RFR: 8294194: [AArch64] Create intrinsics compress and expand [v9]

[Andrew Haley]

On Mon, 16 Jan 2023 14:19:17 GMT, Stuart Monteith <smonteith at openjdk.org> wrote:

>> The java.lang.Long and java.lang.Integer classes have the methods "compress(i, mask)" and "expand(i, mask)". They compile down to 236 assembler instructions. There are no scalar instructions that perform the equivalent functions on aarch64, instead the intrinsics can be implemented with vector instructions included in SVE2; expand with BDEP, compress with BEXT.
>> 
>> Only the first lane of each vector will be used, two MOV instructions will move the inputs from GPRs into temporary vector registers, and another to do the reverse for the result. Autovectorization for this functionality is/will be implemented separately.
>> 
>> Running on an SVE2 enabled system, I ran the following benchmarks:
>> 
>>         org.openjdk.bench.java.lang.Integers
>>         org.openjdk.bench.java.lang.Longs
>> 
>> The time for each operation reduced to 56% to 72% of the original run time:
>> 
>> 
>> Benchmark               Result  error   Unit    % against non-SVE2
>> Integers.expand         2.106   0.011   us/op
>> Integers.expand-SVE     1.431   0.009   us/op   67.95%
>> Longs.expand            2.606   0.006   us/op
>> Longs.expand-SVE        1.46    0.003   us/op   56.02%
>> Integers.compress       1.982   0.004   us/op
>> Integers.compress-SVE   1.427   0.003   us/op   72.00%
>> Longs.compress          2.501   0.002   us/op
>> Longs.compress-SVE      1.441   0.003   us/op   57.62%
>> 
>> 
>> These methods can bed  specifically tested with:
>> `make test TEST="jtreg:compiler/intrinsics/TestBitShuffleOpers.java"`
>
> Stuart Monteith has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Add pattern for for memory operands
>   
>   Adds new patterns to match a memory address and a constant, which are
>   emitted as two loads. C2 expects that constants INTs will always be
>   manifested by immediate loads. However, as these intrinsics are avoiding
>   loading into GPRs and moving them into the vector registers, there is an
>   addition to allow integers to be loaded as a constant.
>   
>   There isn't a pattern to match two LoadI nodes - C2 doesn't handle nodes
>   with two memory nodes for calculating anti-dependencies.
>   
>   Benchmark                  Result  Units  % against non-SVE2
>   Integers.compress          2.009   µs/op
>   Integers.compress-SVE      1.435   µs/op  71.43%
>   Integers.compress-SVE+mem  1.263   µs/op  62.87%
>   Integers.expand            2.129   µs/op
>   Integers.expand-SVE        1.433   µs/op  67.31%
>   Integers.expand-SVE+mem    1.32    µs/op  62.00%
>   Longs.compress             2.504   µs/op
>   Longs.compress-SVE         1.445   µs/op  57.71%
>   Longs.compress-SVE+mem     1.269   µs/op  50.68%
>   Longs.expand               2.614   µs/op
>   Longs.expand-SVE           1.489   µs/op  56.96%
>   Longs.expand-SVE+mem       1.272   µs/op  48.66%
>   
>   Change-Id: I6b78d2fcc11cd00fb6d14e9f9456c0cce55dd9d4
>   CustomizedGitHooks: yes

Marked as reviewed by aph (Reviewer).

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

PR: https://git.openjdk.org/jdk/pull/10537