lsx_vssrarn_w_d((v2i64)_1, (v2i64)_2);
2119	}
2120
2121	extern __inline
2122	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2123	__lsx_vssrarn_bu_h(__m128i _1, __m128i _2) {
2124	return (__m128i)__builtin_lsx_vssrarn_bu_h((v8u16)_1, (v8u16)_2);
2125	}
2126
2127	extern __inline
2128	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2129	__lsx_vssrarn_hu_w(__m128i _1, __m128i _2) {
2130	return (__m128i)__builtin_lsx_vssrarn_hu_w((v4u32)_1, (v4u32)_2);
2131	}
2132
2133	extern __inline
2134	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2135	__lsx_vssrarn_wu_d(__m128i _1, __m128i _2) {
2136	return (__m128i)__builtin_lsx_vssrarn_wu_d((v2u64)_1, (v2u64)_2);
2137	}
2138
2139	extern __inline
2140	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2141	__lsx_vsrln_b_h(__m128i _1, __m128i _2) {
2142	return (__m128i)__builtin_lsx_vsrln_b_h((v8i16)_1, (v8i16)_2);
2143	}
2144
2145	extern __inline
2146	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2147	__lsx_vsrln_h_w(__m128i _1, __m128i _2) {
2148	return (__m128i)__builtin_lsx_vsrln_h_w((v4i32)_1, (v4i32)_2);
2149	}
2150
2151	extern __inline
2152	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2153	__lsx_vsrln_w_d(__m128i _1, __m128i _2) {
2154	return (__m128i)__builtin_lsx_vsrln_w_d((v2i64)_1, (v2i64)_2);
2155	}
2156
2157	extern __inline
2158	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2159	__lsx_vssrln_bu_h(__m128i _1, __m128i _2) {
2160	return (__m128i)__builtin_lsx_vssrln_bu_h((v8u16)_1, (v8u16)_2);
2161	}
2162
2163	extern __inline
2164	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2165	__lsx_vssrln_hu_w(__m128i _1, __m128i _2) {
2166	return (__m128i)__builtin_lsx_vssrln_hu_w((v4u32)_1, (v4u32)_2);
2167	}
2168
2169	extern __inline
2170	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2171	__lsx_vssrln_wu_d(__m128i _1, __m128i _2) {
2172	return (__m128i)__builtin_lsx_vssrln_wu_d((v2u64)_1, (v2u64)_2);
2173	}
2174
2175	extern __inline
2176	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2177	__lsx_vsrlrn_b_h(__m128i _1, __m128i _2) {
2178	return (__m128i)__builtin_lsx_vsrlrn_b_h((v8i16)_1, (v8i16)_2);
2179	}
2180
2181	extern __inline
2182	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2183	__lsx_vsrlrn_h_w(__m128i _1, __m128i _2) {
2184	return (__m128i)__builtin_lsx_vsrlrn_h_w((v4i32)_1, (v4i32)_2);
2185	}
2186
2187	extern __inline
2188	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2189	__lsx_vsrlrn_w_d(__m128i _1, __m128i _2) {
2190	return (__m128i)__builtin_lsx_vsrlrn_w_d((v2i64)_1, (v2i64)_2);
2191	}
2192
2193	extern __inline
2194	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2195	__lsx_vssrlrn_bu_h(__m128i _1, __m128i _2) {
2196	return (__m128i)__builtin_lsx_vssrlrn_bu_h((v8u16)_1, (v8u16)_2);
2197	}
2198
2199	extern __inline
2200	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2201	__lsx_vssrlrn_hu_w(__m128i _1, __m128i _2) {
2202	return (__m128i)__builtin_lsx_vssrlrn_hu_w((v4u32)_1, (v4u32)_2);
2203	}
2204
2205	extern __inline
2206	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2207	__lsx_vssrlrn_wu_d(__m128i _1, __m128i _2) {
2208	return (__m128i)__builtin_lsx_vssrlrn_wu_d((v2u64)_1, (v2u64)_2);
2209	}
2210
2211	#define __lsx_vfrstpi_b(/__m128i/ _1, /__m128i/ _2, /ui5/ _3) \
2212	((__m128i)__builtin_lsx_vfrstpi_b((v16i8)(_1), (v16i8)(_2), (_3)))
2213
2214	#define __lsx_vfrstpi_h(/__m128i/ _1, /__m128i/ _2, /ui5/ _3) \
2215	((__m128i)__builtin_lsx_vfrstpi_h((v8i16)(_1), (v8i16)(_2), (_3)))
2216
2217	extern __inline
2218	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2219	__lsx_vfrstp_b(__m128i _1, __m128i _2, __m128i _3) {
2220	return (__m128i)__builtin_lsx_vfrstp_b((v16i8)_1, (v16i8)_2, (v16i8)_3);
2221	}
2222
2223	extern __inline
2224	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2225	__lsx_vfrstp_h(__m128i _1, __m128i _2, __m128i _3) {
2226	return (__m128i)__builtin_lsx_vfrstp_h((v8i16)_1, (v8i16)_2, (v8i16)_3);
2227	}
2228
2229	#define __lsx_vshuf4i_d(/__m128i/ _1, /__m128i/ _2, /ui8/ _3) \
2230	((__m128i)__builtin_lsx_vshuf4i_d((v2i64)(_1), (v2i64)(_2), (_3)))
2231
2232	#define __lsx_vbsrl_v(/__m128i/ _1, /ui5/ _2) \
2233	((__m128i)__builtin_lsx_vbsrl_v((v16i8)(_1), (_2)))
2234
2235	#define __lsx_vbsll_v(/__m128i/ _1, /ui5/ _2) \
2236	((__m128i)__builtin_lsx_vbsll_v((v16i8)(_1), (_2)))
2237
2238	#define __lsx_vextrins_b(/__m128i/ _1, /__m128i/ _2, /ui8/ _3) \
2239	((__m128i)__builtin_lsx_vextrins_b((v16i8)(_1), (v16i8)(_2), (_3)))
2240
2241	#define __lsx_vextrins_h(/__m128i/ _1, /__m128i/ _2, /ui8/ _3) \
2242	((__m128i)__builtin_lsx_vextrins_h((v8i16)(_1), (v8i16)(_2), (_3)))
2243
2244	#define __lsx_vextrins_w(/__m128i/ _1, /__m128i/ _2, /ui8/ _3) \
2245	((__m128i)__builtin_lsx_vextrins_w((v4i32)(_1), (v4i32)(_2), (_3)))
2246
2247	#define __lsx_vextrins_d(/__m128i/ _1, /__m128i/ _2, /ui8/ _3) \
2248	((__m128i)__builtin_lsx_vextrins_d((v2i64)(_1), (v2i64)(_2), (_3)))
2249
2250	extern __inline
2251	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2252	__lsx_vmskltz_b(__m128i _1) {
2253	return (__m128i)__builtin_lsx_vmskltz_b((v16i8)_1);
2254	}
2255
2256	extern __inline
2257	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2258	__lsx_vmskltz_h(__m128i _1) {
2259	return (__m128i)__builtin_lsx_vmskltz_h((v8i16)_1);
2260	}
2261
2262	extern __inline
2263	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2264	__lsx_vmskltz_w(__m128i _1) {
2265	return (__m128i)__builtin_lsx_vmskltz_w((v4i32)_1);
2266	}
2267
2268	extern __inline
2269	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2270	__lsx_vmskltz_d(__m128i _1) {
2271	return (__m128i)__builtin_lsx_vmskltz_d((v2i64)_1);
2272	}
2273
2274	extern __inline
2275	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2276	__lsx_vsigncov_b(__m128i _1, __m128i _2) {
2277	return (__m128i)__builtin_lsx_vsigncov_b((v16i8)_1, (v16i8)_2);
2278	}
2279
2280	extern __inline
2281	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2282	__lsx_vsigncov_h(__m128i _1, __m128i _2) {
2283	return (__m128i)__builtin_lsx_vsigncov_h((v8i16)_1, (v8i16)_2);
2284	}
2285
2286	extern __inline
2287	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2288	__lsx_vsigncov_w(__m128i _1, __m128i _2) {
2289	return (__m128i)__builtin_lsx_vsigncov_w((v4i32)_1, (v4i32)_2);
2290	}
2291
2292	extern __inline
2293	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2294	__lsx_vsigncov_d(__m128i _1, __m128i _2) {
2295	return (__m128i)__builtin_lsx_vsigncov_d((v2i64)_1, (v2i64)_2);
2296	}
2297
2298	extern __inline
2299	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
2300	__lsx_vfmadd_s(__m128 _1, __m128 _2, __m128 _3) {
2301	return (__m128)__builtin_lsx_vfmadd_s((v4f32)_1, (v4f32)_2, (v4f32)_3);
2302	}
2303
2304	extern __inline
2305	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
2306	__lsx_vfmadd_d(__m128d _1, __m128d _2, __m128d _3) {
2307	return (__m128d)__builtin_lsx_vfmadd_d((v2f64)_1, (v2f64)_2, (v2f64)_3);
2308	}
2309
2310	extern __inline
2311	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
2312	__lsx_vfmsub_s(__m128 _1, __m128 _2, __m128 _3) {
2313	return (__m128)__builtin_lsx_vfmsub_s((v4f32)_1, (v4f32)_2, (v4f32)_3);
2314	}
2315
2316	extern __inline
2317	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
2318	__lsx_vfmsub_d(__m128d _1, __m128d _2, __m128d _3) {
2319	return (__m128d)__builtin_lsx_vfmsub_d((v2f64)_1, (v2f64)_2, (v2f64)_3);
2320	}
2321
2322	extern __inline
2323	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
2324	__lsx_vfnmadd_s(__m128 _1, __m128 _2, __m128 _3) {
2325	return (__m128)__builtin_lsx_vfnmadd_s((v4f32)_1, (v4f32)_2, (v4f32)_3);
2326	}
2327
2328	extern __inline
2329	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
2330	__lsx_vfnmadd_d(__m128d _1, __m128d _2, __m128d _3) {
2331	return (__m128d)__builtin_lsx_vfnmadd_d((v2f64)_1, (v2f64)_2, (v2f64)_3);
2332	}
2333
2334	extern __inline
2335	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
2336	__lsx_vfnmsub_s(__m128 _1, __m128 _2, __m128 _3) {
2337	return (__m128)__builtin_lsx_vfnmsub_s((v4f32)_1, (v4f32)_2, (v4f32)_3);
2338	}
2339
2340	extern __inline
2341	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
2342	__lsx_vfnmsub_d(__m128d _1, __m128d _2, __m128d _3) {
2343	return (__m128d)__builtin_lsx_vfnmsub_d((v2f64)_1, (v2f64)_2, (v2f64)_3);
2344	}
2345
2346	extern __inline
2347	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2348	__lsx_vftintrne_w_s(__m128 _1) {
2349	return (__m128i)__builtin_lsx_vftintrne_w_s((v4f32)_1);
2350	}
2351
2352	extern __inline
2353	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2354	__lsx_vftintrne_l_d(__m128d _1) {
2355	return (__m128i)__builtin_lsx_vftintrne_l_d((v2f64)_1);
2356	}
2357
2358	extern __inline
2359	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2360	__lsx_vftintrp_w_s(__m128 _1) {
2361	return (__m128i)__builtin_lsx_vftintrp_w_s((v4f32)_1);
2362	}
2363
2364	extern __inline
2365	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2366	__lsx_vftintrp_l_d(__m128d _1) {
2367	return (__m128i)__builtin_lsx_vftintrp_l_d((v2f64)_1);
2368	}
2369
2370	extern __inline
2371	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2372	__lsx_vftintrm_w_s(__m128 _1) {
2373	return (__m128i)__builtin_lsx_vftintrm_w_s((v4f32)_1);
2374	}
2375
2376	extern __inline
2377	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2378	__lsx_vftintrm_l_d(__m128d _1) {
2379	return (__m128i)__builtin_lsx_vftintrm_l_d((v2f64)_1);
2380	}
2381
2382	extern __inline
2383	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2384	__lsx_vftint_w_d(__m128d _1, __m128d _2) {
2385	return (__m128i)__builtin_lsx_vftint_w_d((v2f64)_1, (v2f64)_2);
2386	}
2387
2388	extern __inline
2389	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
2390	__lsx_vffint_s_l(__m128i _1, __m128i _2) {
2391	return (__m128)__builtin_lsx_vffint_s_l((v2i64)_1, (v2i64)_2);
2392	}
2393
2394	extern __inline
2395	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2396	__lsx_vftintrz_w_d(__m128d _1, __m128d _2) {
2397	return (__m128i)__builtin_lsx_vftintrz_w_d((v2f64)_1, (v2f64)_2);
2398	}
2399
2400	extern __inline
2401	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2402	__lsx_vftintrp_w_d(__m128d _1, __m128d _2) {
2403	return (__m128i)__builtin_lsx_vftintrp_w_d((v2f64)_1, (v2f64)_2);
2404	}
2405
2406	extern __inline
2407	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2408	__lsx_vftintrm_w_d(__m128d _1, __m128d _2) {
2409	return (__m128i)__builtin_lsx_vftintrm_w_d((v2f64)_1, (v2f64)_2);
2410	}
2411
2412	extern __inline
2413	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2414	__lsx_vftintrne_w_d(__m128d _1, __m128d _2) {
2415	return (__m128i)__builtin_lsx_vftintrne_w_d((v2f64)_1, (v2f64)_2);
2416	}
2417
2418	extern __inline
2419	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2420	__lsx_vftintl_l_s(__m128 _1) {
2421	return (__m128i)__builtin_lsx_vftintl_l_s((v4f32)_1);
2422	}
2423
2424	extern __inline
2425	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2426	__lsx_vftinth_l_s(__m128 _1) {
2427	return (__m128i)__builtin_lsx_vftinth_l_s((v4f32)_1);
2428	}
2429
2430	extern __inline
2431	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
2432	__lsx_vffinth_d_w(__m128i _1) {
2433	return (__m128d)__builtin_lsx_vffinth_d_w((v4i32)_1);
2434	}
2435
2436	extern __inline
2437	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
2438	__lsx_vffintl_d_w(__m128i _1) {
2439	return (__m128d)__builtin_lsx_vffintl_d_w((v4i32)_1);
2440	}
2441
2442	extern __inline
2443	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2444	__lsx_vftintrzl_l_s(__m128 _1) {
2445	return (__m128i)__builtin_lsx_vftintrzl_l_s((v4f32)_1);
2446	}
2447
2448	extern __inline
2449	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2450	__lsx_vftintrzh_l_s(__m128 _1) {
2451	return (__m128i)__builtin_lsx_vftintrzh_l_s((v4f32)_1);
2452	}
2453
2454	extern __inline
2455	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2456	__lsx_vftintrpl_l_s(__m128 _1) {
2457	return (__m128i)__builtin_lsx_vftintrpl_l_s((v4f32)_1);
2458	}
2459
2460	extern __inline
2461	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2462	__lsx_vftintrph_l_s(__m128 _1) {
2463	return (__m128i)__builtin_lsx_vftintrph_l_s((v4f32)_1);
2464	}
2465
2466	extern __inline
2467	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2468	__lsx_vftintrml_l_s(__m128 _1) {
2469	return (__m128i)__builtin_lsx_vftintrml_l_s((v4f32)_1);
2470	}
2471
2472	extern __inline
2473	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2474	__lsx_vftintrmh_l_s(__m128 _1) {
2475	return (__m128i)__builtin_lsx_vftintrmh_l_s((v4f32)_1);
2476	}
2477
2478	extern __inline
2479	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2480	__lsx_vftintrnel_l_s(__m128 _1) {
2481	return (__m128i)__builtin_lsx_vftintrnel_l_s((v4f32)_1);
2482	}
2483
2484	extern __inline
2485	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2486	__lsx_vftintrneh_l_s(__m128 _1) {
2487	return (__m128i)__builtin_lsx_vftintrneh_l_s((v4f32)_1);
2488	}
2489
2490	extern __inline
2491	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
2492	__lsx_vfrintrne_s(__m128 _1) {
2493	return (__m128)__builtin_lsx_vfrintrne_s((v4f32)_1);
2494	}
2495
2496	extern __inline
2497	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
2498	__lsx_vfrintrne_d(__m128d _1) {
2499	return (__m128d)__builtin_lsx_vfrintrne_d((v2f64)_1);
2500	}
2501
2502	extern __inline
2503	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
2504	__lsx_vfrintrz_s(__m128 _1) {
2505	return (__m128)__builtin_lsx_vfrintrz_s((v4f32)_1);
2506	}
2507
2508	extern __inline
2509	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
2510	__lsx_vfrintrz_d(__m128d _1) {
2511	return (__m128d)__builtin_lsx_vfrintrz_d((v2f64)_1);
2512	}
2513
2514	extern __inline
2515	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
2516	__lsx_vfrintrp_s(__m128 _1) {
2517	return (__m128)__builtin_lsx_vfrintrp_s((v4f32)_1);
2518	}
2519
2520	extern __inline
2521	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
2522	__lsx_vfrintrp_d(__m128d _1) {
2523	return (__m128d)__builtin_lsx_vfrintrp_d((v2f64)_1);
2524	}
2525
2526	extern __inline
2527	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
2528	__lsx_vfrintrm_s(__m128 _1) {
2529	return (__m128)__builtin_lsx_vfrintrm_s((v4f32)_1);
2530	}
2531
2532	extern __inline
2533	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
2534	__lsx_vfrintrm_d(__m128d _1) {
2535	return (__m128d)__builtin_lsx_vfrintrm_d((v2f64)_1);
2536	}
2537
2538	#define __lsx_vstelm_b(/__m128i/ _1, /void / _2, /si8/ _3, /idx*/ _4) \
2539	((void)__builtin_lsx_vstelm_b((v16i8)(_1), (void *)(_2), (_3), (_4)))
2540
2541	#define __lsx_vstelm_h(/__m128i/ _1, /void / _2, /si8/ _3, /idx*/ _4) \
2542	((void)__builtin_lsx_vstelm_h((v8i16)(_1), (void *)(_2), (_3), (_4)))
2543
2544	#define __lsx_vstelm_w(/__m128i/ _1, /void / _2, /si8/ _3, /idx*/ _4) \
2545	((void)__builtin_lsx_vstelm_w((v4i32)(_1), (void *)(_2), (_3), (_4)))
2546
2547	#define __lsx_vstelm_d(/__m128i/ _1, /void / _2, /si8/ _3, /idx*/ _4) \
2548	((void)__builtin_lsx_vstelm_d((v2i64)(_1), (void *)(_2), (_3), (_4)))
2549
2550	extern __inline
2551	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2552	__lsx_vaddwev_d_w(__m128i _1, __m128i _2) {
2553	return (__m128i)__builtin_lsx_vaddwev_d_w((v4i32)_1, (v4i32)_2);
2554	}
2555
2556	extern __inline
2557	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2558	__lsx_vaddwev_w_h(__m128i _1, __m128i _2) {
2559	return (__m128i)__builtin_lsx_vaddwev_w_h((v8i16)_1, (v8i16)_2);
2560	}
2561
2562	extern __inline
2563	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2564	__lsx_vaddwev_h_b(__m128i _1, __m128i _2) {
2565	return (__m128i)__builtin_lsx_vaddwev_h_b((v16i8)_1, (v16i8)_2);
2566	}
2567
2568	extern __inline
2569	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2570	__lsx_vaddwod_d_w(__m128i _1, __m128i _2) {
2571	return (__m128i)__builtin_lsx_vaddwod_d_w((v4i32)_1, (v4i32)_2);
2572	}
2573
2574	extern __inline
2575	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2576	__lsx_vaddwod_w_h(__m128i _1, __m128i _2) {
2577	return (__m128i)__builtin_lsx_vaddwod_w_h((v8i16)_1, (v8i16)_2);
2578	}
2579
2580	extern __inline
2581	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2582	__lsx_vaddwod_h_b(__m128i _1, __m128i _2) {
2583	return (__m128i)__builtin_lsx_vaddwod_h_b((v16i8)_1, (v16i8)_2);
2584	}
2585
2586	extern __inline
2587	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2588	__lsx_vaddwev_d_wu(__m128i _1, __m128i _2) {
2589	return (__m128i)__builtin_lsx_vaddwev_d_wu((v4u32)_1, (v4u32)_2);
2590	}
2591
2592	extern __inline
2593	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2594	__lsx_vaddwev_w_hu(__m128i _1, __m128i _2) {
2595	return (__m128i)__builtin_lsx_vaddwev_w_hu((v8u16)_1, (v8u16)_2);
2596	}
2597
2598	extern __inline
2599	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2600	__lsx_vaddwev_h_bu(__m128i _1, __m128i _2) {
2601	return (__m128i)__builtin_lsx_vaddwev_h_bu((v16u8)_1, (v16u8)_2);
2602	}
2603
2604	extern __inline
2605	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2606	__lsx_vaddwod_d_wu(__m128i _1, __m128i _2) {
2607	return (__m128i)__builtin_lsx_vaddwod_d_wu((v4u32)_1, (v4u32)_2);
2608	}
2609
2610	extern __inline
2611	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2612	__lsx_vaddwod_w_hu(__m128i _1, __m128i _2) {
2613	return (__m128i)__builtin_lsx_vaddwod_w_hu((v8u16)_1, (v8u16)_2);
2614	}
2615
2616	extern __inline
2617	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2618	__lsx_vaddwod_h_bu(__m128i _1, __m128i _2) {
2619	return (__m128i)__builtin_lsx_vaddwod_h_bu((v16u8)_1, (v16u8)_2);
2620	}
2621
2622	extern __inline
2623	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2624	__lsx_vaddwev_d_wu_w(__m128i _1, __m128i _2) {
2625	return (__m128i)__builtin_lsx_vaddwev_d_wu_w((v4u32)_1, (v4i32)_2);
2626	}
2627
2628	extern __inline
2629	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2630	__lsx_vaddwev_w_hu_h(__m128i _1, __m128i _2) {
2631	return (__m128i)__builtin_lsx_vaddwev_w_hu_h((v8u16)_1, (v8i16)_2);
2632	}
2633
2634	extern __inline
2635	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2636	__lsx_vaddwev_h_bu_b(__m128i _1, __m128i _2) {
2637	return (__m128i)__builtin_lsx_vaddwev_h_bu_b((v16u8)_1, (v16i8)_2);
2638	}
2639
2640	extern __inline
2641	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2642	__lsx_vaddwod_d_wu_w(__m128i _1, __m128i _2) {
2643	return (__m128i)__builtin_lsx_vaddwod_d_wu_w((v4u32)_1, (v4i32)_2);
2644	}
2645
2646	extern __inline
2647	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2648	__lsx_vaddwod_w_hu_h(__m128i _1, __m128i _2) {
2649	return (__m128i)__builtin_lsx_vaddwod_w_hu_h((v8u16)_1, (v8i16)_2);
2650	}
2651
2652	extern __inline
2653	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2654	__lsx_vaddwod_h_bu_b(__m128i _1, __m128i _2) {
2655	return (__m128i)__builtin_lsx_vaddwod_h_bu_b((v16u8)_1, (v16i8)_2);
2656	}
2657
2658	extern __inline
2659	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2660	__lsx_vsubwev_d_w(__m128i _1, __m128i _2) {
2661	return (__m128i)__builtin_lsx_vsubwev_d_w((v4i32)_1, (v4i32)_2);
2662	}
2663
2664	extern __inline
2665	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2666	__lsx_vsubwev_w_h(__m128i _1, __m128i _2) {
2667	return (__m128i)__builtin_lsx_vsubwev_w_h((v8i16)_1, (v8i16)_2);
2668	}
2669
2670	extern __inline
2671	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2672	__lsx_vsubwev_h_b(__m128i _1, __m128i _2) {
2673	return (__m128i)__builtin_lsx_vsubwev_h_b((v16i8)_1, (v16i8)_2);
2674	}
2675
2676	extern __inline
2677	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2678	__lsx_vsubwod_d_w(__m128i _1, __m128i _2) {
2679	return (__m128i)__builtin_lsx_vsubwod_d_w((v4i32)_1, (v4i32)_2);
2680	}
2681
2682	extern __inline
2683	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2684	__lsx_vsubwod_w_h(__m128i _1, __m128i _2) {
2685	return (__m128i)__builtin_lsx_vsubwod_w_h((v8i16)_1, (v8i16)_2);
2686	}
2687
2688	extern __inline
