RFR: 8331117: [PPC64] secondary_super_cache does not scale well [v6]

Amit Kumar amitkumar at openjdk.org
Fri Jun 14 15:38:20 UTC 2024


On Fri, 14 Jun 2024 12:01:43 GMT, Martin Doerr <mdoerr at openjdk.org> wrote:

>> PPC64 implementation of [JDK-8180450](https://bugs.openjdk.org/browse/JDK-8180450). Please review!
>> I noticed that `r_array_length` is sometimes 0 and I don't see code for that on x86. Any idea? (This has been addressed in the discussion.)
>> How can we verify it? By comparing the performance using the micro benchmarks?
>> 
>> Micro benchmark results without patch (measured on Power10 with 2*8 hardware threads):
>> 
>> Original 
>> SecondarySuperCacheHits: 13.033 ±(99.9%) 0.058 ns/op [Average]
>> SecondarySuperCacheInterContention.test     avgt   15  432.366 ±  8.364  ns/op
>> SecondarySuperCacheInterContention.test:t1  avgt   15  432.310 ±  8.460  ns/op
>> SecondarySuperCacheInterContention.test:t2  avgt   15  432.422 ± 10.819  ns/op
>> SecondarySuperCacheIntraContention.test  avgt   15  355.192 ± 3.597  ns/op
>> SecondarySupersLookup.testNegative00  avgt   15  12.274 ± 0.026  ns/op
>> SecondarySupersLookup.testNegative01  avgt   15  12.300 ± 0.039  ns/op
>> SecondarySupersLookup.testNegative02  avgt   15  12.304 ± 0.034  ns/op
>> SecondarySupersLookup.testNegative03  avgt   15  12.276 ± 0.050  ns/op
>> SecondarySupersLookup.testNegative04  avgt   15  12.235 ± 0.044  ns/op
>> SecondarySupersLookup.testNegative05  avgt   15  12.308 ± 0.156  ns/op
>> SecondarySupersLookup.testNegative06  avgt   15  12.291 ± 0.048  ns/op
>> SecondarySupersLookup.testNegative07  avgt   15  12.307 ± 0.052  ns/op
>> SecondarySupersLookup.testNegative08  avgt   15  12.398 ± 0.075  ns/op
>> SecondarySupersLookup.testNegative09  avgt   15  12.552 ± 0.122  ns/op
>> SecondarySupersLookup.testNegative10  avgt   15  12.490 ± 0.083  ns/op
>> SecondarySupersLookup.testNegative16  avgt   15  12.565 ± 0.092  ns/op
>> SecondarySupersLookup.testNegative20  avgt   15  19.059 ± 0.958  ns/op
>> SecondarySupersLookup.testNegative30  avgt   15  19.268 ± 0.124  ns/op
>> SecondarySupersLookup.testNegative32  avgt   15  20.059 ± 0.114  ns/op
>> SecondarySupersLookup.testNegative40  avgt   15  25.117 ± 0.368  ns/op
>> SecondarySupersLookup.testNegative50  avgt   15  32.735 ± 0.359  ns/op
>> SecondarySupersLookup.testNegative55  avgt   15  34.866 ± 0.152  ns/op
>> SecondarySupersLookup.testNegative56  avgt   15  35.492 ± 0.276  ns/op
>> SecondarySupersLookup.testNegative57  avgt   15  36.620 ± 0.334  ns/op
>> SecondarySupersLookup.testNegative58  avgt   15  37.226 ± 0.180  ns/op
>> SecondarySupersLookup.testNegative59  avgt   15  37.774 ± 0.241  ns/op
>> SecondarySupersLookup.testNegative60  avgt   15  38.627 ± 1.451  ns/op
>> Sec...
>
> Martin Doerr has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Minor improvements according to review suggestions.

src/hotspot/cpu/ppc/macroAssembler_ppc.cpp line 2160:

> 2158:          r_array_length == R5_ARG3                                 && \
> 2159:          (r_array_index == R6_ARG4      || r_array_index == noreg) && \
> 2160:          (r_sub_klass   == R7_ARG5      || r_sub_klass   == noreg) && \

Maybe we can set `r_super_klass = R5` and `r_sub_klass =R7` to keep consistency in `c1_Runtime1_ppc.cpp`:


    case slow_subtype_check_id:
      { // Support for uint StubRoutine::partial_subtype_check( Klass sub, Klass super );
        const Register sub_klass = R5,
                       super_klass = R4,
                       temp1_reg = R6,
                       temp2_reg = R0;
        __ check_klass_subtype_slow_path(sub_klass, super_klass, temp1_reg, temp2_reg); // returns with CR0.eq if successful
        __ crandc(CCR0, Assembler::equal, CCR0, Assembler::equal); // failed: CR0.ne
        __ blr();
      }
      break;


I can see this being done for `aarch64`, `x86` and `risc-v` as well.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/19368#discussion_r1640004253


More information about the hotspot-dev mailing list