[Bug rtl-optimization/55270] [4.8 Regression] ICE in get_loop_body, at cfgloop.c:823

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55270

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> 2013-01-28 14:03:18 UTC ---
When propagate_rhs_into_lhs alters the CFG from

   if ()
    {
     if ()
      exit_loop;
    }

to

   if ()
    {
      exit_loop;
    }

it fails to fixup loop structure:

          /* Propagation into these nodes may make certain edges in
             the CFG unexecutable.  We want to identify them as PHI nodes
             at the destination of those unexecutable edges may become
             degenerates.  */
          else if (gimple_code (use_stmt) == GIMPLE_COND
                   || gimple_code (use_stmt) == GIMPLE_SWITCH
                   || gimple_code (use_stmt) == GIMPLE_GOTO)
            {
...

Now, for full propagation the edge wants marking as non-executable
(which is done by removing it here - somewhat premature, as CFG cleanup
is run afterwards anyway and that handles the situation correctly
by calling fix_loop_structure).

Index: gcc/tree-ssa-dom.c
===================================================================
--- gcc/tree-ssa-dom.c  (revision 195502)
+++ gcc/tree-ssa-dom.c  (working copy)
@@ -3006,7 +3006,14 @@ eliminate_degenerate_phis (void)
     }

   if (cfg_altered)
-    free_dominance_info (CDI_DOMINATORS);
+    {
+      free_dominance_info (CDI_DOMINATORS);
+      if (current_loops)
+       {
+         calculate_dominance_info (CDI_DOMINATORS);
+         fix_loop_structure (NULL);
+       }
+    }

   /* Propagation of const and copies may make some EH edges dead.  Purge
      such edges from the CFG as needed.  */

fixes this (one more reason to fix loop structure at loop_optimizer_init
instead ...)