Bug 81900

Created attachment 42008 [details]
executable testcase

Standalone executable testcase, exits 0 when compiled correctly, aborts or exits 1 otherwise.  Fails for me with gcc-8 on x86_64-linux but works on same with gcc-7, gcc-6, and gcc-5.

Started with r250767.

I've investigated some more, and Mikael's bisection appears to confirm my investigation:

The problem appears to be this chunk:

@@ -2418,7 +2423,9 @@ compute_antic (void)
   inverted_post_order_compute (&postorder);

   auto_sbitmap worklist (last_basic_block_for_fn (cfun) + 1);
-  bitmap_ones (worklist);
+  bitmap_clear (worklist);
+  FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)
+    bitmap_set_bit (worklist, e->src->index);
   while (changed)
     { 
       if (dump_file && (dump_flags & TDF_DETAILS))


It assumes there are always edges from function calls to the exit block, when the function call might exit nonlocally. However, cfganal.c:add_noreturn_fake_exit_edges is never called for functions that call setjmp, so such bbs are, wrongly, never visited by compute_antic.


The brute-force method of disabling the tree/PRE pass for functions that call setjmp() appears to work and fix the issue, but I'm not sure that's the right way to go.

Mine.

So while there is one simple issue with the iteration, fix:

@@ -2119,14 +2170,13 @@ static sbitmap has_abnormal_preds;
 static bool
 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
 {
-  bool changed = false;
   bitmap_set_t S, old, ANTIC_OUT;
   bitmap_iterator bi;
   unsigned int bii;
   edge e;
   edge_iterator ei;
-  bool was_visited = BB_VISITED (block);
 
+  bool changed = ! BB_VISITED (block);
   old = ANTIC_OUT = S = NULL;
   BB_VISITED (block) = 1;
 
@@ -2217,7 +2267,7 @@ compute_antic_aux (basic_block block, bo
   /* clean (ANTIC_IN (block)) is defered to after the iteration converged
      because it can cause non-convergence, see for example PR81181.  */
 
-  if (!was_visited || !bitmap_set_equal (old, ANTIC_IN (block)))
+  if (!bitmap_set_equal (old, ANTIC_IN (block)))
     changed = true;
 
  maybe_dump_sets:

there is another issue as we think *_7 is antic and thus hoist the dereference
before the _7 == 0 check which will optimize that away as not necessary in
a later pass.

Ah ...

@@ -2396,9 +2446,6 @@ compute_antic (void)
        if (e->flags & EDGE_ABNORMAL)
          {
            bitmap_set_bit (has_abnormal_preds, block->index);
-
-           /* We also anticipate nothing.  */
-           BB_VISITED (block) = 1;
            break;
          }

I can confirm that fixes things here, thank you!

Fixed.

Author: rguenth
Date: Mon Aug 21 13:18:35 2017
New Revision: 251226

URL: https://gcc.gnu.org/viewcvs?rev=251226&root=gcc&view=rev
Log:
2017-08-21  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/81900
	* tree-ssa-pre.c (compute_antic_aux): Properly compute changed
	for blocks with abnormal predecessors.
	(compute_antic): Do not set visited flag prematurely.

	* gcc.dg/torture/pr81900.c: New testcase.

Added:
    trunk/gcc/testsuite/gcc.dg/torture/pr81900.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/tree-ssa-pre.c

*** Bug 81954 has been marked as a duplicate of this bug. ***

Author: aldyh
Date: Wed Sep 13 17:14:07 2017
New Revision: 252505

URL: https://gcc.gnu.org/viewcvs?rev=252505&root=gcc&view=rev
Log:
2017-08-21  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/81900
	* tree-ssa-pre.c (compute_antic_aux): Properly compute changed
	for blocks with abnormal predecessors.
	(compute_antic): Do not set visited flag prematurely.

	* gcc.dg/torture/pr81900.c: New testcase.

Added:
    branches/range-gen2/gcc/testsuite/gcc.dg/torture/pr81900.c
Modified:
    branches/range-gen2/gcc/ChangeLog
    branches/range-gen2/gcc/testsuite/ChangeLog
    branches/range-gen2/gcc/tree-ssa-pre.c