[lworld+vector] RFR: 8319945: [lworld+vector] Fix vector api jtreg crash with "-XX:-EnableVectorSupport" [v2]

Jatin Bhateja jbhateja at openjdk.org
Mon Nov 27 10:13:31 UTC 2023


On Mon, 27 Nov 2023 07:38:43 GMT, Xiaohong Gong <xgong at openjdk.org> wrote:

>> Several vector api jtreg tests crash with `"-XX:-EnableVectorSupport"` when running on machines that the operations of vector payload can be vectorized.
>> 
>> Here is the log:
>> 
>> 
>>  Internal Error (valhalla/src/hotspot/share/opto/library_call.cpp:2465), pid=467482, tid=467499
>>  assert(bt == alias_type->basic_type() || bt == T_PRIMITIVE_OBJECT) failed: should match
>> 
>>  JRE version: OpenJDK Runtime Environment (22.0) (fastdebug build 22-internal-git-5b2fa0aab)
>>  Java VM: OpenJDK 64-Bit Server VM (fastdebug 22-internal-git-5b2fa0aab, mixed mode, sharing, compressed oops, compressed class ptrs, g1 gc, linux-aarch64)
>>  Problematic frame:
>>  V [[libjvm.so](http://libjvm.so/)+0x11b1738] LibraryCallKit::inline_unsafe_access(bool, BasicType, LibraryCallKit::AccessKind, bool)+0x13e8
>> 
>> 
>> The reason is the matched field of the given offset in the vector payload is not found when using Usafe's access methods. And the root cause is when the operations of the multifields are vectorizable in C2, only the multifield base is added into the holder klass's non-static fields list in ci stage. Hence when accessing other multifields by Unsafe's get/put methods (which is the fall-back java implementation of current vector operations), the fields cannot be found with the given field offset.
>> 
>> This issue is involved by [JDK-8314980](https://bugs.openjdk.java.net/browse/JDK-8314980) (see [1]), which refactors the multifields parsing in ci stage. The special handling for multifields in `ciInstanceKlass::get_field_by_offset` is removed. Adding it back can fix this issue.
>> 
>> This patch also reverted the special consideration for multifields when printing out the JVM state in a safepoint. It is not needed anymore, because the fields of an `InlineTypeNode` in C2 is always consistent with the holder klass's non-static fields after [JDK-8314980](https://bugs.openjdk.java.net/browse/JDK-8314980).
>> 
>> [1] https://github.com/openjdk/valhalla/pull/918
>> 
>> Verified with Vector API jtreg tests with `-XX:-EnableVectorSupport`, and all tests clean.
>
> Xiaohong Gong has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Add TODO comment in "ciInlineKlass::field_index_by_offset()"

Marked as reviewed by jbhateja (Committer).

-------------

PR Review: https://git.openjdk.org/valhalla/pull/953#pullrequestreview-1750074276



More information about the valhalla-dev mailing list