[PATCH] Fix PR49518

[Richard Guenther]

Handling of negative steps broke one of the many asserts in
the vectorizer.  The following patch drops one that I can't
make sense of.  I think all asserts need comments - especially
this one would, as I can't see why using vf is correct to
test against and not nelements (and why <= vf and not < vf).

Well, ok?

Thanks,
Richard.

2011-07-04  Richard Guenther  <rguent...@suse.de>

        PR tree-optimization/49518
        * tree-vect-data-refs.c (vect_enhance_data_refs_alignment):
        Drop assert.

        * gcc.dg/torture/pr49518.c: New testcase.

Index: gcc/tree-vect-data-refs.c
===================================================================
--- gcc/tree-vect-data-refs.c   (revision 175800)
+++ gcc/tree-vect-data-refs.c   (working copy)
@@ -1552,7 +1552,6 @@ vect_enhance_data_refs_alignment (loop_v
 
               for (j = 0; j < possible_npeel_number; j++)
                 {
-                  gcc_assert (npeel_tmp <= vf);
                   vect_peeling_hash_insert (loop_vinfo, dr, npeel_tmp);
                   npeel_tmp += nelements;
                 }
Index: gcc/testsuite/gcc.dg/torture/pr49518.c
===================================================================
--- gcc/testsuite/gcc.dg/torture/pr49518.c      (revision 0)
+++ gcc/testsuite/gcc.dg/torture/pr49518.c      (revision 0)
@@ -0,0 +1,19 @@
+/* { dg-do compile } */
+
+int a, b;
+struct S { unsigned int s, t, u; } c, d = { 0, 1, 0 };
+
+void
+test (unsigned char z)
+{
+  char e[] = {0, 0, 0, 0, 1};
+  for (c.s = 1; c.s; c.s++)
+    {
+      b = e[c.s];
+      if (a)
+       break;
+      b = z >= c.u;
+      if (d.t)
+       break;
+    }
+}