Request for reviews (S): 7035946: Up to 15% regression on JDK 7 b136 vs b135 on specjvm2008.crypto.rsa on x64
vladimir.kozlov at oracle.com
Fri Sep 9 12:13:33 PDT 2011
Thank you, Tom
Tom Rodriguez wrote:
> Looks good.
> On Sep 9, 2011, at 9:54 AM, Vladimir Kozlov wrote:
>> 7035946: Up to 15% regression on JDK 7 b136 vs b135 on specjvm2008.crypto.rsa on x64
>> It is almost exact anti-delta of my changes in 7008866. I kept replace_parallel_iv() as separate method and added new TraceLoopOpts output.
>> I traced down the most of regression to my changes for "replace parallel induction variables" optimization in 7008866 fix. I replaced new nodes IGVN registration with immediate Ideal transformation. As result Split_If optimization reversed it back by doing spit through Phi and it prevents loop predicates (which follow Split_If) execute RCE. If local transformation in replace_parallel_iv() is not executed the created new code matches RCE shape: iv*ratio+(init2-init*ratio) where (init2-init*ratio) is loop invariant. And Split_If optimization does not execute spit through Phi for AddI nodes in Counted loops. Yes, a lot of hidden dependences.
>> An other part of regression came from always creation of new loop iteration Phi node. I was not able to find why it has regression affect. I returned code back to keep original Phi.
>> I do not see affect on refworkload benchmark after this fix.
More information about the hotspot-compiler-dev