authorgravatar for andrew@ziglang.orgAndrew Kelley <andrew@ziglang.org> 2019-03-22 16:21:56-04:00
committergravatar for noreply@github.comGitHub <noreply@github.com> 2019-03-22 16:21:56-04:00
log6272847917be186fa83f92b0580ed27f459014bd
tree85b35863cc49d4c5c83c05e6d6dffb3f0a8a83a9
parent1ca78e39e4497821df352e0343132228e08894c3
parent02767690e0084ccbad2268c4f4abb0b2d2e8c30a
signaturebadge-question-mark Signed by PGP key 4AEE18F83AFDEB23

Merge pull request #2094 from ziglang/f128-decimal-literal

float literals now parse using musl's 128 bit float code

9 files changed, 1089 insertions(+), 25 deletions(-)

CMakeLists.txt+9-7
...@@ -313,6 +313,7 @@ set(EMBEDDED_SOFTFLOAT_SOURCES...@@ -313,6 +313,7 @@ set(EMBEDDED_SOFTFLOAT_SOURCES
313 "${CMAKE_SOURCE_DIR}/deps/SoftFloat-3e/source/f32_to_f128M.c"313 "${CMAKE_SOURCE_DIR}/deps/SoftFloat-3e/source/f32_to_f128M.c"
314 "${CMAKE_SOURCE_DIR}/deps/SoftFloat-3e/source/f64_to_f128M.c"314 "${CMAKE_SOURCE_DIR}/deps/SoftFloat-3e/source/f64_to_f128M.c"
315 "${CMAKE_SOURCE_DIR}/deps/SoftFloat-3e/source/f64_to_f16.c"315 "${CMAKE_SOURCE_DIR}/deps/SoftFloat-3e/source/f64_to_f16.c"
316 "${CMAKE_SOURCE_DIR}/deps/SoftFloat-3e/source/i32_to_f128M.c"
316 "${CMAKE_SOURCE_DIR}/deps/SoftFloat-3e/source/s_add256M.c"317 "${CMAKE_SOURCE_DIR}/deps/SoftFloat-3e/source/s_add256M.c"
317 "${CMAKE_SOURCE_DIR}/deps/SoftFloat-3e/source/s_addCarryM.c"318 "${CMAKE_SOURCE_DIR}/deps/SoftFloat-3e/source/s_addCarryM.c"
318 "${CMAKE_SOURCE_DIR}/deps/SoftFloat-3e/source/s_addComplCarryM.c"319 "${CMAKE_SOURCE_DIR}/deps/SoftFloat-3e/source/s_addComplCarryM.c"
...@@ -427,11 +428,12 @@ set(ZIG_SOURCES...@@ -427,11 +428,12 @@ set(ZIG_SOURCES
427 "${CMAKE_SOURCE_DIR}/src/range_set.cpp"428 "${CMAKE_SOURCE_DIR}/src/range_set.cpp"
428 "${CMAKE_SOURCE_DIR}/src/target.cpp"429 "${CMAKE_SOURCE_DIR}/src/target.cpp"
429 "${CMAKE_SOURCE_DIR}/src/tokenizer.cpp"430 "${CMAKE_SOURCE_DIR}/src/tokenizer.cpp"
430 "${CMAKE_SOURCE_DIR}/src/util.cpp"
431 "${CMAKE_SOURCE_DIR}/src/translate_c.cpp"431 "${CMAKE_SOURCE_DIR}/src/translate_c.cpp"
432 "${CMAKE_SOURCE_DIR}/src/util.cpp"
432)433)
433set(BLAKE_SOURCES434set(OPTIMIZED_C_SOURCES
434 "${CMAKE_SOURCE_DIR}/src/blake2b.c"435 "${CMAKE_SOURCE_DIR}/src/blake2b.c"
436 "${CMAKE_SOURCE_DIR}/src/parse_f128.c"
435)437)
436set(ZIG_CPP_SOURCES438set(ZIG_CPP_SOURCES
437 "${CMAKE_SOURCE_DIR}/src/zig_llvm.cpp"439 "${CMAKE_SOURCE_DIR}/src/zig_llvm.cpp"
...@@ -6600,7 +6602,7 @@ else()...@@ -6600,7 +6602,7 @@ else()
6600 endif()6602 endif()
6601endif()6603endif()
66026604
6603set(BLAKE_CFLAGS "-std=c99")6605set(OPTIMIZED_C_FLAGS "-std=c99 -O3")
66046606
6605set(EXE_LDFLAGS " ")6607set(EXE_LDFLAGS " ")
6606if(MINGW)6608if(MINGW)
...@@ -6626,9 +6628,9 @@ set_target_properties(zig_cpp PROPERTIES...@@ -6626,9 +6628,9 @@ set_target_properties(zig_cpp PROPERTIES
6626 COMPILE_FLAGS ${EXE_CFLAGS}6628 COMPILE_FLAGS ${EXE_CFLAGS}
6627)6629)
66286630
6629add_library(embedded_blake STATIC ${BLAKE_SOURCES})6631add_library(opt_c_util STATIC ${OPTIMIZED_C_SOURCES})
6630set_target_properties(embedded_blake PROPERTIES6632set_target_properties(opt_c_util PROPERTIES
6631 COMPILE_FLAGS "${BLAKE_CFLAGS} -O3"6633 COMPILE_FLAGS "${OPTIMIZED_C_FLAGS}"
6632)6634)
66336635
6634add_executable(zig ${ZIG_SOURCES})6636add_executable(zig ${ZIG_SOURCES})
...@@ -6639,7 +6641,7 @@ set_target_properties(zig PROPERTIES...@@ -6639,7 +6641,7 @@ set_target_properties(zig PROPERTIES
66396641
6640target_link_libraries(zig LINK_PUBLIC6642target_link_libraries(zig LINK_PUBLIC
6641 zig_cpp6643 zig_cpp
6642 embedded_blake6644 opt_c_util
6643 ${SOFTFLOAT_LIBRARIES}6645 ${SOFTFLOAT_LIBRARIES}
6644 ${CLANG_LIBRARIES}6646 ${CLANG_LIBRARIES}
6645 ${LLD_LIBRARIES}6647 ${LLD_LIBRARIES}
src/bigfloat.cpp+4-7
...@@ -9,6 +9,7 @@...@@ -9,6 +9,7 @@
9#include "bigint.hpp"9#include "bigint.hpp"
10#include "buffer.hpp"10#include "buffer.hpp"
11#include "softfloat.hpp"11#include "softfloat.hpp"
12#include "parse_f128.h"
12#include <stdio.h>13#include <stdio.h>
13#include <math.h>14#include <math.h>
14#include <errno.h>15#include <errno.h>
...@@ -65,22 +66,18 @@ void bigfloat_init_bigint(BigFloat *dest, const BigInt *op) {...@@ -65,22 +66,18 @@ void bigfloat_init_bigint(BigFloat *dest, const BigInt *op) {
65 }66 }
66}67}
6768
68int bigfloat_init_buf_base10(BigFloat *dest, const uint8_t *buf_ptr, size_t buf_len) {69Error bigfloat_init_buf(BigFloat *dest, const uint8_t *buf_ptr, size_t buf_len) {
