On Wed, Oct 19, 2022 at 1:25 AM liuhongt <hongtao....@intel.com> wrote:
>
> Fix unexpected non-canon form from gimple vector selector.
>
> Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}.
> Ok for trunk?
>
> gcc/ChangeLog:
>
> PR target/107271
> * config/i386/i386-expand.cc (ix86_vec_perm_index_canon): New.
> (expand_vec_perm_shufps_shufps): Call
> ix86_vec_perm_index_canon
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/i386/pr107271.c: New test.
OK.
Thanks,
Uros.
> ---
> gcc/config/i386/i386-expand.cc | 17 +++++++++++++++++
> gcc/testsuite/gcc.target/i386/pr107271.c | 16 ++++++++++++++++
> 2 files changed, 33 insertions(+)
> create mode 100644 gcc/testsuite/gcc.target/i386/pr107271.c
>
> diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc
> index 6baff6d0e61..4f121516091 100644
> --- a/gcc/config/i386/i386-expand.cc
> +++ b/gcc/config/i386/i386-expand.cc
> @@ -19604,6 +19604,22 @@ expand_vec_perm_1 (struct expand_vec_perm_d *d)
> return false;
> }
>
> +/* Canonicalize vec_perm index to make the first index
> + always comes from the first index. */
> +static void
> +ix86_vec_perm_index_canon (struct expand_vec_perm_d *d)
> +{
> + unsigned nelt = d->nelt;
> + if (d->perm[0] < nelt)
> + return;
> +
> + for (unsigned i = 0; i != nelt; i++)
> + d->perm[i] = (d->perm[i] + nelt) % (2 * nelt);
> +
> + std::swap (d->op0, d->op1);
> + return;
> +}
> +
> /* A subroutine of ix86_expand_vec_perm_const_1. Try to implement D
> in terms of a pair of shufps+ shufps/pshufd instructions. */
> static bool
> @@ -19621,6 +19637,7 @@ expand_vec_perm_shufps_shufps (struct
> expand_vec_perm_d *d)
> if (d->testing_p)
> return true;
>
> + ix86_vec_perm_index_canon (d);
> for (i = 0; i < 4; ++i)
> count += d->perm[i] > 3 ? 1 : 0;
>
> diff --git a/gcc/testsuite/gcc.target/i386/pr107271.c
> b/gcc/testsuite/gcc.target/i386/pr107271.c
> new file mode 100644
> index 00000000000..fe89c9a5bef
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr107271.c
> @@ -0,0 +1,16 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O0" } */
> +
> +typedef int __attribute__((__vector_size__ (16))) V;
> +
> +static inline __attribute__((__always_inline__)) V
> +bar (V v128u32_0)
> +{
> + return __builtin_shuffle ((V){}, v128u32_0, v128u32_0);
> +}
> +
> +V
> +foo (void)
> +{
> + return bar ((V){7, 4, 4});
> +}
> --
> 2.27.0
>
