RFR: 8256438: AArch64: Implement match rules with ROR shift register value

[Eric Liu]

Hi Andrew,

Thanks for your feedback.

I refined those benchmarks and on two platforms could have a better performance. 
One can get about 100% gain, and 30% for another one. Other platforms were hard to 
see any performance changes.

Before:
  Benchmark                Mode  Cnt     Score   Error  Units
  Rotation.andRotateRight  avgt   15  3860.994 ± 3.409  ns/op
  Rotation.bicRotateRight  avgt   15  3861.247 ± 3.321  ns/op
  Rotation.eonRotateRight  avgt   15  3860.865 ± 3.003  ns/op
  Rotation.ornRotateRight  avgt   15  3860.884 ± 3.260  ns/op
  Rotation.xorRotateRight  avgt   15  3860.886 ± 2.728  ns/op

After:
  Benchmark                 Mode  Cnt     Score    Error  Units
  Rotation.andRotateRight   avgt   15  1933.495 ± 0.263  ns/op
  Rotation.bicRotateRight   avgt   15  1933.436 ± 0.244  ns/op
  Rotation.eonRotateRight   avgt   15  1933.459 ± 0.255  ns/op
  Rotation.ornRotateRight   avgt   15  1933.559 ± 0.316  ns/op
  Rotation.xorRotateRight   avgt   15  1933.467 ± 0.245  ns/op

I would update my patch with the benchmark tests after finishing the whole tests.

--Eric

-----Original Message-----
From: hotspot-compiler-dev <hotspot-compiler-dev-retn at openjdk.java.net> On Behalf Of Andrew Haley
Sent: Saturday, January 30, 2021 12:30 AM
To: hotspot-compiler-dev at openjdk.java.net
Subject: Re: RFR: 8256438: AArch64: Implement match rules with ROR shift register value

On 29/01/2021 10:11, Andrew Haley wrote:
> On 1/29/21 8:35 AM, Eric Liu wrote:
>> I benchmarked on 7 platforms with jmh test[1], most of them were hard to see any performance changes. On one platform the performance become bad, and seems a little unstable.

Your benchmark did not work for me. It did not generate the correct instructions.

Please try with this or similar:

    @Benchmark
    public void xorRotateRight(MyState s, Blackhole blackhole) {
        int x = s.xi;
        int y = s.yi;
        for (int i = 0; i < COUNT; i++) {
            y = x ^ ((y >>> 5) | (y << -5));
            x = y ^ ((x >>> 5) | (x << -5));
        }
        blackhole.consume(x);
    }

I get:

Benchmark                             Mode  Cnt     Score    Error  Units
Rotation.xorRotateRight   (before)    avgt    3  6142.575 ± 15.940  ns/op
Rotation.xorRotateRight   (after)     avgt    3  4081.587 ± 33.904  ns/op

Please integrate the corrected benchmark into your patch.

--
Andrew Haley  (he/him)
Java Platform Lead Engineer
Red Hat UK Ltd. <https://www.redhat.com> https://keybase.io/andrewhaley
EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671