Thanks Aleksey, for the usual convincing benchmark result. You win!
I can imagine a machine that has a native cas instruction, but no native
exchange instruction, but no one ever got fired for optimizing for x86.
public final boolean getAndSet(boolean newValue) {
return (int)VALUE.getAndSet(this, (newValue ? 1 : 0)) != 0;
}
On Thu, Aug 11, 2016 at 1:08 PM, Aleksey Shipilev <
aleksey.shipi...@gmail.com> wrote:
> On 08/11/2016 09:59 PM, Martin Buchholz wrote:
> > On Thu, Aug 11, 2016 at 10:40 AM, Aleksey Shipilev
> >> Why doesn't AtomicBoolean.getAndSet use VarHandle.getAndSet, which is
> >> hopefully intrinsified completely?
> >
> > Good question. This optimization originated in a suggestion for jdk8,
> > where there are no VarHandles.
> > VarHandle.getAndSet does not know there are only two values for the int,
> > so it is not obvious which one wins. I can imagine getAndSet being
> > compiled to a weak cas loop.
>
> getAndSet/getAndAdd have a nice rationale for x86: they compile to
> xchg/xadd, respectively. This is what I meant about "intrinsified
> completely".
>
> >> Any benchmarks backing up this optimization?
> >
> > I measured a reduction in bytecode! Does that count?
>
> I measured x86 instructions count, and you would not believe what it
> told me!
> http://cr.openjdk.java.net/~shade/8162805/BooleanGetAndSet.java
>
> Thanks,
> -Aleksey
>
>