2689	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2690	__lsx_vsubwod_h_b(__m128i _1, __m128i _2) {
2691	return (__m128i)__builtin_lsx_vsubwod_h_b((v16i8)_1, (v16i8)_2);
2692	}
2693
2694	extern __inline
2695	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2696	__lsx_vsubwev_d_wu(__m128i _1, __m128i _2) {
2697	return (__m128i)__builtin_lsx_vsubwev_d_wu((v4u32)_1, (v4u32)_2);
2698	}
2699
2700	extern __inline
2701	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2702	__lsx_vsubwev_w_hu(__m128i _1, __m128i _2) {
2703	return (__m128i)__builtin_lsx_vsubwev_w_hu((v8u16)_1, (v8u16)_2);
2704	}
2705
2706	extern __inline
2707	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2708	__lsx_vsubwev_h_bu(__m128i _1, __m128i _2) {
2709	return (__m128i)__builtin_lsx_vsubwev_h_bu((v16u8)_1, (v16u8)_2);
2710	}
2711
2712	extern __inline
2713	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2714	__lsx_vsubwod_d_wu(__m128i _1, __m128i _2) {
2715	return (__m128i)__builtin_lsx_vsubwod_d_wu((v4u32)_1, (v4u32)_2);
2716	}
2717
2718	extern __inline
2719	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2720	__lsx_vsubwod_w_hu(__m128i _1, __m128i _2) {
2721	return (__m128i)__builtin_lsx_vsubwod_w_hu((v8u16)_1, (v8u16)_2);
2722	}
2723
2724	extern __inline
2725	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2726	__lsx_vsubwod_h_bu(__m128i _1, __m128i _2) {
2727	return (__m128i)__builtin_lsx_vsubwod_h_bu((v16u8)_1, (v16u8)_2);
2728	}
2729
2730	extern __inline
2731	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2732	__lsx_vaddwev_q_d(__m128i _1, __m128i _2) {
2733	return (__m128i)__builtin_lsx_vaddwev_q_d((v2i64)_1, (v2i64)_2);
2734	}
2735
2736	extern __inline
2737	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2738	__lsx_vaddwod_q_d(__m128i _1, __m128i _2) {
2739	return (__m128i)__builtin_lsx_vaddwod_q_d((v2i64)_1, (v2i64)_2);
2740	}
2741
2742	extern __inline
2743	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2744	__lsx_vaddwev_q_du(__m128i _1, __m128i _2) {
2745	return (__m128i)__builtin_lsx_vaddwev_q_du((v2u64)_1, (v2u64)_2);
2746	}
2747
2748	extern __inline
2749	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2750	__lsx_vaddwod_q_du(__m128i _1, __m128i _2) {
2751	return (__m128i)__builtin_lsx_vaddwod_q_du((v2u64)_1, (v2u64)_2);
2752	}
2753
2754	extern __inline
2755	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2756	__lsx_vsubwev_q_d(__m128i _1, __m128i _2) {
2757	return (__m128i)__builtin_lsx_vsubwev_q_d((v2i64)_1, (v2i64)_2);
2758	}
2759
2760	extern __inline
2761	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2762	__lsx_vsubwod_q_d(__m128i _1, __m128i _2) {
2763	return (__m128i)__builtin_lsx_vsubwod_q_d((v2i64)_1, (v2i64)_2);
2764	}
2765
2766	extern __inline
2767	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2768	__lsx_vsubwev_q_du(__m128i _1, __m128i _2) {
2769	return (__m128i)__builtin_lsx_vsubwev_q_du((v2u64)_1, (v2u64)_2);
2770	}
2771
2772	extern __inline
2773	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2774	__lsx_vsubwod_q_du(__m128i _1, __m128i _2) {
2775	return (__m128i)__builtin_lsx_vsubwod_q_du((v2u64)_1, (v2u64)_2);
2776	}
2777
2778	extern __inline
2779	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2780	__lsx_vaddwev_q_du_d(__m128i _1, __m128i _2) {
2781	return (__m128i)__builtin_lsx_vaddwev_q_du_d((v2u64)_1, (v2i64)_2);
2782	}
2783
2784	extern __inline
2785	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2786	__lsx_vaddwod_q_du_d(__m128i _1, __m128i _2) {
2787	return (__m128i)__builtin_lsx_vaddwod_q_du_d((v2u64)_1, (v2i64)_2);
2788	}
2789
2790	extern __inline
2791	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2792	__lsx_vmulwev_d_w(__m128i _1, __m128i _2) {
2793	return (__m128i)__builtin_lsx_vmulwev_d_w((v4i32)_1, (v4i32)_2);
2794	}
2795
2796	extern __inline
2797	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2798	__lsx_vmulwev_w_h(__m128i _1, __m128i _2) {
2799	return (__m128i)__builtin_lsx_vmulwev_w_h((v8i16)_1, (v8i16)_2);
2800	}
2801
2802	extern __inline
2803	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2804	__lsx_vmulwev_h_b(__m128i _1, __m128i _2) {
2805	return (__m128i)__builtin_lsx_vmulwev_h_b((v16i8)_1, (v16i8)_2);
2806	}
2807
2808	extern __inline
2809	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2810	__lsx_vmulwod_d_w(__m128i _1, __m128i _2) {
2811	return (__m128i)__builtin_lsx_vmulwod_d_w((v4i32)_1, (v4i32)_2);
2812	}
2813
2814	extern __inline
2815	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2816	__lsx_vmulwod_w_h(__m128i _1, __m128i _2) {
2817	return (__m128i)__builtin_lsx_vmulwod_w_h((v8i16)_1, (v8i16)_2);
2818	}
2819
2820	extern __inline
2821	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2822	__lsx_vmulwod_h_b(__m128i _1, __m128i _2) {
2823	return (__m128i)__builtin_lsx_vmulwod_h_b((v16i8)_1, (v16i8)_2);
2824	}
2825
2826	extern __inline
2827	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2828	__lsx_vmulwev_d_wu(__m128i _1, __m128i _2) {
2829	return (__m128i)__builtin_lsx_vmulwev_d_wu((v4u32)_1, (v4u32)_2);
2830	}
2831
2832	extern __inline
2833	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2834	__lsx_vmulwev_w_hu(__m128i _1, __m128i _2) {
2835	return (__m128i)__builtin_lsx_vmulwev_w_hu((v8u16)_1, (v8u16)_2);
2836	}
2837
2838	extern __inline
2839	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2840	__lsx_vmulwev_h_bu(__m128i _1, __m128i _2) {
2841	return (__m128i)__builtin_lsx_vmulwev_h_bu((v16u8)_1, (v16u8)_2);
2842	}
2843
2844	extern __inline
2845	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2846	__lsx_vmulwod_d_wu(__m128i _1, __m128i _2) {
2847	return (__m128i)__builtin_lsx_vmulwod_d_wu((v4u32)_1, (v4u32)_2);
2848	}
2849
2850	extern __inline
2851	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2852	__lsx_vmulwod_w_hu(__m128i _1, __m128i _2) {
2853	return (__m128i)__builtin_lsx_vmulwod_w_hu((v8u16)_1, (v8u16)_2);
2854	}
2855
2856	extern __inline
2857	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2858	__lsx_vmulwod_h_bu(__m128i _1, __m128i _2) {
2859	return (__m128i)__builtin_lsx_vmulwod_h_bu((v16u8)_1, (v16u8)_2);
2860	}
2861
2862	extern __inline
2863	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2864	__lsx_vmulwev_d_wu_w(__m128i _1, __m128i _2) {
2865	return (__m128i)__builtin_lsx_vmulwev_d_wu_w((v4u32)_1, (v4i32)_2);
2866	}
2867
2868	extern __inline
2869	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2870	__lsx_vmulwev_w_hu_h(__m128i _1, __m128i _2) {
2871	return (__m128i)__builtin_lsx_vmulwev_w_hu_h((v8u16)_1, (v8i16)_2);
2872	}
2873
2874	extern __inline
2875	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2876	__lsx_vmulwev_h_bu_b(__m128i _1, __m128i _2) {
2877	return (__m128i)__builtin_lsx_vmulwev_h_bu_b((v16u8)_1, (v16i8)_2);
2878	}
2879
2880	extern __inline
2881	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2882	__lsx_vmulwod_d_wu_w(__m128i _1, __m128i _2) {
2883	return (__m128i)__builtin_lsx_vmulwod_d_wu_w((v4u32)_1, (v4i32)_2);
2884	}
2885
2886	extern __inline
2887	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2888	__lsx_vmulwod_w_hu_h(__m128i _1, __m128i _2) {
2889	return (__m128i)__builtin_lsx_vmulwod_w_hu_h((v8u16)_1, (v8i16)_2);
2890	}
2891
2892	extern __inline
2893	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2894	__lsx_vmulwod_h_bu_b(__m128i _1, __m128i _2) {
2895	return (__m128i)__builtin_lsx_vmulwod_h_bu_b((v16u8)_1, (v16i8)_2);
2896	}
2897
2898	extern __inline
2899	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2900	__lsx_vmulwev_q_d(__m128i _1, __m128i _2) {
2901	return (__m128i)__builtin_lsx_vmulwev_q_d((v2i64)_1, (v2i64)_2);
2902	}
2903
2904	extern __inline
2905	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2906	__lsx_vmulwod_q_d(__m128i _1, __m128i _2) {
2907	return (__m128i)__builtin_lsx_vmulwod_q_d((v2i64)_1, (v2i64)_2);
2908	}
2909
2910	extern __inline
2911	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2912	__lsx_vmulwev_q_du(__m128i _1, __m128i _2) {
2913	return (__m128i)__builtin_lsx_vmulwev_q_du((v2u64)_1, (v2u64)_2);
2914	}
2915
2916	extern __inline
2917	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2918	__lsx_vmulwod_q_du(__m128i _1, __m128i _2) {
2919	return (__m128i)__builtin_lsx_vmulwod_q_du((v2u64)_1, (v2u64)_2);
2920	}
2921
2922	extern __inline
2923	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2924	__lsx_vmulwev_q_du_d(__m128i _1, __m128i _2) {
2925	return (__m128i)__builtin_lsx_vmulwev_q_du_d((v2u64)_1, (v2i64)_2);
2926	}
2927
2928	extern __inline
2929	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2930	__lsx_vmulwod_q_du_d(__m128i _1, __m128i _2) {
2931	return (__m128i)__builtin_lsx_vmulwod_q_du_d((v2u64)_1, (v2i64)_2);
2932	}
2933
2934	extern __inline
2935	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2936	__lsx_vhaddw_q_d(__m128i _1, __m128i _2) {
2937	return (__m128i)__builtin_lsx_vhaddw_q_d((v2i64)_1, (v2i64)_2);
2938	}
2939
2940	extern __inline
2941	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2942	__lsx_vhaddw_qu_du(__m128i _1, __m128i _2) {
2943	return (__m128i)__builtin_lsx_vhaddw_qu_du((v2u64)_1, (v2u64)_2);
2944	}
2945
2946	extern __inline
2947	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2948	__lsx_vhsubw_q_d(__m128i _1, __m128i _2) {
2949	return (__m128i)__builtin_lsx_vhsubw_q_d((v2i64)_1, (v2i64)_2);
2950	}
2951
2952	extern __inline
2953	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2954	__lsx_vhsubw_qu_du(__m128i _1, __m128i _2) {
2955	return (__m128i)__builtin_lsx_vhsubw_qu_du((v2u64)_1, (v2u64)_2);
2956	}
2957
2958	extern __inline
2959	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2960	__lsx_vmaddwev_d_w(__m128i _1, __m128i _2, __m128i _3) {
2961	return (__m128i)__builtin_lsx_vmaddwev_d_w((v2i64)_1, (v4i32)_2, (v4i32)_3);
2962	}
2963
2964	extern __inline
2965	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2966	__lsx_vmaddwev_w_h(__m128i _1, __m128i _2, __m128i _3) {
2967	return (__m128i)__builtin_lsx_vmaddwev_w_h((v4i32)_1, (v8i16)_2, (v8i16)_3);
2968	}
2969
2970	extern __inline
2971	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2972	__lsx_vmaddwev_h_b(__m128i _1, __m128i _2, __m128i _3) {
2973	return (__m128i)__builtin_lsx_vmaddwev_h_b((v8i16)_1, (v16i8)_2, (v16i8)_3);
2974	}
2975
2976	extern __inline
2977	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2978	__lsx_vmaddwev_d_wu(__m128i _1, __m128i _2, __m128i _3) {
2979	return (__m128i)__builtin_lsx_vmaddwev_d_wu((v2u64)_1, (v4u32)_2, (v4u32)_3);
2980	}
2981
2982	extern __inline
2983	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2984	__lsx_vmaddwev_w_hu(__m128i _1, __m128i _2, __m128i _3) {
2985	return (__m128i)__builtin_lsx_vmaddwev_w_hu((v4u32)_1, (v8u16)_2, (v8u16)_3);
2986	}
2987
2988	extern __inline
2989	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2990	__lsx_vmaddwev_h_bu(__m128i _1, __m128i _2, __m128i _3) {
2991	return (__m128i)__builtin_lsx_vmaddwev_h_bu((v8u16)_1, (v16u8)_2, (v16u8)_3);
2992	}
2993
2994	extern __inline
2995	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
2996	__lsx_vmaddwod_d_w(__m128i _1, __m128i _2, __m128i _3) {
2997	return (__m128i)__builtin_lsx_vmaddwod_d_w((v2i64)_1, (v4i32)_2, (v4i32)_3);
2998	}
2999
3000	extern __inline
3001	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3002	__lsx_vmaddwod_w_h(__m128i _1, __m128i _2, __m128i _3) {
3003	return (__m128i)__builtin_lsx_vmaddwod_w_h((v4i32)_1, (v8i16)_2, (v8i16)_3);
3004	}
3005
3006	extern __inline
3007	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3008	__lsx_vmaddwod_h_b(__m128i _1, __m128i _2, __m128i _3) {
3009	return (__m128i)__builtin_lsx_vmaddwod_h_b((v8i16)_1, (v16i8)_2, (v16i8)_3);
3010	}
3011
3012	extern __inline
3013	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3014	__lsx_vmaddwod_d_wu(__m128i _1, __m128i _2, __m128i _3) {
3015	return (__m128i)__builtin_lsx_vmaddwod_d_wu((v2u64)_1, (v4u32)_2, (v4u32)_3);
3016	}
3017
3018	extern __inline
3019	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3020	__lsx_vmaddwod_w_hu(__m128i _1, __m128i _2, __m128i _3) {
3021	return (__m128i)__builtin_lsx_vmaddwod_w_hu((v4u32)_1, (v8u16)_2, (v8u16)_3);
3022	}
3023
3024	extern __inline
3025	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3026	__lsx_vmaddwod_h_bu(__m128i _1, __m128i _2, __m128i _3) {
3027	return (__m128i)__builtin_lsx_vmaddwod_h_bu((v8u16)_1, (v16u8)_2, (v16u8)_3);
3028	}
3029
3030	extern __inline
3031	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3032	__lsx_vmaddwev_d_wu_w(__m128i _1, __m128i _2, __m128i _3) {
3033	return (__m128i)__builtin_lsx_vmaddwev_d_wu_w((v2i64)_1, (v4u32)_2,
3034	(v4i32)_3);
3035	}
3036
3037	extern __inline
3038	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3039	__lsx_vmaddwev_w_hu_h(__m128i _1, __m128i _2, __m128i _3) {
3040	return (__m128i)__builtin_lsx_vmaddwev_w_hu_h((v4i32)_1, (v8u16)_2,
3041	(v8i16)_3);
3042	}
3043
3044	extern __inline
3045	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3046	__lsx_vmaddwev_h_bu_b(__m128i _1, __m128i _2, __m128i _3) {
3047	return (__m128i)__builtin_lsx_vmaddwev_h_bu_b((v8i16)_1, (v16u8)_2,
3048	(v16i8)_3);
3049	}
3050
3051	extern __inline
3052	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3053	__lsx_vmaddwod_d_wu_w(__m128i _1, __m128i _2, __m128i _3) {
3054	return (__m128i)__builtin_lsx_vmaddwod_d_wu_w((v2i64)_1, (v4u32)_2,
3055	(v4i32)_3);
3056	}
3057
3058	extern __inline
3059	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3060	__lsx_vmaddwod_w_hu_h(__m128i _1, __m128i _2, __m128i _3) {
3061	return (__m128i)__builtin_lsx_vmaddwod_w_hu_h((v4i32)_1, (v8u16)_2,
3062	(v8i16)_3);
3063	}
3064
3065	extern __inline
3066	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3067	__lsx_vmaddwod_h_bu_b(__m128i _1, __m128i _2, __m128i _3) {
3068	return (__m128i)__builtin_lsx_vmaddwod_h_bu_b((v8i16)_1, (v16u8)_2,
3069	(v16i8)_3);
3070	}
3071
3072	extern __inline
3073	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3074	__lsx_vmaddwev_q_d(__m128i _1, __m128i _2, __m128i _3) {
3075	return (__m128i)__builtin_lsx_vmaddwev_q_d((v2i64)_1, (v2i64)_2, (v2i64)_3);
3076	}
3077
3078	extern __inline
3079	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3080	__lsx_vmaddwod_q_d(__m128i _1, __m128i _2, __m128i _3) {
3081	return (__m128i)__builtin_lsx_vmaddwod_q_d((v2i64)_1, (v2i64)_2, (v2i64)_3);
3082	}
3083
3084	extern __inline
3085	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3086	__lsx_vmaddwev_q_du(__m128i _1, __m128i _2, __m128i _3) {
3087	return (__m128i)__builtin_lsx_vmaddwev_q_du((v2u64)_1, (v2u64)_2, (v2u64)_3);
3088	}
3089
3090	extern __inline
3091	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3092	__lsx_vmaddwod_q_du(__m128i _1, __m128i _2, __m128i _3) {
3093	return (__m128i)__builtin_lsx_vmaddwod_q_du((v2u64)_1, (v2u64)_2, (v2u64)_3);
3094	}
3095
3096	extern __inline
3097	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3098	__lsx_vmaddwev_q_du_d(__m128i _1, __m128i _2, __m128i _3) {
3099	return (__m128i)__builtin_lsx_vmaddwev_q_du_d((v2i64)_1, (v2u64)_2,
3100	(v2i64)_3);
3101	}
3102
3103	extern __inline
3104	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3105	__lsx_vmaddwod_q_du_d(__m128i _1, __m128i _2, __m128i _3) {
3106	return (__m128i)__builtin_lsx_vmaddwod_q_du_d((v2i64)_1, (v2u64)_2,
3107	(v2i64)_3);
3108	}
3109
3110	extern __inline
3111	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3112	__lsx_vrotr_b(__m128i _1, __m128i _2) {
3113	return (__m128i)__builtin_lsx_vrotr_b((v16i8)_1, (v16i8)_2);
3114	}
3115
3116	extern __inline
3117	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3118	__lsx_vrotr_h(__m128i _1, __m128i _2) {
3119	return (__m128i)__builtin_lsx_vrotr_h((v8i16)_1, (v8i16)_2);
3120	}
3121
3122	extern __inline
3123	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3124	__lsx_vrotr_w(__m128i _1, __m128i _2) {
3125	return (__m128i)__builtin_lsx_vrotr_w((v4i32)_1, (v4i32)_2);
3126	}
3127
3128	extern __inline
3129	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3130	__lsx_vrotr_d(__m128i _1, __m128i _2) {
3131	return (__m128i)__builtin_lsx_vrotr_d((v2i64)_1, (v2i64)_2);
3132	}
3133
3134	extern __inline
3135	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3136	__lsx_vadd_q(__m128i _1, __m128i _2) {
3137	return (__m128i)__builtin_lsx_vadd_q((v2i64)_1, (v2i64)_2);
3138	}
3139
3140	extern __inline
3141	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3142	__lsx_vsub_q(__m128i _1, __m128i _2) {
3143	return (__m128i)__builtin_lsx_vsub_q((v2i64)_1, (v2i64)_2);
3144	}
3145
3146	#define __lsx_vldrepl_b(/void / _1, /si12*/ _2) \
3147	((__m128i)__builtin_lsx_vldrepl_b((void const *)(_1), (_2)))
3148
3149	#define __lsx_vldrepl_h(/void / _1, /si11*/ _2) \
3150	((__m128i)__builtin_lsx_vldrepl_h((void const *)(_1), (_2)))
3151
3152	#define __lsx_vldrepl_w(/void / _1, /si10*/ _2) \
3153	((__m128i)__builtin_lsx_vldrepl_w((void const *)(_1), (_2)))
3154
3155	#define __lsx_vldrepl_d(/void / _1, /si9*/ _2) \
3156	((__m128i)__builtin_lsx_vldrepl_d((void const *)(_1), (_2)))
3157
3158	extern __inline
3159	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3160	__lsx_vmskgez_b(__m128i _1) {
3161	return (__m128i)__builtin_lsx_vmskgez_b((v16i8)_1);
3162	}
3163
3164	extern __inline
3165	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3166	__lsx_vmsknz_b(__m128i _1) {
3167	return (__m128i)__builtin_lsx_vmsknz_b((v16i8)_1);
3168	}
3169
3170	extern __inline
3171	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3172	__lsx_vexth_h_b(__m128i _1) {
3173	return (__m128i)__builtin_lsx_vexth_h_b((v16i8)_1);
3174	}
3175
3176	extern __inline
3177	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3178	__lsx_vexth_w_h(__m128i _1) {
3179	return (__m128i)__builtin_lsx_vexth_w_h((v8i16)_1);
3180	}
3181
3182	extern __inline
3183	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3184	__lsx_vexth_d_w(__m128i _1) {
3185	return (__m128i)__builtin_lsx_vexth_d_w((v4i32)_1);
3186	}
3187
3188	extern __inline
3189	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3190	__lsx_vexth_q_d(__m128i _1) {
3191	return (__m128i)__builtin_lsx_vexth_q_d((v2i64)_1);
3192	}
3193
3194	extern __inline
3195	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3196	__lsx_vexth_hu_bu(__m128i _1) {
3197	return (__m128i)__builtin_lsx_vexth_hu_bu((v16u8)_1);
3198	}
3199
3200	extern __inline
3201	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3202	__lsx_vexth_wu_hu(__m128i _1) {
3203	return (__m128i)__builtin_lsx_vexth_wu_hu((v8u16)_1);
3204	}
3205
3206	extern __inline
3207	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3208	__lsx_vexth_du_wu(__m128i _1) {
3209	return (__m128i)__builtin_lsx_vexth_du_wu((v4u32)_1);
3210	}
3211
3212	extern __inline
3213	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3214	__lsx_vexth_qu_du(__m128i _1) {
3215	return (__m128i)__builtin_lsx_vexth_qu_du((v2u64)_1);
3216	}
3217
3218	#define __lsx_vrotri_b(/__m128i/ _1, /ui3/ _2) \
3219	((__m128i)__builtin_lsx_vrotri_b((v16i8)(_1), (_2)))
3220
3221	#define __lsx_vrotri_h(/__m128i/ _1, /ui4/ _2) \
3222	((__m128i)__builtin_lsx_vrotri_h((v8i16)(_1), (_2)))
3223
3224	#define __lsx_vrotri_w(/__m128i/ _1, /ui5/ _2) \
3225	((__m128i)__builtin_lsx_vrotri_w((v4i32)(_1), (_2)))
3226
3227	#define __lsx_vrotri_d(/__m128i/ _1, /ui6/ _2) \
3228	((__m128i)__builtin_lsx_vrotri_d((v2i64)(_1), (_2)))
3229
3230	extern __inline
3231	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3232	__lsx_vextl_q_d(__m128i _1) {
3233	return (__m128i)__builtin_lsx_vextl_q_d((v2i64)_1);
3234	}
3235
3236	#define __lsx_vsrlni_b_h(/__m128i/ _1, /__m128i/ _2, /ui4/ _3) \
3237	((__m128i)__builtin_lsx_vsrlni_b_h((v16i8)(_1), (v16i8)(_2), (_3)))
3238
3239	#define __lsx_vsrlni_h_w(/__m128i/ _1, /__m128i/ _2, /ui5/ _3) \
3240	((__m128i)__builtin_lsx_vsrlni_h_w((v8i16)(_1), (v8i16)(_2), (_3)))
3241
3242	#define __lsx_vsrlni_w_d(/__m128i/ _1, /__m128i/ _2, /ui6/ _3) \
3243	((__m128i)__builtin_lsx_vsrlni_w_d((v4i32)(_1), (v4i32)(_2), (_3)))
3244
3245	#define __lsx_vsrlni_d_q(/__m128i/ _1, /__m128i/ _2, /ui7/ _3) \
