RFR (M) 8186042: Optimize OopMapCache lookup

coleen.phillimore at oracle.com coleen.phillimore at oracle.com
Fri Aug 25 14:05:26 UTC 2017

This was reviewed on hotspot-runtime-dev but it was noticed that the 
traversal of the interpreter's OopMapCache can be done outside of a 
safepoint by the graal compiler and by stack walking.   This would make 
reclamation of the "old" cached entries not safe.  I have disallowed use 
of the interpreter OopMapCache in this change by any calls outside of 
GC.   The OopMapCache is an optimization primarily (now exclusively) for 
GC root walking.

open webrev at http://cr.openjdk.java.net/~coleenp/8186042.02/webrev
bug link https://bugs.openjdk.java.net/browse/JDK-8186042

Retested with jck tests with -Xlog:interpreter+oopmap=debug, hotspot 
tier1 which includes the compiler AOT tests that failed, jvmti tests, 
and stack walking tests in the jdk.


On 8/11/17 11:46 AM, coleen.phillimore at oracle.com wrote:
> Summary: Use lock free access to oopMapCache
> Contributed-by: frederic.parain at oracle.com, coleen.phillimore at oracle.com
> The OopMapCache::lookup() function took out a mutex to protect access 
> between the GC threads that are running concurrently.  See bug for 
> more info.   The function lookup() is run by multiple GC threads 
> concurrently.  If there's a collision in the hashtable, this uses 
> atomic cmpxchg to add the entry to a list to be cleaned up after the 
> safepoint is over.   GC isn't doing lookup at that point.
> This change is contributed by Frederic Parain, with some cleanup and 
> logging from me.
> open webrev at http://cr.openjdk.java.net/~coleenp/8186042.01/webrev
> bug link https://bugs.openjdk.java.net/browse/JDK-8186042
> Tested with RBT equivalent of nightly on linux x64.  Also ran dacapo 
> with -Xint -Xlog:interpreter+oopmap=debug to verify.  This change also 
> removes -XX:+TraceOopMapGeneration (not -XX:+TraceNewOopMapGeneration 
> however) in favor of new logging.  A linked CSR request is pending.
> Thanks,
> Coleen

More information about the hotspot-dev mailing list