I failed to notice that current_loop_nest is modified in
setup_current_loop_nest and used in sel_setup_region_sched_flags. Therefore,
sel_setup_region_sched_flags should be moved with setup_current_loop_nest
after sel_init_bbs as well.
I've committed the following patch under the 'obvious' rule after noticing the
problem on ia64 bootstrap with selective scheduler enabled at -O2 (I normally
do this together with amd64 bootstraps before committing patches, but this
time I made a typo that completely disabled sel-sched for ia64 testing).
Sorry for the noise.
2011-04-08 Alexander Monakov <amona...@ispras.ru>
* sel-sched.c (sel_region_init): Move call to
sel_setup_region_sched_flags after setup_current_loop_nest.
Index: gcc/sel-sched.c
===================================================================
--- gcc/sel-sched.c (revision 172177)
+++ gcc/sel-sched.c (working copy)
@@ -6868,8 +6868,6 @@ sel_region_init (int rgn)
if (current_region_empty_p ())
return true;
- sel_setup_region_sched_flags ();
-
bbs = VEC_alloc (basic_block, heap, current_nr_blocks);
for (i = 0; i < current_nr_blocks; i++)
@@ -6880,6 +6878,8 @@ sel_region_init (int rgn)
if (flag_sel_sched_pipelining)
setup_current_loop_nest (rgn, &bbs);
+ sel_setup_region_sched_flags ();
+
/* Initialize luids and dependence analysis which both sel-sched and haifa
need. */
sched_init_luids (bbs, NULL, NULL, NULL);
