[PATCH] Fall back to SLP reduction discovery when reduction group fails

[Richard Biener]

This helps saving some IVs (though I guess the situation where it
matches in practice is scarce).

Bootstrapped and tested on x86_64-unknown-linux-gnu, applied to trunk.

Richard.

2019-10-24  Richard Biener  <rguent...@suse.de>

        * tree-vect-slp.c (vect_analyze_slp): When reduction group
        SLP discovery fails try to handle the reduction as part
        of SLP reduction discovery.

        * gcc.dg/vect/slp-reduc-9.c: New testcase.

Index: gcc/tree-vect-slp.c
===================================================================
--- gcc/tree-vect-slp.c (revision 277330)
+++ gcc/tree-vect-slp.c (working copy)
@@ -2271,14 +2271,18 @@ vect_analyze_slp (vec_info *vinfo, unsig
              {
                /* Dissolve reduction chain group.  */
                stmt_vec_info vinfo = first_element;
+               stmt_vec_info last = NULL;
                while (vinfo)
                  {
                    stmt_vec_info next = REDUC_GROUP_NEXT_ELEMENT (vinfo);
                    REDUC_GROUP_FIRST_ELEMENT (vinfo) = NULL;
                    REDUC_GROUP_NEXT_ELEMENT (vinfo) = NULL;
+                   last = vinfo;
                    vinfo = next;
                  }
                STMT_VINFO_DEF_TYPE (first_element) = vect_internal_def;
+               /* It can be still vectorized as part of an SLP reduction.  */
+               loop_vinfo->reductions.safe_push (last);
              }
        }
 
Index: gcc/testsuite/gcc.dg/vect/slp-reduc-9.c
===================================================================
--- gcc/testsuite/gcc.dg/vect/slp-reduc-9.c     (nonexistent)
+++ gcc/testsuite/gcc.dg/vect/slp-reduc-9.c     (working copy)
@@ -0,0 +1,25 @@
+/* { dg-do compile } */
+/* { dg-require-effective-target vect_int_mult } */
+
+int
+bar (int *x, int a, int b, int n)
+{
+  x = __builtin_assume_aligned (x, __BIGGEST_ALIGNMENT__);
+  int sum1 = 0;
+  int sum2 = 0;
+  for (int i = 0; i < n; ++i)
+    {
+      /* Reduction chain vectorization fails here because of the
+         different operations but we can still vectorize both
+        reductions as SLP reductions, saving IVs.  */
+      sum1 += x[2*i] - a;
+      sum1 += x[2*i+1] * b;
+      sum2 += x[2*i] - b;
+      sum2 += x[2*i+1] * a;
+    }
+  return sum1 + sum2;
+}
+
+/* { dg-final { scan-tree-dump "Loop contains only SLP stmts" "vect" } } */
+/* { dg-final { scan-tree-dump "vectorizing stmts using SLP" "vect" } } */
+/* { dg-final { scan-tree-dump "vectorized 1 loops" "vect" } } */