http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53695
--- Comment #17 from rguenther at suse dot de <rguenther at suse dot de> 2012-08-23 09:19:04 UTC --- On Thu, 23 Aug 2012, steven at gcc dot gnu.org wrote: > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53695 > > --- Comment #16 from Steven Bosscher <steven at gcc dot gnu.org> 2012-08-23 > 08:53:04 UTC --- > (In reply to comment #15) > > Makes me wonder why the loop isn't recognized in the original test case... > > Ah, maybe because bb3 has an abnormal predecessor and is therefore ignored as > a > potential loop header? > > /* If we have an abnormal predecessor, do not consider the > loop (not worth the problems). */ > if (bb_has_abnormal_pred (header)) > continue; > > Which brings things back to my question why this kind of loop header is > rejected! :-) Because gimple_split_edge doesn't like to split abnormal edges, called via force_single_succ_latches (). So we do definitely not allow abnormal latch -> header edges. Still abnormal loop entries should be fine. So, Index: gcc/cfgloop.c =================================================================== --- gcc/cfgloop.c (revision 190613) +++ gcc/cfgloop.c (working copy) @@ -400,24 +400,21 @@ flow_loops_find (struct loops *loops) { edge_iterator ei; - /* If we have an abnormal predecessor, do not consider the - loop (not worth the problems). */ - if (bb_has_abnormal_pred (header)) - continue; - FOR_EACH_EDGE (e, ei, header->preds) { basic_block latch = e->src; - gcc_assert (!(e->flags & EDGE_ABNORMAL)); - /* Look for back edges where a predecessor is dominated by this block. A natural loop has a single entry node (header) that dominates all the nodes in the loop. It also has single back edge to the header from a latch node. */ if (latch != ENTRY_BLOCK_PTR - && dominated_by_p (CDI_DOMINATORS, latch, header)) + && dominated_by_p (CDI_DOMINATORS, latch, header) + /* We cannot make latches simple by splitting the + latch -> header edge if the latch edge is abnormal. */ + && (single_succ_p (latch) + || !(e->flags & EDGE_ABNORMAL))) { /* Shared headers should be eliminated by now. */ SET_BIT (headers, header->index); should "work". But doesn't fix the testcase (of course). Richard.