Hi!

Even if vect_build_slp_tree on child fails, it might have pushed some
nodes to SLP_TREE_CHILDREN (child) vector before returning false.
If we retry, we need to start with no grandchildren of course.

Bootstrapped/regtested on x86_64-linux and i686-linux, acked by Richard in
the PR, committed to trunk so far.

2015-03-24  Jakub Jelinek  <ja...@redhat.com>

        PR tree-optimization/65533
        * tree-vect-slp.c (vect_build_slp_tree): Before re-trying
        with swapped operands, call vect_free_slp_tree on
        SLP_TREE_CHILDREN of child and truncate the SLP_TREE_CHILDREN
        vector.

        * gcc.dg/pr65533.c: New test.

--- gcc/tree-vect-slp.c.jj      2015-03-23 16:15:30.000000000 +0100
+++ gcc/tree-vect-slp.c 2015-03-24 08:54:51.524143880 +0100
@@ -1035,13 +1035,20 @@ vect_build_slp_tree (loop_vec_info loop_
             behavior.  */
          && *npermutes < 4)
        {
+         unsigned int j;
+         slp_tree grandchild;
+
          /* Roll back.  */
          *max_nunits = old_max_nunits;
          loads->truncate (old_nloads);
+         FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (child), j, grandchild)
+           vect_free_slp_tree (grandchild);
+         SLP_TREE_CHILDREN (child).truncate (0);
+
          /* Swap mismatched definition stmts.  */
          dump_printf_loc (MSG_NOTE, vect_location,
                           "Re-trying with swapped operands of stmts ");
-         for (unsigned j = 0; j < group_size; ++j)
+         for (j = 0; j < group_size; ++j)
            if (!matches[j])
              {
                gimple tem = oprnds_info[0]->def_stmts[j];
--- gcc/testsuite/gcc.dg/pr65533.c.jj   2015-03-24 08:52:24.437577841 +0100
+++ gcc/testsuite/gcc.dg/pr65533.c      2015-03-24 08:51:55.000000000 +0100
@@ -0,0 +1,25 @@
+/* PR tree-optimization/65533 */
+/* { dg-do compile } */
+/* { dg-options "-Ofast -w" } */
+
+struct A { int a[2]; };
+struct B { double b[2]; };
+struct C { double c[4][1]; };
+
+static inline void
+bar (struct B *x, double y, double z)
+{
+  x->b[0] = y;
+  x->b[1] = z;
+}
+
+void baz (struct B *);
+
+void
+foo (struct C *x, struct A *y)
+{
+  struct B d;
+  bar (&d, x->c[1][0] * y->a[0] + x->c[0][1] * y->a[1] + x->c[0][0] * 
x->c[0][1],
+       x->c[0][0] * y->a[0] + x->c[0][1] * y->a[1] + x->c[0][1] * y->a[0] + 
x->c[0][0]);
+  baz (&d);
+}

        Jakub

Reply via email to