[New Bug] Reliable C2 Crash in intrinsifying heap-based MemorySegment Vector store/loads

[Joel Knighton]

Hello,

I'm experiencing a reliable JVM crash from C2 compilation threads that's
reproducible on the latest release of 22 as well as fastdebug builds of the
master branch. This appears to have been introduced by JDK-8319111.
Post-8319111, it is possible to attempt to intrinsify a store/load of a
mismatched vector through inline_vector_mem_operation when the mem_elem_bt
is T_ILLEGAL. This causes the C2 thread to crash in type2aelembytes
(fastdebug) or vector_width_in_bytes (release). Patching to
fail intrinsifying in these cases does eliminate the crash for me, but I'm
not exceedingly familiar with this area of the codebase, so I may be
missing other similar issues.

I've uploaded a small, reliable reproducer here:
https://gist.github.com/jkni/e9b569d406a9f4d4f2766e80f9e904d2
For me, this reliably reproduces the issue on x86 under UseAVX=2 or
UseAVX=3.

I've also attached an excerpt of hs_err (full contents available upon
request, but I expect this should be sufficient):
https://gist.github.com/jkni/d31c5c2a77de0070f7066c093b4efac9

I've been unable to locate any sign of this bug already being filed. Is
there any other additional information I could provide that would be
helpful?

Best,
Joel Knighton
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/panama-dev/attachments/20240401/948a27aa/attachment.htm>