Invokedynamic slower than reflection?

[John Rose]

It was some sort of bitrot.

I pushed a fix for this to mlvm/hotspot.

Thanks for the reports!

-- John

On Jun 5, 2010, at 10:58 AM, Rémi Forax wrote:

> Le 05/06/2010 02:01, John Rose a écrit :
>> Is the call site megamutable?  (Is it linked many times instead of once?)  -- John
>> 
> 
> no !
> The callsite is linked only once.
> 
> Rémi
> 
>> On Jun 4, 2010, at 4:40 AM, Rémi Forax wrote:
>> 
>> 
>>> It's funny, I've found the same error last night.
>>> 
>>> Yes, there is a problem,
>>> if you run with -XX:+PrintCompilation, you will see
>>> lot of "made not entrant" on the same bytecode location.
>>> 
>>> 183%  made not entrant  (2)  Hello::main @ -2 (145 bytes)
>>> 184%      Hello::main @ 6 (145 bytes)
>>> 
>>> It seems that invokedynamic have trouble and
>>> cycle between optimization and deoptimization phases.
>>> 
>>> Rémi
>>> 
>>> Le 04/06/2010 10:43, Eric Bodden a écrit :
>>> 
>>>> Hi all.
>>>> 
>>>> I did some simple micro-benchmarking with the current implementation
>>>> of invokedynamic. In the attached test class, I call a method
>>>> "greeter" 100,000,000 times using invokedynamic and then using
>>>> reflection. Reflection only takes about 1338ms, while invokedynamic
>>>> takes about 12099. (This was taken on OSX 10.6, with build
>>>> "1.7.0-internal-stephen_2010_05_28_19_48-b00"). By the way
>>>> invokedynamic works, I had expected it to be at least as fast as a
>>>> reflective call.
>>>> 
>>>> Can you tell me if this kind of performance is expected in general?
>>>> 
>>>> Best wishes,
>>>> Eric
>>>> 
>>>> --
>>>> Dr. Eric Bodden
>>>> Software Technology Group, Technische Universität Darmstadt, Germany
>>>> Tel: +49 6151 16-5478    Fax: +49 6151 16-5410
>>>> Mailing Address: S2|02 A209, Hochschulstraße 10, 64289 Darmstadt