x86: Respect disableAVX2/512 options and check XCR0 for OS support

[BradleyWood]

Prereq eclipse-omr/omr#7602

Issue: #19408

[BradleyWood]

@0xdaryl See the force-push

[0xdaryl]

Jenkins test sanity xlinux,win,osx jdk21

[0xdaryl]

Windows nodes are offline. As this PR does have a Windows-specific component to it, can you confirm it's been built and tested on Windows?

[BradleyWood]

@0xdaryl can you relaunch?

[0xdaryl]

Jenkins test sanity.functional,sanity.openjdk win jdk21

[0xdaryl]

Just a ping on triaging the Windows CI failures.

[BradleyWood]

I don't think any of the windows tests ran. The only obvious error I see is this:

[2025-01-28T19:51:24.425Z] /bin/sh: -c: line 3: syntax error: unexpected end of file
[2025-01-28T19:51:24.425Z] make[2]: *** [settings.mk:383: setup_testList] Error 1

[0xdaryl]

Jenkins test sanity.functional,sanity.openjdk win jdk21

[0xdaryl]

I sniffed at the CI failures and see at least one JVM crash with an illegal instruction exception:

https://openj9-jenkins.osuosl.org/job/Test_openjdk21_j9_sanity.openjdk_x86-64_windows_Personal_testList_1/65/consoleFull

[BradleyWood]

Probably issue #21000? All the crashes happen in byte array hashcode with -XX:-UseCompressedOops. I will try to confirm soon...

[0xdaryl]

I will try to confirm soon...
Have you confirmed that all the failures are known issues?

[BradleyWood]

@0xdaryl The other issue did not manifest as an illegal instruction. That makes me a little concerned because the issue in #21000 shouldn't cause that. Can you confirm what microarchitecture that test machine has?

[0xdaryl]

Can you confirm what microarchitecture that test machine has?

Not off the top of my head. I suggest you request access to the machine to check. Ask @AdamBrousseau how to do that.

[BradleyWood]

@0xdaryl I got a little concerned because because the crash in jenkins would indicate to me that this does not fix the issue. However, the client had indicated that the fix works. I created a linux vm on an AVX-512 machine and modified the kernel to always disable OPMASK, Hi16_ZMM, and ZMM_Hi256 bits in XCR0 register. I was able to reproduce the SIGILL crash easily without this fix, and was unable make it crash with this fix.

As for windows, I didn't get access to that machine yet, but can't think of a reason this PR wouldn't fix the issue on windows server 2012. I looked through the build history of that machine, and I see the classic manifestation of #21000, wrong result, no SIGILL. That makes no sense as windows server 2012 does not support AVX-512 which is required to reproduce that crash. On the other hand, I found this build, which reproduced the SIGILL (which is what I would expect). So at the very least, I don't think this PR actually makes the issue worse...

Can you manually launch a grinder test on win2012x64-openj9-1? I just need to see jitdump.

On another note, I really wish the crash info would dump 16 bytes from [RIP] to the console to make it easier to see what instruction was generated.

[BradleyWood]

@0xdaryl Can you launch tests again?

[0xdaryl]

Jenkins test sanity.functional,sanity.openjdk win jdk21

[BradleyWood]

Failure is #21000

[mpirvu]

As far as I can tell, _xgetbv(unsigned int ecx) is defined only for Windows, but its usage is not protected by an ifdef.

[BradleyWood]

As far as I can tell, _xgetbv(unsigned int ecx) is defined only for Windows, but its usage is not protected by an ifdef.

It is protected by ifdef? With msvc we use intrinsic, otherwise define this function.

#if defined(OMR_OS_WINDOWS) && defined(TR_TARGET_X86)
#include <intrin.h>
#elif defined(TR_TARGET_X86)
inline unsigned long long _xgetbv(unsigned int ecx)
   {
   unsigned int eax, edx;
   __asm__ __volatile__("xgetbv" : "=a"(eax), "=d"(edx) : "c"(ecx));
   return ((unsigned long long)edx << 32) | eax;
   }
#endif

[0xdaryl]

@BradleyWood : see @mpirvu comment above

[dsouzai]

As far as I can tell, _xgetbv(unsigned int ecx) is defined only for Windows, but its usage is not protected by an ifdef.

It is protected by ifdef? With msvc we use intrinsic, otherwise define this function.

#if defined(OMR_OS_WINDOWS) && defined(TR_TARGET_X86)
#include <intrin.h>
#elif defined(TR_TARGET_X86)
inline unsigned long long _xgetbv(unsigned int ecx)
   {
   unsigned int eax, edx;
   __asm__ __volatile__("xgetbv" : "=a"(eax), "=d"(edx) : "c"(ecx));
   return ((unsigned long long)edx << 32) | eax;
   }
#endif

That means that _xgetbv is only defined for TR_TARGET_X86, so the uses of _xgetbv also need to be appropriately guarded. Currently there are build failures (eg, see #21158).

[BradleyWood]

I'm going to assume there is a good reason that perform specific code in this file is guarded by if statements and not #ifdef so I will just guard the small block of code. See #21160