Arg ordering on the test path method handle for GwT

[Rémi Forax]

On 01/13/2011 07:02 PM, Mark Roos wrote:
>
> > From:
> > Rémi Forax <forax at univ-mlv.fr>
>
> > Also it's better to use directly Java object instead of your own object,
> > e.g. use Object instead of RtObject. If RtObject has methods, you can
> > refactor them has static method and use invokedynamic to call them on
> > a plain old Object.
> > Of course, this doesn't work if you have states in your object.
>
> I have wondered (worried) about that and its impact on performance.  I do
> need to have state but the state is the pointer to the array of method 
> tables
> for that Smalltalk class.  I have about 2500 classes and one thought 
> would be
> to just create 2500 java mirro classes with no methods just the method 
> table instance var.

At least it will be easier to debug with a Java Debugger.

>
> But it seems to me that then my methods will have type conversions 
> everywhere
> effectively making everything invokeGeneric.  So its not clear what I 
> gain by
> having the Java versions of each class.
>
> My test method then would be:
>         convertArgs dropArgs testMH  vs just dropArgs testMH
> And may fast path
>         invokeGeneric doitMH  vs invokeExact doitMH
> So are convertArgs and invokeGeneric expensive vs nothing and 
> invokeExact?

I don't follow you here.
If you convert your MH to a full Object signature before putting them in 
the vtable,
then you can use invokeExact. Or if you don't convert them you have to 
use invokeGeneric
which will do the conversion at each call.

>
> The cost of my current approach may be with integers and floats.  I 
> suspect that
> hot spot knows how to deal with these objects but in my case they are 
> boxed again.
> But I can't extend them to add the ST method array pointer.  It would 
> be the
> same if java was to support ints as a variant of a object pointer. 
>  This would
> require that I test at each callsite for the shape of the object and 
> branch
> to the integer/float handler before testing for the class.In this case 
> each
> callsite would have a required gwt for each special type before it 
> went to the
> normal types.

Correct me if i'm wrong but Smalltalk has no overloading for method,
only for operators. So you need to handle them in a special way
that may use several gwts.

>
> I think all this can be done I just don't know enough about hot spot 
> to make
> the trade-offs yet.
>
> thanks for listening
> mark
>

Rémi
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/mlvm-dev/attachments/20110114/b4909115/attachment.html