RFR: 8264945: Optimize the code-gen for Math.pow(x, 0.5)

[Jie Fu]

On Fri, 9 Apr 2021 22:39:44 GMT, Vladimir Kozlov <kvn at openjdk.org> wrote:

>> Hi all,
>> 
>> I'd like to optimize the code-gen for Math.pow(x, 0.5).
>> And 7x ~ 14x performance improvement is observed by the jmh micro-benchmarks.
>> 
>> While I was optimizing a machine learning program, I found both Math.pow(x, 2) and Math.pow(x, 0.5) are used.
>> To my surprise, C2 just optimizes the case for Math.pow(x, 2) [1], but still not for Math.pow(x, 0.5) yet.
>> 
>> The patch just replace Math.pow(x, 0.5) with Math.sqrt(x).
>> 
>> Before:
>> 
>> Benchmark                     (seed)   Mode  Cnt      Score    Error   Units
>> MathBench.powDouble0Dot5           0  thrpt    8  45525.117 ? 11.686  ops/ms
>> MathBench.powDouble0Dot5Loop       0  thrpt    8      0.031 ?  0.001  ops/ms
>> 
>> Benchmark                     (seed)   Mode  Cnt      Score    Error   Units
>> MathBench.powDouble0Dot5           0  thrpt    8  45509.317 ?  6.581  ops/ms
>> MathBench.powDouble0Dot5Loop       0  thrpt    8      0.031 ?  0.001  ops/ms
>> 
>> 
>> After:
>> 
>> Benchmark                     (seed)   Mode  Cnt       Score     Error   Units
>> MathBench.powDouble0Dot5           0  thrpt    8  343354.892 ? 362.900  ops/ms
>> MathBench.powDouble0Dot5Loop       0  thrpt    8       0.457 ?   0.001  ops/mso
>> 
>> Benchmark                     (seed)   Mode  Cnt       Score    Error   Units
>> MathBench.powDouble0Dot5           0  thrpt    8  343421.559 ? 49.326  ops/ms
>> MathBench.powDouble0Dot5Loop       0  thrpt    8       0.457 ?  0.001  ops/ms
>> 
>> 
>> Testing:
>>   - tier1~3 on Linux/x64
>> 
>> Thanks,
>> Best regards,
>> Jie
>> 
>> [1] https://github.com/openjdk/jdk/blob/master/src/hotspot/share/opto/library_call.cpp#L1680
>
> Please, verify that result is the same when run with -Xint (Interpreter only) and (-XX:TieredStopAtLevel=1) C1 only. May be they need the same optimization.

> Hi @vnkozlov @neliasso and @huishi-hs ,
> 
> Thanks for your review and comments.
> While I was implementing the opt for C1 and interpreter, I found Math.pow(x, 0.5) and Math.sqrt(x) would compute different values for x={-0.0, Double.NEGATIVE_INFINITY} [1].
> Let's hold on this issue until we have a conclusion about that question.
> Thanks.
> 
> [1] https://mail.openjdk.java.net/pipermail/core-libs-dev/2021-April/076195.html

Hi all,

According to the discussion [1], we can still perform pow(x, 0.5) => sqrt(x) for x >= 0.0 or x is +Inf/NaN.

To better help the code review, the whole optimization has been split into JDK-8265325 and JDK-8264945.
 - 1) JDK-8265325: Optimize StubRoutines::dpow() for Math.pow(x, 0.5)
 - 2) JDK-8264945: Optimize the code-gen for Math.pow(x, 0.5)

I'll update this pr once JDK-8265325 is finished since it depends on JDK-8265325.

Thanks.
Best regards,
Jie

[1] https://mail.openjdk.java.net/pipermail/core-libs-dev/2021-April/076220.html

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

PR: https://git.openjdk.java.net/jdk/pull/3404