Hi!

On the following testcase, there is a redundant & 1 and when reassociating,
we remove the redundant & 1.  The outer 2 ops are the same, so changed
is still false when we reach rewrite_expr_tree with the last 2 ops (i.e.
last statement).  But, as we don't keep that statement as is nor just swap
the arguments, that isn't a guarantee that the SSA_NAME has always the same
value as previously (in this case we've dropped the & 1 there), so the
following patch ensures we don't reuse the SSA_NAME unless:
1) we don't change the stmt at all
2) we just swap the two arguments
3) opindex == 0 (reassociate_bb doesn't really expect to see a new SSA_NAME
   if it hasn't called it with changed = true, and in that case the SSA_NAME
   is necessarily always the same as before the optimization, otherwise
   the optimization would be wrong.
This way, we don't preserve bogus VR info, nor generate incorrect debug
info.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2015-02-26  Jakub Jelinek  <ja...@redhat.com>

        PR tree-optimization/65216
        * tree-ssa-reassoc.c (rewrite_expr_tree): Force creation of
        new stmt and new SSA_NAME for lhs whenever the arguments have
        changed and weren't just swapped.  Fix comment typo.

        * gcc.c-torture/execute/pr65216.c: New test.

--- gcc/tree-ssa-reassoc.c.jj   2015-02-14 09:21:56.000000000 +0100
+++ gcc/tree-ssa-reassoc.c      2015-02-26 13:03:24.308614169 +0100
@@ -3532,7 +3532,7 @@ rewrite_expr_tree (gimple stmt, unsigned
 
   /* The final recursion case for this function is that you have
      exactly two operations left.
-     If we had one exactly one op in the entire list to start with, we
+     If we had exactly one op in the entire list to start with, we
      would have never called this function, and the tail recursion
      rewrites them one at a time.  */
   if (opindex + 2 == ops.length ())
@@ -3553,7 +3553,11 @@ rewrite_expr_tree (gimple stmt, unsigned
              print_gimple_stmt (dump_file, stmt, 0, 0);
            }
 
-         if (changed)
+         /* Even when changed is false, reassociation could have e.g. removed
+            some redundant operations, so unless we are just swapping the
+            arguments or unless there is no change at all (then we just
+            return lhs), force creation of a new SSA_NAME.  */
+         if (changed || ((rhs1 != oe2->op || rhs2 != oe1->op) && opindex))
            {
              gimple insert_point = find_insert_point (stmt, oe1->op, oe2->op);
              lhs = make_ssa_name (TREE_TYPE (lhs));
--- gcc/testsuite/gcc.c-torture/execute/pr65216.c.jj    2015-02-26 
13:05:12.199816826 +0100
+++ gcc/testsuite/gcc.c-torture/execute/pr65216.c       2015-02-26 
13:04:53.000000000 +0100
@@ -0,0 +1,20 @@
+/* PR tree-optimization/65216 */
+
+int a, b = 62, e;
+volatile int c, d;
+
+int
+main ()
+{
+  int f = 0;
+  for (a = 0; a < 2; a++)
+    {
+      b &= (8 ^ f) & 1;
+      for (e = 0; e < 6; e++)
+       if (c)
+         f = d;
+    }
+  if (b != 0)
+    __builtin_abort ();
+  return 0;
+}

        Jakub

Reply via email to