RFR: 8308869: C2: use profile data in subtype checks when profile has more than one class [v10]
Tobias Hartmann
thartmann at openjdk.org
Mon Sep 4 08:09:56 UTC 2023
On Wed, 23 Aug 2023 11:57:22 GMT, Roland Westrelin <roland at openjdk.org> wrote:
>> In this simple micro benchmark:
>>
>> https://github.com/franz1981/java-puzzles/blob/main/src/main/java/red/hat/puzzles/polymorphism/RequireNonNullCheckcastScalability.java#L70
>>
>> Performance drops sharply with polluted profile:
>>
>>
>> Benchmark (typePollution) Mode Cnt Score Error Units
>> RequireNonNullCheckcastScalability.isDuplicated1 false thrpt 10 1453.372 ± 24.919 ops/us
>>
>>
>> to:
>>
>>
>> Benchmark (typePollution) Mode Cnt Score Error Units
>> RequireNonNullCheckcastScalability.isDuplicated1 true thrpt 10 28.579 ± 2.280 ops/us
>>
>>
>> The test has 2 type checks to 2 different interfaces so caching with
>> `secondary_super_cache` doesn't help.
>>
>> The micro-benchmark only uses 2 different concrete classes
>> (`DuplicatedContext` and `NonDuplicatedContext`) and they are recorded
>> in profile data at the type checks. But c2 only take advantage of
>> profile data at type checks if they report a single class.
>>
>> What I propose is that the full blown type check expanded in
>> `Phase::gen_subtype_check()` takes advantage of profile data. So in
>> the case of the micro benchmark, before checking the
>> `secondary_super_cache`, generated code checks whether the object
>> being type checked is a `DuplicatedContext` or a
>> `NonDuplicatedContext`.
>>
>> This works fairly well on this micro benchmark:
>>
>>
>> Benchmark (typePollution) Mode Cnt Score Error Units
>> RequireNonNullCheckcastScalability.isDuplicated1 true thrpt 10 871.224 ± 20.750 ops/us
>>
>>
>> It also scales much better if there are multiple threads running the
>> same test (`secondary_super_cache` doesn't scale well: see
>> JDK-8180450).
>>
>> Now if the micro-benchmark is changed according to the comment:
>>
>> https://github.com/franz1981/java-puzzles/blob/d2d60af3d0dfe7a2567807395138edcb1d1c24f5/src/main/java/red/hat/puzzles/polymorphism/RequireNonNullCheckcastScalability.java#L62
>>
>> so the type check hits in the `secondary_super_cache`, the current
>> code performs much better:
>>
>>
>> Benchmark (typePollution) Mode Cnt Score Error Units
>> RequireNonNullCheckcastScalability.isDuplicated1 true thrpt 10 871.224 ± 20.750 ops/us
>>
>>
>> but leveraging profiling as explained above performs even better:
>>
>>
>> Benchmark ...
>
> Roland Westrelin has updated the pull request with a new target base due to a merge or a rebase. The incremental webrev excludes the unrelated changes brought in by the merge/rebase. The pull request contains 19 additional commits since the last revision:
>
> - white spaces
> - test case fix
> - Merge branch 'master' into JDK-8308869
> - riscv support
> - improvements to test
> - Merge branch 'master' into JDK-8308869
> - never common SubTypeCheckNode nodes
> - keep both ways of doing profile
> - whitespace
> - reworked change
> - ... and 9 more: https://git.openjdk.org/jdk/compare/e5a66903...082efa2e
Great work, Roland. The changes look good to me. I just added some minor style suggestions of things I stumbled upon when reading the code.
What's the plan for PPC, S390, 32-bit ARM, Zero support?
src/hotspot/cpu/aarch64/c1_LIRAssembler_aarch64.cpp line 1300:
> 1298: }
> 1299: Label *success_target = success;
> 1300: Label *failure_target = failure;
Suggestion:
Label* success_target = success;
Label* failure_target = failure;
src/hotspot/cpu/aarch64/c1_LIRAssembler_aarch64.cpp line 1429:
> 1427: Label done;
> 1428: Label *success_target = &done;
> 1429: Label *failure_target = stub->entry();
Suggestion:
Label* success_target = &done;
Label* failure_target = stub->entry();
src/hotspot/cpu/riscv/c1_LIRAssembler_riscv.cpp line 1184:
> 1182: }
> 1183: Label *success_target = success;
> 1184: Label *failure_target = failure;
Suggestion:
Label* success_target = success;
Label* failure_target = failure;
src/hotspot/cpu/riscv/c1_LIRAssembler_riscv.cpp line 2133:
> 2131: Label done;
> 2132: Label *success_target = &done;
> 2133: Label *failure_target = stub->entry();
Suggestion:
Label* success_target = &done;
Label* failure_target = stub->entry();
src/hotspot/cpu/x86/c1_LIRAssembler_x86.cpp line 1699:
> 1697: }
> 1698: Label *success_target = success;
> 1699: Label *failure_target = failure;
Suggestion:
Label* success_target = success;
Label* failure_target = failure;
src/hotspot/cpu/x86/c1_LIRAssembler_x86.cpp line 1855:
> 1853: Label done;
> 1854: Label *success_target = &done;
> 1855: Label *failure_target = stub->entry();
Suggestion:
Label* success_target = &done;
Label* failure_target = stub->entry();
src/hotspot/cpu/x86/interp_masm_x86.cpp line 1720:
> 1718: increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
> 1719: } else {
> 1720: int non_profiled_offset = in_bytes(CounterData::count_offset());
Unused?
