RFR: 8366333: AArch64: Enhance SVE subword type implementation of vector compress

[erifan]

On Thu, 11 Sep 2025 06:10:59 GMT, Galder Zamarreño <galder at openjdk.org> wrote:

>> The AArch64 SVE and SVE2 architectures lack an instruction suitable for subword-type `compress` operations. Therefore, the current implementation uses the 32-bit SVE `compact` instruction to compress subword types by first widening the high and low parts to 32 bits, compressing them, and then narrowing them back to their original type. Finally, the high and low parts are merged using the `index + tbl` instructions.
>> 
>> This approach is significantly slower compared to architectures with native support. After evaluating all available AArch64 SVE instructions and experimenting with various implementations—such as looping over the active elements, extraction, and insertion—I confirmed that the existing algorithm is optimal given the instruction set. However, there is still room for optimization in the following two aspects:
>> 1. Merging with `index + tbl` is suboptimal due to the high latency of the `index` instruction.
>> 2. For partial subword types, operations to the highest half are unnecessary because those bits are invalid.
>> 
>> This pull request introduces the following changes:
>> 1. Replaces `index + tbl` with the `whilelt + splice` instructions, which offer lower latency and higher throughput.
>> 2. Eliminates unnecessary compress operations for partial subword type cases.
>> 3. For `sve_compress_byte`, one less temporary register is used to alleviate potential register pressure.
>> 
>> Benchmark results demonstrate that these changes significantly improve performance.
>> 
>> Benchmarks on Nvidia Grace machine with 128-bit SVE:
>> 
>> Benchmark	            Unit	Before	 Error	After	 Error	Uplift
>> Byte128Vector.compress	ops/ms	4846.97	 26.23	6638.56	 31.60	1.36
>> Byte64Vector.compress	ops/ms	2447.69	 12.95	7167.68	 34.49	2.92
>> Short128Vector.compress	ops/ms	7174.88	 40.94	8398.45	 9.48	1.17
>> Short64Vector.compress	ops/ms	3618.72	 3.04	8618.22	 10.91	2.38
>> 
>> 
>> This PR was tested on 128-bit, 256-bit, and 512-bit SVE environments, and all tests passed.
>
> src/hotspot/cpu/aarch64/c2_MacroAssembler_aarch64.cpp line 2292:
> 
>> 2290:   // Return if the vector length is no more than MaxVectorSize/2, since the
>> 2291:   // highest half is invalid.
>> 2292:   if (vector_length_in_bytes <= (MaxVectorSize >> 1)) {
> 
> Couldn't this check be done first thing when the function is called? Then you would avoid unnecessary work?
> 
> I also wonder if this check should be done before `sve_compress_byte` is called, but I think at the very least it should be done first thing in this function.

We need to do the lower half, so I think there's no unnecessary work.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/27188#discussion_r2347805971