Request for reviews (S): 7050280: assert(u->as_Unlock()->is_eliminated()) failed: sanity
Vladimir Kozlov
vladimir.kozlov at oracle.com
Fri Jun 3 14:27:43 PDT 2011
Thank you, Igor
Vladimir
Igor Veresov wrote:
> This seems ok.
>
> igor
>
> On 6/2/11 7:25 PM, Vladimir Kozlov wrote:
>> http://cr.openjdk.java.net/~kvn/7050280/webrev
>>
>> Fixed 7050280: assert(u->as_Unlock()->is_eliminated()) failed: sanity
>>
>> Marking locks for elimination is done in IGVN. Unfortunately the order
>> of IGVN worklist processing may affect this marking. Also during EA obj
>> may point to several objects but after few ideal graph transformations
>> (CCP) it may point to only one non escaping object (but still using
>> phi), corresponding locks and unlocks will be marked for elimination.
>> Later obj could be replaced with a new node (new phi) and will be no
>> escape information about it. And later after some graph reshape other
>> locks and unlocks (which were not marked for elimination before) are
>> connected to this new obj but they still will not be marked for
>> elimination since new obj has no escape information.
>>
>> Fix:
>> Do first round of locks marking during EA after ConnectionGraph is
>> constructed when all nodes have escape information. During Macro nodes
>> expansion move creation of new eliminated BoxLock into new separate
>> method which is called before locks elimination. Use this method to mark
>> all associated (same box and obj) lock and unlock nodes which were not
>> marked before (instead of the assert).
>>
>> Tested with failing case and full CTW.
>
More information about the hotspot-compiler-dev
mailing list