This patch fixes PR52424 by pushing a missing marker onto
const_and_copies_stack before calling dom_thread_across_edge. This bug
has resulted in lost copy propagation opportunities in the presence of
edge threading.
Bootstrapped and tested with no regressions on powerpc64-linux-gnu. I'm
waiting on Jiangning Liu to verify the patch fixes his performance
concern. Provided it does, is this OK for trunk?
Also, this bug has apparently been there for a long time. Should we
consider backporting the fix to 4.6 and/or 4.5?
Thanks,
Bill
2012-02-28 Bill Schmidt <wschm...@linux.vnet.ibm.com>
PR tree-optimization/52424
* tree-ssa-dom.c (dom_opt_leave_block): Push a marker before
calling dom_thread_across_edge.
Index: gcc/tree-ssa-dom.c
===================================================================
--- gcc/tree-ssa-dom.c (revision 184625)
+++ gcc/tree-ssa-dom.c (working copy)
@@ -1779,6 +1779,9 @@ dom_opt_leave_block (struct dom_walk_data *walk_da
&& (single_succ_edge (bb)->flags & EDGE_ABNORMAL) == 0
&& potentially_threadable_block (single_succ (bb)))
{
+ /* Push a marker on the stack, which thread_across_edge expects
+ and will remove. */
+ VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
dom_thread_across_edge (walk_data, single_succ_edge (bb));
}
else if ((last = last_stmt (bb))
