tree-optimization/101083 - fix ICE with SLP reassoc
authorRichard Biener <rguenther@suse.de>
Wed, 16 Jun 2021 06:56:21 +0000 (08:56 +0200)
committerRichard Biener <rguenther@suse.de>
Wed, 16 Jun 2021 06:58:28 +0000 (08:58 +0200)
This makes us pass down the vector type for the two-operand
SLP node build rather than picking that from operand one which,
when constant or external, could be NULL.

2021-06-16  Richard Biener  <rguenther@suse.de>

PR tree-optimization/101083
* tree-vect-slp.c (vect_slp_build_two_operator_nodes): Get
vectype as argument.
(vect_build_slp_tree_2): Adjust.

* gcc.dg/vect/pr97832-4.c: New testcase.

gcc/testsuite/gcc.dg/vect/pr97832-4.c [new file with mode: 0644]
gcc/tree-vect-slp.c

diff --git a/gcc/testsuite/gcc.dg/vect/pr97832-4.c b/gcc/testsuite/gcc.dg/vect/pr97832-4.c
new file mode 100644 (file)
index 0000000..74ae27f
--- /dev/null
@@ -0,0 +1,28 @@
+/* { dg-do compile } */
+/* { dg-additional-options "-Ofast" } */
+/* { dg-require-effective-target vect_double } */
+
+void foo1x1(double* restrict y, const double* restrict x, int clen)
+{
+  int xi = clen & 2;
+  double f_re = x[0+xi+0];
+  double f_im = x[4+xi+0];
+  int clen2 = (clen+xi) * 2;
+#pragma GCC unroll 0
+  for (int c = 0; c < clen2; c += 8) {
+#pragma GCC unroll 4
+    for (int k = 0; k < 4; ++k) {
+      double x_re = x[k];
+      double x_im = x[c+4+k];
+      double y_re = y[c+0+k];
+      double y_im = y[c+4+k];
+      y_re = y_re - x_re * f_re - x_im * f_im;;
+      y_im = y_im + x_re * f_im - x_im * f_re;
+      y[c+0+k] = y_re;
+      y[c+4+k] = y_im;
+    }
+  }
+}
+
+/* { dg-final { scan-tree-dump "vectorizing stmts using SLP" "vect" } } */
+/* { dg-final { scan-tree-dump "Loop contains only SLP stmts" "vect" } } */
index 9ded585..8ec589b 100644 (file)
@@ -1536,13 +1536,12 @@ vect_build_slp_tree (vec_info *vinfo,
 /* Helper for building an associated SLP node chain.  */
 
 static void
-vect_slp_build_two_operator_nodes (slp_tree perm,
+vect_slp_build_two_operator_nodes (slp_tree perm, tree vectype,
                                   slp_tree op0, slp_tree op1,
                                   stmt_vec_info oper1, stmt_vec_info oper2,
                                   vec<std::pair<unsigned, unsigned> > lperm)
 {
   unsigned group_size = SLP_TREE_LANES (op1);
-  tree vectype = SLP_TREE_VECTYPE (op1);
 
   slp_tree child1 = new _slp_tree;
   SLP_TREE_DEF_TYPE (child1) = vect_internal_def;
@@ -2087,7 +2086,7 @@ vect_build_slp_tree_2 (vec_info *vinfo, slp_tree node,
                  for (unsigned lane = 0; lane < group_size; ++lane)
                    lperm.quick_push (std::make_pair
                      (chains[lane][i].code != chains[0][i].code, lane));
-                 vect_slp_build_two_operator_nodes (child, op0, op1,
+                 vect_slp_build_two_operator_nodes (child, vectype, op0, op1,
                                                     (chains[0][i].code == code
                                                      ? op_stmt_info
                                                      : other_op_stmt_info),