Re: [PATCH] Don't check can_vec_perm_const_p for nonlinear iv_init when it's constant.

[Richard Biener <richard.guenther@gmail.com>]

On Wed, Sep 21, 2022 at 1:41 AM liuhongt via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> When init_expr is INTEGER_CST or REAL_CST, can_vec_perm_const_p is not
> necessary since there's no real vec_perm needed, but
> vec_gen_perm_mask_checked will gcc_assert (can_vec_perm_const_p). So
> it's better to use vec_gen_perm_mask_any in
> vect_create_nonlinear_iv_init.

and the VEC_PERM build will fold the permute away?

> Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}.
> Ok for trunk?

OK.

Thanks,
Richard.

> gcc/ChangeLog:
>
>         PR tree-optimization/106963
>         * tree-vect-loop.cc (vect_create_nonlinear_iv_init): Use
>         vec_gen_perm_mask_any instead of vec_gen_perm_mask_check.
>
> gcc/testsuite/ChangeLog:
>
>         * gcc.target/i386/pr106963.c: New test.
> ---
>  gcc/testsuite/gcc.target/i386/pr106963.c | 14 ++++++++++++++
>  gcc/tree-vect-loop.cc                    |  5 ++++-
>  2 files changed, 18 insertions(+), 1 deletion(-)
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106963.c
>
> diff --git a/gcc/testsuite/gcc.target/i386/pr106963.c b/gcc/testsuite/gcc.target/i386/pr106963.c
> new file mode 100644
> index 00000000000..9f2d20e2523
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106963.c
> @@ -0,0 +1,14 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx -mno-avx2" } */
> +
> +void
> +foo_neg_const (int *a)
> +{
> +  int i, b = 1;
> +
> +  for (i = 0; i < 1000; i++)
> +    {
> +      a[i] = b;
> +      b = -b;
> +    }
> +}
> diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> index 9c434b66c5b..aabdc6f2d81 100644
> --- a/gcc/tree-vect-loop.cc
> +++ b/gcc/tree-vect-loop.cc
> @@ -8356,8 +8356,11 @@ vect_create_nonlinear_iv_init (gimple_seq* stmts, tree init_expr,
>             sel[2 * i + 1] = i + nunits;
>           }
>         vec_perm_indices indices (sel, 2, nunits);
> +       /* Don't use vect_gen_perm_mask_checked since can_vec_perm_const_p may
> +          fail when vec_init is const vector. In that situation vec_perm is not
> +          really needed.  */
>         tree perm_mask_even
> -         = vect_gen_perm_mask_checked (vectype, indices);
> +         = vect_gen_perm_mask_any (vectype, indices);
>         vec_init = gimple_build (stmts, VEC_PERM_EXPR,
>                                  vectype,
>                                  vec_init, vec_neg,
> --
> 2.18.1
>