[Bug target/80355] Improve __builtin_shuffle on AVX512F

[Bug target/80355] Improve __builtin_shuffle on AVX512F

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80355

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2021-08-07
             Status|UNCONFIRMED                 |NEW
           Severity|normal                      |enhancement
     Ever confirmed|0                           |1

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed.

[Bug target/80355] Improve __builtin_shuffle on AVX512F

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80355

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|unassigned at gcc dot gnu.org      |jakub at gcc dot gnu.org
             Status|NEW                         |ASSIGNED

[Bug target/80355] Improve __builtin_shuffle on AVX512F

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80355

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 51277
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51277&action=edit
gcc12-pr80355-1.patch

Untested fix.  For 32-byte vectors/AVX512VL we don't need this, we already emit
vperm2i128 or vpermq.

[Bug target/80355] Improve __builtin_shuffle on AVX512F

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80355

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 51278
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51278&action=edit
gcc12-pr80355-2.patch

And this incremental patch makes it handle even similar V32HImode/V64QImode
permutations with -mavx512f -mno-avx512bw.

[Bug target/80355] Improve __builtin_shuffle on AVX512F

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80355

--- Comment #4 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jakub Jelinek <jakub@gcc.gnu.org>:

https://gcc.gnu.org/g:7665af0b1a964b1baae3a59b22fcc420369c63cf

commit r12-2835-g7665af0b1a964b1baae3a59b22fcc420369c63cf
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Tue Aug 10 11:34:53 2021 +0200

    i386: Improve single operand AVX512F permutations [PR80355]

    On the following testcase we emit
            vmovdqa32       .LC0(%rip), %zmm1
            vpermd  %zmm0, %zmm1, %zmm0
    and
            vmovdqa64       .LC1(%rip), %zmm1
            vpermq  %zmm0, %zmm1, %zmm0
    instead of
            vshufi32x4      $78, %zmm0, %zmm0, %zmm0
    and
            vshufi64x2      $78, %zmm0, %zmm0, %zmm0
    we can emit with the patch.  We have patterns that match two argument
    permutations for vshuf[if]*, but for one argument it doesn't trigger.
    Either we can add two patterns for that, or we would need to add another
    routine to i386-expand.c that would transform under certain condition
    these cases to the two argument vshuf*, doing it in sse.md looked simpler.
    We don't need this for 32-byte vectors, we already emit single insn
    permutation that doesn't need memory op there.

    2021-08-10  Jakub Jelinek  <jakub@redhat.com>

            PR target/80355
            * config/i386/sse.md
(*avx512f_shuf_<shuffletype>64x2_1<mask_name>_1,
            *avx512f_shuf_<shuffletype>32x4_1<mask_name>_1): New define_insn
            patterns.

            * gcc.target/i386/avx512f-pr80355-1.c: New test.

[Bug target/80355] Improve __builtin_shuffle on AVX512F

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80355

--- Comment #5 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jakub Jelinek <jakub@gcc.gnu.org>:

https://gcc.gnu.org/g:50b5877925ef5ae8e9f913d6d2b5ce0204ebc588

commit r12-2837-g50b5877925ef5ae8e9f913d6d2b5ce0204ebc588
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Tue Aug 10 12:38:00 2021 +0200

    i386: Allow some V32HImode and V64QImode permutations even without AVX512BW
[PR80355]

    When working on the PR, I've noticed we generate terrible code for
    V32HImode or V64QImode permutations for -mavx512f -mno-avx512bw.
    Generally we can't do much with such permutations, but since PR68655
    we can handle at least some, those expressible using V16SImode or V8DImode
    permutations, but that wasn't reachable, because
ix86_vectorize_vec_perm_const
    didn't even try, it said without TARGET_AVX512BW it can't do anything, and
    with it can do everything, no d.testing_p attempts.

    This patch makes it try it for TARGET_AVX512F && !TARGET_AVX512BW.

    The first hunk is to avoid ICE, expand_vec_perm_even_odd_1 asserts d->vmode
    isn't V32HImode because expand_vec_perm_1 for AVX512BW handles already
    all permutations, but when we let it through without !TARGET_AVX512BW,
    expand_vec_perm_1 doesn't handle it.

    If we want, that hunk can be dropped if we implement in
    expand_vec_perm_even_odd_1 and its helper the even permutation as
    vpmovdw + vpmovdw + vinserti64x4 and odd permutation as
    vpsrld $16 + vpsrld $16 + vpmovdw + vpmovdw + vinserti64x4.

    2021-08-10  Jakub Jelinek  <jakub@redhat.com>

            PR target/80355
            * config/i386/i386-expand.c (expand_vec_perm_even_odd): Return
false
            for V32HImode if !TARGET_AVX512BW.
            (ix86_vectorize_vec_perm_const) <case E_V32HImode, case
E_V64QImode>:
            If !TARGET_AVX512BW and TARGET_AVX512F and d.testing_p, don't fail
            early, but actually check the permutation.

            * gcc.target/i386/avx512f-pr80355-2.c: New test.

[Bug target/80355] Improve __builtin_shuffle on AVX512F

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80355

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
             Status|ASSIGNED                    |RESOLVED

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Fixed.