Request for reviews (L): 7004535: Clone loop predicates when loop is cloned
vladimir.kozlov at oracle.com
Tue Mar 29 09:33:33 PDT 2011
Thank you, Tom
Tom Rodriguez wrote:
> This appears to be unused.
> + bool counted_loop = r->is_CountedLoop();
> ! // endif predicat [clone]
> + bool counted_loop = head->is_CountedLoop();
> How does the code in loopPredicate.cpp compare to what was originally in loopTransform.cpp? Are there changes or is it a straight copy?
No, I added a lot to do cloning. Here are diffs vs loopTransform.cpp:
> On Mar 28, 2011, at 5:06 PM, Vladimir Kozlov wrote:
>> I still need reviews for this. I merged with latest changes and added new method skip_loop_predicates() used in policy_do_remove_empty_loop().
>> Vladimir Kozlov wrote:
>>> Second loop opts changes.
>>> Fixed 7004535: Clone loop predicates when loop is cloned
>>> Currently loop predicates generated during parsing could be separated from loops when loops are cloned. As result such predicates are removed. There could be more optimization opportunities if loop predicates are also cloned in such situations.
>>> Loop predicate code become big so I moved it into new file loopPredicate.cpp.
>>> I separated the cloning code for IdealLoop and IterGVN. The generated Ideal
>>> code is the same but registration of new Ideal nodes is different enough to
>>> have separate methods, I think. I welcome any suggestions to improve this code.
>>> Keep loop predicates after CCP and perform optimizations with them until no more
>>> loop optimizations could be done. After that switch them off and do more loop
>>> VectorNode missed size_of() method as result it was cloned incorrectly.
>>> Added TraceLoopOpts outputs I missed in previous changes.
>>> Tested: ctw, nsk, jprt
More information about the hotspot-compiler-dev