RFR: 8269725: AArch64: Add VectorMask query implementation for NEON [v2]

[Andrew Haley]

On Thu, 8 Jul 2021 07:27:59 GMT, Xiaohong Gong <xgong at openjdk.org> wrote:

>> The VectorMask query (`trueCount, firstTrue, lastTrue`) APIs can be intrinsified after [1] is closed. This patch adds the Arm NEON backend implementation for the new added vector nodes.
>> 
>> Here is the performance comparison data for the three APIs with and without this patch:
>> 
>> Benchmark                                        (bits) (inputs) Before       After      Gain  Units
>> MaskQueryOperationsBenchmark.testFirstTrueByte    128      1    42583.141   103900.253   2.44  ops/ms
>> MaskQueryOperationsBenchmark.testFirstTrueByte    128      2    37158.470   108234.110   2.91  ops/ms
>> MaskQueryOperationsBenchmark.testFirstTrueByte    128      3    42583.584   108235.231   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testFirstTrueInt     128      1    42583.625   108236.859   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testFirstTrueInt     128      2    42583.288   107368.205   2.52  ops/ms
>> MaskQueryOperationsBenchmark.testFirstTrueInt     128      3    42583.673   108232.371   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testFirstTrueLong    128      1    42583.408   108232.617   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testFirstTrueLong    128      2    42583.443   107367.035   2.52  ops/ms
>> MaskQueryOperationsBenchmark.testFirstTrueLong    128      3    42583.111   108236.036   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testFirstTrueShort   128      1    42583.536   108230.365   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testFirstTrueShort   128      2    41231.639   108239.148   2.62  ops/ms
>> MaskQueryOperationsBenchmark.testFirstTrueShort   128      3    42583.630   108238.542   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testLastTrueByte     128      1    42584.067   108238.989   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testLastTrueByte     128      2    36845.596   108234.297   2.94  ops/ms
>> MaskQueryOperationsBenchmark.testLastTrueByte     128      3    42583.759   108237.501   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testLastTrueInt      128      1    42583.319   108236.218   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testLastTrueInt      128      2    42583.112   108234.516   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testLastTrueInt      128      3    42583.340   108238.777   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testLastTrueLong     128      1    42581.004   108233.701   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testLastTrueLong     128      2    42583.266   108238.323   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testLastTrueLong     128      3    42583.542   108234.327   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testLastTrueShort    128      1    42583.552   108238.011   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testLastTrueShort    128      2    41231.142   108237.919   2.63  ops/ms
>> MaskQueryOperationsBenchmark.testLastTrueShort    128      3    44784.270   108238.011   2.42  ops/ms
>> MaskQueryOperationsBenchmark.testTrueCountByte    128      1    37075.556   108233.571   2.92  ops/ms
>> MaskQueryOperationsBenchmark.testTrueCountByte    128      2    37527.370   108233.396   2.88  ops/ms
>> MaskQueryOperationsBenchmark.testTrueCountByte    128      3    36585.788   107372.032   2.93  ops/ms
>> MaskQueryOperationsBenchmark.testTrueCountInt     128      1    42583.608   108233.721   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testTrueCountInt     128      2    42584.733   107369.578   2.52  ops/ms
>> MaskQueryOperationsBenchmark.testTrueCountInt     128      3    42583.623   107367.859   2.52  ops/ms
>> MaskQueryOperationsBenchmark.testTrueCountLong    128      1    42583.671   107368.004   2.52  ops/ms
>> MaskQueryOperationsBenchmark.testTrueCountLong    128      2    42583.661   108233.301   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testTrueCountLong    128      3    42583.015   108232.783   2.54  ops/ms
>> MaskQueryOperationsBenchmark.testTrueCountShort   128      1    41229.280   108233.369   2.63  ops/ms
>> MaskQueryOperationsBenchmark.testTrueCountShort   128      2    41231.914   107366.904   2.60  ops/ms
>> MaskQueryOperationsBenchmark.testTrueCountShort   128      3    41231.734   108233.606   2.63  ops/ms
>> 
>> All VectorAPI jtreg tests pass with patch [2] is applied together.
>> 
>> [1] https://bugs.openjdk.java.net/browse/JDK-8256973
>> [2] https://github.com/openjdk/jdk17/pull/168
>> 
>> Tested tier1 and jdk:tier3.
>
> Xiaohong Gong has updated the pull request with a new target base due to a merge or a rebase. The pull request now contains two commits:
> 
>  - Merge branch 'jdk:master' into JDK-8269725
>  - 8269725: AArch64: Add VectorMask query implementation for NEON

src/hotspot/cpu/aarch64/aarch64_neon_ad.m4 line 2299:

> 2297:   ins_encode %{
> 2298:     // Revert the bits and count the leading zero bytes.
> 2299:     __ negr(as_FloatRegister($tmp$$reg), __ T8B, as_FloatRegister($src$$reg));

Should that be "Reverse the bits?" But in any case, we can see that the code calls rbit then clz, presumably because you want to count the trailing bits. What does the negr do here?

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

PR: https://git.openjdk.java.net/jdk/pull/4699