RFR: 8343377: Performance regression in reflective invocation of native methods

[Chen Liang]

On Fri, 15 Nov 2024 22:17:10 GMT, Chen Liang <liach at openjdk.org> wrote:

> When core reflection was migrated to be implemented by Method Handles, somehow, the method handles are not used for native methods, which are generally linkable by method handles.  This causes significant performance regressions when reflecting native methods, even if their overrides may be non-native methods.  This is evident in `Object.clone` and `Object.hashCode` as shown in the original report.
> 
> I believe the blanket restriction previously placed on the native methods was because of signature polymorphic methods ([JLS 15.12.3](https://docs.oracle.com/javase/specs/jls/se23/html/jls-15.html#jls-15.12.3), [JVMS 2.9.3](https://docs.oracle.com/javase/specs/jvms/se23/html/jvms-2.html#jvms-2.9.3)) for MethodHandle and VarHandle; method handles do not link to the backing implementation that throws UOE while core reflection is required to do so.  I have narrowed the restrictions to be specifically against these methods.
> 
> Additionally, I cleaned up another check for invalid varargs flag.  Together, I clarified the scenarios where native method accessors are used - all to bypass restrictions of java.lang.invoke.
> 
> Testing: tier 1-5 green

Done. Performance numbers:

New:

Benchmark                               Mode  Cnt   Score   Error  Units
NativeMethodInvoke.objectHashCode       avgt   15  41.200 ± 2.783  ns/op
NativeMethodInvoke.threadCurrentThread  avgt   15   2.963 ± 0.143  ns/op


Old:

Benchmark                               Mode  Cnt    Score    Error  Units
NativeMethodInvoke.objectHashCode       avgt   15  533.562 ± 27.083  ns/op
NativeMethodInvoke.threadCurrentThread  avgt   15  108.492 ±  4.101  ns/op

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

PR Comment: https://git.openjdk.org/jdk/pull/22169#issuecomment-2502376813