i386: Allow some V32HImode and V64QImode permutations even without AVX512BW [PR80355]
authorJakub Jelinek <jakub@redhat.com>
Tue, 10 Aug 2021 10:38:00 +0000 (12:38 +0200)
committerJakub Jelinek <jakub@redhat.com>
Tue, 10 Aug 2021 10:38:00 +0000 (12:38 +0200)
commit50b5877925ef5ae8e9f913d6d2b5ce0204ebc588
tree585409f12939f0216c515a7274967044f6e48e6e
parent08aa0e3d4f781fd6a6e293bb06d280365a0bdc1d
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.
gcc/config/i386/i386-expand.c
gcc/testsuite/gcc.target/i386/avx512f-pr80355-2.c [new file with mode: 0644]