On Thu, Aug 12, 2021 at 3:49 PM Jakub Jelinek <ja...@redhat.com> wrote:
>
> Hi!
>
> My patch from yesterday apparently broke some V32HImode permutations
> as the testcase shows.
> The first function assumed it would never be called in d->testing_p mode
> and so went right away into emitting the code.
> And the second one assumed V32HImode would never reach it, which now
> can for the !TARGET_AVX512BW case. We don't have a instruction
> in that case though.
>
> Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
> trunk?
LGTM.
>
> 2021-08-12 Jakub Jelinek <ja...@redhat.com>
>
> PR target/101860
> * config/i386/i386-expand.c (ix86_expand_vec_one_operand_perm_avx512):
> If d->testing_p, return true after performing checks instead of
> actually expanding the insn.
> (expand_vec_perm_broadcast_1): Handle V32HImode - assert
> !TARGET_AVX512BW and return false.
>
> * gcc.target/i386/avx512f-pr101860.c: New test.
>
> --- gcc/config/i386/i386-expand.c.jj 2021-08-10 12:37:53.867159317 +0200
> +++ gcc/config/i386/i386-expand.c 2021-08-11 11:02:03.994908828 +0200
> @@ -18116,6 +18116,9 @@ ix86_expand_vec_one_operand_perm_avx512
> return false;
> }
>
> + if (d->testing_p)
> + return true;
> +
> target = d->target;
> op0 = d->op0;
> for (int i = 0; i < d->nelt; ++i)
> @@ -20481,6 +20484,10 @@ expand_vec_perm_broadcast_1 (struct expa
> gcc_assert (!TARGET_AVX2 || d->perm[0]);
> return false;
>
> + case E_V32HImode:
> + gcc_assert (!TARGET_AVX512BW);
> + return false;
> +
> default:
> gcc_unreachable ();
> }
> --- gcc/testsuite/gcc.target/i386/avx512f-pr101860.c.jj 2021-08-11
> 11:05:28.090072461 +0200
> +++ gcc/testsuite/gcc.target/i386/avx512f-pr101860.c 2021-08-11
> 11:05:17.157224399 +0200
> @@ -0,0 +1,5 @@
> +/* PR target/101860 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx512f -mno-avx512bw" } */
> +
> +#include "../../gcc.dg/torture/vshuf-v32hi.c"
>
> Jakub
>
--
BR,
Hongtao
