Conditional moves vs. branching in unrolled loops

[Vitaly Davidovich]

This is a known issue: https://bugs.openjdk.java.net/browse/JDK-8039104

On Tuesday, January 5, 2016, Paul Sandoz <paul.sandoz at oracle.com> wrote:

> Hi,
>
> Recent investigation comparing for loops with streams exposed what appears
> to be an issue with Math.max and generated code in unrolled loops.
>
> Namely this:
>
>     @Benchmark
>     public int forTest_if() {
>         int[] a = ints;
>         int e = ints.length;
>         int m = Integer.MIN_VALUE;
>         for (int i = 0; i < e; i++)
>             if (a[i] >= m)
>                 m = a[i];
>         return m;
>     }
>
> is faster than this:
>
>     @Benchmark
>     public int forTest_MathMax() {
>         int[] a = ints;
>         int e = ints.length;
>         int m = Integer.MIN_VALUE;
>         for (int i = 0; i < e; i++)
>             m = Math.max(m, a[i]);
>         return m;
>     }
>
> Or this:
>
>   Arrays.stream(ints).reduce(Integer.MIN_VALUE, (a, b) -> a >= b ? a : b);
>
> is faster than this:
>
>   Arrays.stream(ints).reduce(Integer.MIN_VALUE, Math::max);
>
> at least on an x86 i5 processor.
>
> See the following links for more details:
>
>   https://bugs.openjdk.java.net/browse/JDK-8146071
>
> https://bugs.openjdk.java.net/browse/JDK-8146071?focusedCommentId=13883495&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-13883495
>
> For generated code in the for loop cases above see:
>
>
> https://bugs.openjdk.java.net/secure/attachment/56221/mathMax.perfasm.txt
>
> I am not familiar enough with the x86 architecture to fully explain why,
> but i presume branch prediction is trumping the conditional moves, which
> suggests that on certain processors the generated code for the Math.max
> intrinsic (and others) in unrolled loops should not use conditional moves.
>
> Thanks,
> Paul.
>


-- 
Sent from my phone
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/attachments/20160105/ac00a1c6/attachment.html>