3246	((__m128i)__builtin_lsx_vsrlni_d_q((v2i64)(_1), (v2i64)(_2), (_3)))
3247
3248	#define __lsx_vsrlrni_b_h(/__m128i/ _1, /__m128i/ _2, /ui4/ _3) \
3249	((__m128i)__builtin_lsx_vsrlrni_b_h((v16i8)(_1), (v16i8)(_2), (_3)))
3250
3251	#define __lsx_vsrlrni_h_w(/__m128i/ _1, /__m128i/ _2, /ui5/ _3) \
3252	((__m128i)__builtin_lsx_vsrlrni_h_w((v8i16)(_1), (v8i16)(_2), (_3)))
3253
3254	#define __lsx_vsrlrni_w_d(/__m128i/ _1, /__m128i/ _2, /ui6/ _3) \
3255	((__m128i)__builtin_lsx_vsrlrni_w_d((v4i32)(_1), (v4i32)(_2), (_3)))
3256
3257	#define __lsx_vsrlrni_d_q(/__m128i/ _1, /__m128i/ _2, /ui7/ _3) \
3258	((__m128i)__builtin_lsx_vsrlrni_d_q((v2i64)(_1), (v2i64)(_2), (_3)))
3259
3260	#define __lsx_vssrlni_b_h(/__m128i/ _1, /__m128i/ _2, /ui4/ _3) \
3261	((__m128i)__builtin_lsx_vssrlni_b_h((v16i8)(_1), (v16i8)(_2), (_3)))
3262
3263	#define __lsx_vssrlni_h_w(/__m128i/ _1, /__m128i/ _2, /ui5/ _3) \
3264	((__m128i)__builtin_lsx_vssrlni_h_w((v8i16)(_1), (v8i16)(_2), (_3)))
3265
3266	#define __lsx_vssrlni_w_d(/__m128i/ _1, /__m128i/ _2, /ui6/ _3) \
3267	((__m128i)__builtin_lsx_vssrlni_w_d((v4i32)(_1), (v4i32)(_2), (_3)))
3268
3269	#define __lsx_vssrlni_d_q(/__m128i/ _1, /__m128i/ _2, /ui7/ _3) \
3270	((__m128i)__builtin_lsx_vssrlni_d_q((v2i64)(_1), (v2i64)(_2), (_3)))
3271
3272	#define __lsx_vssrlni_bu_h(/__m128i/ _1, /__m128i/ _2, /ui4/ _3) \
3273	((__m128i)__builtin_lsx_vssrlni_bu_h((v16u8)(_1), (v16i8)(_2), (_3)))
3274
3275	#define __lsx_vssrlni_hu_w(/__m128i/ _1, /__m128i/ _2, /ui5/ _3) \
3276	((__m128i)__builtin_lsx_vssrlni_hu_w((v8u16)(_1), (v8i16)(_2), (_3)))
3277
3278	#define __lsx_vssrlni_wu_d(/__m128i/ _1, /__m128i/ _2, /ui6/ _3) \
3279	((__m128i)__builtin_lsx_vssrlni_wu_d((v4u32)(_1), (v4i32)(_2), (_3)))
3280
3281	#define __lsx_vssrlni_du_q(/__m128i/ _1, /__m128i/ _2, /ui7/ _3) \
3282	((__m128i)__builtin_lsx_vssrlni_du_q((v2u64)(_1), (v2i64)(_2), (_3)))
3283
3284	#define __lsx_vssrlrni_b_h(/__m128i/ _1, /__m128i/ _2, /ui4/ _3) \
3285	((__m128i)__builtin_lsx_vssrlrni_b_h((v16i8)(_1), (v16i8)(_2), (_3)))
3286
3287	#define __lsx_vssrlrni_h_w(/__m128i/ _1, /__m128i/ _2, /ui5/ _3) \
3288	((__m128i)__builtin_lsx_vssrlrni_h_w((v8i16)(_1), (v8i16)(_2), (_3)))
3289
3290	#define __lsx_vssrlrni_w_d(/__m128i/ _1, /__m128i/ _2, /ui6/ _3) \
3291	((__m128i)__builtin_lsx_vssrlrni_w_d((v4i32)(_1), (v4i32)(_2), (_3)))
3292
3293	#define __lsx_vssrlrni_d_q(/__m128i/ _1, /__m128i/ _2, /ui7/ _3) \
3294	((__m128i)__builtin_lsx_vssrlrni_d_q((v2i64)(_1), (v2i64)(_2), (_3)))
3295
3296	#define __lsx_vssrlrni_bu_h(/__m128i/ _1, /__m128i/ _2, /ui4/ _3) \
3297	((__m128i)__builtin_lsx_vssrlrni_bu_h((v16u8)(_1), (v16i8)(_2), (_3)))
3298
3299	#define __lsx_vssrlrni_hu_w(/__m128i/ _1, /__m128i/ _2, /ui5/ _3) \
3300	((__m128i)__builtin_lsx_vssrlrni_hu_w((v8u16)(_1), (v8i16)(_2), (_3)))
3301
3302	#define __lsx_vssrlrni_wu_d(/__m128i/ _1, /__m128i/ _2, /ui6/ _3) \
3303	((__m128i)__builtin_lsx_vssrlrni_wu_d((v4u32)(_1), (v4i32)(_2), (_3)))
3304
3305	#define __lsx_vssrlrni_du_q(/__m128i/ _1, /__m128i/ _2, /ui7/ _3) \
3306	((__m128i)__builtin_lsx_vssrlrni_du_q((v2u64)(_1), (v2i64)(_2), (_3)))
3307
3308	#define __lsx_vsrani_b_h(/__m128i/ _1, /__m128i/ _2, /ui4/ _3) \
3309	((__m128i)__builtin_lsx_vsrani_b_h((v16i8)(_1), (v16i8)(_2), (_3)))
3310
3311	#define __lsx_vsrani_h_w(/__m128i/ _1, /__m128i/ _2, /ui5/ _3) \
3312	((__m128i)__builtin_lsx_vsrani_h_w((v8i16)(_1), (v8i16)(_2), (_3)))
3313
3314	#define __lsx_vsrani_w_d(/__m128i/ _1, /__m128i/ _2, /ui6/ _3) \
3315	((__m128i)__builtin_lsx_vsrani_w_d((v4i32)(_1), (v4i32)(_2), (_3)))
3316
3317	#define __lsx_vsrani_d_q(/__m128i/ _1, /__m128i/ _2, /ui7/ _3) \
3318	((__m128i)__builtin_lsx_vsrani_d_q((v2i64)(_1), (v2i64)(_2), (_3)))
3319
3320	#define __lsx_vsrarni_b_h(/__m128i/ _1, /__m128i/ _2, /ui4/ _3) \
3321	((__m128i)__builtin_lsx_vsrarni_b_h((v16i8)(_1), (v16i8)(_2), (_3)))
3322
3323	#define __lsx_vsrarni_h_w(/__m128i/ _1, /__m128i/ _2, /ui5/ _3) \
3324	((__m128i)__builtin_lsx_vsrarni_h_w((v8i16)(_1), (v8i16)(_2), (_3)))
3325
3326	#define __lsx_vsrarni_w_d(/__m128i/ _1, /__m128i/ _2, /ui6/ _3) \
3327	((__m128i)__builtin_lsx_vsrarni_w_d((v4i32)(_1), (v4i32)(_2), (_3)))
3328
3329	#define __lsx_vsrarni_d_q(/__m128i/ _1, /__m128i/ _2, /ui7/ _3) \
3330	((__m128i)__builtin_lsx_vsrarni_d_q((v2i64)(_1), (v2i64)(_2), (_3)))
3331
3332	#define __lsx_vssrani_b_h(/__m128i/ _1, /__m128i/ _2, /ui4/ _3) \
3333	((__m128i)__builtin_lsx_vssrani_b_h((v16i8)(_1), (v16i8)(_2), (_3)))
3334
3335	#define __lsx_vssrani_h_w(/__m128i/ _1, /__m128i/ _2, /ui5/ _3) \
3336	((__m128i)__builtin_lsx_vssrani_h_w((v8i16)(_1), (v8i16)(_2), (_3)))
3337
3338	#define __lsx_vssrani_w_d(/__m128i/ _1, /__m128i/ _2, /ui6/ _3) \
3339	((__m128i)__builtin_lsx_vssrani_w_d((v4i32)(_1), (v4i32)(_2), (_3)))
3340
3341	#define __lsx_vssrani_d_q(/__m128i/ _1, /__m128i/ _2, /ui7/ _3) \
3342	((__m128i)__builtin_lsx_vssrani_d_q((v2i64)(_1), (v2i64)(_2), (_3)))
3343
3344	#define __lsx_vssrani_bu_h(/__m128i/ _1, /__m128i/ _2, /ui4/ _3) \
3345	((__m128i)__builtin_lsx_vssrani_bu_h((v16u8)(_1), (v16i8)(_2), (_3)))
3346
3347	#define __lsx_vssrani_hu_w(/__m128i/ _1, /__m128i/ _2, /ui5/ _3) \
3348	((__m128i)__builtin_lsx_vssrani_hu_w((v8u16)(_1), (v8i16)(_2), (_3)))
3349
3350	#define __lsx_vssrani_wu_d(/__m128i/ _1, /__m128i/ _2, /ui6/ _3) \
3351	((__m128i)__builtin_lsx_vssrani_wu_d((v4u32)(_1), (v4i32)(_2), (_3)))
3352
3353	#define __lsx_vssrani_du_q(/__m128i/ _1, /__m128i/ _2, /ui7/ _3) \
3354	((__m128i)__builtin_lsx_vssrani_du_q((v2u64)(_1), (v2i64)(_2), (_3)))
3355
3356	#define __lsx_vssrarni_b_h(/__m128i/ _1, /__m128i/ _2, /ui4/ _3) \
3357	((__m128i)__builtin_lsx_vssrarni_b_h((v16i8)(_1), (v16i8)(_2), (_3)))
3358
3359	#define __lsx_vssrarni_h_w(/__m128i/ _1, /__m128i/ _2, /ui5/ _3) \
3360	((__m128i)__builtin_lsx_vssrarni_h_w((v8i16)(_1), (v8i16)(_2), (_3)))
3361
3362	#define __lsx_vssrarni_w_d(/__m128i/ _1, /__m128i/ _2, /ui6/ _3) \
3363	((__m128i)__builtin_lsx_vssrarni_w_d((v4i32)(_1), (v4i32)(_2), (_3)))
3364
3365	#define __lsx_vssrarni_d_q(/__m128i/ _1, /__m128i/ _2, /ui7/ _3) \
3366	((__m128i)__builtin_lsx_vssrarni_d_q((v2i64)(_1), (v2i64)(_2), (_3)))
3367
3368	#define __lsx_vssrarni_bu_h(/__m128i/ _1, /__m128i/ _2, /ui4/ _3) \
3369	((__m128i)__builtin_lsx_vssrarni_bu_h((v16u8)(_1), (v16i8)(_2), (_3)))
3370
3371	#define __lsx_vssrarni_hu_w(/__m128i/ _1, /__m128i/ _2, /ui5/ _3) \
3372	((__m128i)__builtin_lsx_vssrarni_hu_w((v8u16)(_1), (v8i16)(_2), (_3)))
3373
3374	#define __lsx_vssrarni_wu_d(/__m128i/ _1, /__m128i/ _2, /ui6/ _3) \
3375	((__m128i)__builtin_lsx_vssrarni_wu_d((v4u32)(_1), (v4i32)(_2), (_3)))
3376
3377	#define __lsx_vssrarni_du_q(/__m128i/ _1, /__m128i/ _2, /ui7/ _3) \
3378	((__m128i)__builtin_lsx_vssrarni_du_q((v2u64)(_1), (v2i64)(_2), (_3)))
3379
3380	#define __lsx_vpermi_w(/__m128i/ _1, /__m128i/ _2, /ui8/ _3) \
3381	((__m128i)__builtin_lsx_vpermi_w((v4i32)(_1), (v4i32)(_2), (_3)))
3382
3383	#define __lsx_vld(/void / _1, /si12*/ _2) \
3384	((__m128i)__builtin_lsx_vld((void const *)(_1), (_2)))
3385
3386	#define __lsx_vst(/__m128i/ _1, /void / _2, /si12*/ _3) \
3387	((void)__builtin_lsx_vst((v16i8)(_1), (void *)(_2), (_3)))
3388
3389	extern __inline
3390	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3391	__lsx_vssrlrn_b_h(__m128i _1, __m128i _2) {
3392	return (__m128i)__builtin_lsx_vssrlrn_b_h((v8i16)_1, (v8i16)_2);
3393	}
3394
3395	extern __inline
3396	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3397	__lsx_vssrlrn_h_w(__m128i _1, __m128i _2) {
3398	return (__m128i)__builtin_lsx_vssrlrn_h_w((v4i32)_1, (v4i32)_2);
3399	}
3400
3401	extern __inline
3402	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3403	__lsx_vssrlrn_w_d(__m128i _1, __m128i _2) {
3404	return (__m128i)__builtin_lsx_vssrlrn_w_d((v2i64)_1, (v2i64)_2);
3405	}
3406
3407	extern __inline
3408	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3409	__lsx_vssrln_b_h(__m128i _1, __m128i _2) {
3410	return (__m128i)__builtin_lsx_vssrln_b_h((v8i16)_1, (v8i16)_2);
3411	}
3412
3413	extern __inline
3414	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3415	__lsx_vssrln_h_w(__m128i _1, __m128i _2) {
3416	return (__m128i)__builtin_lsx_vssrln_h_w((v4i32)_1, (v4i32)_2);
3417	}
3418
3419	extern __inline
3420	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3421	__lsx_vssrln_w_d(__m128i _1, __m128i _2) {
3422	return (__m128i)__builtin_lsx_vssrln_w_d((v2i64)_1, (v2i64)_2);
3423	}
3424
3425	extern __inline
3426	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3427	__lsx_vorn_v(__m128i _1, __m128i _2) {
3428	return (__m128i)__builtin_lsx_vorn_v((v16i8)_1, (v16i8)_2);
3429	}
3430
3431	#define __lsx_vldi(/i13/ _1) ((__m128i)__builtin_lsx_vldi((_1)))
3432
3433	extern __inline
3434	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3435	__lsx_vshuf_b(__m128i _1, __m128i _2, __m128i _3) {
3436	return (__m128i)__builtin_lsx_vshuf_b((v16i8)_1, (v16i8)_2, (v16i8)_3);
3437	}
3438
3439	extern __inline
3440	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3441	__lsx_vldx(void const *_1, long int _2) {
3442	return (__m128i)__builtin_lsx_vldx((void const *)_1, (long int)_2);
3443	}
3444
3445	extern __inline
3446	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) void
3447	__lsx_vstx(__m128i _1, void *_2, long int _3) {
3448	return (void)__builtin_lsx_vstx((v16i8)_1, (void *)_2, (long int)_3);
3449	}
3450
3451	extern __inline
3452	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3453	__lsx_vextl_qu_du(__m128i _1) {
3454	return (__m128i)__builtin_lsx_vextl_qu_du((v2u64)_1);
3455	}
3456
3457	#define __lsx_bnz_b(/__m128i/ _1) ((int)__builtin_lsx_bnz_b((v16u8)(_1)))
3458
3459	#define __lsx_bnz_d(/__m128i/ _1) ((int)__builtin_lsx_bnz_d((v2u64)(_1)))
3460
3461	#define __lsx_bnz_h(/__m128i/ _1) ((int)__builtin_lsx_bnz_h((v8u16)(_1)))
3462
3463	#define __lsx_bnz_v(/__m128i/ _1) ((int)__builtin_lsx_bnz_v((v16u8)(_1)))
3464
3465	#define __lsx_bnz_w(/__m128i/ _1) ((int)__builtin_lsx_bnz_w((v4u32)(_1)))
3466
3467	#define __lsx_bz_b(/__m128i/ _1) ((int)__builtin_lsx_bz_b((v16u8)(_1)))
3468
3469	#define __lsx_bz_d(/__m128i/ _1) ((int)__builtin_lsx_bz_d((v2u64)(_1)))
3470
3471	#define __lsx_bz_h(/__m128i/ _1) ((int)__builtin_lsx_bz_h((v8u16)(_1)))
3472
3473	#define __lsx_bz_v(/__m128i/ _1) ((int)__builtin_lsx_bz_v((v16u8)(_1)))
3474
3475	#define __lsx_bz_w(/__m128i/ _1) ((int)__builtin_lsx_bz_w((v4u32)(_1)))
3476
3477	extern __inline
3478	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3479	__lsx_vfcmp_caf_d(__m128d _1, __m128d _2) {
3480	return (__m128i)__builtin_lsx_vfcmp_caf_d((v2f64)_1, (v2f64)_2);
3481	}
3482
3483	extern __inline
3484	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3485	__lsx_vfcmp_caf_s(__m128 _1, __m128 _2) {
3486	return (__m128i)__builtin_lsx_vfcmp_caf_s((v4f32)_1, (v4f32)_2);
3487	}
3488
3489	extern __inline
3490	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3491	__lsx_vfcmp_ceq_d(__m128d _1, __m128d _2) {
3492	return (__m128i)__builtin_lsx_vfcmp_ceq_d((v2f64)_1, (v2f64)_2);
3493	}
3494
3495	extern __inline
3496	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3497	__lsx_vfcmp_ceq_s(__m128 _1, __m128 _2) {
3498	return (__m128i)__builtin_lsx_vfcmp_ceq_s((v4f32)_1, (v4f32)_2);
3499	}
3500
3501	extern __inline
3502	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3503	__lsx_vfcmp_cle_d(__m128d _1, __m128d _2) {
3504	return (__m128i)__builtin_lsx_vfcmp_cle_d((v2f64)_1, (v2f64)_2);
3505	}
3506
3507	extern __inline
3508	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3509	__lsx_vfcmp_cle_s(__m128 _1, __m128 _2) {
3510	return (__m128i)__builtin_lsx_vfcmp_cle_s((v4f32)_1, (v4f32)_2);
3511	}
3512
3513	extern __inline
3514	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3515	__lsx_vfcmp_clt_d(__m128d _1, __m128d _2) {
3516	return (__m128i)__builtin_lsx_vfcmp_clt_d((v2f64)_1, (v2f64)_2);
3517	}
3518
3519	extern __inline
3520	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3521	__lsx_vfcmp_clt_s(__m128 _1, __m128 _2) {
3522	return (__m128i)__builtin_lsx_vfcmp_clt_s((v4f32)_1, (v4f32)_2);
3523	}
3524
3525	extern __inline
3526	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3527	__lsx_vfcmp_cne_d(__m128d _1, __m128d _2) {
3528	return (__m128i)__builtin_lsx_vfcmp_cne_d((v2f64)_1, (v2f64)_2);
3529	}
3530
3531	extern __inline
3532	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3533	__lsx_vfcmp_cne_s(__m128 _1, __m128 _2) {
3534	return (__m128i)__builtin_lsx_vfcmp_cne_s((v4f32)_1, (v4f32)_2);
3535	}
3536
3537	extern __inline
3538	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3539	__lsx_vfcmp_cor_d(__m128d _1, __m128d _2) {
3540	return (__m128i)__builtin_lsx_vfcmp_cor_d((v2f64)_1, (v2f64)_2);
3541	}
3542
3543	extern __inline
3544	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3545	__lsx_vfcmp_cor_s(__m128 _1, __m128 _2) {
3546	return (__m128i)__builtin_lsx_vfcmp_cor_s((v4f32)_1, (v4f32)_2);
3547	}
3548
3549	extern __inline
3550	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3551	__lsx_vfcmp_cueq_d(__m128d _1, __m128d _2) {
3552	return (__m128i)__builtin_lsx_vfcmp_cueq_d((v2f64)_1, (v2f64)_2);
3553	}
3554
3555	extern __inline
3556	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3557	__lsx_vfcmp_cueq_s(__m128 _1, __m128 _2) {
3558	return (__m128i)__builtin_lsx_vfcmp_cueq_s((v4f32)_1, (v4f32)_2);
3559	}
3560
3561	extern __inline
3562	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3563	__lsx_vfcmp_cule_d(__m128d _1, __m128d _2) {
3564	return (__m128i)__builtin_lsx_vfcmp_cule_d((v2f64)_1, (v2f64)_2);
3565	}
3566
3567	extern __inline
3568	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3569	__lsx_vfcmp_cule_s(__m128 _1, __m128 _2) {
3570	return (__m128i)__builtin_lsx_vfcmp_cule_s((v4f32)_1, (v4f32)_2);
3571	}
3572
3573	extern __inline
3574	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3575	__lsx_vfcmp_cult_d(__m128d _1, __m128d _2) {
3576	return (__m128i)__builtin_lsx_vfcmp_cult_d((v2f64)_1, (v2f64)_2);
3577	}
3578
3579	extern __inline
3580	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3581	__lsx_vfcmp_cult_s(__m128 _1, __m128 _2) {
3582	return (__m128i)__builtin_lsx_vfcmp_cult_s((v4f32)_1, (v4f32)_2);
3583	}
3584
3585	extern __inline
3586	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3587	__lsx_vfcmp_cun_d(__m128d _1, __m128d _2) {
3588	return (__m128i)__builtin_lsx_vfcmp_cun_d((v2f64)_1, (v2f64)_2);
3589	}
3590
3591	extern __inline
3592	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3593	__lsx_vfcmp_cune_d(__m128d _1, __m128d _2) {
3594	return (__m128i)__builtin_lsx_vfcmp_cune_d((v2f64)_1, (v2f64)_2);
3595	}
3596
3597	extern __inline
3598	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3599	__lsx_vfcmp_cune_s(__m128 _1, __m128 _2) {
3600	return (__m128i)__builtin_lsx_vfcmp_cune_s((v4f32)_1, (v4f32)_2);
3601	}
3602
3603	extern __inline
3604	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3605	__lsx_vfcmp_cun_s(__m128 _1, __m128 _2) {
3606	return (__m128i)__builtin_lsx_vfcmp_cun_s((v4f32)_1, (v4f32)_2);
3607	}
3608
3609	extern __inline
3610	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3611	__lsx_vfcmp_saf_d(__m128d _1, __m128d _2) {
3612	return (__m128i)__builtin_lsx_vfcmp_saf_d((v2f64)_1, (v2f64)_2);
3613	}
3614
3615	extern __inline
3616	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3617	__lsx_vfcmp_saf_s(__m128 _1, __m128 _2) {
3618	return (__m128i)__builtin_lsx_vfcmp_saf_s((v4f32)_1, (v4f32)_2);
3619	}
3620
3621	extern __inline
3622	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3623	__lsx_vfcmp_seq_d(__m128d _1, __m128d _2) {
3624	return (__m128i)__builtin_lsx_vfcmp_seq_d((v2f64)_1, (v2f64)_2);
3625	}
3626
3627	extern __inline
3628	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3629	__lsx_vfcmp_seq_s(__m128 _1, __m128 _2) {
3630	return (__m128i)__builtin_lsx_vfcmp_seq_s((v4f32)_1, (v4f32)_2);
3631	}
3632
3633	extern __inline
3634	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3635	__lsx_vfcmp_sle_d(__m128d _1, __m128d _2) {
3636	return (__m128i)__builtin_lsx_vfcmp_sle_d((v2f64)_1, (v2f64)_2);
3637	}
3638
3639	extern __inline
3640	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3641	__lsx_vfcmp_sle_s(__m128 _1, __m128 _2) {
3642	return (__m128i)__builtin_lsx_vfcmp_sle_s((v4f32)_1, (v4f32)_2);
3643	}
3644
3645	extern __inline
3646	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3647	__lsx_vfcmp_slt_d(__m128d _1, __m128d _2) {
3648	return (__m128i)__builtin_lsx_vfcmp_slt_d((v2f64)_1, (v2f64)_2);
3649	}
3650
3651	extern __inline
3652	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3653	__lsx_vfcmp_slt_s(__m128 _1, __m128 _2) {
3654	return (__m128i)__builtin_lsx_vfcmp_slt_s((v4f32)_1, (v4f32)_2);
3655	}
3656
3657	extern __inline
3658	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3659	__lsx_vfcmp_sne_d(__m128d _1, __m128d _2) {
3660	return (__m128i)__builtin_lsx_vfcmp_sne_d((v2f64)_1, (v2f64)_2);
3661	}
3662
3663	extern __inline
3664	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3665	__lsx_vfcmp_sne_s(__m128 _1, __m128 _2) {
3666	return (__m128i)__builtin_lsx_vfcmp_sne_s((v4f32)_1, (v4f32)_2);
3667	}
3668
3669	extern __inline
3670	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3671	__lsx_vfcmp_sor_d(__m128d _1, __m128d _2) {
3672	return (__m128i)__builtin_lsx_vfcmp_sor_d((v2f64)_1, (v2f64)_2);
3673	}
3674
3675	extern __inline
3676	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3677	__lsx_vfcmp_sor_s(__m128 _1, __m128 _2) {
3678	return (__m128i)__builtin_lsx_vfcmp_sor_s((v4f32)_1, (v4f32)_2);
3679	}
3680
3681	extern __inline
3682	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3683	__lsx_vfcmp_sueq_d(__m128d _1, __m128d _2) {
3684	return (__m128i)__builtin_lsx_vfcmp_sueq_d((v2f64)_1, (v2f64)_2);
3685	}
3686
3687	extern __inline
3688	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3689	__lsx_vfcmp_sueq_s(__m128 _1, __m128 _2) {
3690	return (__m128i)__builtin_lsx_vfcmp_sueq_s((v4f32)_1, (v4f32)_2);
3691	}
3692
3693	extern __inline
3694	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3695	__lsx_vfcmp_sule_d(__m128d _1, __m128d _2) {
3696	return (__m128i)__builtin_lsx_vfcmp_sule_d((v2f64)_1, (v2f64)_2);
3697	}
3698
3699	extern __inline
3700	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3701	__lsx_vfcmp_sule_s(__m128 _1, __m128 _2) {
3702	return (__m128i)__builtin_lsx_vfcmp_sule_s((v4f32)_1, (v4f32)_2);
3703	}
3704
3705	extern __inline
3706	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3707	__lsx_vfcmp_sult_d(__m128d _1, __m128d _2) {
3708	return (__m128i)__builtin_lsx_vfcmp_sult_d((v2f64)_1, (v2f64)_2);
3709	}
3710
3711	extern __inline
3712	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3713	__lsx_vfcmp_sult_s(__m128 _1, __m128 _2) {
3714	return (__m128i)__builtin_lsx_vfcmp_sult_s((v4f32)_1, (v4f32)_2);
3715	}
3716
3717	extern __inline
3718	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3719	__lsx_vfcmp_sun_d(__m128d _1, __m128d _2) {
3720	return (__m128i)__builtin_lsx_vfcmp_sun_d((v2f64)_1, (v2f64)_2);
3721	}
3722
3723	extern __inline
3724	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3725	__lsx_vfcmp_sune_d(__m128d _1, __m128d _2) {
3726	return (__m128i)__builtin_lsx_vfcmp_sune_d((v2f64)_1, (v2f64)_2);
3727	}
3728
3729	extern __inline
3730	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3731	__lsx_vfcmp_sune_s(__m128 _1, __m128 _2) {
3732	return (__m128i)__builtin_lsx_vfcmp_sune_s((v4f32)_1, (v4f32)_2);
3733	}
3734
3735	extern __inline
3736	__attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
3737	__lsx_vfcmp_sun_s(__m128 _1, __m128 _2) {
3738	return (__m128i)__builtin_lsx_vfcmp_sun_s((v4f32)_1, (v4f32)_2);
3739	}
3740
3741	#define __lsx_vrepli_b(/si10/ _1) ((__m128i)__builtin_lsx_vrepli_b((_1)))
3742
3743	#define __lsx_vrepli_d(/si10/ _1) ((__m128i)__builtin_lsx_vrepli_d((_1)))
3744
3745	#define __lsx_vrepli_h(/si10/ _1) ((__m128i)__builtin_lsx_vrepli_h((_1)))
3746
3747	#define __lsx_vrepli_w(/si10/ _1) ((__m128i)__builtin_lsx_vrepli_w((_1)))
3748
3749	#endif /* defined(__loongarch_sx) */
3750	#endif /* _LOONGSON_SXINTRIN_H */

