Re: [PATCH] vect: Fix missed gather load opportunity

[Richard Biener <rguenther@suse.de>]

On Fri, 16 Sep 2022, Richard Sandiford wrote:

> While writing a testcase for PR106794, I noticed that we failed
> to vectorise the testcase in the patch for SVE.  The code that
> recognises gather loads tries to optimise the point at which
> the offset is calculated, to avoid unnecessary extensions or
> truncations:
>
> 	  /* Don't include the conversion if the target is happy with
> 	     the current offset type.  */
>
> But breaking only makes sense if we're at an SSA_NAME (which could
> then be vectorised).  We shouldn't break on a conversion embedded
> in a generic expression.
>
> Tested on aarch64-linux-gnu & x86_64-linux-gnu.  OK to install?

OK,

Thanks,
Richard.

> Richard
>
>
> gcc/
> 	* tree-vect-data-refs.cc (vect_check_gather_scatter): Restrict
> 	early-out optimisation to SSA_NAMEs.
>
> gcc/testsuite/
> 	* gcc.dg/vect/vect-gather-5.c: New test.
> ---
> gcc/testsuite/gcc.dg/vect/vect-gather-5.c | 42 +++++++++++++++++++++++
> gcc/tree-vect-data-refs.cc                |  1 +
> 2 files changed, 43 insertions(+)
> create mode 100644 gcc/testsuite/gcc.dg/vect/vect-gather-5.c
>
> diff --git a/gcc/testsuite/gcc.dg/vect/vect-gather-5.c b/gcc/testsuite/gcc.dg/vect/vect-gather-5.c
> new file mode 100644
> index 00000000000..8b5074bba88
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/vect/vect-gather-5.c
> @@ -0,0 +1,42 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-O3" } */
> +
> +#ifdef __aarch64__
> +#pragma GCC target "+sve"
> +#endif
> +
> +long a[100], b[100], c[100];
> +
> +void g1 ()
> +{
> +  for (int i = 0; i < 100; i += 2)
> +    {
> +      c[i] += a[b[i]] + 1;
> +      c[i + 1] += a[b[i + 1]] + 2;
> +    }
> +}
> +
> +long g2 ()
> +{
> +  long res = 0;
> +  for (int i = 0; i < 100; i += 2)
> +    {
> +      res += a[b[i + 1]];
> +      res += a[b[i]];
> +    }
> +  return res;
> +}
> +
> +long g3 ()
> +{
> +  long res = 0;
> +  for (int i = 0; i < 100; i += 2)
> +    {
> +      res += a[b[i]];
> +      res += a[b[i + 1]];
> +    }
> +  return res;
> +}
> +
> +/* { dg-final { scan-tree-dump-times {add new stmt[^\n]*GATHER_LOAD} 3 "vect" { target aarch64*-*-* } } } */
> +/* { dg-final { scan-tree-dump-not {add new stmt[^\n]*VEC_PERM_EXPR} "vect" { target aarch64*-*-* } } } */
> diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
> index b279a82551e..e03b50498d1 100644
> --- a/gcc/tree-vect-data-refs.cc
> +++ b/gcc/tree-vect-data-refs.cc
> @@ -4151,6 +4151,7 @@ vect_check_gather_scatter (stmt_vec_info stmt_info, loop_vec_info loop_vinfo,
> 	  /* Don't include the conversion if the target is happy with
> 	     the current offset type.  */
> 	  if (use_ifn_p
> +	      && TREE_CODE (off) == SSA_NAME
> 	      && !POINTER_TYPE_P (TREE_TYPE (off))
> 	      && vect_gather_scatter_fn_p (loop_vinfo, DR_IS_READ (dr),
> 					   masked_p, vectype, memory_type,
> -- 
> 2.25.1
>
>