From 5b5b04d6282df0364424c6f2c0462e5c1a4394b0 Mon Sep 17 00:00:00 2001 From: Joseph Myers Date: Sat, 3 Nov 2012 19:48:53 +0000 Subject: Make fma use of Dekker and Knuth algorithms use round-to-nearest (bug 14796). --- sysdeps/ieee754/ldbl-96/s_fma.c | 18 ++++++++++++++++-- sysdeps/ieee754/ldbl-96/s_fmal.c | 18 ++++++++++++++++-- 2 files changed, 32 insertions(+), 4 deletions(-) (limited to 'sysdeps/ieee754/ldbl-96') diff --git a/sysdeps/ieee754/ldbl-96/s_fma.c b/sysdeps/ieee754/ldbl-96/s_fma.c index 001d8063d4..bf2d794c9b 100644 --- a/sysdeps/ieee754/ldbl-96/s_fma.c +++ b/sysdeps/ieee754/ldbl-96/s_fma.c @@ -42,6 +42,10 @@ __fma (double x, double y, double z) if (__builtin_expect ((x == 0 || y == 0) && z == 0, 0)) return x * y + z; + fenv_t env; + feholdexcept (&env); + fesetround (FE_TONEAREST); + /* Multiplication m1 + m2 = x * y using Dekker's algorithm. */ #define C ((1ULL << (LDBL_MANT_DIG + 1) / 2) + 1) long double x1 = (long double) x * C; @@ -60,9 +64,19 @@ __fma (double x, double y, double z) t1 = m1 - t1; t2 = z - t2; long double a2 = t1 + t2; + feclearexcept (FE_INEXACT); + + /* If the result is an exact zero, ensure it has the correct + sign. */ + if (a1 == 0 && m2 == 0) + { + feupdateenv (&env); + /* Ensure that round-to-nearest value of z + m1 is not + reused. */ + asm volatile ("" : "=m" (z) : "m" (z)); + return z + m1; + } - fenv_t env; - feholdexcept (&env); fesetround (FE_TOWARDZERO); /* Perform m2 + a2 addition with round to odd. */ a2 = a2 + m2; diff --git a/sysdeps/ieee754/ldbl-96/s_fmal.c b/sysdeps/ieee754/ldbl-96/s_fmal.c index b5fdcba6bc..32e71a18ba 100644 --- a/sysdeps/ieee754/ldbl-96/s_fmal.c +++ b/sysdeps/ieee754/ldbl-96/s_fmal.c @@ -168,6 +168,10 @@ __fmal (long double x, long double y, long double z) if (__builtin_expect ((x == 0 || y == 0) && z == 0, 0)) return x * y + z; + fenv_t env; + feholdexcept (&env); + fesetround (FE_TONEAREST); + /* Multiplication m1 + m2 = x * y using Dekker's algorithm. */ #define C ((1LL << (LDBL_MANT_DIG + 1) / 2) + 1) long double x1 = x * C; @@ -186,9 +190,19 @@ __fmal (long double x, long double y, long double z) t1 = m1 - t1; t2 = z - t2; long double a2 = t1 + t2; + feclearexcept (FE_INEXACT); + + /* If the result is an exact zero, ensure it has the correct + sign. */ + if (a1 == 0 && m2 == 0) + { + feupdateenv (&env); + /* Ensure that round-to-nearest value of z + m1 is not + reused. */ + asm volatile ("" : "=m" (z) : "m" (z)); + return z + m1; + } - fenv_t env; - feholdexcept (&env); fesetround (FE_TOWARDZERO); /* Perform m2 + a2 addition with round to odd. */ u.d = a2 + m2; -- cgit 1.4.1