lib/include/mmintrin.h+7-5

 @@ -22,7 +22,9 @@ typedef short __v4hi __attribute__((__vector_size__(8)));
 typedef char __v8qi __attribute__((__vector_size__(8)));
 /* Define the default attributes for the functions in this file. */
-#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("mmx"), __min_vector_width__(64)))
+#define __DEFAULT_FN_ATTRS                                                     \
+  __attribute__((__always_inline__, __nodebug__, __target__("mmx,no-evex512"), \
+                 __min_vector_width__(64)))
 /// Clears the MMX state by setting the state of the x87 stack registers
 ///    to empty.
 @@ -31,10 +33,10 @@ typedef char __v8qi __attribute__((__vector_size__(8)));
 ///
 /// This intrinsic corresponds to the <c> EMMS </c> instruction.
 ///
-static __inline__ void  __attribute__((__always_inline__, __nodebug__, __target__("mmx")))
+static __inline__ void __attribute__((__always_inline__, __nodebug__,
-_mm_empty(void)
+                                      __target__("mmx,no-evex512")))
-{
+_mm_empty(void) {
-    __builtin_ia32_emms();
+  __builtin_ia32_emms();
 }
 /// Constructs a 64-bit integer vector, setting the lower 32 bits to the

lib/include/module.modulemap+155-2

...	@@ -153,9 +153,162 @@ module _Builtin_intrinsics [system] [extern_c] {	...	@@ -153,9 +153,162 @@ module _Builtin_intrinsics [system] [extern_c] {
153	}	153	}
154	}	154	}
155		155
156	module _Builtin_stddef_max_align_t [system] [extern_c] {	156	// Start -fbuiltin-headers-in-system-modules affected modules
157	header "__stddef_max_align_t.h"	157
		158	// The following modules all ignore their headers when
		159	// -fbuiltin-headers-in-system-modules is passed, and many of
		160	// those headers join system modules when present.
		161
		162	// e.g. if -fbuiltin-headers-in-system-modules is passed, then
		163	// float.h will not be in the _Builtin_float module (that module
		164	// will be empty). If there is a system module that declares
		165	// `header "float.h"`, then the builtin float.h will join
		166	// that module. The system float.h (if present) will be treated
		167	// as a textual header in the sytem module.
		168	module _Builtin_float [system] {
		169	header "float.h"
		170	export *
		171	}
		172
		173	module _Builtin_inttypes [system] {
		174	header "inttypes.h"
		175	export *
		176	}
		177
		178	module _Builtin_iso646 [system] {
		179	header "iso646.h"
		180	export *
		181	}
		182
		183	module _Builtin_limits [system] {
		184	header "limits.h"
		185	export *
		186	}
		187
		188	module _Builtin_stdalign [system] {
		189	header "stdalign.h"
		190	export *
		191	}
		192
		193	module _Builtin_stdarg [system] {
		194	textual header "stdarg.h"
		195
		196	explicit module __gnuc_va_list {
		197	header "__stdarg___gnuc_va_list.h"
		198	export *
		199	}
		200
		201	explicit module __va_copy {
		202	header "__stdarg___va_copy.h"
		203	export *
		204	}
		205
		206	explicit module va_arg {
		207	header "__stdarg_va_arg.h"
		208	export *
		209	}
		210
		211	explicit module va_copy {
		212	header "__stdarg_va_copy.h"
		213	export *
		214	}
		215
		216	explicit module va_list {
		217	header "__stdarg_va_list.h"
		218	export *
		219	}
		220	}
		221
		222	module _Builtin_stdatomic [system] {
		223	header "stdatomic.h"
		224	export *
		225	}
		226
		227	module _Builtin_stdbool [system] {
		228	header "stdbool.h"
		229	export *
		230	}
		231
		232	module _Builtin_stddef [system] {
		233	textual header "stddef.h"
		234
		235	// __stddef_max_align_t.h is always in this module, even if
		236	// -fbuiltin-headers-in-system-modules is passed.
		237	explicit module max_align_t {
		238	header "__stddef_max_align_t.h"
		239	export *
		240	}
		241
		242	explicit module null {
		243	header "__stddef_null.h"
		244	export *
		245	}
		246
		247	explicit module nullptr_t {
		248	header "__stddef_nullptr_t.h"
		249	export *
		250	}
		251
		252	explicit module offsetof {
		253	header "__stddef_offsetof.h"
		254	export *
		255	}
		256
		257	explicit module ptrdiff_t {
		258	header "__stddef_ptrdiff_t.h"
		259	export *
		260	}
		261
		262	explicit module rsize_t {
		263	header "__stddef_rsize_t.h"
		264	export *
		265	}
		266
		267	explicit module size_t {
		268	header "__stddef_size_t.h"
		269	export *
		270	}
		271
		272	explicit module unreachable {
		273	header "__stddef_unreachable.h"
		274	export *
		275	}
		276
		277	explicit module wchar_t {
		278	header "__stddef_wchar_t.h"
		279	export *
		280	}
		281	}
		282
		283	// wint_t is provided by <wchar.h> and not <stddef.h>. It's here
		284	// for compatibility, but must be explicitly requested. Therefore
		285	// __stddef_wint_t.h is not part of _Builtin_stddef. It is always in
		286	// this module even if -fbuiltin-headers-in-system-modules is passed.
		287	module _Builtin_stddef_wint_t [system] {
		288	header "__stddef_wint_t.h"
		289	export *
		290	}
		291
		292	module _Builtin_stdint [system] {
		293	header "stdint.h"
		294	export *
		295	}
		296
		297	module _Builtin_stdnoreturn [system] {
		298	header "stdnoreturn.h"
		299	export *
		300	}
		301
		302	module _Builtin_tgmath [system] {
		303	header "tgmath.h"
		304	export *
		305	}
		306
		307	module _Builtin_unwind [system] {
		308	header "unwind.h"
		309	export *
158	}	310	}
		311	// End -fbuiltin-headers-in-system-modules affected modules
159		312
160	module opencl_c {	313	module opencl_c {
161	requires opencl	314	requires opencl

lib/include/opencl-c-base.h+5

lib/include/openmp_wrappers/cmath+1-1

...	@@ -1,4 +1,4 @@	...	@@ -1,4 +1,4 @@
1	/*===-- __clang_openmp_device_functions.h - OpenMP math declares ------ c++ -===	1	/===-- __clang_openmp_device_functions.h - OpenMP math declares -- c++ -*-===
2	*	2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.	3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.	4	* See https://llvm.org/LICENSE.txt for license information.

lib/include/pmmintrin.h+3-2

 @@ -17,8 +17,9 @@
 #include <emmintrin.h>
 /* Define the default attributes for the functions in this file. */
-#define __DEFAULT_FN_ATTRS \
+#define __DEFAULT_FN_ATTRS                                                     \
-  __attribute__((__always_inline__, __nodebug__, __target__("sse3"), __min_vector_width__(128)))
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("sse3,no-evex512"), __min_vector_width__(128)))
 /// Loads data from an unaligned memory location to elements in a 128-bit
 ///    vector.

lib/include/ppc_wrappers/nmmintrin.h created+26

...	@@ -0,0 +1,26 @@
1	/*===---- nmmintrin.h - Implementation of SSE4 intrinsics on PowerPC -------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	#ifndef NO_WARN_X86_INTRINSICS
11	/* This header is distributed to simplify porting x86_64 code that
12	makes explicit use of Intel intrinsics to powerpc64le.
13	It is the user's responsibility to determine if the results are
14	acceptable and make additional changes as necessary.
15	Note that much code that uses Intel intrinsics can be rewritten in
16	standard C or GNU C extensions, which are more portable and better
17	optimized across multiple targets. */
18	#endif
19
20	#ifndef NMMINTRIN_H_
21	#define NMMINTRIN_H_
22
23	/* We just include SSE4.1 header file. */
24	#include <smmintrin.h>
25
26	#endif /* NMMINTRIN_H_ */

lib/include/ppc_wrappers/smmintrin.h+35-15

 @@ -14,7 +14,7 @@
 #ifndef NO_WARN_X86_INTRINSICS
 /* This header is distributed to simplify porting x86_64 code that
-   makes explicit use of Intel intrinsics to powerp64/powerpc64le.
+   makes explicit use of Intel intrinsics to powerpc64/powerpc64le.
    It is the user's responsibility to determine if the results are
    acceptable and make additional changes as necessary.
 @@ -68,10 +68,10 @@ extern __inline __m128d
     __asm__("mffsce %0" : "=f"(__fpscr_save.__fr));
     __enables_save.__fpscr = __fpscr_save.__fpscr & 0xf8;
 #else
-    __fpscr_save.__fr = __builtin_mffs();
+    __fpscr_save.__fr = __builtin_ppc_mffs();
     __enables_save.__fpscr = __fpscr_save.__fpscr & 0xf8;
     __fpscr_save.__fpscr &= ~0xf8;
-    __builtin_mtfsf(0b00000011, __fpscr_save.__fr);
+    __builtin_ppc_mtfsf(0b00000011, __fpscr_save.__fr);
 #endif
     /* Insert an artificial "read/write" reference to the variable
        read below, to ensure the compiler does not schedule
 @@ -83,10 +83,15 @@ extern __inline __m128d
   switch (__rounding) {
   case _MM_FROUND_TO_NEAREST_INT:
-    __fpscr_save.__fr = __builtin_mffsl();
+#ifdef _ARCH_PWR9
+    __fpscr_save.__fr = __builtin_ppc_mffsl();
+#else
+    __fpscr_save.__fr = __builtin_ppc_mffs();
+    __fpscr_save.__fpscr &= 0x70007f0ffL;
+#endif
     __attribute__((fallthrough));
   case _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC:
-    __builtin_set_fpscr_rn(0b00);
+    __builtin_ppc_set_fpscr_rn(0b00);
     /* Insert an artificial "read/write" reference to the variable
        read below, to ensure the compiler does not schedule
        a read/use of the variable before the FPSCR is modified, above.
 @@ -102,7 +107,7 @@ extern __inline __m128d
        This can be removed if and when GCC PR102783 is fixed.
      */
     __asm__("" : : "wa"(__r));
-    __builtin_set_fpscr_rn(__fpscr_save.__fpscr);
+    __builtin_ppc_set_fpscr_rn(__fpscr_save.__fpscr);
     break;
   case _MM_FROUND_TO_NEG_INF:
   case _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC:
 @@ -128,9 +133,14 @@ extern __inline __m128d
      */
     __asm__("" : : "wa"(__r));
     /* Restore enabled exceptions.  */
-    __fpscr_save.__fr = __builtin_mffsl();
+#ifdef _ARCH_PWR9
+    __fpscr_save.__fr = __builtin_ppc_mffsl();
+#else
+    __fpscr_save.__fr = __builtin_ppc_mffs();
+    __fpscr_save.__fpscr &= 0x70007f0ffL;
+#endif
     __fpscr_save.__fpscr |= __enables_save.__fpscr;
-    __builtin_mtfsf(0b00000011, __fpscr_save.__fr);
+    __builtin_ppc_mtfsf(0b00000011, __fpscr_save.__fr);
   }
   return (__m128d)__r;
 }
 @@ -159,10 +169,10 @@ extern __inline __m128
     __asm__("mffsce %0" : "=f"(__fpscr_save.__fr));
     __enables_save.__fpscr = __fpscr_save.__fpscr & 0xf8;
 #else
-    __fpscr_save.__fr = __builtin_mffs();
+    __fpscr_save.__fr = __builtin_ppc_mffs();
     __enables_save.__fpscr = __fpscr_save.__fpscr & 0xf8;
     __fpscr_save.__fpscr &= ~0xf8;
-    __builtin_mtfsf(0b00000011, __fpscr_save.__fr);
+    __builtin_ppc_mtfsf(0b00000011, __fpscr_save.__fr);
 #endif
     /* Insert an artificial "read/write" reference to the variable
        read below, to ensure the compiler does not schedule
 @@ -174,10 +184,15 @@ extern __inline __m128
   switch (__rounding) {
   case _MM_FROUND_TO_NEAREST_INT:
-    __fpscr_save.__fr = __builtin_mffsl();
+#ifdef _ARCH_PWR9
+    __fpscr_save.__fr = __builtin_ppc_mffsl();
+#else
+    __fpscr_save.__fr = __builtin_ppc_mffs();
+    __fpscr_save.__fpscr &= 0x70007f0ffL;
+#endif
     __attribute__((fallthrough));
   case _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC:
-    __builtin_set_fpscr_rn(0b00);
+    __builtin_ppc_set_fpscr_rn(0b00);
     /* Insert an artificial "read/write" reference to the variable
        read below, to ensure the compiler does not schedule
        a read/use of the variable before the FPSCR is modified, above.
 @@ -193,7 +208,7 @@ extern __inline __m128
        This can be removed if and when GCC PR102783 is fixed.
      */
     __asm__("" : : "wa"(__r));
-    __builtin_set_fpscr_rn(__fpscr_save.__fpscr);
+    __builtin_ppc_set_fpscr_rn(__fpscr_save.__fpscr);
     break;
   case _MM_FROUND_TO_NEG_INF:
   case _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC:
 @@ -219,9 +234,14 @@ extern __inline __m128
      */
     __asm__("" : : "wa"(__r));
     /* Restore enabled exceptions.  */
-    __fpscr_save.__fr = __builtin_mffsl();
+#ifdef _ARCH_PWR9
+    __fpscr_save.__fr = __builtin_ppc_mffsl();
+#else
+    __fpscr_save.__fr = __builtin_ppc_mffs();
+    __fpscr_save.__fpscr &= 0x70007f0ffL;
+#endif
     __fpscr_save.__fpscr |= __enables_save.__fpscr;
-    __builtin_mtfsf(0b00000011, __fpscr_save.__fr);
+    __builtin_ppc_mtfsf(0b00000011, __fpscr_save.__fr);
   }
   return (__m128)__r;
 }

lib/include/riscv_bitmanip.h created+195

...	@@ -0,0 +1,195 @@
1	/===---- riscv_bitmanip.h - RISC-V Zb intrinsics --------------------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	#ifndef __RISCV_BITMANIP_H
11	#define __RISCV_BITMANIP_H
12
13	#include <stdint.h>
14
15	#if defined(__cplusplus)
16	extern "C" {
17	#endif
18
19	#if defined(__riscv_zbb)
20	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
21	__riscv_orc_b_32(uint32_t __x) {
22	return __builtin_riscv_orc_b_32(__x);
23	}
24
25	static __inline__ unsigned __attribute__((__always_inline__, __nodebug__))
26	__riscv_clz_32(uint32_t __x) {
27	return __builtin_riscv_clz_32(__x);
28	}
29
30	static __inline__ unsigned __attribute__((__always_inline__, __nodebug__))
31	__riscv_ctz_32(uint32_t __x) {
32	return __builtin_riscv_ctz_32(__x);
33	}
34
35	static __inline__ unsigned __attribute__((__always_inline__, __nodebug__))
36	__riscv_cpop_32(uint32_t __x) {
37	return __builtin_popcount(__x);
38	}
39
40	#if __riscv_xlen == 64
41	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
42	__riscv_orc_b_64(uint64_t __x) {
43	return __builtin_riscv_orc_b_64(__x);
44	}
45
46	static __inline__ unsigned __attribute__((__always_inline__, __nodebug__))
47	__riscv_clz_64(uint64_t __x) {
48	return __builtin_riscv_clz_64(__x);
49	}
50
51	static __inline__ unsigned __attribute__((__always_inline__, __nodebug__))
52	__riscv_ctz_64(uint64_t __x) {
53	return __builtin_riscv_ctz_64(__x);
54	}
55
56	static __inline__ unsigned __attribute__((__always_inline__, __nodebug__))
57	__riscv_cpop_64(uint64_t __x) {
58	return __builtin_popcountll(__x);
59	}
60	#endif
61	#endif // defined(__riscv_zbb)
62
63	#if defined(__riscv_zbb) \|\| defined(__riscv_zbkb)
64	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
65	__riscv_rev8_32(uint32_t __x) {
66	return __builtin_bswap32(__x);
67	}
68
69	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
70	__riscv_rol_32(uint32_t __x, uint32_t __y) {
71	return __builtin_rotateleft32(__x, __y);
72	}
73
74	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
75	__riscv_ror_32(uint32_t __x, uint32_t __y) {
76	return __builtin_rotateright32(__x, __y);
77	}
78
79	#if __riscv_xlen == 64
80	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
81	__riscv_rev8_64(uint64_t __x) {
82	return __builtin_bswap64(__x);
83	}
84
85	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
86	__riscv_rol_64(uint64_t __x, uint32_t __y) {
87	return __builtin_rotateleft64(__x, __y);
88	}
89
90	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
91	__riscv_ror_64(uint64_t __x, uint32_t __y) {
92	return __builtin_rotateright64(__x, __y);
93	}
94	#endif
95	#endif // defined(__riscv_zbb) \|\| defined(__riscv_zbkb)
96
97	#if defined(__riscv_zbkb)
98	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
99	__riscv_brev8_32(uint32_t __x) {
100	return __builtin_riscv_brev8_32(__x);
101	}
102
103	#if __riscv_xlen == 64
104	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
105	__riscv_brev8_64(uint64_t __x) {
106	return __builtin_riscv_brev8_64(__x);
107	}
108	#endif
109
110	#if __riscv_xlen == 32
111	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
112	__riscv_unzip_32(uint32_t __x) {
113	return __builtin_riscv_unzip_32(__x);
114	}
115
116	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
117	__riscv_zip_32(uint32_t __x) {
118	return __builtin_riscv_zip_32(__x);
119	}
120	#endif
121	#endif // defined(__riscv_zbkb)
122
123	#if defined(__riscv_zbc)
124	#if __riscv_xlen == 32
125	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
126	__riscv_clmulr_32(uint32_t __x, uint32_t __y) {
127	return __builtin_riscv_clmulr_32(__x, __y);
128	}
129	#endif
130
131	#if __riscv_xlen == 64
132	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
133	__riscv_clmulr_64(uint64_t __x, uint64_t __y) {
134	return __builtin_riscv_clmulr_64(__x, __y);
135	}
136	#endif
137	#endif // defined(__riscv_zbc)
138
139	#if defined(__riscv_zbkc) \|\| defined(__riscv_zbc)
140	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
141	__riscv_clmul_32(uint32_t __x, uint32_t __y) {
142	return __builtin_riscv_clmul_32(__x, __y);
143	}
144
145	#if __riscv_xlen == 32
146	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
147	__riscv_clmulh_32(uint32_t __x, uint32_t __y) {
148	return __builtin_riscv_clmulh_32(__x, __y);
149	}
150	#endif
151
152	#if __riscv_xlen == 64
153	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
154	__riscv_clmul_64(uint64_t __x, uint64_t __y) {
155	return __builtin_riscv_clmul_64(__x, __y);
156	}
157
158	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
159	__riscv_clmulh_64(uint64_t __x, uint64_t __y) {
160	return __builtin_riscv_clmulh_64(__x, __y);
161	}
162	#endif
163	#endif // defined(__riscv_zbkc) \|\| defined(__riscv_zbc)
164
165	#if defined(__riscv_zbkx)
166	#if __riscv_xlen == 32
167	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
168	__riscv_xperm4_32(uint32_t __x, uint32_t __y) {
169	return __builtin_riscv_xperm4_32(__x, __y);
170	}
171
172	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
173	__riscv_xperm8_32(uint32_t __x, uint32_t __y) {
174	return __builtin_riscv_xperm8_32(__x, __y);
175	}
176	#endif
177
178	#if __riscv_xlen == 64
179	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
180	__riscv_xperm4_64(uint64_t __x, uint64_t __y) {
181	return __builtin_riscv_xperm4_64(__x, __y);
182	}
183
184	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
185	__riscv_xperm8_64(uint64_t __x, uint64_t __y) {
186	return __builtin_riscv_xperm8_64(__x, __y);
187	}
188	#endif
189	#endif // defined(__riscv_zbkx)
190
191	#if defined(__cplusplus)
192	}
193	#endif
194
195	#endif

lib/include/riscv_crypto.h created+170

...	@@ -0,0 +1,170 @@
1	/===---- riscv_crypto.h - RISC-V Zk intrinsics ---------------------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	#ifndef __RISCV_CRYPTO_H
11	#define __RISCV_CRYPTO_H
12
13	#include <stdint.h>
14
15	#if defined(__cplusplus)
16	extern "C" {
17	#endif
18
19	#if defined(__riscv_zknd)
20	#if __riscv_xlen == 32
21	#define __riscv_aes32dsi(x, y, bs) __builtin_riscv_aes32dsi(x, y, bs)
22	#define __riscv_aes32dsmi(x, y, bs) __builtin_riscv_aes32dsmi(x, y, bs)
23	#endif
24
25	#if __riscv_xlen == 64
26	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
27	__riscv_aes64ds(uint64_t __x, uint64_t __y) {
28	return __builtin_riscv_aes64ds(__x, __y);
29	}
30
31	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
32	__riscv_aes64dsm(uint64_t __x, uint64_t __y) {
33	return __builtin_riscv_aes64dsm(__x, __y);
34	}
35
36	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
37	__riscv_aes64im(uint64_t __x) {
38	return __builtin_riscv_aes64im(__x);
39	}
40	#endif
41	#endif // defined(__riscv_zknd)
42
43	#if defined(__riscv_zkne)
44	#if __riscv_xlen == 32
45	#define __riscv_aes32esi(x, y, bs) __builtin_riscv_aes32esi(x, y, bs)
46	#define __riscv_aes32esmi(x, y, bs) __builtin_riscv_aes32esmi(x, y, bs)
47	#endif
48
49	#if __riscv_xlen == 64
50	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
51	__riscv_aes64es(uint64_t __x, uint64_t __y) {
52	return __builtin_riscv_aes64es(__x, __y);
53	}
54
55	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
56	__riscv_aes64esm(uint64_t __x, uint64_t __y) {
57	return __builtin_riscv_aes64esm(__x, __y);
58	}
59	#endif
60	#endif // defined(__riscv_zkne)
61
62	#if defined(__riscv_zknd) \|\| defined(__riscv_zkne)
63	#if __riscv_xlen == 64
64	#define __riscv_aes64ks1i(x, rnum) __builtin_riscv_aes64ks1i(x, rnum)
65
66	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
67	__riscv_aes64ks2(uint64_t __x, uint64_t __y) {
68	return __builtin_riscv_aes64ks2(__x, __y);
69	}
70	#endif
71	#endif // defined(__riscv_zknd) \|\| defined(__riscv_zkne)
72
73	#if defined(__riscv_zknh)
74	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
75	__riscv_sha256sig0(uint32_t __x) {
76	return __builtin_riscv_sha256sig0(__x);
77	}
78
79	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
80	__riscv_sha256sig1(uint32_t __x) {
81	return __builtin_riscv_sha256sig1(__x);
82	}
83
84	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
85	__riscv_sha256sum0(uint32_t __x) {
86	return __builtin_riscv_sha256sum0(__x);
87	}
88
89	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
90	__riscv_sha256sum1(uint32_t __x) {
91	return __builtin_riscv_sha256sum1(__x);
92	}
93
94	#if __riscv_xlen == 32
95	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
96	__riscv_sha512sig0h(uint32_t __x, uint32_t __y) {
97	return __builtin_riscv_sha512sig0h(__x, __y);
98	}
99
100	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
101	__riscv_sha512sig0l(uint32_t __x, uint32_t __y) {
102	return __builtin_riscv_sha512sig0l(__x, __y);
103	}
104
105	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
106	__riscv_sha512sig1h(uint32_t __x, uint32_t __y) {
107	return __builtin_riscv_sha512sig1h(__x, __y);
108	}
109
110	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
111	__riscv_sha512sig1l(uint32_t __x, uint32_t __y) {
112	return __builtin_riscv_sha512sig1l(__x, __y);
113	}
114
115	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
116	__riscv_sha512sum0r(uint32_t __x, uint32_t __y) {
117	return __builtin_riscv_sha512sum0r(__x, __y);
118	}
119
120	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
121	__riscv_sha512sum1r(uint32_t __x, uint32_t __y) {
122	return __builtin_riscv_sha512sum1r(__x, __y);
123	}
124	#endif
125
126	#if __riscv_xlen == 64
127	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
128	__riscv_sha512sig0(uint64_t __x) {
129	return __builtin_riscv_sha512sig0(__x);
130	}
131
132	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
133	__riscv_sha512sig1(uint64_t __x) {
134	return __builtin_riscv_sha512sig1(__x);
135	}
136
137	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
138	__riscv_sha512sum0(uint64_t __x) {
139	return __builtin_riscv_sha512sum0(__x);
140	}
141
142	static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
143	__riscv_sha512sum1(uint64_t __x) {
144	return __builtin_riscv_sha512sum1(__x);
145	}
146	#endif
147	#endif // defined(__riscv_zknh)
148
149	#if defined(__riscv_zksh)
150	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
151	__riscv_sm3p0(uint32_t __x) {
152	return __builtin_riscv_sm3p0(__x);
153	}
154
155	static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
156	__riscv_sm3p1(uint32_t __x) {
157	return __builtin_riscv_sm3p1(__x);
158	}
159	#endif // defined(__riscv_zksh)
160
161	#if defined(__riscv_zksed)
162	#define __riscv_sm4ed(x, y, bs) __builtin_riscv_sm4ed(x, y, bs);
163	#define __riscv_sm4ks(x, y, bs) __builtin_riscv_sm4ks(x, y, bs);
164	#endif // defined(__riscv_zksed)
165
166	#if defined(__cplusplus)
167	}
168	#endif
169
170	#endif

