On Thu, 19 Sep 2024 14:14:59 GMT, Chen Liang <li...@openjdk.org> wrote:
>> This PR changes proxy code gen to avoid generating
>> `Class.forName("java.lang.Object")`, instead emitting an ldc for the class
>> literal, `ldc(CD_Object)`, java code equivalent `Object.class`.
>>
>> More types could profitably use `ldc(ClassDesc/-Entry)`, taking cues from
>> `InvokerBytecodeGenerator.isStaticallyInvocable`, but just addressing the
>> `Object` methods gets rid of most `Class.forName` emits. It's not terribly
>> important for throughput performance since these are called in the generated
>> `clinit`, so getting a quick win with few additional checks is a good
>> starting point.
>>
>> Added a few unrelated minor refactors/improvements, guided by diagnostic
>> runs of the now fixed microbenchmark.
>
> src/java.base/share/classes/java/lang/reflect/ProxyGenerator.java line 833:
>
>> 831: cob.ldc(objectCE);
>> 832: } else {
>> 833: cob.ldc(cl.getName())
>
> This `Class.forName` is only necessary in a very small number of cases,
> namely when the overridden interface method has an unaccessible parameter
> type, usually a package-private type not accessible to the implementing
> class. Maybe we can always directly ldc if the class is `public`.
Yeah, I'm just not sure what rules are quite right here. Wouldn't want to
inadvertently regress this again, and not sure we have tests for all
eventualities, so I started this off with the most conservative yet most
beneficial improvement.
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/21090#discussion_r1766973606