src/hotspot/cpu/x86/interp_masm_x86.cpp line 1720:
> 1718: increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
> 1719: } else {
> 1720: int non_profiled_offset = in_bytes(CounterData::count_offset());
Unused?
src/hotspot/share/opto/c2_globals.hpp line 780:
> 778: product(intx, TypeProfileSubTypeCheckCommonThreshold, 50, \
> 779: "Use profile data at type check if together profiled types" \
> 780: "account for more than this threshold") \
I would remove "together".
src/hotspot/share/opto/graphKit.cpp line 2773:
> 2771:
> 2772: // Gather the various success & failures here
> 2773: RegionNode *r_not_subtype = new RegionNode(3);
Suggestion:
RegionNode* r_not_subtype = new RegionNode(3);
src/hotspot/share/opto/graphKit.cpp line 2775:
> 2773: RegionNode *r_not_subtype = new RegionNode(3);
> 2774: gvn.record_for_igvn(r_not_subtype);
> 2775: RegionNode *r_ok_subtype = new RegionNode(4);
Suggestion:
RegionNode* r_ok_subtype = new RegionNode(4);
src/hotspot/share/opto/graphKit.cpp line 2786:
> 2784: if (!profile.has_receiver(i)) {
> 2785: break;
> 2786: }
Suggestion:
for (int i = 0; profile.has_receiver(i); ++i) {
src/hotspot/share/opto/graphKit.cpp line 2795:
> 2793: if (!profile.has_receiver(i)) {
> 2794: break;
> 2795: }
Suggestion:
for (int i = 0; profile.has_receiver(i); ++i) {
src/hotspot/share/opto/ifnode.cpp line 1598:
> 1596: }
> 1597: // If the comparison is a subtype check, then SubTypeCheck nodes may have profile data attached to them and may be
> 1598: // different nodes even-though they perform the same subtype check
This comment should go to the check in line 1606 below.
src/hotspot/share/opto/subtypenode.hpp line 65:
> 63: private:
> 64: ciMethod* _method;
> 65: int _bci;
Please add comments describing the purpose of these fields.
-------------
Marked as reviewed by thartmann (Reviewer).
PR Review: https://git.openjdk.org/jdk/pull/14375#pullrequestreview-1608854638
PR Review Comment: https://git.openjdk.org/jdk/pull/14375#discussion_r1314570689
PR Review Comment: https://git.openjdk.org/jdk/pull/14375#discussion_r1314570457
PR Review Comment: https://git.openjdk.org/jdk/pull/14375#discussion_r1314569724
PR Review Comment: https://git.openjdk.org/jdk/pull/14375#discussion_r1314569538
PR Review Comment: https://git.openjdk.org/jdk/pull/14375#discussion_r1314568947
PR Review Comment: https://git.openjdk.org/jdk/pull/14375#discussion_r1314567439
PR Review Comment: https://git.openjdk.org/jdk/pull/14375#discussion_r1314565568
PR Review Comment: https://git.openjdk.org/jdk/pull/14375#discussion_r1314567116
PR Review Comment: https://git.openjdk.org/jdk/pull/14375#discussion_r1314507710
PR Review Comment: https://git.openjdk.org/jdk/pull/14375#discussion_r1314531000
PR Review Comment: https://git.openjdk.org/jdk/pull/14375#discussion_r1314531157
PR Review Comment: https://git.openjdk.org/jdk/pull/14375#discussion_r1314509452
PR Review Comment: https://git.openjdk.org/jdk/pull/14375#discussion_r1314510038
PR Review Comment: https://git.openjdk.org/jdk/pull/14375#discussion_r1314553523
PR Review Comment: https://git.openjdk.org/jdk/pull/14375#discussion_r1314497398
More information about the hotspot-compiler-dev
mailing list