| 1 | /** |
| 2 | * This file has no copyright assigned and is placed in the Public Domain. |
| 3 | * This file is part of the mingw-w64 runtime package. |
| 4 | * No warranty is given; refer to the file DISCLAIMER.PD within this package. |
| 5 | */ |
| 6 | #ifndef _MINGWEX_FASTMATH_H_ |
| 7 | #define _MINGWEX_FASTMATH_H_ |
| 8 | |
| 9 | /* Fast math inlines |
| 10 | No range or domain checks. No setting of errno. No tweaks to |
| 11 | protect precision near range limits. */ |
| 12 | |
| 13 | /* For now this is an internal header with just the functions that |
| 14 | are currently used in building libmingwex.a math components */ |
| 15 | |
| 16 | /* FIXME: We really should get rid of the code duplication using euther |
| 17 | C++ templates or tgmath-type macros. */ |
| 18 | |
| 19 | static __inline__ double __fast_sqrt (double x) |
| 20 | { |
| 21 | double res; |
| 22 | asm __volatile__ ("fsqrt" : "=t" (res) : "0" (x)); |
| 23 | return res; |
| 24 | } |
| 25 | |
| 26 | static __inline__ long double __fast_sqrtl (long double x) |
| 27 | { |
| 28 | long double res; |
| 29 | asm __volatile__ ("fsqrt" : "=t" (res) : "0" (x)); |
| 30 | return res; |
| 31 | } |
| 32 | |
| 33 | static __inline__ float __fast_sqrtf (float x) |
| 34 | { |
| 35 | float res; |
| 36 | asm __volatile__ ("fsqrt" : "=t" (res) : "0" (x)); |
| 37 | return res; |
| 38 | } |
| 39 | |
| 40 | |
| 41 | static __inline__ double __fast_log (double x) |
| 42 | { |
| 43 | double res; |
| 44 | asm __volatile__ |
| 45 | ("fldln2\n\t" |
| 46 | "fxch\n\t" |
| 47 | "fyl2x" |
| 48 | : "=t" (res) : "0" (x) : "st(1)"); |
| 49 | return res; |
| 50 | } |
| 51 | |
| 52 | static __inline__ long double __fast_logl (long double x) |
| 53 | { |
| 54 | long double res; |
| 55 | asm __volatile__ |
| 56 | ("fldln2\n\t" |
| 57 | "fxch\n\t" |
| 58 | "fyl2x" |
| 59 | : "=t" (res) : "0" (x) : "st(1)"); |
| 60 | return res; |
| 61 | } |
| 62 | |
| 63 | |
| 64 | static __inline__ float __fast_logf (float x) |
| 65 | { |
| 66 | float res; |
| 67 | asm __volatile__ |
| 68 | ("fldln2\n\t" |
| 69 | "fxch\n\t" |
| 70 | "fyl2x" |
| 71 | : "=t" (res) : "0" (x) : "st(1)"); |
| 72 | return res; |
| 73 | } |
| 74 | |
| 75 | static __inline__ double __fast_log1p (double x) |
| 76 | { |
| 77 | double res; |
| 78 | /* fyl2xp1 accurate only for |x| <= 1.0 - 0.5 * sqrt (2.0) */ |
| 79 | if (fabs (x) >= 1.0 - 0.5 * 1.41421356237309504880) |
| 80 | res = __fast_log (1.0 + x); |
| 81 | else |
| 82 | asm __volatile__ |
| 83 | ("fldln2\n\t" |
| 84 | "fxch\n\t" |
| 85 | "fyl2xp1" |
| 86 | : "=t" (res) : "0" (x) : "st(1)"); |
| 87 | return res; |
| 88 | } |
| 89 | |
| 90 | static __inline__ long double __fast_log1pl (long double x) |
| 91 | { |
| 92 | long double res; |
| 93 | /* fyl2xp1 accurate only for |x| <= 1.0 - 0.5 * sqrt (2.0) */ |
| 94 | if (fabsl (x) >= 1.0L - 0.5L * 1.41421356237309504880L) |
| 95 | res = __fast_logl (1.0L + x); |
| 96 | else |
| 97 | asm __volatile__ |
| 98 | ("fldln2\n\t" |
| 99 | "fxch\n\t" |
| 100 | "fyl2xp1" |
| 101 | : "=t" (res) : "0" (x) : "st(1)"); |
| 102 | return res; |
| 103 | } |
| 104 | |
| 105 | static __inline__ float __fast_log1pf (float x) |
| 106 | { |
| 107 | float res; |
| 108 | /* fyl2xp1 accurate only for |x| <= 1.0 - 0.5 * sqrt (2.0) */ |
| 109 | if (fabsf (x) >= 1.0 - 0.5 * 1.41421356237309504880) |
| 110 | res = __fast_logf (1.0 + x); |
| 111 | else |
| 112 | asm __volatile__ |
| 113 | ("fldln2\n\t" |
| 114 | "fxch\n\t" |
| 115 | "fyl2xp1" |
| 116 | : "=t" (res) : "0" (x) : "st(1)"); |
| 117 | return res; |
| 118 | } |
| 119 | |
| 120 | #endif |