https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96487
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Last reconfirmed| |2020-08-06
Ever confirmed|0 |1
CC| |hubicka at gcc dot gnu.org,
| |rguenth at gcc dot gnu.org
Status|UNCONFIRMED |NEW
Keywords| |missed-optimization
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
I think the issue is that control dependence looks different - for C the latch
is control dependent on itself while for C++ it ends up being control dependent
on the outgoing edges from the conditional that ends up being marked necessary.
I guess both answers are somewhat "wrong" but it boils down to the
testcase not having any edges to exit and thus post-dominance being somewhat
ill-defined. Doing connect_infinite_loops_to_exit () doesn't help because
that as well makes different decisions based on the BB order.
So the question is whether DCE should do sth different than marking
control dependent edges of each non-finites loop latch, especially in
the case where the loop has no exit (if there are exits we should get
to the "first" exit test - that's the whole idea here). That is,
if there's no exit maybe simply mark the latch as bb_contains_live_stmts?
diff --git a/gcc/tree-ssa-dce.c b/gcc/tree-ssa-dce.c
index fae5ae72340..a66fc34c9e5 100644
--- a/gcc/tree-ssa-dce.c
+++ b/gcc/tree-ssa-dce.c
@@ -450,7 +450,10 @@ find_obviously_necessary_stmts (bool aggressive)
{
if (dump_file)
fprintf (dump_file, "cannot prove finiteness of loop %i\n",
loop->num);
- mark_control_dependent_edges_necessary (loop->latch, false);
+ if (loop->exits->next->e)
+ mark_control_dependent_edges_necessary (loop->latch, false);
+ else
+ bitmap_set_bit (bb_contains_live_stmts, loop->latch->index);
}
}
}
