Revert "Preserve ymm/zmm registers on x" #15999
Conversation
|
This will require an explanation for why this is necessary. Can you add the rationale to the commit message please? |
|
Jenkins test sanity xlinux,osx,win jdk17 In the meantime, I'll launch PR testing. |
|
For the record lest it get lost in the next push, JDK17 builds and sanity testing passed on all platforms. |
BradleyWood
force-pushed
the
revert-14632
branch
from
047e145
to
1541109
Compare
I have updated the commit message. |
|
Can you also add: "A previous attempt at disabling just the extended vector register preservation did not recoup the performance, but only completely reverting this PR restored the regression. This suggests there is something with the way the helpers have been restructured that is causing the performance degradation. This PR will be reverted while the cause is investigated." |
BradleyWood
force-pushed
the
revert-14632
branch
from
1541109
to
34b4da1
Compare
|
Is this something we'll want in the 0.35 release? |
runtime/jilgen/jilconsts.c
Outdated
| @@ -1,5 +1,5 @@ | |||
| /******************************************************************************* | |||
| * Copyright (c) 1991, 2022 IBM Corp. and others | |||
| * Copyright (c) 1991, 2021 IBM Corp. and others | |||
There was a problem hiding this comment.
This copyright shouldn't be reverted to 2021, the file is modified by other commits in 2022.
runtime/oti/xhelpers.m4
Outdated
| @@ -1,4 +1,4 @@ | |||
| dnl Copyright (c) 1991, 2022 IBM Corp. and others | |||
| dnl Copyright (c) 1991, 2021 IBM Corp. and others | |||
There was a problem hiding this comment.
This copyright shouldn't be reverted to 2021, the file is modified by other commits in 2022.
This reverts PR eclipse-openj9#14632. A 10%+ performance regression (issue eclipse-openj9#15716) reported on JDK 8 was isolated to eclipse-openj9#14632. Reverting this PR should not have any function change, it was only intended to support the Vector API JEPs which have not yet been enabled on x86 in OpenJ9. A previous attempt at disabling just the extended vector register preservation did not recoup the performance, but only completely reverting this PR restored the regression. This suggests there is something with the way the helpers have been restructured that is causing the performance degradation. This PR will be reverted while the cause is investigated.
BradleyWood
force-pushed
the
revert-14632
branch
from
34b4da1
to
b5f288f
Compare
Yes, if it is not too late and this PR has a chance to soak in at least one test cycle. @BradleyWood : could you prepare the 0.35 PR and have it ready for us to merge please? |
This reverts PR #14632. A 10%+ performance
regression (issue #15716) reported on JDK 8
was isolated to #14632. Reverting this PR
should not have any function change, it
was only intended to support the Vector
API JEPs which have not yet been enabled
on x86 in OpenJ9. A previous attempt at
disabling just the extended vector register
preservation did not recoup the performance,
but only completely reverting this PR
restored the regression. This suggests
there is something with the way the helpers
have been restructured that is causing the
performance degradation. This PR will be
reverted while the cause is investigated.