On Wed, 5 Feb 2025, Xi Ruoyao wrote:

> With things like
> 
>   // signed char a_14, a_16;
>   a.0_4 = (unsigned char) a_14;
>   _5 = (int) a.0_4;
>   b.1_6 = (unsigned char) b_16;
>   _7 = (int) b.1_6;
>   c_17 = _5 - _7;
>   _8 = ABS_EXPR <c_17>;
>   r_18 = _8 + r_23;
> 
> An ABD pattern will be recognized for _8:
> 
>   patt_31 = .ABD (a.0_4, b.1_6);
> 
> It's still correct.  But then when the SAD pattern is recognized:
> 
>   patt_29 = SAD_EXPR <a_14, b_16, r_23>;
> 
> This is not correct.  This only happens for targets with both uabd and
> sabd but not vec_widen_{s,u}abd, currently LoongArch is the only target
> affected.
> 
> The problem is vect_look_through_possible_promotion will throw away a
> series of conversions if the effect is equivalent to a sign change and a
> promotion, but here the sign change is definitely relevant, and the
> promotion is also relevant for "mixed sign" cases like
> r += abs((unsigned int)(unsigned char) a - (signed int)(signed char) b
> (we need to promote to HImode as the difference can exceed the range of
> QImode).
> 
> If there were any redundant promotion, it should have been stripped in
> vect_recog_abd_pattern (i.e. when patt_31 = .ABD (a.0_4, b.1_6) is
> recognized) instead of in vect_recog_sad_pattern, or we'd have a
> missed-optimization if the ABD output is not summerized.  So anyway
> vect_recog_sad_pattern is just not a proper location to call
> vect_look_through_possible_promotion for the ABD inputs, remove the
> calls to fix the issue.

OK.

Thanks,
Richard.

> gcc/ChangeLog:
> 
>       PR tree-optimization/118727
>       * tree-vect-patterns.cc (vect_recog_sad_pattern): Don't call
>       vect_look_through_possible_promotion on ABD inputs.
> 
> gcc/testsuite/ChangeLog:
> 
>       PR tree-optimization/118727
>       * gcc.dg/pr108692.c: Mention PR 118727 in the comment.
>       * gcc.dg/pr118727.c: New test case.
> ---
> 
> Bootstrapped and regtested on loongarch64-linux-gnu.  Ok for trunk?
> 
>  gcc/testsuite/gcc.dg/pr108692.c |  1 +
>  gcc/testsuite/gcc.dg/pr118727.c | 32 ++++++++++++++++++++++++++++++++
>  gcc/tree-vect-patterns.cc       | 11 ++---------
>  3 files changed, 35 insertions(+), 9 deletions(-)
>  create mode 100644 gcc/testsuite/gcc.dg/pr118727.c
> 
> diff --git a/gcc/testsuite/gcc.dg/pr108692.c b/gcc/testsuite/gcc.dg/pr108692.c
> index 13a27496ad9..eeee2203281 100644
> --- a/gcc/testsuite/gcc.dg/pr108692.c
> +++ b/gcc/testsuite/gcc.dg/pr108692.c
> @@ -1,4 +1,5 @@
>  /* PR tree-optimization/108692 */
> +/* PR tree-optimization/118727 */
>  /* { dg-do run } */
>  /* { dg-options "-O2 -ftree-vectorize" } */
>  
> diff --git a/gcc/testsuite/gcc.dg/pr118727.c b/gcc/testsuite/gcc.dg/pr118727.c
> new file mode 100644
> index 00000000000..2ee5fa78236
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/pr118727.c
> @@ -0,0 +1,32 @@
> +/* PR tree-optimization/118727 */
> +/* { dg-do run } */
> +/* { dg-options "-O2 -ftree-vectorize" } */
> +
> +__attribute__((noipa)) int
> +foo (signed char *x, signed char *y, int n)
> +{
> +  int i, r = 0;
> +  signed char a, b;
> +  for (i = 0; i < n; i++)
> +    {
> +      a = x[i];
> +      b = y[i];
> +      /* Slightly twisted from pr108692.c.  */
> +      int c = (unsigned int)(unsigned char) a - (signed int)(signed char) b;
> +      r = r + (c < 0 ? -c : c);
> +    }
> +  return r;
> +}
> +
> +int
> +main ()
> +{
> +  signed char x[64] = {}, y[64] = {};
> +  if (__CHAR_BIT__ != 8 || __SIZEOF_INT__ != 4)
> +    return 0;
> +  x[32] = -1;
> +  y[32] = -128;
> +  if (foo (x, y, 64) != 383)
> +    __builtin_abort ();
> +  return 0;
> +}
> diff --git a/gcc/tree-vect-patterns.cc b/gcc/tree-vect-patterns.cc
> index 5aebf950548..6fc97d1b6ef 100644
> --- a/gcc/tree-vect-patterns.cc
> +++ b/gcc/tree-vect-patterns.cc
> @@ -1405,15 +1405,8 @@ vect_recog_sad_pattern (vec_info *vinfo,
>        tree abd_oprnd0 = gimple_call_arg (abd_stmt, 0);
>        tree abd_oprnd1 = gimple_call_arg (abd_stmt, 1);
>  
> -      if (gimple_call_internal_fn (abd_stmt) == IFN_ABD)
> -     {
> -       if (!vect_look_through_possible_promotion (vinfo, abd_oprnd0,
> -                                                  &unprom[0])
> -           || !vect_look_through_possible_promotion (vinfo, abd_oprnd1,
> -                                                     &unprom[1]))
> -         return NULL;
> -     }
> -      else if (gimple_call_internal_fn (abd_stmt) == IFN_VEC_WIDEN_ABD)
> +      if (gimple_call_internal_fn (abd_stmt) == IFN_ABD
> +       || gimple_call_internal_fn (abd_stmt) == IFN_VEC_WIDEN_ABD)
>       {
>         unprom[0].op = abd_oprnd0;
>         unprom[0].type = TREE_TYPE (abd_oprnd0);
> 

-- 
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