[foreign-memaccess+abi] RFR: 8268266: Investigate way to lazily customize upcall lambda forms

[Maurizio Cimadamore]

On Thu, 10 Jun 2021 10:02:48 GMT, Jorn Vernee <jvernee at openjdk.org> wrote:

> Hi,
> 
> This patch makes upcall method handle customization lazy, which helps a lot in reducing the number of lambda form classes generated during for intstance the TestUpcall test, making it complete a lot faster.
> 
> This adds a very slight overhead per call, since there's another indirection, but I think the tradeoff is acceptable.
> 
> Since the patch was otherwise really small, this also adds a qsort benchmark that is also used the measure the cost of the extra indirection. Results are as follows:
> 
> 
> QSort without lazy customization:
> 
> Benchmark                         Mode  Cnt      Score     Error  Units
> QSort.jni_upcall_qsort_naive      avgt   30  21096.871 � 299.669  ns/op
> QSort.jni_upcall_qsort_optimized  avgt   30   3603.495 �  16.464  ns/op
> QSort.native_qsort                avgt   30     96.236 �   0.713  ns/op
> QSort.panama_upcall_qsort         avgt   30    937.675 �  26.160  ns/op
> 
> QSort with lazy customization:
> 
> Benchmark                         Mode  Cnt      Score     Error  Units
> QSort.jni_upcall_qsort_naive      avgt   30  21454.476 � 306.135  ns/op
> QSort.jni_upcall_qsort_optimized  avgt   30   3691.395 �  35.671  ns/op
> QSort.native_qsort                avgt   30     97.440 �   0.894  ns/op
> QSort.panama_upcall_qsort         avgt   30   1072.929 �  14.948  ns/op
> 
> 
> So, while there is some regression from this change, it's not super bad (not integer factor), and I think the tradeoff is worth it to generate fewer lambda form classes, which of course has it's own overhead not visible in this benchmark.
> 
> Thanks,
> Jorn

Looks good - what is the cost for downcalls?

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

Marked as reviewed by mcimadamore (Committer).

PR: https://git.openjdk.java.net/panama-foreign/pull/553