69 char *str_begin = (char *)buf_ptr;70 char *str_begin = (char *)buf_ptr;
70 char *str_end;71 char *str_end;
7172
72 errno = 0;73 errno = 0;
73 double value = strtod(str_begin, &str_end); // TODO actual f128 parsing74 dest->value = parse_f128(str_begin, &str_end);
74 if (errno) {75 if (errno) {
75 return ErrorOverflow;76 return ErrorOverflow;
76 }77 }
7778
78 float64_t value_f64;
79 memcpy(&value_f64, &value, sizeof(double));
80 f64_to_f128M(value_f64, &dest->value);
81
82 assert(str_end <= ((char*)buf_ptr) + buf_len);79 assert(str_end <= ((char*)buf_ptr) + buf_len);
83 return 0;80 return ErrorNone;
84}81}
8582
86void bigfloat_add(BigFloat *dest, const BigFloat *op1, const BigFloat *op2) {83void bigfloat_add(BigFloat *dest, const BigFloat *op1, const BigFloat *op2) {
src/bigfloat.hpp+1-1
...@@ -28,7 +28,7 @@ void bigfloat_init_64(BigFloat *dest, double x);...@@ -28,7 +28,7 @@ void bigfloat_init_64(BigFloat *dest, double x);
28void bigfloat_init_128(BigFloat *dest, float128_t x);28void bigfloat_init_128(BigFloat *dest, float128_t x);
29void bigfloat_init_bigfloat(BigFloat *dest, const BigFloat *x);29void bigfloat_init_bigfloat(BigFloat *dest, const BigFloat *x);
30void bigfloat_init_bigint(BigFloat *dest, const BigInt *op);30void bigfloat_init_bigint(BigFloat *dest, const BigInt *op);
31int bigfloat_init_buf_base10(BigFloat *dest, const uint8_t *buf_ptr, size_t buf_len);31Error bigfloat_init_buf(BigFloat *dest, const uint8_t *buf_ptr, size_t buf_len);
3232
33float16_t bigfloat_to_f16(const BigFloat *bigfloat);33float16_t bigfloat_to_f16(const BigFloat *bigfloat);
34float bigfloat_to_f32(const BigFloat *bigfloat);34float bigfloat_to_f32(const BigFloat *bigfloat);
src/parse_f128.c created+1039
...@@ -0,0 +1,1039 @@
1// Code ported from musl libc 8f12c4e110acb3bbbdc8abfb3a552c3ced718039
2// and then modified to use softfloat and to assume f128 for everything
3
4#include "parse_f128.h"
5#include "softfloat.h"
6#include <stddef.h>
7#include <sys/types.h>
8#include <errno.h>
9#include <limits.h>
10#include <string.h>
11#include <math.h>
12
13#define shcnt(f) ((f)->shcnt + ((f)->rpos - (f)->buf))
14#define shlim(f, lim) __shlim((f), (lim))
15#define shgetc(f) (((f)->rpos != (f)->shend) ? *(f)->rpos++ : __shgetc(f))
16#define shunget(f) ((f)->shlim>=0 ? (void)(f)->rpos-- : (void)0)
17
18#define sh_fromstring(f, s) \
19 ((f)->buf = (f)->rpos = (void *)(s), (f)->rend = (void*)-1)
20
21#define LD_B1B_DIG 4
22#define LD_B1B_MAX 10384593, 717069655, 257060992, 658440191
23#define KMAX 2048
24
25#define MASK (KMAX-1)
26
27#define CONCAT2(x,y) x ## y
28#define CONCAT(x,y) CONCAT2(x,y)
29
30#define F_PERM 1
31#define F_NORD 4
32#define F_NOWR 8
33#define F_EOF 16
34#define F_ERR 32
35#define F_SVB 64
36#define F_APP 128
37
38#define EOF (-1)
39
40#define LDBL_MANT_DIG 113
41#define LDBL_MIN_EXP (-16381)
42#define LDBL_MAX_EXP 16384
43
44#define LDBL_DIG 33
45#define LDBL_MIN_10_EXP (-4931)
46#define LDBL_MAX_10_EXP 4932
47
48#define DECIMAL_DIG 36
49
50
51#if __BYTE_ORDER == __LITTLE_ENDIAN
52union ldshape {
53 float128_t f;
54 struct {
55 uint64_t lo;
56 uint32_t mid;
57 uint16_t top;
58 uint16_t se;
59 } i;
60 struct {
61 uint64_t lo;
62 uint64_t hi;
63 } i2;
64};
65#elif __BYTE_ORDER == __BIG_ENDIAN
66union ldshape {
67 float128_t f;
68 struct {
69 uint16_t se;
70 uint16_t top;
71 uint32_t mid;
72 uint64_t lo;
73 } i;
74 struct {
75 uint64_t hi;
76 uint64_t lo;
77 } i2;
78};
79#error Unsupported endian
80#endif
81
82struct MuslFILE {
83 unsigned flags;
84 unsigned char *rpos, *rend;
85 int (*close)(struct MuslFILE *);
86 unsigned char *wend, *wpos;
87 unsigned char *mustbezero_1;
88 unsigned char *wbase;
89 size_t (*read)(struct MuslFILE *, unsigned char *, size_t);
90 size_t (*write)(struct MuslFILE *, const unsigned char *, size_t);
91 off_t (*seek)(struct MuslFILE *, off_t, int);
92 unsigned char *buf;
93 size_t buf_size;
94 struct MuslFILE *prev, *next;
95 int fd;
96 int pipe_pid;
97 long lockcount;
98 int mode;
99 volatile int lock;
100 int lbf;
101 void *cookie;
102 off_t off;
103 char *getln_buf;
104 void *mustbezero_2;
105 unsigned char *shend;
106 off_t shlim, shcnt;
107 struct MuslFILE *prev_locked, *next_locked;
108 struct __locale_struct *locale;
109};
110
111static void __shlim(struct MuslFILE *f, off_t lim)
112{
113 f->shlim = lim;
114 f->shcnt = f->buf - f->rpos;
