On Wed, 2 Sep 2026, liuhongt wrote:
> The average fallback is restricted to unsigned types, but it also works for
> signed types. Arithmetic right shifts round down, and the low-bit carry
> adjusts the result in the same way.
>
> Remove the unsigned restriction. Use target_supports_op_p rather than
> optab_for_tree_code to check whether the fallback operations are available.
>
> Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}.
> Ok for trunk?
OK.
Thanks,
Richard.
> gcc/ChangeLog:
>
> PR tree-optimization/89007
> * tree-vect-patterns.cc (vect_recog_average_pattern): Allow signed
> types in the fallback. Check whether the fallback operations are
> supported.
>
> gcc/testsuite/ChangeLog:
>
> PR tree-optimization/89007
> * gcc.target/i386/pr89007.c: New test.
> ---
> gcc/testsuite/gcc.target/i386/pr89007.c | 20 ++++++++++++++++++++
> gcc/tree-vect-patterns.cc | 24 ++++++++++++------------
> 2 files changed, 32 insertions(+), 12 deletions(-)
> create mode 100644 gcc/testsuite/gcc.target/i386/pr89007.c
>
> diff --git a/gcc/testsuite/gcc.target/i386/pr89007.c
> b/gcc/testsuite/gcc.target/i386/pr89007.c
> new file mode 100644
> index 00000000000..854fb54700a
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr89007.c
> @@ -0,0 +1,20 @@
> +/* PR tree-optimization/89007 */
> +/* { dg-do compile } */
> +/* { dg-options "-O3 -mavx512bw -mavx512vl -mprefer-vector-width=512" } */
> +
> +void
> +avg_floor (short *restrict d, short *restrict a, short *restrict b, int n)
> +{
> + for (int i = 0; i < n; i++)
> + d[i] = (a[i] + b[i]) >> 1;
> +}
> +
> +void
> +avg_ceil (short *restrict d, short *restrict a, short *restrict b, int n)
> +{
> + for (int i = 0; i < n; i++)
> + d[i] = (a[i] + b[i] + 1) >> 1;
> +}
> +
> +/* { dg-final { scan-assembler "vpsraw\[ \t\]" } } */
> +/* { dg-final { scan-assembler-not "vpsrad\[ \t\]" } } */
> diff --git a/gcc/tree-vect-patterns.cc b/gcc/tree-vect-patterns.cc
> index 406f81b95c8..15046a1cea9 100644
> --- a/gcc/tree-vect-patterns.cc
> +++ b/gcc/tree-vect-patterns.cc
> @@ -3581,17 +3581,15 @@ vect_recog_average_pattern (vec_info *vinfo,
> if (!new_vectype)
> return NULL;
>
> - bool fallback_p = false;
> -
> - if (direct_internal_fn_supported_p (ifn, new_vectype, OPTIMIZE_FOR_SPEED))
> - ;
> - else if (TYPE_UNSIGNED (new_type)
> - && optab_for_tree_code (RSHIFT_EXPR, new_vectype, optab_scalar)
> - && optab_for_tree_code (PLUS_EXPR, new_vectype, optab_default)
> - && optab_for_tree_code (BIT_IOR_EXPR, new_vectype, optab_default)
> - && optab_for_tree_code (BIT_AND_EXPR, new_vectype, optab_default))
> - fallback_p = true;
> - else
> + bool fallback_p = !direct_internal_fn_supported_p (ifn, new_vectype,
> + OPTIMIZE_FOR_SPEED);
> + if (fallback_p
> + && (!target_supports_op_p (new_vectype, RSHIFT_EXPR, optab_scalar)
> + || !target_supports_op_p (new_vectype, PLUS_EXPR, optab_default)
> + || !target_supports_op_p (new_vectype, BIT_AND_EXPR, optab_default)
> + || (ifn == IFN_AVG_CEIL
> + && !target_supports_op_p (new_vectype, BIT_IOR_EXPR,
> + optab_default))))
> return NULL;
>
> /* The IR requires a valid vector type for the cast result, even though
> @@ -3615,7 +3613,9 @@ vect_recog_average_pattern (vec_info *vinfo,
> unmasked_carry = new_ops[0] and/or new_ops[1];
> carry = unmasked_carry & 1;
> new_var = sum_of_shifted + carry;
> - */
> +
> + For signed types, arithmetic shifts round down and the carry is one
> + when both operands are odd (or either operand for IFN_AVG_CEIL). */
>
> tree one_cst = build_one_cst (new_type);
> gassign *g;
>
--
Richard Biener <[email protected]>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Jochen Jaser, Andrew McDonald, Abhinav Puri; (HRB 36809, AG Nuernberg)
