RFR: 8273409: Receiver type narrowed by CCP does not always trigger post-parse call devirtualization [v2]

[Nils Eliasson]

On Wed, 8 Sep 2021 15:00:48 GMT, Tobias Hartmann <thartmann at openjdk.org> wrote:

>> While working on [JDK-8273323](https://bugs.openjdk.java.net/browse/JDK-8273323), I noticed that even if CCP is able to narrow the receiver type of a virtual call to an exact type, post-parse call devirtualization is not always triggered. The problem is that CCP only adds nodes to the IGVN worklist if their types have been improved:
>> https://github.com/openjdk/jdk/blob/4023646ed1bcb821b1d18f7e5104f04995e8171d/src/hotspot/share/opto/phaseX.cpp#L1961-L1965
>> 
>> And in turn, IGVN only adds user nodes to the worklist if the type of the parent node could be further improved. As a result, a `CallNode` is not always added to the worklist if the type of its receiver node has been improved. Often, we are lucky and macro expansion or another optimization phase after CCP adds the `CallNode` to the worklist. However, as `testDynamicCallWithCCP` shows, that's not always the case.
>> 
>> The fix is to explicitly add `CallStaticJavaNode` and `CallDynamicJava` to the worklist after CCP to give call devirtualization a chance to run:
>> https://github.com/openjdk/jdk/blob/4023646ed1bcb821b1d18f7e5104f04995e8171d/src/hotspot/share/opto/callnode.cpp#L1054-L1057
>> https://github.com/openjdk/jdk/blob/4023646ed1bcb821b1d18f7e5104f04995e8171d/src/hotspot/share/opto/callnode.cpp#L1137-L1140
>> 
>> I also considered moving the optimization from `Ideal` to `Value` (which in contrast to `Ideal` is already executed **during** CCP) but we can't trust receiver types during CCP (because CCP goes from optimistic types to pessimistic ones). Also, the current implementation relies on setting the `_generator` field which is not possible from `Value` which is declared `const`:
>> https://github.com/openjdk/jdk/blob/4023646ed1bcb821b1d18f7e5104f04995e8171d/src/hotspot/share/opto/callnode.cpp#L1169-L1172
>> 
>> All the newly added tests fail IR verification if post-parse call devirtualization is disabled (`-XX:-IncrementalInlineVirtual -XX:-IncrementalInlineMH`). Without the fix, `testDynamicCallWithCCP` always fails.
>> 
>> I filed [JDK-8273496](https://bugs.openjdk.java.net/browse/JDK-8273496) to improve how CCP handles this.
>> 
>> Thanks,
>> Tobias
>
> Tobias Hartmann has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Added missing IgnoreUnrecognizedVMOptions

Looks good!

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

Marked as reviewed by neliasso (Reviewer).

PR: https://git.openjdk.java.net/jdk/pull/5418