115 /* If lim is nonzero, rend must be a valid pointer. */
116 if (lim && f->rend - f->rpos > lim)
117 f->shend = f->rpos + lim;
118 else
119 f->shend = f->rend;
120}
121
122static int __toread(struct MuslFILE *f)
123{
124 f->mode |= f->mode-1;
125 if (f->wpos != f->wbase) f->write(f, 0, 0);
126 f->wpos = f->wbase = f->wend = 0;
127 if (f->flags & F_NORD) {
128 f->flags |= F_ERR;
129 return EOF;
130 }
131 f->rpos = f->rend = f->buf + f->buf_size;
132 return (f->flags & F_EOF) ? EOF : 0;
133}
134
135static int __uflow(struct MuslFILE *f)
136{
137 unsigned char c;
138 if (!__toread(f) && f->read(f, &c, 1)==1) return c;
139 return EOF;
140}
141
142static int __shgetc(struct MuslFILE *f)
143{
144 int c;
145 off_t cnt = shcnt(f);
146 if (f->shlim && cnt >= f->shlim || (c=__uflow(f)) < 0) {
147 f->shcnt = f->buf - f->rpos + cnt;
148 f->shend = f->rpos;
149 f->shlim = -1;
150 return EOF;
151 }
152 cnt++;
153 if (f->shlim && f->rend - f->rpos > f->shlim - cnt)
154 f->shend = f->rpos + (f->shlim - cnt);
155 else
156 f->shend = f->rend;
157 f->shcnt = f->buf - f->rpos + cnt;
158 if (f->rpos[-1] != c) f->rpos[-1] = c;
159 return c;
160}
161
162static long long scanexp(struct MuslFILE *f, int pok)
163{
164 int c;
165 int x;
166 long long y;
167 int neg = 0;
168
169 c = shgetc(f);
170 if (c=='+' || c=='-') {
171 neg = (c=='-');
172 c = shgetc(f);
173 if (c-'0'>=10U && pok) shunget(f);
174 }
175 if (c-'0'>=10U) {
176 shunget(f);
177 return LLONG_MIN;
178 }
179 for (x=0; c-'0'<10U && x<INT_MAX/10; c = shgetc(f))
180 x = 10*x + c-'0';
181 for (y=x; c-'0'<10U && y<LLONG_MAX/100; c = shgetc(f))
182 y = 10*y + c-'0';
183 for (; c-'0'<10U; c = shgetc(f));
184 shunget(f);
185 return neg ? -y : y;
186}
187
188static float128_t copysignf128(float128_t x, float128_t y)
189{
190 union ldshape ux = {x}, uy = {y};
191 ux.i.se &= 0x7fff;
192 ux.i.se |= uy.i.se & 0x8000;
193 return ux.f;
194}
195
196static void mul_eq_f128_float(float128_t *x, float op_float) {
197 //x *= 0x1p120f;
198 float32_t op_f32;
199 memcpy(&op_f32, &op_float, sizeof(float));
200 float128_t op_f128;
201 f32_to_f128M(op_f32, &op_f128);
202 float128_t new_value;
203 f128M_mul(x, &op_f128, &new_value);
204 *x = new_value;
205}
206
207static float128_t dbl_to_f128(double x) {
208 float64_t x_f64;
209 memcpy(&x_f64, &x, sizeof(double));
210 float128_t result;
211 f64_to_f128M(x_f64, &result);
212 return result;
213}
214
215static float128_t fmodf128(float128_t x, float128_t y)
216{
217 union ldshape ux = {x}, uy = {y};
218 int ex = ux.i.se & 0x7fff;
219 int ey = uy.i.se & 0x7fff;
220 int sx = ux.i.se & 0x8000;
221
222 float128_t zero;
223 ui32_to_f128M(0, &zero);
224 // if (y == 0 || isnan(y) || ex == 0x7fff)
225 if (f128M_eq(&y, &zero) || f128M_isSignalingNaN(&y) || ex == 0x7fff) {
226 //return (x*y)/(x*y);
227 float128_t x_times_y;
228 f128M_mul(&x, &y, &x_times_y);
229 float128_t result;
230 f128M_div(&x_times_y, &x_times_y, &result);
231 return result;
232 }
233 ux.i.se = ex;
234 uy.i.se = ey;
235 //if (ux.f <= uy.f) {
236 if (f128M_le(&ux.f, &uy.f)) {
237 //if (ux.f == uy.f) {
238 if (f128M_eq(&ux.f, &uy.f)) {
239 //return 0*x;
240 float128_t result;
241 f128M_mul(&zero, &x, &result);
242 return result;
243 }
244 return x;
245 }
246
247 /* normalize x and y */
248 if (!ex) {
249 //ux.f *= 0x1p120f;
250 mul_eq_f128_float(&ux.f, 0x1p120f);
251
252 ex = ux.i.se - 120;
253 }
254 if (!ey) {
255 //uy.f *= 0x1p120f;
256 mul_eq_f128_float(&uy.f, 0x1p120f);
257
258 ey = uy.i.se - 120;
259 }
260
261 /* x mod y */
262 uint64_t hi, lo, xhi, xlo, yhi, ylo;
263 xhi = (ux.i2.hi & -1ULL>>16) | 1ULL<<48;
264 yhi = (uy.i2.hi & -1ULL>>16) | 1ULL<<48;
265 xlo = ux.i2.lo;
266 ylo = uy.i2.lo;
267 for (; ex > ey; ex--) {
268 hi = xhi - yhi;
269 lo = xlo - ylo;
270 if (xlo < ylo)
271 hi -= 1;
272 if (hi >> 63 == 0) {
273 if ((hi|lo) == 0) {
274 //return 0*x;
275 float128_t result;
276 f128M_mul(&zero, &x, &result);
277 return result;
278 }
279 xhi = 2*hi + (lo>>63);
280 xlo = 2*lo;
281 } else {
282 xhi = 2*xhi + (xlo>>63);
283 xlo = 2*xlo;
284 }
285 }
286 hi = xhi - yhi;
287 lo = xlo - ylo;
288 if (xlo < ylo)
289 hi -= 1;
290 if (hi >> 63 == 0) {
291 if ((hi|lo) == 0) {
292 //return 0*x;
293 float128_t result;
294 f128M_mul(&zero, &x, &result);
295 return result;
296 }
297 xhi = hi;
298 xlo = lo;
299 }
300 for (; xhi >> 48 == 0; xhi = 2*xhi + (xlo>>63), xlo = 2*xlo, ex--);
301 ux.i2.hi = xhi;
