RFR: 8354242: VectorAPI: combine vector not operation with compare [v6]

[Xiaohong Gong]

On Wed, 28 May 2025 12:26:31 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

>> erifan 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 10 additional commits since the last revision:
>> 
>>  - Refactor the JTReg tests for compare.xor(maskAll)
>>    
>>    Also made a bit change to support pattern `VectorMask.fromLong()`.
>>  - Merge branch 'master' into JDK-8354242
>>  - Refactor code
>>    
>>    Add a new function XorVNode::Ideal_XorV_VectorMaskCmp to do this
>>    optimization, making the code more modular.
>>  - Merge branch 'master' into JDK-8354242
>>  - Update the jtreg test
>>  - Merge branch 'master' into JDK-8354242
>>  - Addressed some review comments
>>    
>>    1. Call VectorNode::Ideal() only once in XorVNode::Ideal.
>>    2. Improve code comments.
>>  - Merge branch 'master' into JDK-8354242
>>  - Merge branch 'master' into JDK-8354242
>>  - 8354242: VectorAPI: combine vector not operation with compare
>>    
>>    This patch optimizes the following patterns:
>>    For integer types:
>>    ```
>>    (XorV (VectorMaskCmp src1 src2 cond) (Replicate -1))
>>        => (VectorMaskCmp src1 src2 ncond)
>>    (XorVMask (VectorMaskCmp src1 src2 cond) (MaskAll m1))
>>        => (VectorMaskCmp src1 src2 ncond)
>>    ```
>>    cond can be eq, ne, le, ge, lt, gt, ule, uge, ult and ugt, ncond is the
>>    negative comparison of cond.
>>    
>>    For float and double types:
>>    ```
>>    (XorV (VectorMaskCast (VectorMaskCmp src1 src2 cond)) (Replicate -1))
>>        => (VectorMaskCast (VectorMaskCmp src1 src2 ncond))
>>    (XorVMask (VectorMaskCast (VectorMaskCmp src1 src2 cond)) (MaskAll m1))
>>        => (VectorMaskCast (VectorMaskCmp src1 src2 ncond))
>>    ```
>>    cond can be eq or ne.
>>    
>>    Benchmarks on Nvidia Grace machine with 128-bit SVE2:
>>    With option `-XX:UseSVE=2`:
>>    ```
>>    Benchmark			Unit	Before		Score Error	After		Score Error	Uplift
>>    testCompareEQMaskNotByte	ops/s	7912127.225	2677.289518	10266136.26	8955.008548	1.29
>>    testCompareEQMaskNotDouble	ops/s	884737.6799	446.963779	1179760.772	448.031844	1.33
>>    testCompareEQMaskNotFloat	ops/s	1765045.787	682.332214	2359520.803	896.305743	1.33
>>    testCompareEQMaskNotInt		ops/s	1787221.411	977.743935	2353952.519	960.069976	1.31
>>    testCompareEQMaskNotLong	ops/s	895297.1974	673.44808	1178449.02	323.804205	1.31
>>    testCompareEQMaskNotShort	ops/s	3339987.002	3415.2226	4712761.965	2110.862053	1.41
>>    testCompareGEMaskNotByte	ops/s	7907615.16	4...
>
> test/micro/org/openjdk/bench/jdk/incubator/vector/MaskCompareNotBenchmark.java line 49:
> 
>> 47:     private static final VectorSpecies<Long> L_SPECIES = LongVector.SPECIES_MAX;
>> 48:     private static final VectorSpecies<Float> F_SPECIES = FloatVector.SPECIES_MAX;
>> 49:     private static final VectorSpecies<Double> D_SPECIES = DoubleVector.SPECIES_MAX;
> 
> Are you taking `SPECIES_MAX` on purpose here, or could we take `SPECIES_PREFERRED` instead?
> @jatin-bhateja What is the best to do in these tests? I suppose best would be to test with all vector lengths...

Thanks for pointing out this @eme64 ! Per my understanding, `SPECIES_MAX` is almost the same with `SPECIES_PREFERRED` in this case which are all specified to the max vector size of a hardware. Since the max vector size is different on different architectures, not all vector lengths are supported to be intrinsified on a specified architecture like AArch64, especially the SVE arch with different vector register size. Hence, just testing the max species makes sense to me as this is a mid-end common transformation.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/24674#discussion_r2113020416