speed of invokeExact

[Christian Thalinger]

On May 9, 2013, at 2:26 AM, Jochen Theodorou <blackdrag at gmx.org> wrote:

> Am 08.05.2013 23:12, schrieb John Rose:
>> On May 8, 2013, at 5:33 AM, Jochen Theodorou <blackdrag at GMX.ORG> wrote:
>> 
>>> Am 07.05.2013 19:31, schrieb Christian Thalinger:
>>> [...]
>>>> Do you have any numbers?The problem is that if the MH is not constant
>>>> we can't do any inlining  and it will be an out-of-line call (with a trampoline in between). Is
>>>> your DMH a static or virtual?
>>> 
>>> arg... looks like I made a mistake. It is not the invokeExact that is
>>> slow, it is my preparation of the handle that is slow.
>> 
>> So it's an insertArgs and asType that's slow.  Can you distill the operations into a micro-benchmark for us?
> 
> to reduce it to only asType:
> 
> compare the times of this:
> 
>>        MethodHandle mh = MethodHandles.lookup().findVirtual(String.class, "toString", MethodType.methodType(String.class));
>>        int tmax = 1_000_000;
>>        for (int k=0; k<20; k++) {
>>            long t1 = System.nanoTime();
>>            for (int i=0; i<tmax; i++) {
>>                String s = (String) mh.invokeExact(str);
>>            }
>>            long t2 = System.nanoTime();
>>            System.out.println((t2-t1)/tmax);
>>        }
> 
> with this:
> 
>>        MethodHandle mh = MethodHandles.lookup().findVirtual(String.class, "toString", MethodType.methodType(String.class));
>>        MethodType invokeType = MethodType.methodType(Object.class, String.class);
>>        int tmax = 1_000_000;
>>        for (int k=0; k<20; k++) {
>>            long t1 = System.nanoTime();
>>            for (int i=0; i<tmax; i++) {
>>                MethodHandle invoker = mh.asType(invokeType);
>>                Object o = invoker.invokeExact(str);
>>            }
>>            long t2 = System.nanoTime();
>>            System.out.println((t2-t1)/tmax);
>>        }
> 
> The first one has times of about 9-15 on my machine, the later one 
> 800-900. I am aware that this is a test in which hotspot should throw 
> out the invocation completely, but it does not.

That's because your method handle is not constant and so the compiler cannot inline the call.

-- Chris

> Even more interesing is 
> when you reduce the test even further:
> 
>>        MethodHandle mh = MethodHandles.lookup().findVirtual(String.class, "toString", MethodType.methodType(String.class));
>>        MethodType invokeType = MethodType.methodType(Object.class, String.class);
>>        int tmax = 1_000_000;
>>        for (int k=0; k<20; k++) {
>>            long t1 = System.nanoTime();
>>            for (int i=0; i<tmax; i++) {
>>                MethodHandle invoker = mh.asType(invokeType);
>>            }
>>            long t2 = System.nanoTime();
>>            System.out.println((t2-t1)/tmax);
>>        }
> 
> Now this should really be optimized away completely by hotspot, but is 
> not and takes about the same time as the second case... which is not 
> surprising since the asType call is totally dominating the "benchmark".
> 
> btw, the test was with a jdk 1.8.0 ea with tiered compilation turned 
> off. The tests contain no warmup, but that's what why the test is 
> repeated 20 times to see the development of the values instead of doing 
> a one time measurement. And the values in this test go up and down a lot too
> 
> bye blackdrag
> 
> -- 
> Jochen "blackdrag" Theodorou - Groovy Project Tech Lead
> blog: http://blackdragsview.blogspot.com/
> german groovy discussion newsgroup: de.comp.lang.misc
> For Groovy programming sources visit http://groovy-lang.org
> 
> _______________________________________________
> mlvm-dev mailing list
> mlvm-dev at openjdk.java.net
> http://mail.openjdk.java.net/mailman/listinfo/mlvm-dev