302 ux.i2.lo = xlo;
303
304 /* scale result */
305 if (ex <= 0) {
306 ux.i.se = (ex+120)|sx;
307 //ux.f *= 0x1p-120f;
308 mul_eq_f128_float(&ux.f, 0x1p-120f);
309 } else
310 ux.i.se = ex|sx;
311 return ux.f;
312}
313
314static float128_t int_mul_f128_cast_u32(int sign, uint32_t x0) {
315 float128_t x0_f128;
316 ui32_to_f128M(x0, &x0_f128);
317 float128_t sign_f128;
318 i32_to_f128M(sign, &sign_f128);
319 float128_t result;
320 f128M_mul(&sign_f128, &x0_f128, &result);
321 return result;
322}
323
324static float128_t triple_divide(int sign, uint32_t x0, int p10s) {
325 float128_t part1 = int_mul_f128_cast_u32(sign, x0);
326 float128_t p10s_f128;
327 i32_to_f128M(p10s, &p10s_f128);
328 float128_t result;
329 f128M_div(&part1, &p10s_f128, &result);
330 return result;
331}
332
333static float128_t triple_multiply(int sign, uint32_t x0, int p10s) {
334 float128_t part1 = int_mul_f128_cast_u32(sign, x0);
335 float128_t p10s_f128;
336 i32_to_f128M(p10s, &p10s_f128);
337 float128_t result;
338 f128M_mul(&part1, &p10s_f128, &result);
339 return result;
340}
341
342static void mul_eq_f128_int(float128_t *y, int sign) {
343 float128_t sign_f128;
344 i32_to_f128M(sign, &sign_f128);
345 float128_t new_value;
346 f128M_mul(y, &sign_f128, &new_value);
347 *y = new_value;
348}
349
350static float128_t make_f128(uint64_t hi, uint64_t lo) {
351 union ldshape ux;
352 ux.i2.hi = hi;
353 ux.i2.lo = lo;
354 return ux.f;
355}
356
357static void mul_eq_f128_f128(float128_t *a, float128_t b) {
358 float128_t new_value;
359 f128M_mul(a, &b, &new_value);
360 *a = new_value;
361}
362
363static void add_eq_f128_dbl(float128_t *a, double b) {
364 float64_t b_f64;
365 memcpy(&b_f64, &b, sizeof(double));
366
367 float128_t b_f128;
368 f64_to_f128M(b_f64, &b_f128);
369
370 float128_t new_value;
371 f128M_add(a, &b_f128, &new_value);
372 *a = new_value;
373}
374
375static float128_t scalbnf128(float128_t x, int n)
376{
377 union ldshape u;
378
379 if (n > 16383) {
380 //x *= 0x1p16383q;
381 mul_eq_f128_f128(&x, make_f128(0x7ffe000000000000, 0x0000000000000000));
382 n -= 16383;
383 if (n > 16383) {
384 //x *= 0x1p16383q;
385 mul_eq_f128_f128(&x, make_f128(0x7ffe000000000000, 0x0000000000000000));
386 n -= 16383;
387 if (n > 16383)
388 n = 16383;
389 }
390 } else if (n < -16382) {
391 //x *= 0x1p-16382q * 0x1p113q;
392 {
393 float128_t mul_result;
394 float128_t a = make_f128(0x0001000000000000, 0x0000000000000000);
395 float128_t b = make_f128(0x4070000000000000, 0x0000000000000000);
396 f128M_mul(&a, &b, &mul_result);
397 mul_eq_f128_f128(&x, mul_result);
398 }
399 n += 16382 - 113;
400 if (n < -16382) {
401 //x *= 0x1p-16382q * 0x1p113q;
402 {
403 float128_t mul_result;
404 float128_t a = make_f128(0x0001000000000000, 0x0000000000000000);
405 float128_t b = make_f128(0x4070000000000000, 0x0000000000000000);
406 f128M_mul(&a, &b, &mul_result);
407 mul_eq_f128_f128(&x, mul_result);
408 }
409 n += 16382 - 113;
410 if (n < -16382)
411 n = -16382;
412 }
413 }
414 //u.f = 1.0;
415 ui32_to_f128M(1, &u.f);
416 u.i.se = 0x3fff + n;
417 mul_eq_f128_f128(&x, u.f);
418 return x;
419}
420
421static float128_t fabsf128(float128_t x)
422{
423 union ldshape u = {x};
424
425 u.i.se &= 0x7fff;
426 return u.f;
427}
428
429static float128_t decfloat(struct MuslFILE *f, int c, int bits, int emin, int sign, int pok)
430{
431 uint32_t x[KMAX];
432 static const uint32_t th[] = { LD_B1B_MAX };
433 int i, j, k, a, z;
434 long long lrp=0, dc=0;
435 long long e10=0;
436 int lnz = 0;
437 int gotdig = 0, gotrad = 0;
438 int rp;
439 int e2;
440 int emax = -emin-bits+3;
441 int denormal = 0;
442 float128_t y;
443 float128_t zero;
444 ui32_to_f128M(0, &zero);
445 float128_t frac=zero;
446 float128_t bias=zero;
447 static const int p10s[] = { 10, 100, 1000, 10000,
448 100000, 1000000, 10000000, 100000000 };
449
450 j=0;
451 k=0;
452
453 /* Don't let leading zeros consume buffer space */
454 for (; c=='0'; c = shgetc(f)) gotdig=1;
455 if (c=='.') {
456 gotrad = 1;
457 for (c = shgetc(f); c=='0'; c = shgetc(f)) gotdig=1, lrp--;
458 }
459
460 x[0] = 0;
461 for (; c-'0'<10U || c=='.'; c = shgetc(f)) {
462 if (c == '.') {
463 if (gotrad) break;
464 gotrad = 1;
465 lrp = dc;
466 } else if (k < KMAX-3) {
467 dc++;
468 if (c!='0') lnz = dc;
469 if (j) x[k] = x[k]*10 + c-'0';
470 else x[k] = c-'0';
471 if (++j==9) {
472 k++;
473 j=0;
474 }
475 gotdig=1;
476 } else {
477 dc++;
