Re: [PATCH]middle-end: refactory vect_recog_absolute_difference to simplify flow [PR114769]

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

On Fri, Apr 19, 2024 at 3:29 PM Tamar Christina <tamar.christina@arm.com> wrote:
>
> Hi All,
>
> As the reporter in PR114769 points out the control flow for the abd detection
> is hard to follow.  This is because vect_recog_absolute_difference has two
> different ways it can return true.
>
> 1. It can return true when the widening operation is matched, in which case
>    unprom is set, half_type is not NULL and diff_stmt is not set.
>
> 2. It can return true when the widening operation is not matched, but the stmt
>    being checked is a minus.  In this case unprom is not set, half_type is set
>    to NULL and diff_stmt is set.  This because to get to diff_stmt you have to
>    dig through the abs statement and any possible promotions.
>
> This however leads to complicated uses of the function at the call sites as the
> exact semantic needs to be known to use it safely.
>
> vect_recog_absolute_difference has two callers:
>
> 1. vect_recog_sad_pattern where if you return true with unprom not set, then
>    *half_type will be NULL.  The call to vect_supportable_direct_optab_p will
>    always reject it since there's no vector mode for NULL.  Note that if looking
>    at the dump files, the convention in the dump files have always been that we
>    first indicate that a pattern could possibly be recognize and then check that
>    it's supported.
>
>    This change somewhat incorrectly makes the diagnostic message get printed for
>    "invalid" patterns.
>
> 2. vect_recog_abd_pattern, where if half_type is NULL, it then uses diff_stmt to
>    set them.
>
> So while the note in the dump file is misleading, the code is safe.
>
> This refactors the code, it now only has 1 success condition, and diff_stmt is
> always set to the minus statement in the abs if there is one.
>
> The function now only returns success if the widening minus is found, in which
> case unprom and half_type set.
>
> This then leaves it up to the caller to decide if they want to do anything with
> diff_stmt.
>
> Bootstrapped Regtested on aarch64-none-linux-gnu and no issues.
>
> Ok for master?

OK.

Thanks,
Richard.

> Thanks,
> Tamar
>
> gcc/ChangeLog:
>
>         PR tree-optimization/114769
>         * tree-vect-patterns.cc:
>         (vect_recog_absolute_difference): Have only one success condition.
>         (vect_recog_abd_pattern): Handle further checks if
>         vect_recog_absolute_difference fails.
>
> ---
> diff --git a/gcc/tree-vect-patterns.cc b/gcc/tree-vect-patterns.cc
> index 4f491c6b8336f8710c3519dec1fa7e0f49387d2b..87c2acff386d91d22a3b2d6e6443d1f2f2326ea6 100644
> --- a/gcc/tree-vect-patterns.cc
> +++ b/gcc/tree-vect-patterns.cc
> @@ -797,8 +797,7 @@ vect_split_statement (vec_info *vinfo, stmt_vec_info stmt2_info, tree new_rhs,
>     HALF_TYPE and UNPROM will be set should the statement be found to
>     be a widened operation.
>     DIFF_STMT will be set to the MINUS_EXPR
> -   statement that precedes the ABS_STMT unless vect_widened_op_tree
> -   succeeds.
> +   statement that precedes the ABS_STMT if it is a MINUS_EXPR..
>   */
>  static bool
>  vect_recog_absolute_difference (vec_info *vinfo, gassign *abs_stmt,
> @@ -843,6 +842,12 @@ vect_recog_absolute_difference (vec_info *vinfo, gassign *abs_stmt,
>    if (!diff_stmt_vinfo)
>      return false;
>
> +  gassign *diff = dyn_cast <gassign *> (STMT_VINFO_STMT (diff_stmt_vinfo));
> +  if (diff_stmt && diff
> +      && gimple_assign_rhs_code (diff) == MINUS_EXPR
> +      && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (abs_oprnd)))
> +    *diff_stmt = diff;
> +
>    /* FORNOW.  Can continue analyzing the def-use chain when this stmt in a phi
>       inside the loop (in case we are analyzing an outer-loop).  */
>    if (vect_widened_op_tree (vinfo, diff_stmt_vinfo,
> @@ -850,17 +855,6 @@ vect_recog_absolute_difference (vec_info *vinfo, gassign *abs_stmt,
>                             false, 2, unprom, half_type))
>      return true;
>
> -  /* Failed to find a widen operation so we check for a regular MINUS_EXPR.  */
> -  gassign *diff = dyn_cast <gassign *> (STMT_VINFO_STMT (diff_stmt_vinfo));
> -  if (diff_stmt && diff
> -      && gimple_assign_rhs_code (diff) == MINUS_EXPR
> -      && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (abs_oprnd)))
> -    {
> -      *diff_stmt = diff;
> -      *half_type = NULL_TREE;
> -      return true;
> -    }
> -
>    return false;
>  }
>
> @@ -1499,27 +1493,22 @@ vect_recog_abd_pattern (vec_info *vinfo,
>    tree out_type = TREE_TYPE (gimple_assign_lhs (last_stmt));
>
>    vect_unpromoted_value unprom[2];
> -  gassign *diff_stmt;
> -  tree half_type;
> -  if (!vect_recog_absolute_difference (vinfo, last_stmt, &half_type,
> +  gassign *diff_stmt = NULL;
> +  tree abd_in_type;
> +  if (!vect_recog_absolute_difference (vinfo, last_stmt, &abd_in_type,
>                                        unprom, &diff_stmt))
> -    return NULL;
> -
> -  tree abd_in_type, abd_out_type;
> -
> -  if (half_type)
> -    {
> -      abd_in_type = half_type;
> -      abd_out_type = abd_in_type;
> -    }
> -  else
>      {
> +      /* We cannot try further without having a non-widening MINUS.  */
> +      if (!diff_stmt)
> +       return NULL;
> +
>        unprom[0].op = gimple_assign_rhs1 (diff_stmt);
>        unprom[1].op = gimple_assign_rhs2 (diff_stmt);
>        abd_in_type = signed_type_for (out_type);
> -      abd_out_type = abd_in_type;
>      }
>
> +  tree abd_out_type = abd_in_type;
> +
>    tree vectype_in = get_vectype_for_scalar_type (vinfo, abd_in_type);
>    if (!vectype_in)
>      return NULL;
>
>
>
>
> --