[PATCH] x86-64: Don't use SSE resolvers for ISA level 3 or above

[Sunil Pandey]

On Fri, Mar 1, 2024 at 2:23 PM Richard Henderson <
richard.henderson@linaro.org> wrote:

> On 2/29/24 16:40, Sunil Pandey wrote:
> >
> >
> > On Thu, Feb 29, 2024 at 12:55 PM Richard Henderson <
> richard.henderson@linaro.org
> > <mailto:richard.henderson@linaro.org>> wrote:
> >
> >     On 2/28/24 07:51, H.J. Lu wrote:
> >      > --- a/sysdeps/x86/cpu-features.c
> >      > +++ b/sysdeps/x86/cpu-features.c
> >      > @@ -18,6 +18,7 @@
> >      >
> >      >   #include <dl-hwcap.h>
> >      >   #include <libc-pointer-arith.h>
> >      > +#include <isa-level.h>
> >      >   #include <get-isa-level.h>
> >      >   #include <cacheinfo.h>
> >      >   #include <dl-cacheinfo.h>
> >      > @@ -1195,7 +1196,9 @@ no_cpuid:
> >      >              TUNABLE_CALLBACK (set_x86_shstk));
> >      >   #endif
> >      >
> >      > +#if MINIMUM_X86_ISA_LEVEL < AVX_X86_ISA_LEVEL
> >      >     if (GLRO(dl_x86_cpu_features).xsave_state_size != 0)
> >      > +#endif
> >      >       {
> >      >         if (CPU_FEATURE_USABLE_P (cpu_features, XSAVEC))
> >      >       {
> >
> >     Surely this can be written with IF and not IFDEF:
> >
> >         if (MINIMUM_X86_ISA_LEVEL >= AVX_X86_ISA_LEVEL
> >             || GLRO(dl_x86_cpu_features).xsave_state_size != 0)
> >
> >
> > For sure it can be written like IF, but using IFDEF has the following
> perf and code size
> > advantage.
> >
> > For MINIMUM_X86_ISA_LEVEL >= AVX_X86_ISA_LEVEL case
> >
> > it will remove conditional
> >
> > if (GLRO(dl_x86_cpu_features).xsave_state_size != 0)
> >
> > and else block code of this conditional during preprocessing.
>
> You miss the point -- both of these are constants and will be folded by
> the compiler.
> There is no perf or code size advantage for cpp.
>
>
Got it. Will submit a patch to fix it.

--Sunil
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://sourceware.org/pipermail/libc-alpha/attachments/20240301/ea0e94e9/attachment.htm>