478 if (c!='0') {
479 lnz = (KMAX-4)*9;
480 x[KMAX-4] |= 1;
481 }
482 }
483 }
484 if (!gotrad) lrp=dc;
485
486 if (gotdig && (c|32)=='e') {
487 e10 = scanexp(f, pok);
488 if (e10 == LLONG_MIN) {
489 if (pok) {
490 shunget(f);
491 } else {
492 shlim(f, 0);
493 return zero;
494 }
495 e10 = 0;
496 }
497 lrp += e10;
498 } else if (c>=0) {
499 shunget(f);
500 }
501 if (!gotdig) {
502 errno = EINVAL;
503 shlim(f, 0);
504 return zero;
505 }
506
507 /* Handle zero specially to avoid nasty special cases later */
508 if (!x[0]) {
509 //return sign * 0.0;
510 return dbl_to_f128(sign * 0.0);
511 }
512
513 /* Optimize small integers (w/no exponent) and over/under-flow */
514 if (lrp==dc && dc<10 && (bits>30 || x[0]>>bits==0)) {
515 //return sign * (float128_t)x[0];
516 float128_t sign_f128;
517 i32_to_f128M(sign, &sign_f128);
518 float128_t x0_f128;
519 ui32_to_f128M(x[0], &x0_f128);
520 float128_t result;
521 f128M_mul(&sign_f128, &x0_f128, &result);
522 return result;
523 }
524 if (lrp > -emin/2) {
525 errno = ERANGE;
526 //return sign * LDBL_MAX * LDBL_MAX;
527 return zero;
528 }
529 if (lrp < emin-2*LDBL_MANT_DIG) {
530 errno = ERANGE;
531 //return sign * LDBL_MIN * LDBL_MIN;
532 return zero;
533 }
534
535 /* Align incomplete final B1B digit */
536 if (j) {
537 for (; j<9; j++) x[k]*=10;
538 k++;
539 j=0;
540 }
541
542 a = 0;
543 z = k;
544 e2 = 0;
545 rp = lrp;
546
547 /* Optimize small to mid-size integers (even in exp. notation) */
548 if (lnz<9 && lnz<=rp && rp < 18) {
549 if (rp == 9) {
550 //return sign * (float128_t)(x[0]);
551 return int_mul_f128_cast_u32(sign, x[0]);
552 }
553 if (rp < 9) {
554 //return sign * (float128_t)(x[0]) / p10s[8-rp];
555 return triple_divide(sign, x[0], p10s[8-rp]);
556 }
557 int bitlim = bits-3*(int)(rp-9);
558 if (bitlim>30 || x[0]>>bitlim==0)
559 //return sign * (float128_t)(x[0]) * p10s[rp-10];
560 return triple_multiply(sign, x[0], p10s[rp-10]);
561 }
562
563 /* Drop trailing zeros */
564 for (; !x[z-1]; z--);
565
566 /* Align radix point to B1B digit boundary */
567 if (rp % 9) {
568 int rpm9 = rp>=0 ? rp%9 : rp%9+9;
569 int p10 = p10s[8-rpm9];
570 uint32_t carry = 0;
571 for (k=a; k!=z; k++) {
572 uint32_t tmp = x[k] % p10;
573 x[k] = x[k]/p10 + carry;
574 carry = 1000000000/p10 * tmp;
575 if (k==a && !x[k]) {
576 a = (a+1 & MASK);
577 rp -= 9;
578 }
579 }
580 if (carry) x[z++] = carry;
581 rp += 9-rpm9;
582 }
583
584 /* Upscale until desired number of bits are left of radix point */
585 while (rp < 9*LD_B1B_DIG || (rp == 9*LD_B1B_DIG && x[a]<th[0])) {
586 uint32_t carry = 0;
587 e2 -= 29;
588 for (k=(z-1 & MASK); ; k=(k-1 & MASK)) {
589 uint64_t tmp = ((uint64_t)x[k] << 29) + carry;
590 if (tmp > 1000000000) {
591 carry = tmp / 1000000000;
592 x[k] = tmp % 1000000000;
593 } else {
594 carry = 0;
595 x[k] = tmp;
596 }
597 if (k==(z-1 & MASK) && k!=a && !x[k]) z = k;
598 if (k==a) break;
599 }
600 if (carry) {
601 rp += 9;
602 a = (a-1 & MASK);
603 if (a == z) {
604 z = (z-1 & MASK);
605 x[z-1 & MASK] |= x[z];
606 }
607 x[a] = carry;
608 }
609 }
610
611 /* Downscale until exactly number of bits are left of radix point */
612 for (;;) {
613 uint32_t carry = 0;
614 int sh = 1;
615 for (i=0; i<LD_B1B_DIG; i++) {
616 k = (a+i & MASK);
617 if (k == z || x[k] < th[i]) {
618 i=LD_B1B_DIG;
619 break;
620 }
621 if (x[a+i & MASK] > th[i]) break;
622 }
623 if (i==LD_B1B_DIG && rp==9*LD_B1B_DIG) break;
624 /* FIXME: find a way to compute optimal sh */
625 if (rp > 9+9*LD_B1B_DIG) sh = 9;
626 e2 += sh;
627 for (k=a; k!=z; k=(k+1 & MASK)) {
628 uint32_t tmp = x[k] & (1<<sh)-1;
629 x[k] = (x[k]>>sh) + carry;
630 carry = (1000000000>>sh) * tmp;
631 if (k==a && !x[k]) {
632 a = (a+1 & MASK);
633 i--;
634 rp -= 9;
635 }
636 }
637 if (carry) {
638 if ((z+1 & MASK) != a) {
639 x[z] = carry;
640 z = (z+1 & MASK);
641 } else x[z-1 & MASK] |= 1;
642 }
643 }
644
645 /* Assemble desired bits into floating point variable */
646 for (y=zero,i=0; i<LD_B1B_DIG; i++) {
647 if ((a+i & MASK)==z) x[(z=(z+1 & MASK))-1] = 0;
648 //y = 1000000000.0L * y + x[a+i & MASK];
649 float128_t const_f128;
650 ui64_to_f128M(1000000000, &const_f128);
651 float128_t mul_y;
652 f128M_mul(&const_f128, &y, &mul_y);
653 float128_t x_f128;
654 ui32_to_f128M(x[a+i & MASK], &x_f128);
655 f128M_add(&mul_y, &x_f128, &y);
656 }
657
658 //y *= sign;
659 mul_eq_f128_int(&y, sign);
660
661 /* Limit precision for denormal results */
