A simple optimization proposal

[Krystal Mok]

Hi John,

Nice to hear from you!

I ran a couple of tests on the patch, one of them is:

public class ArrayRangeCheck {
  private static Object foo(Object[] a, int x) throws Exception {
    if (a.length == 0) throw new Exception();
    return a[x & (a.length - 1)];
  }

  public static void main(String[] args) throws Exception {
    Object[] a = new Object[8];
    foo(a, 6);
    foo(a, 6);
    System.in.read();
  }
}

Ran with:

$ java -XX:CompileCommand="compileonly ArrayRangeCheck foo"
-XX:-TieredCompilation -XX:CICompilerCount=1 -XX:CompileThreshold=1
-XX:PrintIdealGraphLevel=4 -XX:PrintIdealGraphFile=ideal.xml
-XX:+PrintCompilation -XX:+PrintAssembly ArrayRangeCheck

And confirmed that in this case the range check is indeed elided: there's
only one a.length != 0 check dominating the actual element load, and the
else branch goes to the exception throwing code.

Thanks,
Kris


On Wed, Feb 12, 2014 at 9:47 PM, John Rose <john.r.rose at oracle.com> wrote:

> One point behind the bug report is to give a little reward to Java coders
> who write dominating tests that exclude a.length==0, as:
>
>  if (a.length == 0) goAway();
>  else return a[i & a.length-1];
>
> Kris, can your patch do this?  The logic in IfNode can probably elide the
> duplicate dominating test, if the right normalizations occur.  ��
>
> – John
>
> On Feb 12, 2014, at 6:05 PM, Martin Grajcar <maaartinus at gmail.com> wrote:
>
> That's what I've meant with goAway. My point was the jump using already
> computed flags
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/attachments/20140213/9def06c7/attachment.html