RFR: 8342393: Promote commutative vector IR node sharing [v9]

[Jatin Bhateja]

On Thu, 16 Jan 2025 12:01:13 GMT, Jatin Bhateja <jbhateja at openjdk.org> wrote:

>> Patch promotes the sharing of commutative vector IR with the same inputs but different input ordering.
>> Unlike scalar IR where we perform edge swapping by [sorting inputs](https://github.com/openjdk/jdk/blob/master/src/hotspot/share/opto/addnode.cpp#L122) based on node indices during IR idealization, for vector IR we chose a simpler approach to decorate commutative operations with a special node-level flag during IR construction thus
>> obviating any dependency on explicit idealization routines. This flag is later used during GVN hashing to enable node sharing.  
>> 
>> Following are the performance stats for JMH micro included with the patch.
>> 
>> 
>> Granite Rapids (P-core Xeon Server)
>> Baseline : 
>> Benchmark                                                                (size)   Mode  Cnt      Score   Error   Units
>> VectorCommutativeOperSharingBenchmark.commutativeByteOperationShairing     1024  thrpt    2   8982.549          ops/ms
>> VectorCommutativeOperSharingBenchmark.commutativeIntOperationShairing      1024  thrpt    2   6072.773          ops/ms
>> VectorCommutativeOperSharingBenchmark.commutativeLongOperationShairing     1024  thrpt    2   2368.856          ops/ms
>> VectorCommutativeOperSharingBenchmark.commutativeShortOperationShairing    1024  thrpt    2  15215.087          ops/ms
>> 
>> Withopt:
>> Benchmark                                                                (size)   Mode  Cnt      Score   Error   Units
>> VectorCommutativeOperSharingBenchmark.commutativeByteOperationShairing     1024  thrpt    2  11963.554          ops/ms
>> VectorCommutativeOperSharingBenchmark.commutativeIntOperationShairing      1024  thrpt    2   7036.088          ops/ms
>> VectorCommutativeOperSharingBenchmark.commutativeLongOperationShairing     1024  thrpt    2   2906.731          ops/ms
>> VectorCommutativeOperSharingBenchmark.commutativeShortOperationShairing    1024  thrpt    2  17148.131          ops/ms
>> 
>> Sierra Forest (E-core Xeon Server)
>> Baseline:
>> Benchmark                                                                (size)   Mode  Cnt     Score   Error   Units
>> VectorCommutativeOperSharingBenchmark.commutativeByteOperationShairing     1024  thrpt    2  2444.359          ops/ms
>> VectorCommutativeOperSharingBenchmark.commutativeIntOperationShairing      1024  thrpt    2  1710.256          ops/ms
>> VectorCommutativeOperSharingBenchmark.commutativeLongOperationShairing     1024  thrpt    2   308.766          ops/ms
>> VectorCommutativeOperSharingBenc...
>
> Jatin Bhateja has refreshed the contents of this pull request, and previous commits have been removed. The incremental views will show differences compared to the previous content of the PR. The pull request contains one new commit since the last revision:
> 
>   Generalizing vector size constraints covering different AVX levels and KNLSetting

> for vector IR we chose a simpler approach to decorate commutative operations with a special node-level flag during IR construction thus
> obviating any dependency on explicit idealization routines.

Hi @eme64 ,  Thanks,  to me marking a node as commutative right at construction time looks more appealing rather than banking on edge swapping during idealization. Patch is important for x86 where we can save bulky backend implementations through commoning, especially for sub-word vector operations where in absence of direct instruction we first upcast to word or double wordl vector, perform operation and then downcast it back to original type,  there is an order of magnitude performance improvement. I am ok to do the scaler part in follow up RFE as we have some working solution for it in place.

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

PR Comment: https://git.openjdk.org/jdk/pull/22863#issuecomment-2595969696