Re: Fix lto-symtab ICE during Ada LTO bootstrap

[Eric Botcazou <ebotcazou@adacore.com>]

[-- Attachment #1: Type: text/plain, Size: 2530 bytes --]

> I suppose the CFG verifier should also catch this.  I wonder how this can
> lead to wrong code as opossed to infinite loop?
> I can imagine DCE being confused about non-control-flow stmt and conclude
> the abnormal path as the path leaving the loop.  I will look into the
> testcase more.

   <bb 4063>:
   # DEBUG id => e_186
   _11422 = atree__unchecked_access__node4.localalias.3007 (e_186);
 
   <bb 4064>:
   # DEBUG id => NULL
   # DEBUG n => _11422
   # DEBUG n => NULL
   if (_11422 == 0)
     goto <bb 4097> (<L255>);
   else
     goto <bb 4065>;

The next block is:

  <bb 4065>:
  # DEBUG id => e_186
  goto <bb 4046>;

and has the stalled ABNORMAL flag.  This causes the latch edge to be split.

When the PHI node consuming _11422 is processed:

processing: e_186 = PHI <e_7741(4046), _11422(4164)>

the following code is invoked:

          if (aggressive && !degenerate_phi_p (stmt))
            {
              for (k = 0; k < gimple_phi_num_args (stmt); k++)
                {
                  basic_block arg_bb = gimple_phi_arg_edge (phi, k)->src;

          if (gimple_bb (stmt)
              != get_immediate_dominator (CDI_POST_DOMINATORS, arg_bb))
                    {
                      if (!bitmap_bit_p (last_stmt_necessary, arg_bb->index))
                        mark_last_stmt_necessary (arg_bb);
                    }
                  else if (arg_bb != ENTRY_BLOCK_PTR_FOR_FN (cfun)
                           && !bitmap_bit_p (visited_control_parents,
                                         arg_bb->index))
                    mark_control_dependent_edges_necessary (arg_bb, true);
                }
            }

arg_bb is the immediate postdominator of gimple_bb (stmt) so the first 
condition is false.  And the second condition is also false because arg_bb was 
already marked in visited_control_parents from:

      FOR_EACH_LOOP (loop, 0)
        if (!finite_loop_p (loop))
          {
            if (dump_file)
             fprintf (dump_file, "can not prove finiteness of loop %i\n", 
loop->num);
            mark_control_dependent_edges_necessary (loop->latch, false);
          }

I'm not quite sure where the logic goes wrong, but the comment just above the 
first quoted block of code makes one think it is a bit fragile.

In any case, the fixlet I posted was slightly off, so here is the patch I have 
installed as obvious after testing on x86-64/Linux.


	PR tree-optimization/65337
	* tree-ssa-pre.c (eliminate): Also clean up abnormal edges if need be.

-- 
Eric Botcazou

[-- Attachment #2: p.diff --]
[-- Type: text/x-patch, Size: 471 bytes --]

Index: tree-ssa-pre.c
===================================================================
--- tree-ssa-pre.c	(revision 231856)
+++ tree-ssa-pre.c	(working copy)
@@ -4499,6 +4499,8 @@ eliminate (bool do_pre)
 	  unlink_stmt_vdef (stmt);
 	  if (gsi_remove (&gsi, true))
 	    bitmap_set_bit (need_eh_cleanup, bb->index);
+	  if (is_gimple_call (stmt) && stmt_can_make_abnormal_goto (stmt))
+	    bitmap_set_bit (need_ab_cleanup, bb->index);
 	  release_defs (stmt);
 	}