x64 JIT support for virtual threads (Loom) #16002
Conversation
* thisThreadGetOwnedMonitorCountOffset() * thisThreadGetCallOutCountOffset() Signed-off-by: Daryl Maier <maier@ca.ibm.com>
* callOutCount * ownedMonitorCount Signed-off-by: Daryl Maier <maier@ca.ibm.com>
Required for virtual thread support. Signed-off-by: Daryl Maier <maier@ca.ibm.com>
Only required for hard realtime support, which has been long deprecated. Signed-off-by: Daryl Maier <maier@ca.ibm.com>
Signed-off-by: Daryl Maier <maier@ca.ibm.com>
Required for virtual thread support. Signed-off-by: Daryl Maier <maier@ca.ibm.com>
|
This is essentially the same PR as #15552. The main difference is that the refactoring of one of the monitor enter/exit out of line sequences to remove a seemingly unnecessary label is not done. It turns out that label anchors a register dependency that is required to ensure correct register assignment on the OOL sequence. It was removed in the earlier PR to produce a better instruction sequence. This PR leaves it in and adds a monitor counter update instruction to the OOL path. This PR will remain a WIP until internal testing is complete. |
|
Jenkins test sanity all jdk17,jdk19 |
|
IBM SDK8 testing passed on 32-bit/64-bit Windows and Linux. |
|
We had some random failures that I assume were caused by the original changes. #15977 - sanity.functional jdk17 Windows |
|
jenkins test sanity.openjdk xmac,xlinux,win,win32 jdk8,jdk11,jdk17 |
|
FYI last nightly build was fine, but there are other changes merged since then in the PR build. https://openj9-jenkins.osuosl.org/job/Pipeline-Build-Test-JDK19/164/ |
|
OMR acceptance is also hanging, so the hangs are not caused by this PR. |
|
Created #16011 |
|
Internally I ran sanity.functional, extended.functional, and sanity.system on Linux, macOS, and Windows on JDK 17 and JDK 19 and all were clean. |
0xdaryl
changed the title
|
Given the internal testing, and what passed in PR testing, I have no concerns having this merged. |
|
@joransiu : Sorry for the churn. The previous PR that you reviewed and merged was reverted due to some issues being found on 32-bit. This PR is largely the same as that, except it avoids some seemingly benign out-of-line instruction refactoring for monitor enter/exit sequences to make placement of the Loom counter updates easier. It turns out there are some register assignment subtleties that rely on the shape of the OOL sequence so in this PR I left the the OOL sequence largely as-is but did add a counter bump for Loom. Full testing on 32-bit and JDK 17 and 19 all looks good. Failures found in this PR testing are attributed to #16011, as Pete mentions above. Would you mind reviewing and merging again please? Thanks. |
|
Given the testing mentioned above, and failures are unrelated to this issue, I'll go ahead and merge. |
ownedMonitorCountJ9VMThread field on monitor enter/exitcallOutCountJ9VMThread field for JNI dispatchIssue: #15175