Hi!
On the (undefined behavior) testcase below, we end up with
then_bb ending with __builtin_unreachable () at the tree level, therefore
no successor at the RTL level, and else_bb being EXIT_BLOCK_PTR (i.e.
conditional return before a bb with undefined behavior at the end).
Trying to optimize that into a conditional execution of the then_bb insns
doesn't work, we can't merge the else_bb with then_bb and test_bb in this
case, plus it doesn't look like something that would be desirable to do
(conditional return is surely better).
Fixed thusly, ok for trunk/4.8?
2013-04-02 Jakub Jelinek <ja...@redhat.com>
PR rtl-optimization/56745
* ifcvt.c (cond_exec_find_if_block): Don't try to optimize
if then_bb has no successors and else_bb is EXIT_BLOCK_PTR.
* gcc.c-torture/compile/pr56745.c: New test.
--- gcc/ifcvt.c.jj 2013-03-05 23:25:15.000000000 +0100
+++ gcc/ifcvt.c 2013-04-02 12:31:19.476859094 +0200
@@ -3473,7 +3473,7 @@ cond_exec_find_if_block (struct ce_if_bl
code processing. ??? we should fix this in the future. */
if (EDGE_COUNT (then_bb->succs) == 0)
{
- if (single_pred_p (else_bb))
+ if (single_pred_p (else_bb) && else_bb != EXIT_BLOCK_PTR)
{
rtx last_insn = BB_END (then_bb);
--- gcc/testsuite/gcc.c-torture/compile/pr56745.c.jj 2013-04-02
12:38:54.718258527 +0200
+++ gcc/testsuite/gcc.c-torture/compile/pr56745.c 2013-04-02
12:33:22.000000000 +0200
@@ -0,0 +1,15 @@
+/* PR rtl-optimization/56745 */
+
+unsigned char a[6];
+
+void
+foo ()
+{
+ int i;
+ for (i = 5; i >= 0; i++)
+ {
+ if (++a[i] != 0)
+ break;
+ ++a[i];
+ }
+}
Jakub