662 if (bits > LDBL_MANT_DIG+e2-emin) {
663 bits = LDBL_MANT_DIG+e2-emin;
664 if (bits<0) bits=0;
665 denormal = 1;
666 }
667
668 /* Calculate bias term to force rounding, move out lower bits */
669 if (bits < LDBL_MANT_DIG) {
670 bias = copysignf128(dbl_to_f128(scalbn(1, 2*LDBL_MANT_DIG-bits-1)), y);
671 frac = fmodf128(y, dbl_to_f128(scalbn(1, LDBL_MANT_DIG-bits)));
672 //y -= frac;
673 {
674 float128_t new_value;
675 f128M_sub(&y, &frac, &new_value);
676 y = new_value;
677 }
678 //y += bias;
679 {
680 float128_t new_value;
681 f128M_add(&y, &frac, &new_value);
682 y = new_value;
683 }
684 }
685
686 /* Process tail of decimal input so it can affect rounding */
687 if ((a+i & MASK) != z) {
688 uint32_t t = x[a+i & MASK];
689 if (t < 500000000 && (t || (a+i+1 & MASK) != z)) {
690 //frac += 0.25*sign;
691 add_eq_f128_dbl(&frac, 0.25*sign);
692 } else if (t > 500000000) {
693 //frac += 0.75*sign;
694 add_eq_f128_dbl(&frac, 0.75*sign);
695 } else if (t == 500000000) {
696 if ((a+i+1 & MASK) == z) {
697 //frac += 0.5*sign;
698 add_eq_f128_dbl(&frac, 0.5*sign);
699 } else {
700 //frac += 0.75*sign;
701 add_eq_f128_dbl(&frac, 0.75*sign);
702 }
703 }
704 //if (LDBL_MANT_DIG-bits >= 2 && !fmodf128(frac, 1))
705 if (LDBL_MANT_DIG-bits >= 2) {
706 float128_t one;
707 ui32_to_f128M(1, &one);
708 float128_t mod_result = fmodf128(frac, one);
709 if (f128M_eq(&mod_result, &zero)) {
710 //frac++;
711 add_eq_f128_dbl(&frac, 1.0);
712 }
713 }
714 }
715
716 //y += frac;
717 {
718 float128_t new_value;
719 f128M_add(&y, &frac, &new_value);
720 y = new_value;
721 }
722 //y -= bias;
723 {
724 float128_t new_value;
725 f128M_sub(&y, &bias, &new_value);
726 y = new_value;
727 }
728
729 if ((e2+LDBL_MANT_DIG & INT_MAX) > emax-5) {
730 //if (fabsf128(y) >= 0x1p113)
731 float128_t abs_y = fabsf128(y);
732 float128_t mant_f128 = make_f128(0x4070000000000000, 0x0000000000000000);
733 if (!f128M_lt(&abs_y, &mant_f128)) {
734 if (denormal && bits==LDBL_MANT_DIG+e2-emin)
735 denormal = 0;
736 //y *= 0.5;
737 {
738 float128_t point_5 = dbl_to_f128(0.5);
739 float128_t new_value;
740 f128M_mul(&y, &point_5, &new_value);
741 y = new_value;
742 }
743
744 e2++;
745 }
746 if (e2+LDBL_MANT_DIG>emax || (denormal && !f128M_eq(&frac, &zero)))
747 errno = ERANGE;
748 }
749
750 return scalbnf128(y, e2);
751}
752
753static float128_t hexfloat(struct MuslFILE *f, int bits, int emin, int sign, int pok)
754{
755 float128_t zero;
756 ui32_to_f128M(0, &zero);
757 float128_t one;
758 ui32_to_f128M(1, &one);
759 float128_t sixteen;
760 ui32_to_f128M(16, &sixteen);
761 float128_t point_5 = dbl_to_f128(0.5);
762
763 uint32_t x = 0;
764 float128_t y = zero;
765 float128_t scale = one;
766 float128_t bias = zero;
767 int gottail = 0, gotrad = 0, gotdig = 0;
768 long long rp = 0;
769 long long dc = 0;
770 long long e2 = 0;
771 int d;
772 int c;
773
774 c = shgetc(f);
775
776 /* Skip leading zeros */
777 for (; c=='0'; c = shgetc(f)) gotdig = 1;
778
779 if (c=='.') {
780 gotrad = 1;
781 c = shgetc(f);
782 /* Count zeros after the radix point before significand */
783 for (rp=0; c=='0'; c = shgetc(f), rp--) gotdig = 1;
784 }
785
786 for (; c-'0'<10U || (c|32)-'a'<6U || c=='.'; c = shgetc(f)) {
787 if (c=='.') {
788 if (gotrad) break;
789 rp = dc;
790 gotrad = 1;
791 } else {
792 gotdig = 1;
793 if (c > '9') d = (c|32)+10-'a';
794 else d = c-'0';
795 if (dc<8) {
796 x = x*16 + d;
797 } else if (dc < LDBL_MANT_DIG/4+1) {
798 //y += d*(scale/=16);
799 {
800 float128_t divided;
801 f128M_div(&scale, &sixteen, &divided);
802 scale = divided;
803 float128_t d_f128;
804 i32_to_f128M(d, &d_f128);
805 float128_t add_op;
806 f128M_mul(&d_f128, &scale, &add_op);
807 float128_t new_y;
808 f128M_add(&y, &add_op, &new_y);
809 y = new_y;
810 }
811 } else if (d && !gottail) {
812 //y += 0.5*scale;
813 {
814 float128_t add_op;
815 f128M_mul(&point_5, &scale, &add_op);
816 float128_t new_y;
817 f128M_add(&y, &add_op, &new_y);
818 y = new_y;
819 }
820 gottail = 1;
821 }
822 dc++;
823 }
824 }
825 if (!gotdig) {
826 shunget(f);
827 if (pok) {
828 shunget(f);
829 if (gotrad) shunget(f);
830 } else {
831 shlim(f, 0);
832 }
833 //return sign * 0.0;
834 return dbl_to_f128(sign * 0.0);
835 }
836 if (!gotrad) rp = dc;
837 while (dc<8) x *= 16, dc++;
838 if ((c|32)=='p') {
839 e2 = scanexp(f, pok);
840 if (e2 == LLONG_MIN) {
841 if (pok) {
842 shunget(f);
843 } else {
844 shlim(f, 0);
845 return zero;
