Request for reviews (S): 6686791: Side effect in NumberFormat tests with -server -Xcomp
Vladimir Kozlov
Vladimir.Kozlov at Sun.COM
Fri Apr 11 16:17:57 PDT 2008
http://webrev.invokedynamic.info/kvn/6686791/index.html
Fixed 6686791: Side effect in NumberFormat tests with -server -Xcomp
Problem:
New optimization in CmpPNode::sub() (6667580) removed the valid compare
instruction because of false positive answer from detect_dominating_control()
since one of CmpP's inputs (LoadP) has a control edge above the second input
(Allocate).
Solution:
Check control edges of all 'dom' node's inputs instead of checking
only the control edge.
Remove the fixed (20) steps limit, use NodeList to check
visited Region nodes instead.
New code gives 4 times more correct positive answers based on
jbb2005 and jvm98 runs. It doesn't affect compilation time noticeably.
Benchmarks do not show regression.
Reviewed by:
Fix verified (y/n): y
Other testing:
JPRT, refworkload
More information about the hotspot-compiler-dev
mailing list