[Bug ipa/60518] [4.9 Regression] ICE: in verify_loop_structure, at cfgloop.c:1647

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60518

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ok, so it's branch_prob()s call of split_block that splits the common latch of
loops 1 and 2


 <bb 3>:  (header loop 1)

 <bb 4>:  (header loop 2, latch loop 2 and 1)
  _2 = fn1 ();
  if (_2 != 0)
    goto <bb 4>;
  else
    goto <bb 3>;

the cfghook adjusts the latch of loop 2 properly (bb4s loop father is 2)
but fails to notice that bb4 is also the latch of loop 1 ... so this
CFG hook loop handling only handles this correctly if we have simple
latches (because there the latch always belongs to its loop).

Fix:


Index: gcc/cfghooks.c
===================================================================
--- gcc/cfghooks.c      (revision 208536)
+++ gcc/cfghooks.c      (working copy)
@@ -510,9 +510,13 @@ split_block (basic_block bb, void *i)

   if (current_loops != NULL)
     {
+      edge_iterator ei;
+      edge e;
       add_bb_to_loop (new_bb, bb->loop_father);
-      if (bb->loop_father->latch == bb)
-       bb->loop_father->latch = new_bb;
+      /* Identify all loops bb may have been the latch of and adjust them.  */
+      FOR_EACH_EDGE (e, ei, new_bb->succs)
+       if (e->dest->loop_father->latch == bb)
+         e->dest->loop_father->latch = new_bb;
     }

   res = make_single_succ_edge (bb, new_bb, EDGE_FALLTHRU);