RFR: 8240220: IdealLoopTree::dump_head predicate printing is broken
Stefan Karlsson
stefan.karlsson at oracle.com
Fri Feb 28 10:49:01 UTC 2020
Hi all,
Please review this patch to fix the loop predicate printing in
IdealLoopTree::dump_head.
https://cr.openjdk.java.net/~stefank/8240220/webrev.01/
https://bugs.openjdk.java.net/browse/JDK-8240220
In GraphKit::add_predicate you can see the order in which the
predicates are added:
add_predicate_impl(Deoptimization::Reason_predicate, nargs);
add_predicate_impl(Deoptimization::Reason_profile_predicate, nargs);
add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
When IdealLoopTree::dump_head looks for these predicates, the code has
incorrectly swapped the 'predicate' and 'profile_predicate':
2541 Node* predicate =
PhaseIdealLoop::find_predicate_insertion_point(entry,
Deoptimization::Reason_loop_limit_check);
2547 entry = PhaseIdealLoop::find_predicate_insertion_point(entry,
Deoptimization::Reason_predicate);
2554 entry = PhaseIdealLoop::find_predicate_insertion_point(entry,
Deoptimization::Reason_profile_predicate);
(Note that the predicates are added top-down, while dump_head walks them
bottom-up)
The code also uses wrong node when calling skip_loop_predicates. Compare
the skipping of 'limit_check':
2541 Node* predicate =
PhaseIdealLoop::find_predicate_insertion_point(entry,
Deoptimization::Reason_loop_limit_check);
2542 if (predicate != NULL ) {
2543 tty->print(" limit_check");
2544 entry = PhaseIdealLoop::skip_loop_predicates(entry);
and see how the skipping of 'predicate' sets entry on line 2547;
2547 entry = PhaseIdealLoop::find_predicate_insertion_point(entry,
Deoptimization::Reason_predicate);
2548 if (entry != NULL) {
2549 tty->print(" predicated");
2550 entry = PhaseIdealLoop::skip_loop_predicates(entry);
2551 }
Before the patch, the 'profile_predicated' and 'predicated' items were
missing:
Loop: N344/N304 limit_check counted [0,int),+1 (-1 iters) has_sfpt
After the patch, the lines look like this:
Loop: N344/N304 limit_check profile_predicated predicated counted
[0,int),+1 (-1 iters) has_sfpt
I've only tested this by visually inspecting the output. I don't know if
there are any existing tests for this kind of code in the Compiler.
Thanks,
StefanK
More information about the hotspot-compiler-dev
mailing list