On Sat, 2 Nov 2024, Robin Dapp wrote:

> From: Robin Dapp <rd...@ventanamicro.com>
> 
> When predicating a load we implicitly assume that the else value is
> zero.  This matters in case the loaded value is padded (like e.g.
> a Bool) and we must ensure that the padding bytes are zero on targets
> that don't implicitly zero inactive elements.
> 
> A former version of this patch still had this handling in ifcvt but
> the latest version defers it to the vectorizer.

This version is OK.

Thanks,
Richard.

> gcc/ChangeLog:
> 
>       * tree-if-conv.cc (predicate_load_or_store): Add zero else
>       operand and comment.
> ---
>  gcc/tree-if-conv.cc | 12 ++++++++++--
>  1 file changed, 10 insertions(+), 2 deletions(-)
> 
> diff --git a/gcc/tree-if-conv.cc b/gcc/tree-if-conv.cc
> index eb981642bae..f1a1f8fd0d3 100644
> --- a/gcc/tree-if-conv.cc
> +++ b/gcc/tree-if-conv.cc
> @@ -2555,9 +2555,17 @@ predicate_load_or_store (gimple_stmt_iterator *gsi, 
> gassign *stmt, tree mask)
>                  ref);
>    if (TREE_CODE (lhs) == SSA_NAME)
>      {
> +      /* Get a zero else value.  This might not be what a target actually 
> uses
> +      but we cannot be sure about which vector mode the vectorizer will
> +      choose.  Therefore, leave the decision whether we need to force the
> +      inactive elements to zero to the vectorizer.  */
> +      tree els = vect_get_mask_load_else (MASK_LOAD_ELSE_ZERO,
> +                                       TREE_TYPE (lhs));
> +
>        new_stmt
> -     = gimple_build_call_internal (IFN_MASK_LOAD, 3, addr,
> -                                   ptr, mask);
> +     = gimple_build_call_internal (IFN_MASK_LOAD, 4, addr,
> +                                   ptr, mask, els);
> +
>        gimple_call_set_lhs (new_stmt, lhs);
>        gimple_set_vuse (new_stmt, gimple_vuse (stmt));
>      }
> 

-- 
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)

Reply via email to