RFR: 8332587: RISC-V: secondary_super_cache does not scale well

[Gui Cao]

On Mon, 3 Jun 2024 08:27:50 GMT, Hamlin Li <mli at openjdk.org> wrote:

>> Implementation of subtype checking [JDK-8180450](https://bugs.openjdk.org/browse/JDK-8180450) for linux-riscv64.
>> This optimization depends on availability of the Zbb extension which has the cpop instruction.
>> 
>> ### Correctness testing:
>> 
>> - [x]  Run tier1-3, hotspot:tier4 tests on SOPHON SG2042 (release)
>> - [x] Run tier1-3 tests with -XX:+UseRVV on qemu 8.1.0 (fastdebug)
>> - [x] Run all of tier1 with `-XX:+VerifySecondarySupers`
>> 
>> ### JMH tested on Banana Pi BPI-F3 board (has Zbb) and Enable UseZbb, Not Enable UseZba, UseZbs
>> Original:
>> 
>> Benchmark                             Mode  Cnt    Score   Error  Units
>> SecondarySuperCacheHits.test  avgt   15  11.375 ± 0.071  ns/op
>> SecondarySuperCacheInterContention.test     avgt   15  646.087 ± 32.587  ns/op
>> SecondarySuperCacheInterContention.test:t1  avgt   15  600.090 ± 83.779  ns/op
>> SecondarySuperCacheInterContention.test:t2  avgt   15  692.084 ± 73.218  ns/op
>> SecondarySupersLookup.testNegative00  avgt   15   16.420 ± 0.239  ns/op
>> SecondarySupersLookup.testNegative01  avgt   15   18.307 ± 0.260  ns/op
>> SecondarySupersLookup.testNegative02  avgt   15   21.695 ± 0.458  ns/op
>> SecondarySupersLookup.testNegative03  avgt   15   24.855 ± 0.664  ns/op
>> SecondarySupersLookup.testNegative04  avgt   15   27.305 ± 0.522  ns/op
>> SecondarySupersLookup.testNegative05  avgt   15   29.719 ± 0.385  ns/op
>> SecondarySupersLookup.testNegative06  avgt   15   32.231 ± 0.498  ns/op
>> SecondarySupersLookup.testNegative07  avgt   15   33.747 ± 0.603  ns/op
>> SecondarySupersLookup.testNegative08  avgt   15   35.856 ± 0.629  ns/op
>> SecondarySupersLookup.testNegative09  avgt   15   37.077 ± 0.546  ns/op
>> SecondarySupersLookup.testNegative10  avgt   15   39.408 ± 0.465  ns/op
>> SecondarySupersLookup.testNegative16  avgt   15   51.041 ± 0.547  ns/op
>> SecondarySupersLookup.testNegative20  avgt   15   58.722 ± 0.922  ns/op
>> SecondarySupersLookup.testNegative30  avgt   15   77.310 ± 0.654  ns/op
>> SecondarySupersLookup.testNegative32  avgt   15   81.116 ± 0.854  ns/op
>> SecondarySupersLookup.testNegative40  avgt   15   96.311 ± 0.840  ns/op
>> SecondarySupersLookup.testNegative50  avgt   15  115.427 ± 0.838  ns/op
>> SecondarySupersLookup.testNegative55  avgt   15  124.371 ± 1.076  ns/op
>> SecondarySupersLookup.testNegative56  avgt   15  126.796 ± 0.916  ns/op
>> SecondarySupersLookup.testNegative57  avgt   15  127.952 ± 1.202  ns/op
>> SecondarySupersLookup.testNegative58  avgt   15  131.956 ± 4.515  ns/op
>> Seco...
>
> Hey, is there performance comparison data when zbb is not supported? Just want to know if it can also get performance gain in that situation, then this optimization is a more generic one.
> 
> The reason for ask this question is that seems JDK-8180450 is an optimization for cache line, it should or could bring benefit even if zbb is not supported?

@Hamlin-Li : Hi Hamlin, I've fixed all of the above comments. I've enabled Zba, Zbb, Zbs and tested JMH before and the performance has improved a bit compared to just enabling Zbb, I'll be re-testing the JMH with Zba, Zbb,Zbs enabled today and I'll update here when I'm done with it.

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

PR Comment: https://git.openjdk.org/jdk/pull/19320#issuecomment-2146441422