RFR: 8300865: C2: product reduction in ProdRed_Double is not vectorized [v4]

[Sandhya Viswanathan]

On Wed, 31 May 2023 00:47:56 GMT, Sandhya Viswanathan <sviswanathan at openjdk.org> wrote:

>> This PR fixes the problem with double reduction on x86_64. 
>> 
>> In the test compiler.loopopts.superword.ProdRed_Double, the product reduction loop in prodReductionImplement() was not getting vectorized when run as follows:
>> jtreg -XX:CompileCommand=PrintAssembly,compiler.loopopts.superword.ProdRed_Double::prodReductionImplement compiler/loopopts/superword/ProdRed_Double.java
>> The print assembly generated in the pid-xxx.log output in JTwork/scratch directory was not showing any vector_reduction_double node.
>> 
>> This was happening as the ReductionNode::implemented was passed a vector size of one element. For the vector reduction implemented we need to check with at least vector size of two elements.
>> 
>> With this PR the vector_reduction_double node is generated.
>> 
>> Please review.
>> 
>> Best Regards,
>> Sandhya
>
> Sandhya Viswanathan has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Add jmh test case

The performance numbers on my desktop are:
Base runs, no vectorization happens with superword:
Benchmark                              (COUNT)  (seed)  Mode  Cnt    Score   Error  Units
VectorReduction.NoSuperword.mulRedD        512       0  avgt    4  435.795 ± 0.082  ns/op
VectorReduction.WithSuperword.mulRedD      512       0  avgt    4  434.154 ± 0.042  ns/op

With the PR reduction succeeds and vectorization of the loop happens when superword is enabled: 
Benchmark                              (COUNT)  (seed)  Mode  Cnt    Score   Error  Units
VectorReduction.NoSuperword.mulRedD        512       0  avgt    4  435.897 ± 0.137  ns/op
VectorReduction.WithSuperword.mulRedD      512       0  avgt    4  405.479 ± 1.896  ns/op

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

PR Comment: https://git.openjdk.org/jdk/pull/14065#issuecomment-1569336052