lib/include/riscv_ntlh.h+3-5

 @@ -21,8 +21,6 @@ enum {
   __RISCV_NTLH_ALL
 };
-#define __riscv_ntl_load(PTR, DOMAIN) __builtin_riscv_ntl_load((PTR), (DOMAIN))
-#define __riscv_ntl_store(PTR, VAL, DOMAIN)                                    \
-  __builtin_riscv_ntl_store((PTR), (VAL), (DOMAIN))
-#endif
+#define __riscv_ntl_load __builtin_riscv_ntl_load
+#define __riscv_ntl_store __builtin_riscv_ntl_store
+#endif

lib/include/riscv_vector.h+31

...	@@ -392,6 +392,37 @@ typedef __rvv_float64m2x4_t vfloat64m2x4_t;	...	@@ -392,6 +392,37 @@ typedef __rvv_float64m2x4_t vfloat64m2x4_t;
392	typedef __rvv_float64m4_t vfloat64m4_t;	392	typedef __rvv_float64m4_t vfloat64m4_t;
393	typedef __rvv_float64m4x2_t vfloat64m4x2_t;	393	typedef __rvv_float64m4x2_t vfloat64m4x2_t;
394	typedef __rvv_float64m8_t vfloat64m8_t;	394	typedef __rvv_float64m8_t vfloat64m8_t;
		395	typedef __rvv_bfloat16mf4_t vbfloat16mf4_t;
		396	typedef __rvv_bfloat16mf4x2_t vbfloat16mf4x2_t;
		397	typedef __rvv_bfloat16mf4x3_t vbfloat16mf4x3_t;
		398	typedef __rvv_bfloat16mf4x4_t vbfloat16mf4x4_t;
		399	typedef __rvv_bfloat16mf4x5_t vbfloat16mf4x5_t;
		400	typedef __rvv_bfloat16mf4x6_t vbfloat16mf4x6_t;
		401	typedef __rvv_bfloat16mf4x7_t vbfloat16mf4x7_t;
		402	typedef __rvv_bfloat16mf4x8_t vbfloat16mf4x8_t;
		403	typedef __rvv_bfloat16mf2_t vbfloat16mf2_t;
		404	typedef __rvv_bfloat16mf2x2_t vbfloat16mf2x2_t;
		405	typedef __rvv_bfloat16mf2x3_t vbfloat16mf2x3_t;
		406	typedef __rvv_bfloat16mf2x4_t vbfloat16mf2x4_t;
		407	typedef __rvv_bfloat16mf2x5_t vbfloat16mf2x5_t;
		408	typedef __rvv_bfloat16mf2x6_t vbfloat16mf2x6_t;
		409	typedef __rvv_bfloat16mf2x7_t vbfloat16mf2x7_t;
		410	typedef __rvv_bfloat16mf2x8_t vbfloat16mf2x8_t;
		411	typedef __rvv_bfloat16m1_t vbfloat16m1_t;
		412	typedef __rvv_bfloat16m1x2_t vbfloat16m1x2_t;
		413	typedef __rvv_bfloat16m1x3_t vbfloat16m1x3_t;
		414	typedef __rvv_bfloat16m1x4_t vbfloat16m1x4_t;
		415	typedef __rvv_bfloat16m1x5_t vbfloat16m1x5_t;
		416	typedef __rvv_bfloat16m1x6_t vbfloat16m1x6_t;
		417	typedef __rvv_bfloat16m1x7_t vbfloat16m1x7_t;
		418	typedef __rvv_bfloat16m1x8_t vbfloat16m1x8_t;
		419	typedef __rvv_bfloat16m2_t vbfloat16m2_t;
		420	typedef __rvv_bfloat16m2x2_t vbfloat16m2x2_t;
		421	typedef __rvv_bfloat16m2x3_t vbfloat16m2x3_t;
		422	typedef __rvv_bfloat16m2x4_t vbfloat16m2x4_t;
		423	typedef __rvv_bfloat16m4_t vbfloat16m4_t;
		424	typedef __rvv_bfloat16m4x2_t vbfloat16m4x2_t;
		425	typedef __rvv_bfloat16m8_t vbfloat16m8_t;
395	#define __riscv_v_intrinsic_overloading 1	426	#define __riscv_v_intrinsic_overloading 1
396		427
397	#ifdef __cplusplus	428	#ifdef __cplusplus

lib/include/smmintrin.h+3-3

...	@@ -18,8 +18,8 @@	...	@@ -18,8 +18,8 @@
18		18
19	/* Define the default attributes for the functions in this file. */	19	/* Define the default attributes for the functions in this file. */
20	#define __DEFAULT_FN_ATTRS \	20	#define __DEFAULT_FN_ATTRS \
21	__attribute__((__always_inline__, __nodebug__, __target__("sse4.1"), \	21	__attribute__((__always_inline__, __nodebug__, \
22	__min_vector_width__(128)))	22	__target__("sse4.1,no-evex512"), __min_vector_width__(128)))
23		23
24	/* SSE4 Rounding macros. */	24	/* SSE4 Rounding macros. */
25	#define _MM_FROUND_TO_NEAREST_INT 0x00	25	#define _MM_FROUND_TO_NEAREST_INT 0x00
...	@@ -645,7 +645,7 @@ static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_mul_epi32(__m128i __V1,	...	@@ -645,7 +645,7 @@ static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_mul_epi32(__m128i __V1,
645	/// \returns A 128-bit integer vector containing the data stored at the	645	/// \returns A 128-bit integer vector containing the data stored at the
646	/// specified memory location.	646	/// specified memory location.
647	static __inline__ __m128i __DEFAULT_FN_ATTRS	647	static __inline__ __m128i __DEFAULT_FN_ATTRS
648	_mm_stream_load_si128(__m128i const *__V) {	648	_mm_stream_load_si128(const void *__V) {
649	return (__m128i)__builtin_nontemporal_load((const __v2di *)__V);	649	return (__m128i)__builtin_nontemporal_load((const __v2di *)__V);
650	}	650	}
651		651

lib/include/stdalign.h+1-3

 @@ -10,10 +10,8 @@
 #ifndef __STDALIGN_H
 #define __STDALIGN_H
-/* FIXME: This is using the placeholder dates Clang produces for these macros
-   in C2x mode; switch to the correct values once they've been published. */
 #if defined(__cplusplus) ||                                                    \
-    (defined(__STDC_VERSION__) && __STDC_VERSION__ < 202000L)
+    (defined(__STDC_VERSION__) && __STDC_VERSION__ < 202311L)
 #ifndef __cplusplus
 #define alignas _Alignas
 #define alignof _Alignof

lib/include/stdarg.h+59-31

...	@@ -7,45 +7,73 @@	...	@@ -7,45 +7,73 @@
7	*===-----------------------------------------------------------------------===	7	*===-----------------------------------------------------------------------===
8	*/	8	*/
9		9
10	#ifndef __STDARG_H	10	/*
11		11	* This header is designed to be included multiple times. If any of the __need_
12	#ifndef __GNUC_VA_LIST	12	* macros are defined, then only that subset of interfaces are provided. This
13	#define __GNUC_VA_LIST	13	* can be useful for POSIX headers that need to not expose all of stdarg.h, but
14	typedef __builtin_va_list __gnuc_va_list;	14	* need to use some of its interfaces. Otherwise this header provides all of
15	#endif	15	* the expected interfaces.
		16	*
		17	* When clang modules are enabled, this header is a textual header. It ignores
		18	* its header guard so that multiple submodules can export its interfaces.
		19	* Take module SM with submodules A and B, whose headers both include stdarg.h
		20	* When SM.A builds, __STDARG_H will be defined. When SM.B builds, the
		21	* definition from SM.A will leak when building without local submodule
		22	* visibility. stdarg.h wouldn't include any of its implementation headers, and
		23	* SM.B wouldn't import any of the stdarg modules, and SM.B's `export *`
		24	* wouldn't export any stdarg interfaces as expected. However, since stdarg.h
		25	* ignores its header guard when building with modules, it all works as
		26	* expected.
		27	*
		28	* When clang modules are not enabled, the header guards can function in the
		29	* normal simple fashion.
		30	*/
		31	#if !defined(__STDARG_H) \|\| __has_feature(modules) \|\| \
		32	defined(__need___va_list) \|\| defined(__need_va_list) \|\| \
		33	defined(__need_va_arg) \|\| defined(__need___va_copy) \|\| \
		34	defined(__need_va_copy)
16		35
17	#ifdef __need___va_list	36	#if !defined(__need___va_list) && !defined(__need_va_list) && \
18	#undef __need___va_list	37	!defined(__need_va_arg) && !defined(__need___va_copy) && \
19	#else	38	!defined(__need_va_copy)
20	#define __STDARG_H	39	#define __STDARG_H
21	#ifndef _VA_LIST	40	#define __need___va_list
22	typedef __builtin_va_list va_list;	41	#define __need_va_list
23	#define _VA_LIST	42	#define __need_va_arg
24	#endif	43	#define __need___va_copy
25
26	/* FIXME: This is using the placeholder dates Clang produces for these macros
27	in C2x mode; switch to the correct values once they've been published. */
28	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
29	/* C2x does not require the second parameter for va_start. */
30	#define va_start(ap, ...) __builtin_va_start(ap, 0)
31	#else
32	/* Versions before C2x do require the second parameter. */
33	#define va_start(ap, param) __builtin_va_start(ap, param)
34	#endif
35	#define va_end(ap) __builtin_va_end(ap)
36	#define va_arg(ap, type) __builtin_va_arg(ap, type)
37
38	/* GCC always defines __va_copy, but does not define va_copy unless in c99 mode	44	/* GCC always defines __va_copy, but does not define va_copy unless in c99 mode
39	* or -ansi is not specified, since it was not part of C90.	45	* or -ansi is not specified, since it was not part of C90.
40	*/	46	*/
41	#define __va_copy(d,s) __builtin_va_copy(d,s)
42
43	#if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L) \|\| \	47	#if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L) \|\| \
44	(defined(__cplusplus) && __cplusplus >= 201103L) \|\| \	48	(defined(__cplusplus) && __cplusplus >= 201103L) \|\| \
45	!defined(__STRICT_ANSI__)	49	!defined(__STRICT_ANSI__)
46	#define va_copy(dest, src) __builtin_va_copy(dest, src)	50	#define __need_va_copy
		51	#endif
47	#endif	52	#endif
48		53
49	#endif /* __STDARG_H */	54	#ifdef __need___va_list
		55	#include <__stdarg___gnuc_va_list.h>
		56	#undef __need___va_list
		57	#endif /* defined(__need___va_list) */
		58
		59	#ifdef __need_va_list
		60	#include <__stdarg_va_list.h>
		61	#undef __need_va_list
		62	#endif /* defined(__need_va_list) */
		63
		64	#ifdef __need_va_arg
		65	#include <__stdarg_va_arg.h>
		66	#undef __need_va_arg
		67	#endif /* defined(__need_va_arg) */
		68
		69	#ifdef __need___va_copy
		70	#include <__stdarg___va_copy.h>
		71	#undef __need___va_copy
		72	#endif /* defined(__need___va_copy) */
50		73
51	#endif /* not __STDARG_H */	74	#ifdef __need_va_copy
		75	#include <__stdarg_va_copy.h>
		76	#undef __need_va_copy
		77	#endif /* defined(__need_va_copy) */
		78
		79	#endif

lib/include/stdatomic.h+3-5

 @@ -45,16 +45,14 @@ extern "C" {
 #define ATOMIC_POINTER_LOCK_FREE    __CLANG_ATOMIC_POINTER_LOCK_FREE
 /* 7.17.2 Initialization */
-/* FIXME: This is using the placeholder dates Clang produces for these macros
+#if (defined(__STDC_VERSION__) && __STDC_VERSION__ < 202311L) ||               \
-   in C2x mode; switch to the correct values once they've been published. */
-#if (defined(__STDC_VERSION__) && __STDC_VERSION__ < 202000L) ||               \
     defined(__cplusplus)
-/* ATOMIC_VAR_INIT was removed in C2x, but still remains in C++23. */
+/* ATOMIC_VAR_INIT was removed in C23, but still remains in C++23. */
 #define ATOMIC_VAR_INIT(value) (value)
 #endif
 #if ((defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201710L &&              \
-      __STDC_VERSION__ < 202000L) ||                                           \
+      __STDC_VERSION__ < 202311L) ||                                           \
      (defined(__cplusplus) && __cplusplus >= 202002L)) &&                      \
     !defined(_CLANG_DISABLE_CRT_DEPRECATION_WARNINGS)
 /* ATOMIC_VAR_INIT was deprecated in C17 and C++20. */

lib/include/stdckdint.h created+42

...	@@ -0,0 +1,42 @@
1	/*===---- stdckdint.h - Standard header for checking integer----------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9
10	#ifndef __STDCKDINT_H
11	#define __STDCKDINT_H
12
13	/* If we're hosted, fall back to the system's stdckdint.h. FreeBSD, for
14	* example, already has a Clang-compatible stdckdint.h header.
15	*
16	* The `stdckdint.h` header requires C 23 or newer.
17	*/
18	#if __STDC_HOSTED__ && __has_include_next(<stdckdint.h>)
19	#include_next <stdckdint.h>
20	#else
21
22	/* C23 7.20.1 Defines several macros for performing checked integer arithmetic*/
23
24	#define __STDC_VERSION_STDCKDINT_H__ 202311L
25
26	// Both A and B shall be any integer type other than "plain" char, bool, a bit-
27	// precise integer type, or an enumerated type, and they need not be the same.
28
29	// R shall be a modifiable lvalue of any integer type other than "plain" char,
30	// bool, a bit-precise integer type, or an enumerated type. It shouldn't be
31	// short type, either. Otherwise, it may be unable to hold two the result of
32	// operating two 'int's.
33
34	// A diagnostic message will be produced if A or B are not suitable integer
35	// types, or if R is not a modifiable lvalue of a suitable integer type or R
36	// is short type.
37	#define ckd_add(R, A, B) __builtin_add_overflow((A), (B), (R))
38	#define ckd_sub(R, A, B) __builtin_sub_overflow((A), (B), (R))
39	#define ckd_mul(R, A, B) __builtin_mul_overflow((A), (B), (R))
40
41	#endif /* __STDC_HOSTED__ */
42	#endif /* __STDCKDINT_H */

lib/include/stddef.h+78-88

...	@@ -7,126 +7,116 @@	...	@@ -7,126 +7,116 @@
7	*===-----------------------------------------------------------------------===	7	*===-----------------------------------------------------------------------===
8	*/	8	*/
9		9
10	#if !defined(__STDDEF_H) \|\| defined(__need_ptrdiff_t) \|\| \	10	/*
11	defined(__need_size_t) \|\| defined(__need_wchar_t) \|\| \	11	* This header is designed to be included multiple times. If any of the __need_
12	defined(__need_NULL) \|\| defined(__need_wint_t)	12	* macros are defined, then only that subset of interfaces are provided. This
		13	* can be useful for POSIX headers that need to not expose all of stddef.h, but
		14	* need to use some of its interfaces. Otherwise this header provides all of
		15	* the expected interfaces.
		16	*
		17	* When clang modules are enabled, this header is a textual header. It ignores
		18	* its header guard so that multiple submodules can export its interfaces.
		19	* Take module SM with submodules A and B, whose headers both include stddef.h
		20	* When SM.A builds, __STDDEF_H will be defined. When SM.B builds, the
		21	* definition from SM.A will leak when building without local submodule
		22	* visibility. stddef.h wouldn't include any of its implementation headers, and
		23	* SM.B wouldn't import any of the stddef modules, and SM.B's `export *`
		24	* wouldn't export any stddef interfaces as expected. However, since stddef.h
		25	* ignores its header guard when building with modules, it all works as
		26	* expected.
		27	*
		28	* When clang modules are not enabled, the header guards can function in the
		29	* normal simple fashion.
		30	*/
		31	#if !defined(__STDDEF_H) \|\| __has_feature(modules) \|\| \
		32	(defined(__STDC_WANT_LIB_EXT1__) && __STDC_WANT_LIB_EXT1__ >= 1) \|\| \
		33	defined(__need_ptrdiff_t) \|\| defined(__need_size_t) \|\| \
		34	defined(__need_rsize_t) \|\| defined(__need_wchar_t) \|\| \
		35	defined(__need_NULL) \|\| defined(__need_nullptr_t) \|\| \
		36	defined(__need_unreachable) \|\| defined(__need_max_align_t) \|\| \
		37	defined(__need_offsetof) \|\| defined(__need_wint_t)
13		38
14	#if !defined(__need_ptrdiff_t) && !defined(__need_size_t) && \	39	#if !defined(__need_ptrdiff_t) && !defined(__need_size_t) && \
15	!defined(__need_wchar_t) && !defined(__need_NULL) && \	40	!defined(__need_rsize_t) && !defined(__need_wchar_t) && \
16	!defined(__need_wint_t)	41	!defined(__need_NULL) && !defined(__need_nullptr_t) && \
17	/* Always define miscellaneous pieces when modules are available. */	42	!defined(__need_unreachable) && !defined(__need_max_align_t) && \
18	#if !__has_feature(modules)	43	!defined(__need_offsetof) && !defined(__need_wint_t)
19	#define __STDDEF_H	44	#define __STDDEF_H
20	#endif
21	#define __need_ptrdiff_t	45	#define __need_ptrdiff_t
22	#define __need_size_t	46	#define __need_size_t
		47	/* ISO9899:2011 7.20 (C11 Annex K): Define rsize_t if __STDC_WANT_LIB_EXT1__ is
		48	* enabled. */
		49	#if defined(__STDC_WANT_LIB_EXT1__) && __STDC_WANT_LIB_EXT1__ >= 1
		50	#define __need_rsize_t
		51	#endif
23	#define __need_wchar_t	52	#define __need_wchar_t
24	#define __need_NULL	53	#define __need_NULL
25	#define __need_STDDEF_H_misc	54	#if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L) \|\| \
26	/* __need_wint_t is intentionally not defined here. */	55	defined(__cplusplus)
		56	#define __need_nullptr_t
27	#endif	57	#endif
28		58	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
29	#if defined(__need_ptrdiff_t)	59	#define __need_unreachable
30	#if !defined(_PTRDIFF_T) \|\| __has_feature(modules)
31	/* Always define ptrdiff_t when modules are available. */
32	#if !__has_feature(modules)
33	#define _PTRDIFF_T
34	#endif	60	#endif
35	typedef __PTRDIFF_TYPE__ ptrdiff_t;	61	#if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L) \|\| \
		62	(defined(__cplusplus) && __cplusplus >= 201103L)
		63	#define __need_max_align_t
36	#endif	64	#endif
		65	#define __need_offsetof
		66	/* wint_t is provided by <wchar.h> and not <stddef.h>. It's here
		67	* for compatibility, but must be explicitly requested. Therefore
		68	* __need_wint_t is intentionally not defined here. */
		69	#endif
		70
		71	#if defined(__need_ptrdiff_t)
		72	#include <__stddef_ptrdiff_t.h>
37	#undef __need_ptrdiff_t	73	#undef __need_ptrdiff_t
38	#endif /* defined(__need_ptrdiff_t) */	74	#endif /* defined(__need_ptrdiff_t) */
39		75
40	#if defined(__need_size_t)	76	#if defined(__need_size_t)
41	#if !defined(_SIZE_T) \|\| __has_feature(modules)	77	#include <__stddef_size_t.h>
42	/* Always define size_t when modules are available. */
43	#if !__has_feature(modules)
44	#define _SIZE_T
45	#endif
46	typedef __SIZE_TYPE__ size_t;
47	#endif
48	#undef __need_size_t	78	#undef __need_size_t
49	#endif /defined(__need_size_t) /	79	#endif /defined(__need_size_t) /
50		80
51	#if defined(__need_STDDEF_H_misc)	81	#if defined(__need_rsize_t)
52	/* ISO9899:2011 7.20 (C11 Annex K): Define rsize_t if __STDC_WANT_LIB_EXT1__ is	82	#include <__stddef_rsize_t.h>
53	* enabled. */	83	#undef __need_rsize_t
54	#if (defined(__STDC_WANT_LIB_EXT1__) && __STDC_WANT_LIB_EXT1__ >= 1 && \	84	#endif /* defined(__need_rsize_t) */
55	!defined(_RSIZE_T)) \|\| __has_feature(modules)
56	/* Always define rsize_t when modules are available. */
57	#if !__has_feature(modules)
58	#define _RSIZE_T
59	#endif
60	typedef __SIZE_TYPE__ rsize_t;
61	#endif
62	#endif /* defined(__need_STDDEF_H_misc) */
63		85
64	#if defined(__need_wchar_t)	86	#if defined(__need_wchar_t)
65	#if !defined(__cplusplus) \|\| (defined(_MSC_VER) && !_NATIVE_WCHAR_T_DEFINED)	87	#include <__stddef_wchar_t.h>
66	/* Always define wchar_t when modules are available. */
67	#if !defined(_WCHAR_T) \|\| __has_feature(modules)
68	#if !__has_feature(modules)
69	#define _WCHAR_T
70	#if defined(_MSC_EXTENSIONS)
71	#define _WCHAR_T_DEFINED
72	#endif
73	#endif
74	typedef __WCHAR_TYPE__ wchar_t;
75	#endif
76	#endif
77	#undef __need_wchar_t	88	#undef __need_wchar_t
78	#endif /* defined(__need_wchar_t) */	89	#endif /* defined(__need_wchar_t) */
79		90
80	#if defined(__need_NULL)	91	#if defined(__need_NULL)
81	#undef NULL	92	#include <__stddef_null.h>
82	#ifdef __cplusplus
83	# if !defined(__MINGW32__) && !defined(_MSC_VER)
84	# define NULL __null
85	# else
86	# define NULL 0
87	# endif
88	#else
89	# define NULL ((void*)0)
90	#endif
91	#ifdef __cplusplus
92	#if defined(_MSC_EXTENSIONS) && defined(_NATIVE_NULLPTR_SUPPORTED)
93	namespace std { typedef decltype(nullptr) nullptr_t; }
94	using ::std::nullptr_t;
95	#endif
96	#endif
97	#undef __need_NULL	93	#undef __need_NULL
98	#endif /* defined(__need_NULL) */	94	#endif /* defined(__need_NULL) */
99		95
100	/* FIXME: This is using the placeholder dates Clang produces for these macros	96	#if defined(__need_nullptr_t)
101	in C2x mode; switch to the correct values once they've been published. */	97	#include <__stddef_nullptr_t.h>
102	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L	98	#undef __need_nullptr_t
103	typedef typeof(nullptr) nullptr_t;	99	#endif /* defined(__need_nullptr_t) */
104	#endif /* defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L */
105		100
106	#if defined(__need_STDDEF_H_misc) && defined(__STDC_VERSION__) && \	101	#if defined(__need_unreachable)
107	__STDC_VERSION__ >= 202000L	102	#include <__stddef_unreachable.h>
108	#define unreachable() __builtin_unreachable()	103	#undef __need_unreachable
109	#endif /* defined(__need_STDDEF_H_misc) && >= C23 */	104	#endif /* defined(__need_unreachable) */
110		105
111	#if defined(__need_STDDEF_H_misc)	106	#if defined(__need_max_align_t)
112	#if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L) \|\| \	107	#include <__stddef_max_align_t.h>
113	(defined(__cplusplus) && __cplusplus >= 201103L)	108	#undef __need_max_align_t
114	#include "__stddef_max_align_t.h"	109	#endif /* defined(__need_max_align_t) */
115	#endif	110
116	#define offsetof(t, d) __builtin_offsetof(t, d)	111	#if defined(__need_offsetof)
117	#undef __need_STDDEF_H_misc	112	#include <__stddef_offsetof.h>
118	#endif /* defined(__need_STDDEF_H_misc) */	113	#undef __need_offsetof
		114	#endif /* defined(__need_offsetof) */
119		115
120	/* Some C libraries expect to see a wint_t here. Others (notably MinGW) will use	116	/* Some C libraries expect to see a wint_t here. Others (notably MinGW) will use
121	__WINT_TYPE__ directly; accommodate both by requiring __need_wint_t */	117	__WINT_TYPE__ directly; accommodate both by requiring __need_wint_t */
122	#if defined(__need_wint_t)	118	#if defined(__need_wint_t)
123	/* Always define wint_t when modules are available. */	119	#include <__stddef_wint_t.h>
124	#if !defined(_WINT_T) \|\| __has_feature(modules)
125	#if !__has_feature(modules)
126	#define _WINT_T
127	#endif
128	typedef __WINT_TYPE__ wint_t;
129	#endif
130	#undef __need_wint_t	120	#undef __need_wint_t
131	#endif /* __need_wint_t */	121	#endif /* __need_wint_t */
132		122