846 }
847 e2 = 0;
848 }
849 } else {
850 shunget(f);
851 }
852 e2 += 4*rp - 32;
853
854 if (!x) {
855 //return sign * 0.0;
856 return dbl_to_f128(sign * 0.0);
857 }
858 if (e2 > -emin) {
859 errno = ERANGE;
860 //return sign * LDBL_MAX * LDBL_MAX;
861 return zero;
862 }
863 if (e2 < emin-2*LDBL_MANT_DIG) {
864 errno = ERANGE;
865 //return sign * LDBL_MIN * LDBL_MIN;
866 return zero;
867 }
868
869 while (x < 0x80000000) {
870 //if (y>=0.5)
871 if (!f128M_lt(&y, &point_5)) {
872 x += x + 1;
873 //y += y - 1;
874 {
875 float128_t minus_one;
876 f128M_sub(&y, &one, &minus_one);
877 float128_t new_y;
878 f128M_add(&y, &minus_one, &new_y);
879 y = new_y;
880 }
881 } else {
882 x += x;
883 //y += y;
884 {
885 float128_t new_y;
886 f128M_add(&y, &y, &new_y);
887 y = new_y;
888 }
889 }
890 e2--;
891 }
892
893 if (bits > 32+e2-emin) {
894 bits = 32+e2-emin;
895 if (bits<0) bits=0;
896 }
897
898 if (bits < LDBL_MANT_DIG) {
899 float128_t sign_f128;
900 i32_to_f128M(sign, &sign_f128);
901 bias = copysignf128(dbl_to_f128(scalbn(1, 32+LDBL_MANT_DIG-bits-1)), sign_f128);
902 }
903
904 //if (bits<32 && y && !(x&1)) x++, y=0;
905 if (bits<32 && !f128M_eq(&y, &zero) && !(x&1)) x++, y=zero;
906
907 //y = bias + sign*(float128_t)x + sign*y;
908 {
909 float128_t x_f128;
910 ui32_to_f128M(x, &x_f128);
911 float128_t sign_f128;
912 i32_to_f128M(sign, &sign_f128);
913 float128_t sign_mul_x;
914 f128M_mul(&sign_f128, &x_f128, &sign_mul_x);
915 float128_t sign_mul_y;
916 f128M_mul(&sign_f128, &y, &sign_mul_y);
917 float128_t bias_op;
918 f128M_add(&bias, &sign_mul_x, &bias_op);
919 float128_t new_y;
920 f128M_add(&bias_op, &sign_mul_y, &new_y);
921 y = new_y;
922 }
923 //y -= bias;
924 {
925 float128_t new_y;
926 f128M_sub(&y, &bias, &new_y);
927 y = new_y;
928 }
929
930 if (f128M_eq(&y, &zero)) errno = ERANGE;
931
932 return scalbnf128(y, e2);
933}
934
935static int isspace(int c)
936{
937 return c == ' ' || (unsigned)c-'\t' < 5;
938}
939
940static inline float128_t makeInf128() {
941 union ldshape ux;
942 ux.i2.hi = 0x7fff000000000000UL;
943 ux.i2.lo = 0x0UL;
944 return ux.f;
945}
946
947static inline float128_t makeNaN128() {
948 uint64_t rand = 0UL;
949 union ldshape ux;
950 ux.i2.hi = 0x7fff000000000000UL | (rand & 0xffffffffffffUL);
951 ux.i2.lo = 0x0UL;
952 return ux.f;
953}
954
955float128_t __floatscan(struct MuslFILE *f, int prec, int pok)
956{
957 int sign = 1;
958 size_t i;
959 int bits = LDBL_MANT_DIG;
960 int emin = LDBL_MIN_EXP-bits;
961 int c;
962
963 while (isspace((c=shgetc(f))));
964
965 if (c=='+' || c=='-') {
966 sign -= 2*(c=='-');
967 c = shgetc(f);
968 }
969
970 for (i=0; i<8 && (c|32)=="infinity"[i]; i++)
971 if (i<7) c = shgetc(f);
972 if (i==3 || i==8 || (i>3 && pok)) {
973 if (i!=8) {
974 shunget(f);
975 if (pok) for (; i>3; i--) shunget(f);
976 }
977 //return sign * INFINITY;
978 float128_t sign_f128;
979 i32_to_f128M(sign, &sign_f128);
980 float128_t infinity_f128 = makeInf128();
981 float128_t result;
982 f128M_mul(&sign_f128, &infinity_f128, &result);
983 return result;
984 }
985 if (!i) for (i=0; i<3 && (c|32)=="nan"[i]; i++)
986 if (i<2) c = shgetc(f);
987 if (i==3) {
988 if (shgetc(f) != '(') {
989 shunget(f);
990 return makeNaN128();
991 }
992 for (i=1; ; i++) {
993 c = shgetc(f);
994 if (c-'0'<10U || c-'A'<26U || c-'a'<26U || c=='_')
995 continue;
996 if (c==')') return makeNaN128();
997 shunget(f);
998 if (!pok) {
999 errno = EINVAL;
1000 shlim(f, 0);
1001 float128_t zero;
1002 ui32_to_f128M(0, &zero);
1003 return zero;
1004 }
1005 while (i--) shunget(f);
1006 return makeNaN128();
1007 }
1008 return makeNaN128();
1009 }
1010
1011 if (i) {
1012 shunget(f);
1013 errno = EINVAL;
1014 shlim(f, 0);
1015 float128_t zero;
1016 ui32_to_f128M(0, &zero);
1017 return zero;
1018 }
1019
1020 if (c=='0') {
1021 c = shgetc(f);
1022 if ((c|32) == 'x')
1023 return hexfloat(f, bits, emin, sign, pok);
1024 shunget(f);
1025 c = '0';
1026 }
1027
1028 return decfloat(f, c, bits, emin, sign, pok);
1029}
1030
1031float128_t parse_f128(const char *s, char **p) {
1032 struct MuslFILE f;
1033 sh_fromstring(&f, s);
1034 shlim(&f, 0);
1035 float128_t y = __floatscan(&f, 2, 1);
1036 off_t cnt = shcnt(&f);
1037 if (p) *p = cnt ? (char *)s + cnt : (char *)s;
1038 return y;
1039}
src/parse_f128.h created+21
...@@ -0,0 +1,21 @@
1/*
2 * Copyright (c) 2015 Andrew Kelley
3 *
4 * This file is part of zig, which is MIT licensed.
