RFR: 8367333: C2: Vector math operation intrinsification failure [v2]

[Vladimir Ivanov]

On Tue, 16 Sep 2025 20:05:48 GMT, Vladimir Ivanov <vlivanov at openjdk.org> wrote:

>> As part of [JDK-8353786](https://bugs.openjdk.org/browse/JDK-8353786), C2 support for operations backed by the vector math library was completely removed. On JDK side, there is a special dispatching logic added to avoid intrinsic calls in `jdk.internal.vm.vector.VectorSupport`. But it's still possible to observe such paradoxical situations (intrinsic calls with obsolete operation IDs) when processing effectively dead code.
>> 
>> Consider `FloatVector::lanewiseTemplate`:
>> 
>>     FloatVector lanewiseTemplate(VectorOperators.Unary op) {
>>         if (opKind(op, VO_SPECIAL)) {
>>             ...                             
>>             else if (opKind(op, VO_MATHLIB)) {
>>                 return unaryMathOp(op);
>>             }
>>         }
>>         int opc = opCode(op);
>>         return VectorSupport.unaryOp(opc, ...);
>>     }
>> 
>> 
>> At runtime, `unaryMathOp` is unconditionally invoked, but during compilation it's possible to end up with an intrinsification attempt of `VectorSupport.unaryOp()` before `opKind(op, VO_SPECIAL)` is inlined. 
>> 
>> It can be reliably reproduced `-XX:+StressIncrementalInlining` flag.
>> 
>> The fix is to fail-fast intrinsification rather than crashing the VM.
>> 
>> Testing: tier1 - tier4
>
> Vladimir Ivanov has updated the pull request incrementally with one additional commit since the last revision:
> 
>   review feedback

Thanks for the reviews.

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

PR Review: https://git.openjdk.org/jdk/pull/27263#pullrequestreview-3231511953