Hi, I have ran into this issue again, this time on resolute.
The problematic code:
- The loop in void src/hotspot/share/opto/callnode.cpp:
CallJavaNode::copy_call_debug_info(PhaseIterGVN* phase,
for (JVMState *jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) {
jvms->set_map(this);
jvms->set_locoff(jvms->locoff()+jvms_adj);
jvms->set_stkoff(jvms->stkoff()+jvms_adj);
jvms->set_monoff(jvms->monoff()+jvms_adj);
jvms->set_scloff(jvms->scloff()+jvms_adj);
jvms->set_endoff(jvms->endoff()+jvms_adj);
}
and the similiar loop in
src/hotspot/share/opto/callnode.cpp:void JVMState::adapt_position(int delta)
for (JVMState* jvms = this; jvms != nullptr; jvms = jvms->caller()) {
jvms->set_locoff(jvms->locoff() + delta);
jvms->set_stkoff(jvms->stkoff() + delta);
jvms->set_monoff(jvms->monoff() + delta);
jvms->set_scloff(jvms->scloff() + delta);
jvms->set_endoff(jvms->endoff() + delta);
}
The generated code contains vector instructions
```
00000000003bb5d0 <JVMState::adapt_position(int)>:
3bb5d0: e7 03 00 03 20 22 vlvgf %v0,%r3,3
3bb5d6: e7 00 00 03 20 4d vrepf %v0,%v0,3
3bb5dc: b3 c1 00 4b ldgr %f4,%r11
3bb5e0: b9 04 00 bf lgr %r11,%r15
3bb5e4: e7 20 20 0c 00 06 vl %v2,12(%r2)
3bb5ea: 18 13 lr %r1,%r3
3bb5ec: 5a 10 20 1c a %r1,28(%r2)
3bb5f0: e7 62 00 00 20 f3 vaf %v6,%v2,%v0
3bb5f6: e7 60 20 0c 00 0e vst %v6,12(%r2)
3bb5fc: 50 10 20 1c st %r1,28(%r2)
3bb600: e3 20 20 00 00 02 ltg %r2,0(%r2)
3bb606: a7 74 ff ef jne 3bb5e4
<JVMState::adapt_position(int)+0x14>
3bb60a: b3 cd 00 b4 lgdr %r11,%f4
3bb60e: 07 fe br %r14
```
and instead of uniform adjustment we get one of the values only
partially changed:
work: locoff 10 stkoff 21 monoff 21 scloff 21 endoff 21 adj -4
result: locoff 6 stkoff 17 monoff 19 scloff 17 endoff 17
monoff: 21 - 4 = 19
I wonder if the root cause is vector register store/load somewhere in
kernel during the context switch because it is very prevalent on
jammy/noble with 'gs' flag enabled and I have seen only one instance in
pipeline in resolute.
I have worked it around by adding -fno-slp-tree-vectorize in openjdk
configure args, but it would be nice to understand/fix the root cause.
--
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2095441
Title:
S390x build of openjdk crashes with the segmentation fault
To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu-z-systems/+bug/2095441/+subscriptions
--
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