5 * See http://opensource.org/licenses/MIT
6 */
7
8#ifndef ZIG_PARSE_F128_H
9#define ZIG_PARSE_F128_H
10
11#include "softfloat_types.h"
12
13#ifdef __cplusplus
14#define ZIG_EXTERN_C extern "C"
15#else
16#define ZIG_EXTERN_C
17#endif
18
19ZIG_EXTERN_C float128_t parse_f128(const char *s, char **p);
20
21#endif
src/tokenizer.cpp+2-2
...@@ -293,10 +293,10 @@ static void cancel_token(Tokenize *t) {...@@ -293,10 +293,10 @@ static void cancel_token(Tokenize *t) {
293}293}
294294
295static void end_float_token(Tokenize *t) {295static void end_float_token(Tokenize *t) {
296 if (t->radix == 10) {296 if (t->radix == 10 || t->radix == 16) {
297 uint8_t *ptr_buf = (uint8_t*)buf_ptr(t->buf) + t->cur_tok->start_pos;297 uint8_t *ptr_buf = (uint8_t*)buf_ptr(t->buf) + t->cur_tok->start_pos;
298 size_t buf_len = t->cur_tok->end_pos - t->cur_tok->start_pos;298 size_t buf_len = t->cur_tok->end_pos - t->cur_tok->start_pos;
299 if (bigfloat_init_buf_base10(&t->cur_tok->data.float_lit.bigfloat, ptr_buf, buf_len)) {299 if (bigfloat_init_buf(&t->cur_tok->data.float_lit.bigfloat, ptr_buf, buf_len)) {
300 t->cur_tok->data.float_lit.overflow = true;300 t->cur_tok->data.float_lit.overflow = true;
301 }301 }
302 return;302 return;
test/compile_errors.zig+2-2
...@@ -4774,7 +4774,7 @@ pub fn addCases(cases: *tests.CompileErrorContext) void {...@@ -4774,7 +4774,7 @@ pub fn addCases(cases: *tests.CompileErrorContext) void {
4774 cases.add(4774 cases.add(
4775 "float literal too large error",4775 "float literal too large error",
4776 \\comptime {4776 \\comptime {
4777 \\ const a = 0x1.0p16384;4777 \\ const a = 0x1.0p18495;
4778 \\}4778 \\}
4779 ,4779 ,
4780 "tmp.zig:2:15: error: float literal out of range of any type",4780 "tmp.zig:2:15: error: float literal out of range of any type",
...@@ -4783,7 +4783,7 @@ pub fn addCases(cases: *tests.CompileErrorContext) void {...@@ -4783,7 +4783,7 @@ pub fn addCases(cases: *tests.CompileErrorContext) void {
4783 cases.add(4783 cases.add(
4784 "float literal too small error (denormal)",4784 "float literal too small error (denormal)",
4785 \\comptime {4785 \\comptime {
4786 \\ const a = 0x1.0p-16384;4786 \\ const a = 0x1.0p-19000;
4787 \\}4787 \\}
4788 ,4788 ,
4789 "tmp.zig:2:15: error: float literal out of range of any type",4789 "tmp.zig:2:15: error: float literal out of range of any type",
test/stage1/behavior/eval.zig+4-4
...@@ -385,10 +385,10 @@ test "@setEvalBranchQuota" {...@@ -385,10 +385,10 @@ test "@setEvalBranchQuota" {
385 }385 }
386}386}
387387
388// TODO test "float literal at compile time not lossy" {388test "float literal at compile time not lossy" {
389// TODO expect(16777216.0 + 1.0 == 16777217.0);389 expect(16777216.0 + 1.0 == 16777217.0);
390// TODO expect(9007199254740992.0 + 1.0 == 9007199254740993.0);390 expect(9007199254740992.0 + 1.0 == 9007199254740993.0);
391// TODO }391}
392392
393test "f32 at compile time is lossy" {393test "f32 at compile time is lossy" {
394 expect(f32(1 << 24) + 1 == 1 << 24);394 expect(f32(1 << 24) + 1 == 1 << 24);
test/stage1/behavior/math.zig+7-2
...@@ -324,11 +324,11 @@ test "quad hex float literal parsing accurate" {...@@ -324,11 +324,11 @@ test "quad hex float literal parsing accurate" {
324 }324 }
325 {325 {
326 var f: f128 = 0x1.353e45674d89abacc3a2ebf3ff4ffp-50;326 var f: f128 = 0x1.353e45674d89abacc3a2ebf3ff4ffp-50;
327 expect(@bitCast(u128, f) == 0x3fcd353e45674d89abacc3a2ebf3ff4f);327 expect(@bitCast(u128, f) == 0x3fcd353e45674d89abacc3a2ebf3ff50);
328 }328 }
329 {329 {
330 var f: f128 = 0x1.ed8764648369535adf4be3214567fp-9;330 var f: f128 = 0x1.ed8764648369535adf4be3214567fp-9;
331 expect(@bitCast(u128, f) == 0x3ff6ed8764648369535adf4be3214567);331 expect(@bitCast(u128, f) == 0x3ff6ed8764648369535adf4be3214568);
332 }332 }
333 const exp2ft = []f64{333 const exp2ft = []f64{
334 0x1.6a09e667f3bcdp-1,334 0x1.6a09e667f3bcdp-1,
...@@ -597,3 +597,8 @@ test "vector integer addition" {...@@ -597,3 +597,8 @@ test "vector integer addition" {
597 S.doTheTest();597 S.doTheTest();
598 comptime S.doTheTest();598 comptime S.doTheTest();
599}599}
600
601test "binary and octal float literals" {
602 expect(0b10100.00010e0 == 0x1.4100000000000p+4);
603 expect(0o10700.00010e0 == 0x1.1c00010000000p+12);
604}