From: Richard Biener <rguenther@suse.de>
Date: Thu, 13 Oct 2022 12:56:01 +0000 (+0200)
Subject: tree-optimization/107247 - reduce SLP reduction accumulator
X-Git-Tag: basepoints/gcc-14~3931
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=e5139d18dfb8130876ea59178e8471fb1b34bb80;p=thirdparty%2Fgcc.git

tree-optimization/107247 - reduce SLP reduction accumulator

The following makes sure to reduce a multi-vector SLP reduction
accumulator to a single vector using vector operations if
easily possible (if the number of lanes in the vector type is
a multiple of the number of scalar accumulators).

	PR tree-optimization/107247
	* tree-vect-loop.cc (vect_create_epilog_for_reduction):
	Reduce multi vector SLP reduction accumulators.  Check
	the adjusted number of accumulator vectors against
	one for the re-use in the epilogue.
---

diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
index b1442a935819..98a943d8a4ba 100644
--- a/gcc/tree-vect-loop.cc
+++ b/gcc/tree-vect-loop.cc
@@ -5642,9 +5642,21 @@ vect_create_epilog_for_reduction (loop_vec_info loop_vinfo,
      we may end up with more than one vector result.  Here we reduce them
      to one vector.
 
+     The same is true for a SLP reduction, e.g.,
+     # a1 = phi <a2, a0>
+     # b1 = phi <b2, b0>
+     a2 = operation (a1)
+     b2 = operation (a2),
+
+     where we can end up with more than one vector as well.  We can
+     easily accumulate vectors when the number of vector elements is
+     a multiple of the SLP group size.
+
      The same is true if we couldn't use a single defuse cycle.  */
   if (REDUC_GROUP_FIRST_ELEMENT (stmt_info)
       || direct_slp_reduc
+      || (slp_reduc
+	  && constant_multiple_p (TYPE_VECTOR_SUBPARTS (vectype), group_size))
       || ncopies > 1)
     {
       gimple_seq stmts = NULL;
@@ -6233,7 +6245,7 @@ vect_create_epilog_for_reduction (loop_vec_info loop_vinfo,
 
   /* Record this operation if it could be reused by the epilogue loop.  */
   if (STMT_VINFO_REDUC_TYPE (reduc_info) == TREE_CODE_REDUCTION
-      && vec_num == 1)
+      && reduc_inputs.length () == 1)
     loop_vinfo->reusable_accumulators.put (scalar_results[0],
 					   { orig_reduc_input, reduc_info });