Emilio G. Cota <c...@braap.org> writes:
> On Thu, Dec 20, 2018 at 11:10:08 +0000, Alex Bennée wrote: > (snip) >> +#if defined(__GLIBC__) && (__GLIBC__ == 2 && __GLIBC_MINOR__ <= 12) >> +#define QEMU_HARDFLOAT_USE_FMA 0 >> +#else >> +#define QEMU_HARDFLOAT_USE_FMA 1 >> +#endif >> +#else >> +#define QEMU_HARDFLOAT_USE_FMA 1 >> +#endif >> + >> /* >> * QEMU_HARDFLOAT_USE_ISINF chooses whether to use isinf() over >> * float{32,64}_is_infinity when !USE_FP. >> @@ -1551,6 +1570,9 @@ float32_muladd(float32 xa, float32 xb, float32 xc, int >> flags, float_status *s) >> ub.s = xb; >> uc.s = xc; >> >> + if (!QEMU_HARDFLOAT_USE_FMA) { >> + goto soft; >> + } > > I don't think this should be a compile-time check; if the QEMU binary > is run on a system with a newer, fixed glibc (or any other libc), then > we'll have disabled fma hardfloat unnecessarily. > > What do you think about the following? > > Laurent: if you want to test the below, you can pull it from > https://github.com/cota/qemu/tree/fma-fix > > Thanks, > > Emilio > --- > commit ddeec29a2c33550c5d018aeea05d45a23579ae1b > Author: Emilio G. Cota <c...@braap.org> > Date: Fri Dec 21 14:08:57 2018 -0500 > > softfloat: enforce softfloat if the host's FMA is broken > > The added branch is marked as unlikely and therefore its impact > on performance (measured with fp-bench) is within the noise range > when measured on an Intel(R) Xeon(R) Gold 6142 CPU @ 2.60GHz. > > Laurent Desnogues <laurent.desnog...@gmail.com> > Signed-off-by: Emilio G. Cota <c...@braap.org> I've applied b8cc3928cee7b1d91bf39c86bec4801b9dc612e1 from your tree to fpu/next although the numbers look a bit odd. I see: Before: 143.83 MFlops 89.34 MFlops After: 150.20 MFlops 85.73 MFlops On my i7-4770 where as your commit seems to show a big jump in performance which is odd as this is preventing a bug not enabling FMA. > + > +static void __attribute__((constructor)) softfloat_init(void) > +{ > + union_float64 ua, ub, uc, ur; > + > + if (QEMU_NO_HARDFLOAT) { > + return; > + } > + > + /* > + * Test that the host's FMA is not obviously broken. For example, > + * glibc < 2.23 can perform an incorrect FMA on certain hosts; see > + * https://sourceware.org/bugzilla/show_bug.cgi?id=13304 > + */ > + ua.s = 0x0020000000000001; > + ub.s = 0x3ca0000000000000; > + uc.s = 0x0020000000000000; > + ur.h = fma(ua.h, ub.h, uc.h); > + if (ur.s != 0x0020000000000001) { > + host_fma_is_broken = true; > + } > +} I'm fine with the cpuid stuff at the bottom of softfloat for now. We can move it later if the other micro-architectures want to get in on the detecting features game. -- Alex Bennée