lib/include/stdint.h+19-48

...	@@ -499,9 +499,8 @@ typedef __UINTMAX_TYPE__ uintmax_t;	...	@@ -499,9 +499,8 @@ typedef __UINTMAX_TYPE__ uintmax_t;
499	# define INT64_MAX INT64_C( 9223372036854775807)	499	# define INT64_MAX INT64_C( 9223372036854775807)
500	# define INT64_MIN (-INT64_C( 9223372036854775807)-1)	500	# define INT64_MIN (-INT64_C( 9223372036854775807)-1)
501	# define UINT64_MAX UINT64_C(18446744073709551615)	501	# define UINT64_MAX UINT64_C(18446744073709551615)
502	/* FIXME: This is using the placeholder dates Clang produces for these macros	502
503	in C2x mode; switch to the correct values once they've been published. */	503	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
504	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
505	# define UINT64_WIDTH 64	504	# define UINT64_WIDTH 64
506	# define INT64_WIDTH UINT64_WIDTH	505	# define INT64_WIDTH UINT64_WIDTH
507		506
...	@@ -545,9 +544,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;	...	@@ -545,9 +544,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;
545	# define INT_FAST64_MAX __INT_LEAST64_MAX	544	# define INT_FAST64_MAX __INT_LEAST64_MAX
546	# define UINT_FAST64_MAX __UINT_LEAST64_MAX	545	# define UINT_FAST64_MAX __UINT_LEAST64_MAX
547		546
548	/* FIXME: This is using the placeholder dates Clang produces for these macros	547	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
549	in C2x mode; switch to the correct values once they've been published. */
550	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
551	# define UINT_LEAST64_WIDTH __UINT_LEAST64_WIDTH	548	# define UINT_LEAST64_WIDTH __UINT_LEAST64_WIDTH
552	# define INT_LEAST64_WIDTH UINT_LEAST64_WIDTH	549	# define INT_LEAST64_WIDTH UINT_LEAST64_WIDTH
553	# define UINT_FAST64_WIDTH __UINT_LEAST64_WIDTH	550	# define UINT_FAST64_WIDTH __UINT_LEAST64_WIDTH
...	@@ -586,9 +583,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;	...	@@ -586,9 +583,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;
586	# undef __UINT_LEAST8_MAX	583	# undef __UINT_LEAST8_MAX
587	# define __UINT_LEAST8_MAX UINT56_MAX	584	# define __UINT_LEAST8_MAX UINT56_MAX
588		585
589	/* FIXME: This is using the placeholder dates Clang produces for these macros	586	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
590	in C2x mode; switch to the correct values once they've been published. */
591	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
592	# define UINT56_WIDTH 56	587	# define UINT56_WIDTH 56
593	# define INT56_WIDTH UINT56_WIDTH	588	# define INT56_WIDTH UINT56_WIDTH
594	# define UINT_LEAST56_WIDTH UINT56_WIDTH	589	# define UINT_LEAST56_WIDTH UINT56_WIDTH
...	@@ -635,9 +630,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;	...	@@ -635,9 +630,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;
635	# undef __UINT_LEAST8_MAX	630	# undef __UINT_LEAST8_MAX
636	# define __UINT_LEAST8_MAX UINT48_MAX	631	# define __UINT_LEAST8_MAX UINT48_MAX
637		632
638	/* FIXME: This is using the placeholder dates Clang produces for these macros	633	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
639	in C2x mode; switch to the correct values once they've been published. */
640	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
641	#define UINT48_WIDTH 48	634	#define UINT48_WIDTH 48
642	#define INT48_WIDTH UINT48_WIDTH	635	#define INT48_WIDTH UINT48_WIDTH
643	#define UINT_LEAST48_WIDTH UINT48_WIDTH	636	#define UINT_LEAST48_WIDTH UINT48_WIDTH
...	@@ -684,9 +677,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;	...	@@ -684,9 +677,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;
684	# undef __UINT_LEAST8_MAX	677	# undef __UINT_LEAST8_MAX
685	# define __UINT_LEAST8_MAX UINT40_MAX	678	# define __UINT_LEAST8_MAX UINT40_MAX
686		679
687	/* FIXME: This is using the placeholder dates Clang produces for these macros	680	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
688	in C2x mode; switch to the correct values once they've been published. */
689	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
690	# define UINT40_WIDTH 40	681	# define UINT40_WIDTH 40
691	# define INT40_WIDTH UINT40_WIDTH	682	# define INT40_WIDTH UINT40_WIDTH
692	# define UINT_LEAST40_WIDTH UINT40_WIDTH	683	# define UINT_LEAST40_WIDTH UINT40_WIDTH
...	@@ -727,9 +718,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;	...	@@ -727,9 +718,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;
727	# undef __UINT_LEAST8_MAX	718	# undef __UINT_LEAST8_MAX
728	# define __UINT_LEAST8_MAX UINT32_MAX	719	# define __UINT_LEAST8_MAX UINT32_MAX
729		720
730	/* FIXME: This is using the placeholder dates Clang produces for these macros	721	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
731	in C2x mode; switch to the correct values once they've been published. */
732	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
733	# define UINT32_WIDTH 32	722	# define UINT32_WIDTH 32
734	# define INT32_WIDTH UINT32_WIDTH	723	# define INT32_WIDTH UINT32_WIDTH
735	# undef __UINT_LEAST32_WIDTH	724	# undef __UINT_LEAST32_WIDTH
...	@@ -749,9 +738,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;	...	@@ -749,9 +738,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;
749	# define INT_FAST32_MAX __INT_LEAST32_MAX	738	# define INT_FAST32_MAX __INT_LEAST32_MAX
750	# define UINT_FAST32_MAX __UINT_LEAST32_MAX	739	# define UINT_FAST32_MAX __UINT_LEAST32_MAX
751		740
752	/* FIXME: This is using the placeholder dates Clang produces for these macros	741	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
753	in C2x mode; switch to the correct values once they've been published. */
754	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
755	# define UINT_LEAST32_WIDTH __UINT_LEAST32_WIDTH	742	# define UINT_LEAST32_WIDTH __UINT_LEAST32_WIDTH
756	# define INT_LEAST32_WIDTH UINT_LEAST32_WIDTH	743	# define INT_LEAST32_WIDTH UINT_LEAST32_WIDTH
757	# define UINT_FAST32_WIDTH __UINT_LEAST32_WIDTH	744	# define UINT_FAST32_WIDTH __UINT_LEAST32_WIDTH
...	@@ -784,9 +771,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;	...	@@ -784,9 +771,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;
784	# undef __UINT_LEAST8_MAX	771	# undef __UINT_LEAST8_MAX
785	# define __UINT_LEAST8_MAX UINT24_MAX	772	# define __UINT_LEAST8_MAX UINT24_MAX
786		773
787	/* FIXME: This is using the placeholder dates Clang produces for these macros	774	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
788	in C2x mode; switch to the correct values once they've been published. */
789	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
790	# define UINT24_WIDTH 24	775	# define UINT24_WIDTH 24
791	# define INT24_WIDTH UINT24_WIDTH	776	# define INT24_WIDTH UINT24_WIDTH
792	# define UINT_LEAST24_WIDTH UINT24_WIDTH	777	# define UINT_LEAST24_WIDTH UINT24_WIDTH
...	@@ -819,9 +804,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;	...	@@ -819,9 +804,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;
819	# undef __UINT_LEAST8_MAX	804	# undef __UINT_LEAST8_MAX
820	# define __UINT_LEAST8_MAX UINT16_MAX	805	# define __UINT_LEAST8_MAX UINT16_MAX
821		806
822	/* FIXME: This is using the placeholder dates Clang produces for these macros	807	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
823	in C2x mode; switch to the correct values once they've been published. */
824	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
825	# define UINT16_WIDTH 16	808	# define UINT16_WIDTH 16
826	# define INT16_WIDTH UINT16_WIDTH	809	# define INT16_WIDTH UINT16_WIDTH
827	# undef __UINT_LEAST16_WIDTH	810	# undef __UINT_LEAST16_WIDTH
...	@@ -839,9 +822,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;	...	@@ -839,9 +822,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;
839	# define INT_FAST16_MAX __INT_LEAST16_MAX	822	# define INT_FAST16_MAX __INT_LEAST16_MAX
840	# define UINT_FAST16_MAX __UINT_LEAST16_MAX	823	# define UINT_FAST16_MAX __UINT_LEAST16_MAX
841		824
842	/* FIXME: This is using the placeholder dates Clang produces for these macros	825	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
843	in C2x mode; switch to the correct values once they've been published. */
844	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
845	# define UINT_LEAST16_WIDTH __UINT_LEAST16_WIDTH	826	# define UINT_LEAST16_WIDTH __UINT_LEAST16_WIDTH
846	# define INT_LEAST16_WIDTH UINT_LEAST16_WIDTH	827	# define INT_LEAST16_WIDTH UINT_LEAST16_WIDTH
847	# define UINT_FAST16_WIDTH __UINT_LEAST16_WIDTH	828	# define UINT_FAST16_WIDTH __UINT_LEAST16_WIDTH
...	@@ -862,9 +843,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;	...	@@ -862,9 +843,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;
862	# undef __UINT_LEAST8_MAX	843	# undef __UINT_LEAST8_MAX
863	# define __UINT_LEAST8_MAX UINT8_MAX	844	# define __UINT_LEAST8_MAX UINT8_MAX
864		845
865	/* FIXME: This is using the placeholder dates Clang produces for these macros	846	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
866	in C2x mode; switch to the correct values once they've been published. */
867	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
868	# define UINT8_WIDTH 8	847	# define UINT8_WIDTH 8
869	# define INT8_WIDTH UINT8_WIDTH	848	# define INT8_WIDTH UINT8_WIDTH
870	# undef __UINT_LEAST8_WIDTH	849	# undef __UINT_LEAST8_WIDTH
...	@@ -880,9 +859,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;	...	@@ -880,9 +859,7 @@ typedef __UINTMAX_TYPE__ uintmax_t;
880	# define INT_FAST8_MAX __INT_LEAST8_MAX	859	# define INT_FAST8_MAX __INT_LEAST8_MAX
881	# define UINT_FAST8_MAX __UINT_LEAST8_MAX	860	# define UINT_FAST8_MAX __UINT_LEAST8_MAX
882		861
883	/* FIXME: This is using the placeholder dates Clang produces for these macros	862	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
884	in C2x mode; switch to the correct values once they've been published. */
885	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
886	# define UINT_LEAST8_WIDTH __UINT_LEAST8_WIDTH	863	# define UINT_LEAST8_WIDTH __UINT_LEAST8_WIDTH
887	# define INT_LEAST8_WIDTH UINT_LEAST8_WIDTH	864	# define INT_LEAST8_WIDTH UINT_LEAST8_WIDTH
888	# define UINT_FAST8_WIDTH __UINT_LEAST8_WIDTH	865	# define UINT_FAST8_WIDTH __UINT_LEAST8_WIDTH
...	@@ -907,10 +884,8 @@ typedef __UINTMAX_TYPE__ uintmax_t;	...	@@ -907,10 +884,8 @@ typedef __UINTMAX_TYPE__ uintmax_t;
907	#define PTRDIFF_MAX __PTRDIFF_MAX__	884	#define PTRDIFF_MAX __PTRDIFF_MAX__
908	#define SIZE_MAX __SIZE_MAX__	885	#define SIZE_MAX __SIZE_MAX__
909		886
910	/* C2x 7.20.2.4 Width of integer types capable of holding object pointers. */	887	/* C23 7.22.2.4 Width of integer types capable of holding object pointers. */
911	/* FIXME: This is using the placeholder dates Clang produces for these macros	888	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
912	in C2x mode; switch to the correct values once they've been published. */
913	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
914	/* NB: The C standard requires that these be the same value, but the compiler	889	/* NB: The C standard requires that these be the same value, but the compiler
915	exposes separate internal width macros. */	890	exposes separate internal width macros. */
916	#define INTPTR_WIDTH __INTPTR_WIDTH__	891	#define INTPTR_WIDTH __INTPTR_WIDTH__
...	@@ -928,10 +903,8 @@ typedef __UINTMAX_TYPE__ uintmax_t;	...	@@ -928,10 +903,8 @@ typedef __UINTMAX_TYPE__ uintmax_t;
928	#define INTMAX_MAX __INTMAX_MAX__	903	#define INTMAX_MAX __INTMAX_MAX__
929	#define UINTMAX_MAX __UINTMAX_MAX__	904	#define UINTMAX_MAX __UINTMAX_MAX__
930		905
931	/* C2x 7.20.2.5 Width of greatest-width integer types. */	906	/* C23 7.22.2.5 Width of greatest-width integer types. */
932	/* FIXME: This is using the placeholder dates Clang produces for these macros	907	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
933	in C2x mode; switch to the correct values once they've been published. */
934	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
935	/* NB: The C standard requires that these be the same value, but the compiler	908	/* NB: The C standard requires that these be the same value, but the compiler
936	exposes separate internal width macros. */	909	exposes separate internal width macros. */
937	#define INTMAX_WIDTH __INTMAX_WIDTH__	910	#define INTMAX_WIDTH __INTMAX_WIDTH__
...	@@ -964,10 +937,8 @@ typedef __UINTMAX_TYPE__ uintmax_t;	...	@@ -964,10 +937,8 @@ typedef __UINTMAX_TYPE__ uintmax_t;
964	#define INTMAX_C(v) __int_c(v, __INTMAX_C_SUFFIX__)	937	#define INTMAX_C(v) __int_c(v, __INTMAX_C_SUFFIX__)
965	#define UINTMAX_C(v) __int_c(v, __UINTMAX_C_SUFFIX__)	938	#define UINTMAX_C(v) __int_c(v, __UINTMAX_C_SUFFIX__)
966		939
967	/* C2x 7.20.3.x Width of other integer types. */	940	/* C23 7.22.3.x Width of other integer types. */
968	/* FIXME: This is using the placeholder dates Clang produces for these macros	941	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
969	in C2x mode; switch to the correct values once they've been published. */
970	#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202000L
971	#define PTRDIFF_WIDTH __PTRDIFF_WIDTH__	942	#define PTRDIFF_WIDTH __PTRDIFF_WIDTH__
972	#define SIG_ATOMIC_WIDTH __SIG_ATOMIC_WIDTH__	943	#define SIG_ATOMIC_WIDTH __SIG_ATOMIC_WIDTH__
973	#define SIZE_WIDTH __SIZE_WIDTH__	944	#define SIZE_WIDTH __SIZE_WIDTH__

lib/include/stdnoreturn.h+2-2

...	@@ -15,8 +15,8 @@	...	@@ -15,8 +15,8 @@
15		15
16	#if (defined(__STDC_VERSION__) && __STDC_VERSION__ > 201710L) && \	16	#if (defined(__STDC_VERSION__) && __STDC_VERSION__ > 201710L) && \
17	!defined(_CLANG_DISABLE_CRT_DEPRECATION_WARNINGS)	17	!defined(_CLANG_DISABLE_CRT_DEPRECATION_WARNINGS)
18	/* The noreturn macro is deprecated in C2x. We do not mark it as such because	18	/* The noreturn macro is deprecated in C23. We do not mark it as such because
19	including the header file in C2x is also deprecated and we do not want to	19	including the header file in C23 is also deprecated and we do not want to
20	issue a confusing diagnostic for code which includes <stdnoreturn.h>	20	issue a confusing diagnostic for code which includes <stdnoreturn.h>
21	followed by code that writes [[noreturn]]. The issue with such code is not	21	followed by code that writes [[noreturn]]. The issue with such code is not
22	with the attribute, or the use of 'noreturn', but the inclusion of the	22	with the attribute, or the use of 'noreturn', but the inclusion of the

lib/include/tmmintrin.h+7-2

 @@ -17,8 +17,13 @@
 #include <pmmintrin.h>
 /* Define the default attributes for the functions in this file. */
-#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("ssse3"), __min_vector_width__(64)))
+#define __DEFAULT_FN_ATTRS                                                     \
-#define __DEFAULT_FN_ATTRS_MMX __attribute__((__always_inline__, __nodebug__, __target__("mmx,ssse3"), __min_vector_width__(64)))
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("ssse3,no-evex512"), __min_vector_width__(64)))
+#define __DEFAULT_FN_ATTRS_MMX                                                 \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("mmx,ssse3,no-evex512"),                           \
+                 __min_vector_width__(64)))
 /// Computes the absolute value of each of the packed 8-bit signed
 ///    integers in the source operand and stores the 8-bit unsigned integer

lib/include/usermsrintrin.h created+51

...	@@ -0,0 +1,51 @@
1	/*===--------------- usermsrintrin.h - USERMSR intrinsics -----------------===
2	*
3	* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	* See https://llvm.org/LICENSE.txt for license information.
5	* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	*
7	*===-----------------------------------------------------------------------===
8	*/
9	#ifndef __X86GPRINTRIN_H
10	#error "Never use <usermsrintrin.h> directly; include <x86gprintrin.h> instead."
11	#endif // __X86GPRINTRIN_H
12
13	#ifndef __USERMSRINTRIN_H
14	#define __USERMSRINTRIN_H
15	#ifdef __x86_64__
16
17	/// Reads the contents of a 64-bit MSR specified in \a __A into \a dst.
18	///
19	/// This intrinsic corresponds to the <c> URDMSR </c> instruction.
20	/// \param __A
21	/// An unsigned long long.
22	///
23	/// \code{.operation}
24	/// DEST := MSR[__A]
25	/// \endcode
26	static __inline__ unsigned long long
27	__attribute__((__always_inline__, __nodebug__, __target__("usermsr")))
28	_urdmsr(unsigned long long __A) {
29	return __builtin_ia32_urdmsr(__A);
30	}
31
32	/// Writes the contents of \a __B into the 64-bit MSR specified in \a __A.
33	///
34	/// This intrinsic corresponds to the <c> UWRMSR </c> instruction.
35	///
36	/// \param __A
37	/// An unsigned long long.
38	/// \param __B
39	/// An unsigned long long.
40	///
41	/// \code{.operation}
42	/// MSR[__A] := __B
43	/// \endcode
44	static __inline__ void
45	__attribute__((__always_inline__, __nodebug__, __target__("usermsr")))
46	_uwrmsr(unsigned long long __A, unsigned long long __B) {
47	return __builtin_ia32_uwrmsr(__A, __B);
48	}
49
50	#endif // __x86_64__
51	#endif // __USERMSRINTRIN_H

lib/include/vaesintrin.h+4-2

...	@@ -18,8 +18,10 @@	...	@@ -18,8 +18,10 @@
18	#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("vaes"), __min_vector_width__(256)))	18	#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("vaes"), __min_vector_width__(256)))
19		19
20	/* Default attributes for ZMM forms. */	20	/* Default attributes for ZMM forms. */
21	#define __DEFAULT_FN_ATTRS_F __attribute__((__always_inline__, __nodebug__, __target__("avx512f,vaes"), __min_vector_width__(512)))	21	#define __DEFAULT_FN_ATTRS_F \
22		22	__attribute__((__always_inline__, __nodebug__, \
		23	__target__("avx512f,evex512,vaes"), \
		24	__min_vector_width__(512)))
23		25
24	static __inline__ __m256i __DEFAULT_FN_ATTRS	26	static __inline__ __m256i __DEFAULT_FN_ATTRS
25	_mm256_aesenc_epi128(__m256i __A, __m256i __B)	27	_mm256_aesenc_epi128(__m256i __A, __m256i __B)

