X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=src%2Fmath%2Ffmaf.c;h=aa57feb69567c67a79aae02fc7ea18e07d22fd12;hb=6739b13a172aad9c01572c04cadacc99c7041811;hp=cc427fb5ac987b15cad113d02183145f7e3c1c68;hpb=2e77dc13f8bc2053642fcb136996f5f36c88c775;p=musl diff --git a/src/math/fmaf.c b/src/math/fmaf.c index cc427fb5..aa57feb6 100644 --- a/src/math/fmaf.c +++ b/src/math/fmaf.c @@ -26,7 +26,8 @@ */ #include -#include "libm.h" +#include +#include /* * Fused multiply-add: Compute x * y + z with a single rounding error. @@ -37,18 +38,40 @@ */ float fmaf(float x, float y, float z) { + #pragma STDC FENV_ACCESS ON double xy, result; - uint32_t hr, lr; + union {double f; uint64_t i;} u; + int e; xy = (double)x * y; result = xy + z; - EXTRACT_WORDS(hr, lr, result); + u.f = result; + e = u.i>>52 & 0x7ff; /* Common case: The double precision result is fine. */ - if ((lr & 0x1fffffff) != 0x10000000 || /* not a halfway case */ - (hr & 0x7ff00000) == 0x7ff00000 || /* NaN */ + if ((u.i & 0x1fffffff) != 0x10000000 || /* not a halfway case */ + e == 0x7ff || /* NaN */ result - xy == z || /* exact */ fegetround() != FE_TONEAREST) /* not round-to-nearest */ - return (result); + { + /* + underflow may not be raised correctly, example: + fmaf(0x1p-120f, 0x1p-120f, 0x1p-149f) + */ +#if defined(FE_INEXACT) && defined(FE_UNDERFLOW) + if (e < 0x3ff-126 && e >= 0x3ff-149 && fetestexcept(FE_INEXACT)) { + feclearexcept(FE_INEXACT); + /* TODO: gcc and clang bug workaround */ + volatile float vz = z; + result = xy + vz; + if (fetestexcept(FE_INEXACT)) + feraiseexcept(FE_UNDERFLOW); + else + feraiseexcept(FE_INEXACT); + } +#endif + z = result; + return z; + } /* * If result is inexact, and exactly halfway between two float values, @@ -60,7 +83,11 @@ float fmaf(float x, float y, float z) volatile double vxy = xy; /* XXX work around gcc CSE bug */ double adjusted_result = vxy + z; fesetround(FE_TONEAREST); - if (result == adjusted_result) - SET_LOW_WORD(adjusted_result, lr + 1); - return (adjusted_result); + if (result == adjusted_result) { + u.f = adjusted_result; + u.i++; + adjusted_result = u.f; + } + z = adjusted_result; + return z; }