RFR: 8188044: We need Math.unsignedMultiplyHigh [v2]

[Andrew Haley]

On Fri, 2 Jul 2021 13:47:40 GMT, Andrew Haley <aph at openjdk.org> wrote:

>> You can also do that branchlessly which might prove better
>> 
>>          long result = Math.multiplyHigh(x, y);
>>          result += (y & (x >> 63));
>>          result += (x & (y >> 63));
>>          return result;
>
>> You can also do that branchlessly which might prove better
>> 
>> ```
>>      long result = Math.multiplyHigh(x, y);
>>      result += (y & (x >> 63));
>>      result += (x & (y >> 63));
>>      return result;
>> ```
> I doubt very much that it would be better, because these days branch prediction is excellent, and we also have conditional select instructions. Exposing the condition helps C2 to eliminate it if the range of args is known. The `if` code is easier to understand.
> 
> Benchmark results, with one of the operands changing signs every iteration, 1000 iterations:
> 
> 
> Benchmark                  Mode  Cnt     Score    Error  Units
> MulHiTest.mulHiTest1   (aph)     avgt    3  1570.587 ± 16.602  ns/op
> MulHiTest.mulHiTest2   (adinn)   avgt    3  2237.637 ±  4.740  ns/op
> 
> In any case, note that with this optimization the unsigned mulHi is in the nanosecond range, so Good Enough. IMO.

But weirdly, it's the other way around on AArch64, but there's little in it:


Benchmark             Mode  Cnt     Score   Error  Units
MulHiTest.mulHiTest1  avgt    3  1492.108 ± 0.301  ns/op
MulHiTest.mulHiTest2  avgt    3  1219.521 ± 1.516  ns/op


but this is only in the case where we have unpredictable branches. Go with simple and easy to understand; it doesn't much matter.

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

PR: https://git.openjdk.java.net/jdk/pull/4644