RFR(S) 8142436: [JVMCI] fix management of nmethod::_installed_code field
mikael.gerdin at oracle.com
Wed Nov 11 16:34:54 UTC 2015
On 2015-11-11 17:19, Tom Rodriguez wrote:
>> Additionally, you can run with +G1VerifyHeapRegionCodeRoots to enable
>> verification of the heap region to nmethod mapping in order to detect
>> other issues which don't cause a crash right away.
> I didn’t know that existed. That probably would have told me exactly
> what I wanted to know. Why isn’t that on by default? Too expensive? I
> was getting crashes and VerifyBeforeGC didn’t find anything so I added
> verify logic that checked that no zombie methods are in the code root
> set, which helped me track down the bug.
> http://hg.openjdk.java.net/graal/graal-jvmci-8/rev/2d21c13d7829 Having
> at least minimal verification enabled all the time would be nicer, at
> least it would have been for me. Failing this verify could suggest
> enabling G1VerifyHeapRegionCodeRoots.
I don't know why it's disabled by default, G1's verification is
expensive enough that enabling it should not make a large impact I
think. It's probably just a case of "if I don't enable it by default I
don't have to measure its impact"
More information about the hotspot-compiler-dev