RFR: 8267663: AArch64: Add unsigned comparison operators on AArch64
Andrew Haley
aph at openjdk.java.net
Sat Jun 5 08:06:56 UTC 2021
On Sat, 5 Jun 2021 07:59:53 GMT, Andrew Haley <aph at openjdk.org> wrote:
>> This patch implements unsigned vector comparison on AArch64. The
>> performance of unsigned comparison improves about 4x~5x in my local with
>> Byte128Vector.java[1].
>>
>> Before:
>> Benchmark Score(op/ms) Error
>> Byte128Vector.UNSIGNED_GE#size(1024) 99.953 6.17
>> Byte128Vector.UNSIGNED_GT#size(1024) 95.334 8.865
>> Byte128Vector.UNSIGNED_LE#size(1024) 76.908 24.332
>> Byte128Vector.UNSIGNED_LT#size(1024) 78.362 23.507
>>
>> After:
>> Benchmark Score(op/ms) Error
>> Byte128Vector.UNSIGNED_GE#size(1024) 421.809 25.57
>> Byte128Vector.UNSIGNED_GT#size(1024) 420.653 26.779
>> Byte128Vector.UNSIGNED_LE#size(1024) 316.754 92.889
>> Byte128Vector.UNSIGNED_LT#size(1024) 423.683 26.508
>>
>> [Test]
>> - All vector API test cases passed without new failure. 8265312[2] has
>> been implemented this on x86 and supplied sufficient test cases for
>> all kinds of vector.
>> - No performance regression for other comparisons.
>> - libjvm.so drops off about 200KB after this patch by combining those
>> vector compare rules.
>>
>> [1] https://github.com/openjdk/panama-vector/blob/vectorIntrinsics/test/jdk/jdk/incubator/vector/benchmark/src/main/java/benchmark/jdk/incubator/vector/Byte128Vector.java#L1198
>> [2] https://github.com/openjdk/panama-vector/pull/68
>
> src/hotspot/cpu/aarch64/aarch64_neon_ad.m4 line 889:
>
>> 887:
>> 888: static void neon_compare(C2_MacroAssembler masm, FloatRegister dst, BasicType bt,
>> 889: FloatRegister src1, FloatRegister src2, int cond, bool isX) {
>
> Passing masm by value, rather than reference, is unconventional. There's no reason to do that.
Come to think about it, why isn't this an Assembler method?
-------------
PR: https://git.openjdk.java.net/jdk/pull/4358
More information about the hotspot-compiler-dev
mailing list