RFR: 8284578: Relax InterpreterCodelet stub alignment [v2]

[Aleksey Shipilev]

On Wed, 13 Apr 2022 09:15:14 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

>> `InterpreterCodelet` is aligned by `CodeEntryAlignment` (`CAE`) twice. First, the entire stub is aligned, which aligns its data section. Then, the code section in the stub is aligned. Since `CAE` is usually larger than the size of `InterpreterCodelet`, we are wasting quite a bit of space for each codelet. In the extreme cases, like PPC that defaults to `CAE=128`, we have 16 bytes of codelet data effectively taking 128 bytes!
>> 
>> This can be made better by relaxing the `InterpreterCodelet` stub alignment to `HeapWordSize`, while leaving its code section alignment the same.
>> 
>> This tangentially touches the only other user for `StubQueue`: `ICStub`. Unfortunately, we cannot do the same kind of relaxation there, because there is a reverse lookup function that needs to reach data section from the code section, which forces us to keep the same alignment for both.
>> 
>> Interpreter sizes on Linux x86_64 release:
>> 
>> 
>> # Baseline, CEA=32 (default)
>> code size        =     94K bytes
>> avg codelet size =    356  bytes
>> 
>> # Baseline, CEA=128 (PPC-like)
>> code size        =    133K bytes
>> avg codelet size =    501  bytes
>> 
>> # Patched, CEA=32 (default)
>> code size        =     89K bytes
>> avg codelet size =    338  bytes
>> 
>> # Patched, CEA=128 (PPC-like)
>> code size        =    100K bytes
>> avg codelet size =    380  bytes
>> 
>> 
>> Point performance runs (SPECjvm2008:serial with `-Xint` on Linux x86_64 release):
>> 
>> 
>> Benchmark     Mode  Cnt   Score   Error  Units
>> 
>> # Baseline, CEA=32
>> Serial.test  thrpt    9  73.427 ± 0.152  ops/s
>> 
>> # Baseline, CEA=128
>> Serial.test  thrpt    9  70.999 ± 0.246  ops/s
>> 
>> # Patched, CEA=32
>> Serial.test  thrpt    9  73.991 ± 0.860  ops/s
>> 
>> # Patched, CEA=128
>> Serial.test  thrpt    9  72.981 ± 0.301  ops/s
>> 
>> 
>> Additional testing:
>>  - [x] Linux x86_64 fastdebug `tier1`
>>  - [x] Linux x86_64 fastdebug `tier2`
>>  - [x] Linux x86_64 fastdebug `tier3`
>
> Aleksey Shipilev 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 three additional commits since the last revision:
> 
>  - Relax slack space a bit
>  - Merge branch 'master' into JDK-8284578-intcodelet-align
>  - Initial implementation

Thanks!

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

PR: https://git.openjdk.java.net/jdk/pull/8159