On Wed, Aug 10, 2005 at 07:09:04AM -0700, H. J. Lu wrote:
> On Tue, Aug 09, 2005 at 02:58:51PM -0700, Richard Henderson wrote:
> > On Tue, Aug 09, 2005 at 02:30:46PM -0700, H. J. Lu wrote:
> > > There is a minor problem. How can I add crtfastmath.o for SSE targets
> > > only?
> >
> > You don't. You either add code to detect sse, or you make the
> > spec depend on -mfpmath=sse.
> >
>
> Here is the patch to enable FTZ/DAZ for SSE via fast math. There are
> no regressions on Linux/x86_64 nor Linux/ia32. The performance of one
> FP benchmark on EM64T is more than doubled with -ffast-math.
Not all i?86 CPUs support cpuid instruction.
Please look at
gcc/testsuite/gcc.dg/i386-cpuid.h
for the ugly details.
> +static void __attribute__((constructor))
> +set_fast_math (void)
> +{
> + /* Check if SSE is available. */
> + unsigned int eax, ebx, ecx, edx;
> + asm volatile ("xchgl %%ebx, %1; cpuid; xchgl %%ebx, %1"
> + : "=a" (eax), "=r" (ebx), "=c" (ecx), "=d" (edx)
> + : "0" (1));
> +
> + if (edx & (1 << 25))
> + {
> + unsigned int mxcsr = __builtin_ia32_stmxcsr ();
> + mxcsr |= MXCSR_DAZ | MXCSR_FTZ;
> + __builtin_ia32_ldmxcsr (mxcsr);
> + }
> +}
Jakub
