Re: x86: Prepare `strrchr-evex` and `strrchr-evex512` for AVX10

[Noah Goldstein <goldstein.w.n@gmail.com>]

On Wed, Nov 1, 2023 at 4:11 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
>
> On Wed, Nov 1, 2023 at 4:04 PM Florian Weimer <fw@deneb.enyo.de> wrote:
> >
> > * Florian Weimer:
> >
> > > * Noah Goldstein:
> > >
> > >> This commit refactors `strrchr-evex` and `strrchr-evex512` to use a
> > >> common implementation: `strrchr-evex-base.S`.
> > >>
> > >> The motivation is `strrchr-evex` needed to be refactored to not use
> > >> 64-bit masked registers in preperation for AVX10.
> > >>
> > >> Once vec-width masked register combining was removed, the EVEX and
> > >> EVEX512 implementations can easily be implemented in the same file
> > >> without any major overhead.
> > >>
> > >> The net result is performance improvements (measured on TGL) for both
> > >> `strrchr-evex` and `strrchr-evex512`. Although, note there are some
> > >> regressions in the test suite and it may be many of the cases that
> > >> make the total-geomean of improvement/regression across bench-strrchr
> > >> are cold. The point of the performance measurement is to show there
> > >> are no major regressions, but the primary motivation is preperation
> > >> for AVX10.
> > >>
> > >> Benchmarks where taken on TGL:
> > >> https://www.intel.com/content/www/us/en/products/sku/213799/intel-core-i711850h-processor-24m-cache-up-to-4-80-ghz/specifications.html
> > >>
> > >> EVEX geometric_mean(N=5) of all benchmarks New / Original   : 0.74
> > >> EVEX512 geometric_mean(N=5) of all benchmarks New / Original: 0.87
> > >>
> > >> Full check passes on x86.
> > >
> > > I believe this caused some sort of regression because when we upgraded
> > > glibc in the Fedora rawhide buildroot, a lot of things started failing:
> > >
> > >   glibc-2.38.9000-13.fc40 broke rawhide buildroot on x86_64
> > >   <https://bugzilla.redhat.com/show_bug.cgi?id=2244688>
> > >
> > > The list of changes relative to the previous version is rather short:
> > >
> > > - stdlib: fix grouping verification with multi-byte thousands separator (bug 30964)
> > > - build-many-glibcs: Check for required system tools
> > > - x86: Prepare `strrchr-evex` and `strrchr-evex512` for AVX10
> > > - aarch64: Optimise vecmath logs
> > > - aarch64: Cosmetic change in SVE exp routines
> > > - aarch64: Optimize SVE cos & cosf
> > > - aarch64: Improve vecmath sin routines
> > > - nss: Get rid of alloca usage in makedb's write_output.
> > > - debug: Add regression tests for BZ 30932
> > > - Fix FORTIFY_SOURCE false positive
> > > - nss: Rearrange and sort Makefile variables
> > > - inet: Rearrange and sort Makefile variables
> > > - Fix off-by-one OOB write in iconv/tst-iconv-mt
> > >
> > > And this patch is the most likely one to cause issues.  I will try to
> > > revert the patch and see if it fixes the observed issues.
> >
> > We did the revert and the issues were gone.  So I think this commit is
> > faulty.
>
> Bah, didn't see your last email.
> Thank you for reverting. Will look into the issue.

Okay bug is missing VBMI2 check. But the VBMI2 stuff
isn't really needed so will update and repost w/ fixed ISA.