------- Comment #4 from rguenth at gcc dot gnu dot org 2009-03-06 14:16 -------
Scheduling another pass_phi_only_cprop after VRP removes the single-argument
PHI nodes (I think that really cfg-cleanup should do this, as they usually
result from edge removal). While this is reasonably cheap it doesn't get
rid of the conditionals but that would require a forwprop run which is not
that cheap (it requires only the forward_propagate_into_gimple_cond () pieces,
but has to run after removing single-argument PHIs, thus cannot run at
VRP substitution time).
Thus, the following would fix it:
Index: passes.c
===================================================================
--- passes.c (revision 144665)
+++ passes.c (working copy)
@@ -611,6 +611,8 @@ init_optimization_passes (void)
NEXT_PASS (pass_copy_prop);
NEXT_PASS (pass_merge_phi);
NEXT_PASS (pass_vrp);
+ NEXT_PASS (pass_phi_only_cprop);
+ NEXT_PASS (pass_forwprop);
NEXT_PASS (pass_dce);
NEXT_PASS (pass_cselim);
NEXT_PASS (pass_tree_ifcombine);
alternatively only adding pass_phi_only_cprop and calling
forward_propagate_into_gimple_cond from pass_tree_ifcombine.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39390
