On Sun, 30 Apr 2023 06:14:28 GMT, Chen Liang <li...@openjdk.org> wrote:
> Try to intrinsify MethodHandles.arrayConstructor for java.lang, > java.lang.invoke, and java.util types, which InvokerBytecodeGenerator can > refer to directly in a class constant. > > In fact, the performance difference of Array.newInstance and anewarray is > quite negligible per benchmark: > > Benchmark Mode > Cnt Score Error Units > MethodHandlesArrayConstructorCall.baselineConstruct avgt > 5 3.467 ± 0.163 ns/op > MethodHandlesArrayConstructorCall.constantFoldConstruct avgt > 5 3.403 ± 0.070 ns/op > MethodHandlesArrayConstructorCall.constantFoldNonInvocableConstruct avgt > 5 3.445 ± 0.231 ns/op > MethodHandlesArrayConstructorCall.instanceConstruct avgt > 5 4.034 ± 0.168 ns/op > MethodHandlesArrayConstructorCall.instanceNonInvocableConstruct avgt > 5 4.038 ± 0.104 ns/op > MethodHandlesArrayConstructorColdStart.testCreateInvocable ss > 10 1425.991 ± 308.216 us/op > MethodHandlesArrayConstructorColdStart.testCreateNonInvocable ss > 10 1401.480 ± 258.402 us/op > MethodHandlesArrayConstructorColdStart.testCreateObject ss > 10 1213.018 ± 166.522 us/op > MethodHandlesArrayConstructorColdStart.testCreatePrimitive ss > 10 1163.461 ± 178.549 us/op `Array::newInstance` is already intrinsified. It behaves pretty much the same as `anewarray` if the array type is constant ([1], [2]). So, I'm not sure I see how intrinsifying this could improve performance. What kind of performance improvement are you seeing? Could you add a comparison with a method handle that calls `Array::newInstance` bound to the array type? (that should have the same performance as the current implementation). [1]: https://github.com/openjdk/jdk/blob/b3dbf28bc0614bee2f7137af95389134155c9511/src/hotspot/share/opto/library_call.cpp#L4051 [2]: https://github.com/openjdk/jdk/blob/b3dbf28bc0614bee2f7137af95389134155c9511/src/hotspot/share/opto/parse3.cpp#L282 ------------- PR Comment: https://git.openjdk.org/jdk/pull/13735#issuecomment-1529144414