lib/include/vecintrin.h+217-199

 @@ -1543,7 +1543,7 @@ vec_load_len(const double *__ptr, unsigned int __len) {
 #if __ARCH__ >= 12
 static inline __ATTRS_ai __vector unsigned char
 vec_load_len_r(const unsigned char *__ptr, unsigned int __len) {
-  return (__vector unsigned char)__builtin_s390_vlrl(__len, __ptr);
+  return (__vector unsigned char)__builtin_s390_vlrlr(__len, __ptr);
 }
 #endif
 @@ -1617,7 +1617,7 @@ vec_store_len(__vector double __vec, double *__ptr,
 static inline __ATTRS_ai void
 vec_store_len_r(__vector unsigned char __vec, unsigned char *__ptr,
                 unsigned int __len) {
-  __builtin_s390_vstrl((__vector signed char)__vec, __len, __ptr);
+  __builtin_s390_vstrlr((__vector signed char)__vec, __len, __ptr);
 }
 #endif
 @@ -2689,7 +2689,8 @@ vec_cmplt(__vector double __a, __vector double __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector signed char __a, __vector signed char __b) {
   int __cc;
-  __builtin_s390_vceqbs(__a, __b, &__cc);
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
+                        (__vector unsigned char)__b, &__cc);
   return __cc == 0;
 }
 @@ -2697,7 +2698,8 @@ vec_all_eq(__vector signed char __a, __vector signed char __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector signed char __a, __vector __bool char __b) {
   int __cc;
-  __builtin_s390_vceqbs(__a, (__vector signed char)__b, &__cc);
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
+                        (__vector unsigned char)__b, &__cc);
   return __cc == 0;
 }
 @@ -2705,15 +2707,15 @@ vec_all_eq(__vector signed char __a, __vector __bool char __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector __bool char __a, __vector signed char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a, __b, &__cc);
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
+                        (__vector unsigned char)__b, &__cc);
   return __cc == 0;
 }
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector unsigned char __a, __vector unsigned char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a,
+  __builtin_s390_vceqbs(__a, __b, &__cc);
-                        (__vector signed char)__b, &__cc);
   return __cc == 0;
 }
 @@ -2721,8 +2723,7 @@ vec_all_eq(__vector unsigned char __a, __vector unsigned char __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector unsigned char __a, __vector __bool char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a,
+  __builtin_s390_vceqbs(__a, (__vector unsigned char)__b, &__cc);
-                        (__vector signed char)__b, &__cc);
   return __cc == 0;
 }
 @@ -2730,23 +2731,23 @@ vec_all_eq(__vector unsigned char __a, __vector __bool char __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector __bool char __a, __vector unsigned char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a,
+  __builtin_s390_vceqbs((__vector unsigned char)__a, __b, &__cc);
-                        (__vector signed char)__b, &__cc);
   return __cc == 0;
 }
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector __bool char __a, __vector __bool char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a,
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
-                        (__vector signed char)__b, &__cc);
+                        (__vector unsigned char)__b, &__cc);
   return __cc == 0;
 }
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector signed short __a, __vector signed short __b) {
   int __cc;
-  __builtin_s390_vceqhs(__a, __b, &__cc);
+  __builtin_s390_vceqhs((__vector unsigned short)__a,
+                        (__vector unsigned short)__b, &__cc);
   return __cc == 0;
 }
 @@ -2754,7 +2755,8 @@ vec_all_eq(__vector signed short __a, __vector signed short __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector signed short __a, __vector __bool short __b) {
   int __cc;
-  __builtin_s390_vceqhs(__a, (__vector signed short)__b, &__cc);
+  __builtin_s390_vceqhs((__vector unsigned short)__a,
+                        (__vector unsigned short)__b, &__cc);
   return __cc == 0;
 }
 @@ -2762,15 +2764,15 @@ vec_all_eq(__vector signed short __a, __vector __bool short __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector __bool short __a, __vector signed short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a, __b, &__cc);
+  __builtin_s390_vceqhs((__vector unsigned short)__a,
+                        (__vector unsigned short)__b, &__cc);
   return __cc == 0;
 }
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector unsigned short __a, __vector unsigned short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a,
+  __builtin_s390_vceqhs(__a, __b, &__cc);
-                        (__vector signed short)__b, &__cc);
   return __cc == 0;
 }
 @@ -2778,8 +2780,7 @@ vec_all_eq(__vector unsigned short __a, __vector unsigned short __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector unsigned short __a, __vector __bool short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a,
+  __builtin_s390_vceqhs(__a, (__vector unsigned short)__b, &__cc);
-                        (__vector signed short)__b, &__cc);
   return __cc == 0;
 }
 @@ -2787,23 +2788,23 @@ vec_all_eq(__vector unsigned short __a, __vector __bool short __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector __bool short __a, __vector unsigned short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a,
+  __builtin_s390_vceqhs((__vector unsigned short)__a, __b, &__cc);
-                        (__vector signed short)__b, &__cc);
   return __cc == 0;
 }
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector __bool short __a, __vector __bool short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a,
+  __builtin_s390_vceqhs((__vector unsigned short)__a,
-                        (__vector signed short)__b, &__cc);
+                        (__vector unsigned short)__b, &__cc);
   return __cc == 0;
 }
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector signed int __a, __vector signed int __b) {
   int __cc;
-  __builtin_s390_vceqfs(__a, __b, &__cc);
+  __builtin_s390_vceqfs((__vector unsigned int)__a,
+                        (__vector unsigned int)__b, &__cc);
   return __cc == 0;
 }
 @@ -2811,7 +2812,8 @@ vec_all_eq(__vector signed int __a, __vector signed int __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector signed int __a, __vector __bool int __b) {
   int __cc;
-  __builtin_s390_vceqfs(__a, (__vector signed int)__b, &__cc);
+  __builtin_s390_vceqfs((__vector unsigned int)__a,
+                        (__vector unsigned int)__b, &__cc);
   return __cc == 0;
 }
 @@ -2819,15 +2821,15 @@ vec_all_eq(__vector signed int __a, __vector __bool int __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector __bool int __a, __vector signed int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a, __b, &__cc);
+  __builtin_s390_vceqfs((__vector unsigned int)__a,
+                        (__vector unsigned int)__b, &__cc);
   return __cc == 0;
 }
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector unsigned int __a, __vector unsigned int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a,
+  __builtin_s390_vceqfs(__a, __b, &__cc);
-                        (__vector signed int)__b, &__cc);
   return __cc == 0;
 }
 @@ -2835,8 +2837,7 @@ vec_all_eq(__vector unsigned int __a, __vector unsigned int __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector unsigned int __a, __vector __bool int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a,
+  __builtin_s390_vceqfs(__a, (__vector unsigned int)__b, &__cc);
-                        (__vector signed int)__b, &__cc);
   return __cc == 0;
 }
 @@ -2844,23 +2845,23 @@ vec_all_eq(__vector unsigned int __a, __vector __bool int __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector __bool int __a, __vector unsigned int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a,
+  __builtin_s390_vceqfs((__vector unsigned int)__a, __b, &__cc);
-                        (__vector signed int)__b, &__cc);
   return __cc == 0;
 }
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector __bool int __a, __vector __bool int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a,
+  __builtin_s390_vceqfs((__vector unsigned int)__a,
-                        (__vector signed int)__b, &__cc);
+                        (__vector unsigned int)__b, &__cc);
   return __cc == 0;
 }
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector signed long long __a, __vector signed long long __b) {
   int __cc;
-  __builtin_s390_vceqgs(__a, __b, &__cc);
+  __builtin_s390_vceqgs((__vector unsigned long long)__a,
+                        (__vector unsigned long long)__b, &__cc);
   return __cc == 0;
 }
 @@ -2868,7 +2869,8 @@ vec_all_eq(__vector signed long long __a, __vector signed long long __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector signed long long __a, __vector __bool long long __b) {
   int __cc;
-  __builtin_s390_vceqgs(__a, (__vector signed long long)__b, &__cc);
+  __builtin_s390_vceqgs((__vector unsigned long long)__a,
+                        (__vector unsigned long long)__b, &__cc);
   return __cc == 0;
 }
 @@ -2876,15 +2878,15 @@ vec_all_eq(__vector signed long long __a, __vector __bool long long __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector __bool long long __a, __vector signed long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a, __b, &__cc);
+  __builtin_s390_vceqgs((__vector unsigned long long)__a,
+                        (__vector unsigned long long)__b, &__cc);
   return __cc == 0;
 }
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector unsigned long long __a, __vector unsigned long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a,
+  __builtin_s390_vceqgs(__a, __b, &__cc);
-                        (__vector signed long long)__b, &__cc);
   return __cc == 0;
 }
 @@ -2892,8 +2894,7 @@ vec_all_eq(__vector unsigned long long __a, __vector unsigned long long __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector unsigned long long __a, __vector __bool long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a,
+  __builtin_s390_vceqgs(__a, (__vector unsigned long long)__b, &__cc);
-                        (__vector signed long long)__b, &__cc);
   return __cc == 0;
 }
 @@ -2901,16 +2902,15 @@ vec_all_eq(__vector unsigned long long __a, __vector __bool long long __b) {
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector __bool long long __a, __vector unsigned long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a,
+  __builtin_s390_vceqgs((__vector unsigned long long)__a, __b, &__cc);
-                        (__vector signed long long)__b, &__cc);
   return __cc == 0;
 }
 static inline __ATTRS_o_ai int
 vec_all_eq(__vector __bool long long __a, __vector __bool long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a,
+  __builtin_s390_vceqgs((__vector unsigned long long)__a,
-                        (__vector signed long long)__b, &__cc);
+                        (__vector unsigned long long)__b, &__cc);
   return __cc == 0;
 }
 @@ -2935,7 +2935,8 @@ vec_all_eq(__vector double __a, __vector double __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector signed char __a, __vector signed char __b) {
   int __cc;
-  __builtin_s390_vceqbs(__a, __b, &__cc);
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
+                        (__vector unsigned char)__b, &__cc);
   return __cc == 3;
 }
 @@ -2943,7 +2944,8 @@ vec_all_ne(__vector signed char __a, __vector signed char __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector signed char __a, __vector __bool char __b) {
   int __cc;
-  __builtin_s390_vceqbs(__a, (__vector signed char)__b, &__cc);
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
+                        (__vector unsigned char)__b, &__cc);
   return __cc == 3;
 }
 @@ -2951,15 +2953,16 @@ vec_all_ne(__vector signed char __a, __vector __bool char __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector __bool char __a, __vector signed char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a, __b, &__cc);
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
+                        (__vector unsigned char)__b, &__cc);
   return __cc == 3;
 }
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector unsigned char __a, __vector unsigned char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a,
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
-                        (__vector signed char)__b, &__cc);
+                        (__vector unsigned char)__b, &__cc);
   return __cc == 3;
 }
 @@ -2967,8 +2970,7 @@ vec_all_ne(__vector unsigned char __a, __vector unsigned char __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector unsigned char __a, __vector __bool char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a,
+  __builtin_s390_vceqbs(__a, (__vector unsigned char)__b, &__cc);
-                        (__vector signed char)__b, &__cc);
   return __cc == 3;
 }
 @@ -2976,23 +2978,23 @@ vec_all_ne(__vector unsigned char __a, __vector __bool char __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector __bool char __a, __vector unsigned char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a,
+  __builtin_s390_vceqbs((__vector unsigned char)__a, __b, &__cc);
-                        (__vector signed char)__b, &__cc);
   return __cc == 3;
 }
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector __bool char __a, __vector __bool char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a,
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
-                        (__vector signed char)__b, &__cc);
+                        (__vector unsigned char)__b, &__cc);
   return __cc == 3;
 }
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector signed short __a, __vector signed short __b) {
   int __cc;
-  __builtin_s390_vceqhs(__a, __b, &__cc);
+  __builtin_s390_vceqhs((__vector unsigned short)__a,
+                        (__vector unsigned short)__b, &__cc);
   return __cc == 3;
 }
 @@ -3000,7 +3002,8 @@ vec_all_ne(__vector signed short __a, __vector signed short __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector signed short __a, __vector __bool short __b) {
   int __cc;
-  __builtin_s390_vceqhs(__a, (__vector signed short)__b, &__cc);
+  __builtin_s390_vceqhs((__vector unsigned short)__a,
+                        (__vector unsigned short)__b, &__cc);
   return __cc == 3;
 }
 @@ -3008,15 +3011,15 @@ vec_all_ne(__vector signed short __a, __vector __bool short __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector __bool short __a, __vector signed short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a, __b, &__cc);
+  __builtin_s390_vceqhs((__vector unsigned short)__a,
+                        (__vector unsigned short)__b, &__cc);
   return __cc == 3;
 }
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector unsigned short __a, __vector unsigned short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a,
+  __builtin_s390_vceqhs(__a, __b, &__cc);
-                        (__vector signed short)__b, &__cc);
   return __cc == 3;
 }
 @@ -3024,8 +3027,7 @@ vec_all_ne(__vector unsigned short __a, __vector unsigned short __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector unsigned short __a, __vector __bool short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a,
+  __builtin_s390_vceqhs(__a, (__vector unsigned short)__b, &__cc);
-                        (__vector signed short)__b, &__cc);
   return __cc == 3;
 }
 @@ -3033,23 +3035,23 @@ vec_all_ne(__vector unsigned short __a, __vector __bool short __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector __bool short __a, __vector unsigned short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a,
+  __builtin_s390_vceqhs((__vector unsigned short)__a, __b, &__cc);
-                        (__vector signed short)__b, &__cc);
   return __cc == 3;
 }
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector __bool short __a, __vector __bool short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a,
+  __builtin_s390_vceqhs((__vector unsigned short)__a,
-                        (__vector signed short)__b, &__cc);
+                        (__vector unsigned short)__b, &__cc);
   return __cc == 3;
 }
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector signed int __a, __vector signed int __b) {
   int __cc;
-  __builtin_s390_vceqfs(__a, __b, &__cc);
+  __builtin_s390_vceqfs((__vector unsigned int)__a,
+                        (__vector unsigned int)__b, &__cc);
   return __cc == 3;
 }
 @@ -3057,7 +3059,8 @@ vec_all_ne(__vector signed int __a, __vector signed int __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector signed int __a, __vector __bool int __b) {
   int __cc;
-  __builtin_s390_vceqfs(__a, (__vector signed int)__b, &__cc);
+  __builtin_s390_vceqfs((__vector unsigned int)__a,
+                        (__vector unsigned int)__b, &__cc);
   return __cc == 3;
 }
 @@ -3065,15 +3068,15 @@ vec_all_ne(__vector signed int __a, __vector __bool int __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector __bool int __a, __vector signed int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a, __b, &__cc);
+  __builtin_s390_vceqfs((__vector unsigned int)__a,
+                        (__vector unsigned int)__b, &__cc);
   return __cc == 3;
 }
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector unsigned int __a, __vector unsigned int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a,
+  __builtin_s390_vceqfs(__a, __b, &__cc);
-                        (__vector signed int)__b, &__cc);
   return __cc == 3;
 }
 @@ -3081,8 +3084,7 @@ vec_all_ne(__vector unsigned int __a, __vector unsigned int __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector unsigned int __a, __vector __bool int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a,
+  __builtin_s390_vceqfs(__a, (__vector unsigned int)__b, &__cc);
-                        (__vector signed int)__b, &__cc);
   return __cc == 3;
 }
 @@ -3090,23 +3092,23 @@ vec_all_ne(__vector unsigned int __a, __vector __bool int __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector __bool int __a, __vector unsigned int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a,
+  __builtin_s390_vceqfs((__vector unsigned int)__a, __b, &__cc);
-                        (__vector signed int)__b, &__cc);
   return __cc == 3;
 }
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector __bool int __a, __vector __bool int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a,
+  __builtin_s390_vceqfs((__vector unsigned int)__a,
-                        (__vector signed int)__b, &__cc);
+                        (__vector unsigned int)__b, &__cc);
   return __cc == 3;
 }
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector signed long long __a, __vector signed long long __b) {
   int __cc;
-  __builtin_s390_vceqgs(__a, __b, &__cc);
+  __builtin_s390_vceqgs((__vector unsigned long long)__a,
+                        (__vector unsigned long long)__b, &__cc);
   return __cc == 3;
 }
 @@ -3114,7 +3116,8 @@ vec_all_ne(__vector signed long long __a, __vector signed long long __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector signed long long __a, __vector __bool long long __b) {
   int __cc;
-  __builtin_s390_vceqgs(__a, (__vector signed long long)__b, &__cc);
+  __builtin_s390_vceqgs((__vector unsigned long long)__a,
+                        (__vector unsigned long long)__b, &__cc);
   return __cc == 3;
 }
 @@ -3122,15 +3125,15 @@ vec_all_ne(__vector signed long long __a, __vector __bool long long __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector __bool long long __a, __vector signed long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a, __b, &__cc);
+  __builtin_s390_vceqgs((__vector unsigned long long)__a,
+                        (__vector unsigned long long)__b, &__cc);
   return __cc == 3;
 }
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector unsigned long long __a, __vector unsigned long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a,
+  __builtin_s390_vceqgs(__a, __b, &__cc);
-                        (__vector signed long long)__b, &__cc);
   return __cc == 3;
 }
 @@ -3138,8 +3141,7 @@ vec_all_ne(__vector unsigned long long __a, __vector unsigned long long __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector unsigned long long __a, __vector __bool long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a,
+  __builtin_s390_vceqgs(__a, (__vector unsigned long long)__b, &__cc);
-                        (__vector signed long long)__b, &__cc);
   return __cc == 3;
 }
 @@ -3147,16 +3149,15 @@ vec_all_ne(__vector unsigned long long __a, __vector __bool long long __b) {
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector __bool long long __a, __vector unsigned long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a,
+  __builtin_s390_vceqgs((__vector unsigned long long)__a, __b, &__cc);
-                        (__vector signed long long)__b, &__cc);
   return __cc == 3;
 }
 static inline __ATTRS_o_ai int
 vec_all_ne(__vector __bool long long __a, __vector __bool long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a,
+  __builtin_s390_vceqgs((__vector unsigned long long)__a,
-                        (__vector signed long long)__b, &__cc);
+                        (__vector unsigned long long)__b, &__cc);
   return __cc == 3;
 }
 @@ -4241,7 +4242,8 @@ vec_all_numeric(__vector double __a) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector signed char __a, __vector signed char __b) {
   int __cc;
-  __builtin_s390_vceqbs(__a, __b, &__cc);
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
+                        (__vector unsigned char)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4249,7 +4251,8 @@ vec_any_eq(__vector signed char __a, __vector signed char __b) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector signed char __a, __vector __bool char __b) {
   int __cc;
-  __builtin_s390_vceqbs(__a, (__vector signed char)__b, &__cc);
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
+                        (__vector unsigned char)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4257,15 +4260,15 @@ vec_any_eq(__vector signed char __a, __vector __bool char __b) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector __bool char __a, __vector signed char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a, __b, &__cc);
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
+                        (__vector unsigned char)__b, &__cc);
   return __cc <= 1;
 }
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector unsigned char __a, __vector unsigned char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a,
+  __builtin_s390_vceqbs(__a, __b, &__cc);
-                        (__vector signed char)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4273,8 +4276,7 @@ vec_any_eq(__vector unsigned char __a, __vector unsigned char __b) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector unsigned char __a, __vector __bool char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a,
+  __builtin_s390_vceqbs(__a, (__vector unsigned char)__b, &__cc);
-                        (__vector signed char)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4282,23 +4284,23 @@ vec_any_eq(__vector unsigned char __a, __vector __bool char __b) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector __bool char __a, __vector unsigned char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a,
+  __builtin_s390_vceqbs((__vector unsigned char)__a, __b, &__cc);
-                        (__vector signed char)__b, &__cc);
   return __cc <= 1;
 }
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector __bool char __a, __vector __bool char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a,
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
-                        (__vector signed char)__b, &__cc);
+                        (__vector unsigned char)__b, &__cc);
   return __cc <= 1;
 }
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector signed short __a, __vector signed short __b) {
   int __cc;
-  __builtin_s390_vceqhs(__a, __b, &__cc);
+  __builtin_s390_vceqhs((__vector unsigned short)__a,
+                        (__vector unsigned short)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4306,7 +4308,8 @@ vec_any_eq(__vector signed short __a, __vector signed short __b) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector signed short __a, __vector __bool short __b) {
   int __cc;
-  __builtin_s390_vceqhs(__a, (__vector signed short)__b, &__cc);
+  __builtin_s390_vceqhs((__vector unsigned short)__a,
+                        (__vector unsigned short)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4314,15 +4317,15 @@ vec_any_eq(__vector signed short __a, __vector __bool short __b) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector __bool short __a, __vector signed short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a, __b, &__cc);
+  __builtin_s390_vceqhs((__vector unsigned short)__a,
+                        (__vector unsigned short)__b, &__cc);
   return __cc <= 1;
 }
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector unsigned short __a, __vector unsigned short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a,
+  __builtin_s390_vceqhs(__a, __b, &__cc);
-                        (__vector signed short)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4330,8 +4333,7 @@ vec_any_eq(__vector unsigned short __a, __vector unsigned short __b) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector unsigned short __a, __vector __bool short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a,
+  __builtin_s390_vceqhs(__a, (__vector unsigned short)__b, &__cc);
-                        (__vector signed short)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4339,23 +4341,23 @@ vec_any_eq(__vector unsigned short __a, __vector __bool short __b) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector __bool short __a, __vector unsigned short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a,
+  __builtin_s390_vceqhs((__vector unsigned short)__a, __b, &__cc);
-                        (__vector signed short)__b, &__cc);
   return __cc <= 1;
 }
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector __bool short __a, __vector __bool short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a,
+  __builtin_s390_vceqhs((__vector unsigned short)__a,
-                        (__vector signed short)__b, &__cc);
+                        (__vector unsigned short)__b, &__cc);
   return __cc <= 1;
 }
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector signed int __a, __vector signed int __b) {
   int __cc;
-  __builtin_s390_vceqfs(__a, __b, &__cc);
+  __builtin_s390_vceqfs((__vector unsigned int)__a,
+                        (__vector unsigned int)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4363,7 +4365,8 @@ vec_any_eq(__vector signed int __a, __vector signed int __b) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector signed int __a, __vector __bool int __b) {
   int __cc;
-  __builtin_s390_vceqfs(__a, (__vector signed int)__b, &__cc);
+  __builtin_s390_vceqfs((__vector unsigned int)__a,
+                        (__vector unsigned int)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4371,15 +4374,15 @@ vec_any_eq(__vector signed int __a, __vector __bool int __b) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector __bool int __a, __vector signed int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a, __b, &__cc);
+  __builtin_s390_vceqfs((__vector unsigned int)__a,
+                        (__vector unsigned int)__b, &__cc);
   return __cc <= 1;
 }
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector unsigned int __a, __vector unsigned int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a,
+  __builtin_s390_vceqfs(__a, __b, &__cc);
-                        (__vector signed int)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4387,8 +4390,7 @@ vec_any_eq(__vector unsigned int __a, __vector unsigned int __b) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector unsigned int __a, __vector __bool int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a,
+  __builtin_s390_vceqfs(__a, (__vector unsigned int)__b, &__cc);
-                        (__vector signed int)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4396,23 +4398,23 @@ vec_any_eq(__vector unsigned int __a, __vector __bool int __b) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector __bool int __a, __vector unsigned int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a,
+  __builtin_s390_vceqfs((__vector unsigned int)__a, __b, &__cc);
-                        (__vector signed int)__b, &__cc);
   return __cc <= 1;
 }
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector __bool int __a, __vector __bool int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a,
+  __builtin_s390_vceqfs((__vector unsigned int)__a,
-                        (__vector signed int)__b, &__cc);
+                        (__vector unsigned int)__b, &__cc);
   return __cc <= 1;
 }
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector signed long long __a, __vector signed long long __b) {
   int __cc;
-  __builtin_s390_vceqgs(__a, __b, &__cc);
+  __builtin_s390_vceqgs((__vector unsigned long long)__a,
+                        (__vector unsigned long long)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4420,7 +4422,8 @@ vec_any_eq(__vector signed long long __a, __vector signed long long __b) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector signed long long __a, __vector __bool long long __b) {
   int __cc;
-  __builtin_s390_vceqgs(__a, (__vector signed long long)__b, &__cc);
+  __builtin_s390_vceqgs((__vector unsigned long long)__a,
+                        (__vector unsigned long long)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4428,15 +4431,15 @@ vec_any_eq(__vector signed long long __a, __vector __bool long long __b) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector __bool long long __a, __vector signed long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a, __b, &__cc);
+  __builtin_s390_vceqgs((__vector unsigned long long)__a,
+                        (__vector unsigned long long)__b, &__cc);
   return __cc <= 1;
 }
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector unsigned long long __a, __vector unsigned long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a,
+  __builtin_s390_vceqgs(__a, __b, &__cc);
-                        (__vector signed long long)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4444,8 +4447,7 @@ vec_any_eq(__vector unsigned long long __a, __vector unsigned long long __b) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector unsigned long long __a, __vector __bool long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a,
+  __builtin_s390_vceqgs(__a, (__vector unsigned long long)__b, &__cc);
-                        (__vector signed long long)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4453,16 +4455,15 @@ vec_any_eq(__vector unsigned long long __a, __vector __bool long long __b) {
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector __bool long long __a, __vector unsigned long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a,
+  __builtin_s390_vceqgs((__vector unsigned long long)__a, __b, &__cc);
-                        (__vector signed long long)__b, &__cc);
   return __cc <= 1;
 }
 static inline __ATTRS_o_ai int
 vec_any_eq(__vector __bool long long __a, __vector __bool long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a,
+  __builtin_s390_vceqgs((__vector unsigned long long)__a,
-                        (__vector signed long long)__b, &__cc);
+                        (__vector unsigned long long)__b, &__cc);
   return __cc <= 1;
 }
 @@ -4487,7 +4488,8 @@ vec_any_eq(__vector double __a, __vector double __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector signed char __a, __vector signed char __b) {
   int __cc;
-  __builtin_s390_vceqbs(__a, __b, &__cc);
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
+                        (__vector unsigned char)__b, &__cc);
   return __cc != 0;
 }
 @@ -4495,7 +4497,8 @@ vec_any_ne(__vector signed char __a, __vector signed char __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector signed char __a, __vector __bool char __b) {
   int __cc;
-  __builtin_s390_vceqbs(__a, (__vector signed char)__b, &__cc);
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
+                        (__vector unsigned char)__b, &__cc);
   return __cc != 0;
 }
 @@ -4503,15 +4506,15 @@ vec_any_ne(__vector signed char __a, __vector __bool char __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector __bool char __a, __vector signed char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a, __b, &__cc);
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
+                        (__vector unsigned char)__b, &__cc);
   return __cc != 0;
 }
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector unsigned char __a, __vector unsigned char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a,
+  __builtin_s390_vceqbs(__a, __b, &__cc);
-                        (__vector signed char)__b, &__cc);
   return __cc != 0;
 }
 @@ -4519,8 +4522,7 @@ vec_any_ne(__vector unsigned char __a, __vector unsigned char __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector unsigned char __a, __vector __bool char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a,
+  __builtin_s390_vceqbs(__a, (__vector unsigned char)__b, &__cc);
-                        (__vector signed char)__b, &__cc);
   return __cc != 0;
 }
 @@ -4528,23 +4530,23 @@ vec_any_ne(__vector unsigned char __a, __vector __bool char __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector __bool char __a, __vector unsigned char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a,
+  __builtin_s390_vceqbs((__vector unsigned char)__a, __b, &__cc);
-                        (__vector signed char)__b, &__cc);
   return __cc != 0;
 }
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector __bool char __a, __vector __bool char __b) {
   int __cc;
-  __builtin_s390_vceqbs((__vector signed char)__a,
+  __builtin_s390_vceqbs((__vector unsigned char)__a,
-                        (__vector signed char)__b, &__cc);
+                        (__vector unsigned char)__b, &__cc);
   return __cc != 0;
 }
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector signed short __a, __vector signed short __b) {
   int __cc;
-  __builtin_s390_vceqhs(__a, __b, &__cc);
+  __builtin_s390_vceqhs((__vector unsigned short)__a,
+                        (__vector unsigned short)__b, &__cc);
   return __cc != 0;
 }
 @@ -4552,7 +4554,8 @@ vec_any_ne(__vector signed short __a, __vector signed short __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector signed short __a, __vector __bool short __b) {
   int __cc;
-  __builtin_s390_vceqhs(__a, (__vector signed short)__b, &__cc);
+  __builtin_s390_vceqhs((__vector unsigned short)__a,
+                        (__vector unsigned short)__b, &__cc);
   return __cc != 0;
 }
 @@ -4560,15 +4563,15 @@ vec_any_ne(__vector signed short __a, __vector __bool short __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector __bool short __a, __vector signed short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a, __b, &__cc);
+  __builtin_s390_vceqhs((__vector unsigned short)__a,
+                        (__vector unsigned short)__b, &__cc);
   return __cc != 0;
 }
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector unsigned short __a, __vector unsigned short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a,
+  __builtin_s390_vceqhs(__a, __b, &__cc);
-                        (__vector signed short)__b, &__cc);
   return __cc != 0;
 }
 @@ -4576,8 +4579,7 @@ vec_any_ne(__vector unsigned short __a, __vector unsigned short __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector unsigned short __a, __vector __bool short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a,
+  __builtin_s390_vceqhs(__a, (__vector unsigned short)__b, &__cc);
-                        (__vector signed short)__b, &__cc);
   return __cc != 0;
 }
 @@ -4585,23 +4587,23 @@ vec_any_ne(__vector unsigned short __a, __vector __bool short __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector __bool short __a, __vector unsigned short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a,
+  __builtin_s390_vceqhs((__vector unsigned short)__a, __b, &__cc);
-                        (__vector signed short)__b, &__cc);
   return __cc != 0;
 }
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector __bool short __a, __vector __bool short __b) {
   int __cc;
-  __builtin_s390_vceqhs((__vector signed short)__a,
+  __builtin_s390_vceqhs((__vector unsigned short)__a,
-                        (__vector signed short)__b, &__cc);
+                        (__vector unsigned short)__b, &__cc);
   return __cc != 0;
 }
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector signed int __a, __vector signed int __b) {
   int __cc;
-  __builtin_s390_vceqfs(__a, __b, &__cc);
+  __builtin_s390_vceqfs((__vector unsigned int)__a,
+                        (__vector unsigned int)__b, &__cc);
   return __cc != 0;
 }
 @@ -4609,7 +4611,8 @@ vec_any_ne(__vector signed int __a, __vector signed int __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector signed int __a, __vector __bool int __b) {
   int __cc;
-  __builtin_s390_vceqfs(__a, (__vector signed int)__b, &__cc);
+  __builtin_s390_vceqfs((__vector unsigned int)__a,
+                        (__vector unsigned int)__b, &__cc);
   return __cc != 0;
 }
 @@ -4617,15 +4620,15 @@ vec_any_ne(__vector signed int __a, __vector __bool int __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector __bool int __a, __vector signed int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a, __b, &__cc);
+  __builtin_s390_vceqfs((__vector unsigned int)__a,
+                        (__vector unsigned int)__b, &__cc);
   return __cc != 0;
 }
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector unsigned int __a, __vector unsigned int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a,
+  __builtin_s390_vceqfs(__a, __b, &__cc);
-                        (__vector signed int)__b, &__cc);
   return __cc != 0;
 }
 @@ -4633,8 +4636,7 @@ vec_any_ne(__vector unsigned int __a, __vector unsigned int __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector unsigned int __a, __vector __bool int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a,
+  __builtin_s390_vceqfs(__a, (__vector unsigned int)__b, &__cc);
-                        (__vector signed int)__b, &__cc);
   return __cc != 0;
 }
 @@ -4642,23 +4644,23 @@ vec_any_ne(__vector unsigned int __a, __vector __bool int __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector __bool int __a, __vector unsigned int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a,
+  __builtin_s390_vceqfs((__vector unsigned int)__a, __b, &__cc);
-                        (__vector signed int)__b, &__cc);
   return __cc != 0;
 }
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector __bool int __a, __vector __bool int __b) {
   int __cc;
-  __builtin_s390_vceqfs((__vector signed int)__a,
+  __builtin_s390_vceqfs((__vector unsigned int)__a,
-                        (__vector signed int)__b, &__cc);
+                        (__vector unsigned int)__b, &__cc);
   return __cc != 0;
 }
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector signed long long __a, __vector signed long long __b) {
   int __cc;
-  __builtin_s390_vceqgs(__a, __b, &__cc);
+  __builtin_s390_vceqgs((__vector unsigned long long)__a,
+                        (__vector unsigned long long)__b, &__cc);
   return __cc != 0;
 }
 @@ -4666,7 +4668,8 @@ vec_any_ne(__vector signed long long __a, __vector signed long long __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector signed long long __a, __vector __bool long long __b) {
   int __cc;
-  __builtin_s390_vceqgs(__a, (__vector signed long long)__b, &__cc);
+  __builtin_s390_vceqgs((__vector unsigned long long)__a,
+                        (__vector unsigned long long)__b, &__cc);
   return __cc != 0;
 }
 @@ -4674,15 +4677,15 @@ vec_any_ne(__vector signed long long __a, __vector __bool long long __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector __bool long long __a, __vector signed long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a, __b, &__cc);
+  __builtin_s390_vceqgs((__vector unsigned long long)__a,
+                        (__vector unsigned long long)__b, &__cc);
   return __cc != 0;
 }
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector unsigned long long __a, __vector unsigned long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a,
+  __builtin_s390_vceqgs(__a, __b, &__cc);
-                        (__vector signed long long)__b, &__cc);
   return __cc != 0;
 }
 @@ -4690,8 +4693,7 @@ vec_any_ne(__vector unsigned long long __a, __vector unsigned long long __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector unsigned long long __a, __vector __bool long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a,
+  __builtin_s390_vceqgs(__a, (__vector unsigned long long)__b, &__cc);
-                        (__vector signed long long)__b, &__cc);
   return __cc != 0;
 }
 @@ -4699,16 +4701,15 @@ vec_any_ne(__vector unsigned long long __a, __vector __bool long long __b) {
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector __bool long long __a, __vector unsigned long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a,
+  __builtin_s390_vceqgs((__vector unsigned long long)__a, __b, &__cc);
-                        (__vector signed long long)__b, &__cc);
   return __cc != 0;
 }
 static inline __ATTRS_o_ai int
 vec_any_ne(__vector __bool long long __a, __vector __bool long long __b) {
   int __cc;
-  __builtin_s390_vceqgs((__vector signed long long)__a,
+  __builtin_s390_vceqgs((__vector unsigned long long)__a,
-                        (__vector signed long long)__b, &__cc);
+                        (__vector unsigned long long)__b, &__cc);
   return __cc != 0;
 }
 @@ -6565,45 +6566,45 @@ vec_rl(__vector unsigned long long __a, __vector unsigned long long __b) {
 static inline __ATTRS_o_ai __vector signed char
 vec_rli(__vector signed char __a, unsigned long __b) {
   return (__vector signed char)__builtin_s390_verllb(
-    (__vector unsigned char)__a, (int)__b);
+    (__vector unsigned char)__a, (unsigned char)__b);
 }
 static inline __ATTRS_o_ai __vector unsigned char
 vec_rli(__vector unsigned char __a, unsigned long __b) {
-  return __builtin_s390_verllb(__a, (int)__b);
+  return __builtin_s390_verllb(__a, (unsigned char)__b);
 }
 static inline __ATTRS_o_ai __vector signed short
 vec_rli(__vector signed short __a, unsigned long __b) {
   return (__vector signed short)__builtin_s390_verllh(
-    (__vector unsigned short)__a, (int)__b);
+    (__vector unsigned short)__a, (unsigned char)__b);
 }
 static inline __ATTRS_o_ai __vector unsigned short
 vec_rli(__vector unsigned short __a, unsigned long __b) {
-  return __builtin_s390_verllh(__a, (int)__b);
+  return __builtin_s390_verllh(__a, (unsigned char)__b);
 }
 static inline __ATTRS_o_ai __vector signed int
 vec_rli(__vector signed int __a, unsigned long __b) {
   return (__vector signed int)__builtin_s390_verllf(
-    (__vector unsigned int)__a, (int)__b);
+    (__vector unsigned int)__a, (unsigned char)__b);
 }
 static inline __ATTRS_o_ai __vector unsigned int
 vec_rli(__vector unsigned int __a, unsigned long __b) {
-  return __builtin_s390_verllf(__a, (int)__b);
+  return __builtin_s390_verllf(__a, (unsigned char)__b);
 }
 static inline __ATTRS_o_ai __vector signed long long
 vec_rli(__vector signed long long __a, unsigned long __b) {
   return (__vector signed long long)__builtin_s390_verllg(
-    (__vector unsigned long long)__a, (int)__b);
+    (__vector unsigned long long)__a, (unsigned char)__b);
 }
 static inline __ATTRS_o_ai __vector unsigned long long
 vec_rli(__vector unsigned long long __a, unsigned long __b) {
-  return __builtin_s390_verllg(__a, (int)__b);
+  return __builtin_s390_verllg(__a, (unsigned char)__b);
 }
 /*-- vec_rl_mask ------------------------------------------------------------*/
 @@ -8358,7 +8359,7 @@ vec_min(__vector double __a, __vector double __b) {
 static inline __ATTRS_ai __vector unsigned char
 vec_add_u128(__vector unsigned char __a, __vector unsigned char __b) {
-  return __builtin_s390_vaq(__a, __b);
+  return (__vector unsigned char)((__int128)__a + (__int128)__b);
 }
 /*-- vec_addc ---------------------------------------------------------------*/
 @@ -8387,7 +8388,8 @@ vec_addc(__vector unsigned long long __a, __vector unsigned long long __b) {
 static inline __ATTRS_ai __vector unsigned char
 vec_addc_u128(__vector unsigned char __a, __vector unsigned char __b) {
-  return __builtin_s390_vaccq(__a, __b);
+  return (__vector unsigned char)
+         __builtin_s390_vaccq((unsigned __int128)__a, (unsigned __int128)__b);
 }
 /*-- vec_adde_u128 ----------------------------------------------------------*/
 @@ -8395,7 +8397,9 @@ vec_addc_u128(__vector unsigned char __a, __vector unsigned char __b) {
 static inline __ATTRS_ai __vector unsigned char
 vec_adde_u128(__vector unsigned char __a, __vector unsigned char __b,
               __vector unsigned char __c) {
-  return __builtin_s390_vacq(__a, __b, __c);
+  return (__vector unsigned char)
+         __builtin_s390_vacq((unsigned __int128)__a, (unsigned __int128)__b,
+                             (unsigned __int128)__c);
 }
 /*-- vec_addec_u128 ---------------------------------------------------------*/
 @@ -8403,7 +8407,9 @@ vec_adde_u128(__vector unsigned char __a, __vector unsigned char __b,
 static inline __ATTRS_ai __vector unsigned char
 vec_addec_u128(__vector unsigned char __a, __vector unsigned char __b,
                __vector unsigned char __c) {
-  return __builtin_s390_vacccq(__a, __b, __c);
+  return (__vector unsigned char)
+         __builtin_s390_vacccq((unsigned __int128)__a, (unsigned __int128)__b,
+                               (unsigned __int128)__c);
 }
 /*-- vec_avg ----------------------------------------------------------------*/
 @@ -8477,7 +8483,7 @@ vec_gfmsum(__vector unsigned int __a, __vector unsigned int __b) {
 static inline __ATTRS_o_ai __vector unsigned char
 vec_gfmsum_128(__vector unsigned long long __a,
                __vector unsigned long long __b) {
-  return __builtin_s390_vgfmg(__a, __b);
+  return (__vector unsigned char)__builtin_s390_vgfmg(__a, __b);
 }
 /*-- vec_gfmsum_accum -------------------------------------------------------*/
 @@ -8506,7 +8512,8 @@ static inline __ATTRS_o_ai __vector unsigned char
 vec_gfmsum_accum_128(__vector unsigned long long __a,
                      __vector unsigned long long __b,
                      __vector unsigned char __c) {
-  return __builtin_s390_vgfmag(__a, __b, __c);
+  return (__vector unsigned char)
+         __builtin_s390_vgfmag(__a, __b, (unsigned __int128)__c);
 }
 /*-- vec_mladd --------------------------------------------------------------*/
 @@ -8796,15 +8803,21 @@ vec_mulo(__vector unsigned int __a, __vector unsigned int __b) {
 /*-- vec_msum_u128 ----------------------------------------------------------*/
 #if __ARCH__ >= 12
+extern __ATTRS_o __vector unsigned char
+vec_msum_u128(__vector unsigned long long __a, __vector unsigned long long __b,
+              __vector unsigned char __c, int __d)
+  __constant_range(__d, 0, 15);
 #define vec_msum_u128(X, Y, Z, W) \
-  ((__vector unsigned char)__builtin_s390_vmslg((X), (Y), (Z), (W)));
+  ((__typeof__((vec_msum_u128)((X), (Y), (Z), (W)))) \
+   __builtin_s390_vmslg((X), (Y), (unsigned __int128)(Z), (W)))
 #endif
 /*-- vec_sub_u128 -----------------------------------------------------------*/
 static inline __ATTRS_ai __vector unsigned char
 vec_sub_u128(__vector unsigned char __a, __vector unsigned char __b) {
-  return __builtin_s390_vsq(__a, __b);
+  return (__vector unsigned char)((__int128)__a - (__int128)__b);
 }
 /*-- vec_subc ---------------------------------------------------------------*/
 @@ -8833,7 +8846,8 @@ vec_subc(__vector unsigned long long __a, __vector unsigned long long __b) {
 static inline __ATTRS_ai __vector unsigned char
 vec_subc_u128(__vector unsigned char __a, __vector unsigned char __b) {
-  return __builtin_s390_vscbiq(__a, __b);
+  return (__vector unsigned char)
+         __builtin_s390_vscbiq((unsigned __int128)__a, (unsigned __int128)__b);
 }
 /*-- vec_sube_u128 ----------------------------------------------------------*/
 @@ -8841,7 +8855,9 @@ vec_subc_u128(__vector unsigned char __a, __vector unsigned char __b) {
 static inline __ATTRS_ai __vector unsigned char
 vec_sube_u128(__vector unsigned char __a, __vector unsigned char __b,
               __vector unsigned char __c) {
-  return __builtin_s390_vsbiq(__a, __b, __c);
+  return (__vector unsigned char)
+         __builtin_s390_vsbiq((unsigned __int128)__a, (unsigned __int128)__b,
+                              (unsigned __int128)__c);
 }
 /*-- vec_subec_u128 ---------------------------------------------------------*/
 @@ -8849,7 +8865,9 @@ vec_sube_u128(__vector unsigned char __a, __vector unsigned char __b,
 static inline __ATTRS_ai __vector unsigned char
 vec_subec_u128(__vector unsigned char __a, __vector unsigned char __b,
                __vector unsigned char __c) {
-  return __builtin_s390_vsbcbiq(__a, __b, __c);
+  return (__vector unsigned char)
+         __builtin_s390_vsbcbiq((unsigned __int128)__a, (unsigned __int128)__b,
+                                (unsigned __int128)__c);
 }
 /*-- vec_sum2 ---------------------------------------------------------------*/
 @@ -8868,12 +8886,12 @@ vec_sum2(__vector unsigned int __a, __vector unsigned int __b) {
 static inline __ATTRS_o_ai __vector unsigned char
 vec_sum_u128(__vector unsigned int __a, __vector unsigned int __b) {
-  return __builtin_s390_vsumqf(__a, __b);
+  return (__vector unsigned char)__builtin_s390_vsumqf(__a, __b);
 }
 static inline __ATTRS_o_ai __vector unsigned char
 vec_sum_u128(__vector unsigned long long __a, __vector unsigned long long __b) {
-  return __builtin_s390_vsumqg(__a, __b);
+  return (__vector unsigned char)__builtin_s390_vsumqg(__a, __b);
 }
 /*-- vec_sum4 ---------------------------------------------------------------*/

