This is an issue where clear_bb_flags clears BB_IRREDUCIBLE_LOOP and thus
wrecks loop info. I'm not entirely sure what kind of flags we want
to clear (all uses would need to be audited I guess), the function isn't
used much either. The following follows the function documentation
and adds BB_IRREDUCIBLE_LOOP to the preserved set if it has to be
preserved (rather than unconditionally). From looking at cfg-flags.def
possibly BB_MODIFIED when df is active and BB_IN_TRANSACTIOn after
compute_transaction_bits () would need to be preserved as well.
Bootstrap & regtest running on x86_64-unknown-linux-gnu.
Does this look ok?
Thanks,
Richard.
2018-04-19 Richard Biener <rguent...@suse.de>
PR middle-end/85455
* cfg.c (clear_bb_flags): When loop state says we have
marked irreducible regions also preserve BB_IRREDUCIBLE_LOOP.
* gcc.dg/pr85455.c: New testcase.
Index: gcc/cfg.c
===================================================================
--- gcc/cfg.c (revision 259486)
+++ gcc/cfg.c (working copy)
@@ -386,9 +386,13 @@ void
clear_bb_flags (void)
{
basic_block bb;
+ int flags_to_preserve = BB_FLAGS_TO_PRESERVE;
+ if (current_loops
+ && loops_state_satisfies_p (cfun, LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS))
+ flags_to_preserve |= BB_IRREDUCIBLE_LOOP;
FOR_ALL_BB_FN (bb, cfun)
- bb->flags &= BB_FLAGS_TO_PRESERVE;
+ bb->flags &= flags_to_preserve;
}
/* Check the consistency of profile information. We can't do that
Index: gcc/testsuite/gcc.dg/pr85455.c
===================================================================
--- gcc/testsuite/gcc.dg/pr85455.c (nonexistent)
+++ gcc/testsuite/gcc.dg/pr85455.c (working copy)
@@ -0,0 +1,22 @@
+/* { dg-do compile } */
+/* { dg-options "-O1 -fthread-jumps -fno-tree-dominator-opts -fno-tree-reassoc
-fno-tree-sink -fno-tree-slsr" } */
+
+void
+ty (void);
+
+void
+um (void);
+
+void
+au (int qj)
+{
+ if (qj < 1)
+ {
+vq:
+ ty ();
+ }
+
+ um ();
+
+ goto vq;
+}
