deeply nested MethodHandles, performance?

[Garcia Gutierrez Miguel Alfredo]

Hi,

I've been experimenting with MethodHandles for runtime code generation [1] by combining MHs as one would build an AST, e.g. using the "application" combinator :

    public static MethodHandle application(MethodHandle target, MethodHandle[] args) {
        MethodHandle[] invokers = new MethodHandle[args.length];
        for(int idx = 0; idx < args.length; idx++) {
            invokers[idx] = exactInvoker(args[idx].type());
        }
        MethodHandle applied = filterArguments(target, 0, invokers);
        for(int idx = 0; idx < args.length; idx++) {
            applied = applied.bindTo(args[idx]);
        }
        return applied;
    }

Similarly for "assignment", "block expression", and the rest of control flow.

With those building blocks, one can generate, say, a loop to add all numbers in an array [2] (which might not be the most efficient way to compute a sum-total) but which begs the question: How much of simplification can HotSpot perform on those MH trees ?

Miguel
http://lampwww.epfl.ch/~magarcia/ScalaCompilerCornerReloaded<http://lampwww.epfl.ch/%7Emagarcia/ScalaCompilerCornerReloaded>


References
----------

[1] http://lamp.epfl.ch/~magarcia/ScalaCompilerCornerReloaded/2012Q2/RuntimeMP.pdf<http://lamp.epfl.ch/%7Emagarcia/ScalaCompilerCornerReloaded/2012Q2/RuntimeMP.pdf>

[2] https://github.com/magarciaEPFL/MethodHandleUtils/blob/master/src/bt4/SummingUp.java

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/hotspot-dev/attachments/20120518/619259c6/attachment.html