lib/include/x86gprintrin.h+5

 @@ -20,6 +20,11 @@
 #include <uintrintrin.h>
 #endif
+#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
+    defined(__USERMSR__)
+#include <usermsrintrin.h>
+#endif
 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
     defined(__CRC32__)
 #include <crc32intrin.h>

lib/include/xmmintrin.h+9-5

 @@ -32,8 +32,12 @@ typedef unsigned int __v4su __attribute__((__vector_size__(16)));
 #endif
 /* Define the default attributes for the functions in this file. */
-#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("sse"), __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS                                                     \
-#define __DEFAULT_FN_ATTRS_MMX __attribute__((__always_inline__, __nodebug__, __target__("mmx,sse"), __min_vector_width__(64)))
+  __attribute__((__always_inline__, __nodebug__, __target__("sse,no-evex512"), \
+                 __min_vector_width__(128)))
+#define __DEFAULT_FN_ATTRS_MMX                                                 \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("mmx,sse,no-evex512"), __min_vector_width__(64)))
 /// Adds the 32-bit float values in the low-order bits of the operands.
 ///
 @@ -2121,9 +2125,9 @@ _mm_storer_ps(float *__p, __m128 __a)
 /// \param __a
 ///    A 64-bit integer containing the value to be stored.
 static __inline__ void __DEFAULT_FN_ATTRS_MMX
-_mm_stream_pi(__m64 *__p, __m64 __a)
+_mm_stream_pi(void *__p, __m64 __a)
 {
-  __builtin_ia32_movntq(__p, __a);
+  __builtin_ia32_movntq((__m64 *)__p, __a);
 }
 /// Moves packed float values from a 128-bit vector of [4 x float] to a
 @@ -2140,7 +2144,7 @@ _mm_stream_pi(__m64 *__p, __m64 __a)
 /// \param __a
 ///    A 128-bit vector of [4 x float] containing the values to be moved.
 static __inline__ void __DEFAULT_FN_ATTRS
-_mm_stream_ps(float *__p, __m128 __a)
+_mm_stream_ps(void *__p, __m128 __a)
 {
   __builtin_nontemporal_store((__v4sf)__a, (__v4sf*)__p);
 }

...	@@ -36,7 +36,7 @@	...	@@ -36,7 +36,7 @@
36	// because the OpenMP overlay requires constexpr functions here but prior to	36	// because the OpenMP overlay requires constexpr functions here but prior to
37	// c++14 void return functions could not be constexpr.	37	// c++14 void return functions could not be constexpr.
38	#pragma push_macro("__DEVICE_VOID__")	38	#pragma push_macro("__DEVICE_VOID__")
39	#ifdef __OPENMP_NVPTX__ && defined(__cplusplus) && __cplusplus < 201402L	39	#if defined(__OPENMP_NVPTX__) && defined(__cplusplus) && __cplusplus < 201402L
40	#define __DEVICE_VOID__ static __attribute__((always_inline, nothrow))	40	#define __DEVICE_VOID__ static __attribute__((always_inline, nothrow))
41	#else	41	#else
42	#define __DEVICE_VOID__ __DEVICE__	42	#define __DEVICE_VOID__ __DEVICE__
...	@@ -45,9 +45,9 @@	...	@@ -45,9 +45,9 @@
45	// libdevice provides fast low precision and slow full-recision implementations	45	// libdevice provides fast low precision and slow full-recision implementations
46	// for some functions. Which one gets selected depends on	46	// for some functions. Which one gets selected depends on
47	// __CLANG_CUDA_APPROX_TRANSCENDENTALS__ which gets defined by clang if	47	// __CLANG_CUDA_APPROX_TRANSCENDENTALS__ which gets defined by clang if
48	// -ffast-math or -fcuda-approx-transcendentals are in effect.	48	// -ffast-math or -fgpu-approx-transcendentals are in effect.
49	#pragma push_macro("__FAST_OR_SLOW")	49	#pragma push_macro("__FAST_OR_SLOW")
50	#if defined(__CLANG_CUDA_APPROX_TRANSCENDENTALS__)	50	#if defined(__CLANG_GPU_APPROX_TRANSCENDENTALS__)
51	#define __FAST_OR_SLOW(fast, slow) fast	51	#define __FAST_OR_SLOW(fast, slow) fast
52	#else	52	#else
53	#define __FAST_OR_SLOW(fast, slow) slow	53	#define __FAST_OR_SLOW(fast, slow) slow

...	@@ -50,11 +50,11 @@ typedef __bf16 __m128bh __attribute__((__vector_size__(16), __aligned__(16)));	...	@@ -50,11 +50,11 @@ typedef __bf16 __m128bh __attribute__((__vector_size__(16), __aligned__(16)));
50		50
51	/* Define the default attributes for the functions in this file. */	51	/* Define the default attributes for the functions in this file. */
52	#define __DEFAULT_FN_ATTRS \	52	#define __DEFAULT_FN_ATTRS \
53	__attribute__((__always_inline__, __nodebug__, __target__("sse2"), \	53	__attribute__((__always_inline__, __nodebug__, \
54	__min_vector_width__(128)))	54	__target__("sse2,no-evex512"), __min_vector_width__(128)))
55	#define __DEFAULT_FN_ATTRS_MMX \	55	#define __DEFAULT_FN_ATTRS_MMX \
56	__attribute__((__always_inline__, __nodebug__, __target__("mmx,sse2"), \	56	__attribute__((__always_inline__, __nodebug__, \
57	__min_vector_width__(64)))	57	__target__("mmx,sse2,no-evex512"), __min_vector_width__(64)))
58		58
59	/// Adds lower double-precision values in both operands and returns the	59	/// Adds lower double-precision values in both operands and returns the
60	/// sum in the lower 64 bits of the result. The upper 64 bits of the result	60	/// sum in the lower 64 bits of the result. The upper 64 bits of the result
...	@@ -3945,7 +3945,7 @@ static __inline__ void __DEFAULT_FN_ATTRS _mm_storel_epi64(__m128i_u *__p,	...	@@ -3945,7 +3945,7 @@ static __inline__ void __DEFAULT_FN_ATTRS _mm_storel_epi64(__m128i_u *__p,
3945	/// A pointer to the 128-bit aligned memory location used to store the value.	3945	/// A pointer to the 128-bit aligned memory location used to store the value.
3946	/// \param __a	3946	/// \param __a
3947	/// A vector of [2 x double] containing the 64-bit values to be stored.	3947	/// A vector of [2 x double] containing the 64-bit values to be stored.
3948	static __inline__ void __DEFAULT_FN_ATTRS _mm_stream_pd(double *__p,	3948	static __inline__ void __DEFAULT_FN_ATTRS _mm_stream_pd(void *__p,
3949	__m128d __a) {	3949	__m128d __a) {
3950	__builtin_nontemporal_store((__v2df)__a, (__v2df *)__p);	3950	__builtin_nontemporal_store((__v2df)__a, (__v2df *)__p);
3951	}	3951	}
...	@@ -3963,7 +3963,7 @@ static __inline__ void __DEFAULT_FN_ATTRS _mm_stream_pd(double *__p,	...	@@ -3963,7 +3963,7 @@ static __inline__ void __DEFAULT_FN_ATTRS _mm_stream_pd(double *__p,
3963	/// A pointer to the 128-bit aligned memory location used to store the value.	3963	/// A pointer to the 128-bit aligned memory location used to store the value.
3964	/// \param __a	3964	/// \param __a
3965	/// A 128-bit integer vector containing the values to be stored.	3965	/// A 128-bit integer vector containing the values to be stored.
3966	static __inline__ void __DEFAULT_FN_ATTRS _mm_stream_si128(__m128i *__p,	3966	static __inline__ void __DEFAULT_FN_ATTRS _mm_stream_si128(void *__p,
3967	__m128i __a) {	3967	__m128i __a) {
3968	__builtin_nontemporal_store((__v2di)__a, (__v2di *)__p);	3968	__builtin_nontemporal_store((__v2di)__a, (__v2di *)__p);
3969	}	3969	}
...	@@ -3983,8 +3983,8 @@ static __inline__ void __DEFAULT_FN_ATTRS _mm_stream_si128(__m128i *__p,	...	@@ -3983,8 +3983,8 @@ static __inline__ void __DEFAULT_FN_ATTRS _mm_stream_si128(__m128i *__p,
3983	/// A 32-bit integer containing the value to be stored.	3983	/// A 32-bit integer containing the value to be stored.
3984	static __inline__ void	3984	static __inline__ void
3985	__attribute__((__always_inline__, __nodebug__, __target__("sse2")))	3985	__attribute__((__always_inline__, __nodebug__, __target__("sse2")))
3986	_mm_stream_si32(int *__p, int __a) {	3986	_mm_stream_si32(void *__p, int __a) {
3987	__builtin_ia32_movnti(__p, __a);	3987	__builtin_ia32_movnti((int *)__p, __a);
3988	}	3988	}
3989		3989
3990	#ifdef __x86_64__	3990	#ifdef __x86_64__
...	@@ -4003,8 +4003,8 @@ static __inline__ void	...	@@ -4003,8 +4003,8 @@ static __inline__ void
4003	/// A 64-bit integer containing the value to be stored.	4003	/// A 64-bit integer containing the value to be stored.
4004	static __inline__ void	4004	static __inline__ void
4005	__attribute__((__always_inline__, __nodebug__, __target__("sse2")))	4005	__attribute__((__always_inline__, __nodebug__, __target__("sse2")))
4006	_mm_stream_si64(long long *__p, long long __a) {	4006	_mm_stream_si64(void *__p, long long __a) {
4007	__builtin_ia32_movnti64(__p, __a);	4007	__builtin_ia32_movnti64((long long *)__p, __a);
4008	}	4008	}
4009	#endif	4009	#endif
4010		4010

...	@@ -45,6 +45,7 @@	...	@@ -45,6 +45,7 @@
45	#define __opencl_c_ext_fp32_local_atomic_add 1	45	#define __opencl_c_ext_fp32_local_atomic_add 1
46	#define __opencl_c_ext_fp32_global_atomic_min_max 1	46	#define __opencl_c_ext_fp32_global_atomic_min_max 1
47	#define __opencl_c_ext_fp32_local_atomic_min_max 1	47	#define __opencl_c_ext_fp32_local_atomic_min_max 1
		48	#define __opencl_c_ext_image_raw10_raw12 1
48		49
49	#endif // defined(__SPIR__) \|\| defined(__SPIRV__)	50	#endif // defined(__SPIR__) \|\| defined(__SPIRV__)
50	#endif // (defined(__OPENCL_CPP_VERSION__) \|\| __OPENCL_C_VERSION__ >= 200)	51	#endif // (defined(__OPENCL_CPP_VERSION__) \|\| __OPENCL_C_VERSION__ >= 200)
...	@@ -477,6 +478,10 @@ typedef enum memory_order	...	@@ -477,6 +478,10 @@ typedef enum memory_order
477	#if __OPENCL_C_VERSION__ >= CL_VERSION_3_0	478	#if __OPENCL_C_VERSION__ >= CL_VERSION_3_0
478	#define CLK_UNORM_INT_101010_2 0x10E0	479	#define CLK_UNORM_INT_101010_2 0x10E0
479	#endif // __OPENCL_C_VERSION__ >= CL_VERSION_3_0	480	#endif // __OPENCL_C_VERSION__ >= CL_VERSION_3_0
		481	#ifdef __opencl_c_ext_image_raw10_raw12
		482	#define CLK_UNSIGNED_INT_RAW10_EXT 0x10E3
		483	#define CLK_UNSIGNED_INT_RAW12_EXT 0x10E4
		484	#endif // __opencl_c_ext_image_raw10_raw12
480		485
481	// Channel order, numbering must be aligned with cl_channel_order in cl.h	486	// Channel order, numbering must be aligned with cl_channel_order in cl.h
482	//	487	//