[Bug tree-optimization/84670] [8 Regression] ICE: in compute_antic_aux, at tree-ssa-pre.c:2148 with -O2 -fno-tree-dominator-opts

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84670

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |needs-reduction

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Gerald Pfeifer from comment #6)
> Created attachment 43559 [details]
> A testcase from Wine (fairly bigger)
> 
> This is a testcase from a full Wine build.  The others probably are
> better to nail this down, but for verification just in case a first
> fix addresses those but might not cover this one as well.

A fix for the original testcase in this bug also fixes this one (I can
reproduce the ICE with -O2 -m32 on x86_64-linux).

The other two testcases are unrelated - further reduction would be nice
with inclusion of the following to avoid reducing to a testcase for the
issue I have a patch for.  Note that in the end I'm going to remove
the assert again - none of the testcases exhibit the infinite antic-in
computation issues.

Index: gcc/tree-ssa-pre.c
===================================================================
--- gcc/tree-ssa-pre.c  (revision 258124)
+++ gcc/tree-ssa-pre.c  (working copy)
@@ -1393,8 +1393,12 @@ phi_translate_1 (pre_expr expr, bitmap_s
                    else
                      {
                        unsigned value_id = get_expr_value_id (constant);
-                       constant = find_leader_in_sets (value_id, set1, set2,
-                                                       AVAIL_OUT (pred));
+                       /* We want a leader in ANTIC_OUT to be consistent
+                          with the rest of phi_translate but we don't have
+                          that available (and it isn't complete yet) so
+                          we resort to AVAIL_OUT here.  */
+                       constant = bitmap_find_leader (AVAIL_OUT (pred),
+                                                      value_id);
                        if (constant)
                          return constant;
                      }

The original use of set1/set2 is clearly bogus here, as the comment now
states we'd like to use ANTIC_OUT (pred) here but we are computing that
right now.  AVAIL_OUT is at least correct even if not 100% in spirit of
this function.  The other testcases show that phi_translate is effectively
removing things from the set which is what causes the issue.