The following code shows that the rounding mode of the sse2 FPU is altered by a call to fmaf(). #include <stdio.h> #include <float.h> #include <math.h> static float res0, res1, res2, res3; static float half = 0.5f; void chk( void ){ res2 = res0 + half; /* should round down to even */ if( res2 == res0 ){ puts("OK1"); }else{ puts("Bad1"); } res3 = res1 + half; /* should round up to even */ if( res3 != res1 ){ puts("OK2"); }else{ puts("Bad2"); } } int main(void){ res0 = FLT_EPSILON; res0++; res0--; res0 = 1.f / res0; /* even; 1/EPSILON is a magic number */ res1 = res0 + 1.f; /* odd */ chk(); res2 = sinf( 1.f ); /* sse2 rounding is not altered */ chk(); res2 = fmaf( 1.f, 1.f, 1.f ); /* this messes up sse2 rounding */ chk(); return 0; } Hardware: Intel Core 2 Duo in 32-bit mode O.S. : Fedora Core 17 in 32-bit mode Compiler: gcc 4.7.0-2 Library : glibc 2.15-32 Compiler options: -std=gnu11 -O0 -mfpmath=sse -msse2 -fno-builtin -frounding-math -mieee-fp -ffloat-store -fexcess-precision=standard
This looks like a glibc issue, doesn't it? Or do you see something wrong with the code gcc produces for this example?
Indeed.