Duby dynamic dispatch has landed!

[Attila Szegedi]

On 2010.03.25., at 3:17, Charles Oliver Nutter wrote:

> For Attila: I had to remove a spreadArguments handle you used for
> re-binding the method...not sure why. Here's the diff:

Turns out, now the MethodHandle.invokeVarargs() actually does the whole convert-to-generic-and-invoke-as-vararg, so I committed the definitive solution for it now, which is simply:

Index: DynamicLinkerImpl.java
===================================================================
--- DynamicLinkerImpl.java	(revision 233)
+++ DynamicLinkerImpl.java	(revision 228)
@@ -111,6 +111,12 @@
         // Invoke the method. Note we bypass the guard, as the assumption is 
         // that the current arguments will pass the guard (and there actually
         // might be no guard at all).
-        return guardedInvocation.getInvocation().invokeVarargs(arguments);
+        final MethodHandle invocation = guardedInvocation.getInvocation();
+        final MethodType genericType = invocation.type().generic();
+        final MethodHandle genericizedInvocation = 
+            MethodHandles.convertArguments(invocation, genericType);
+        final MethodHandle spreadInvocation = MethodHandles.spreadArguments(
+                genericizedInvocation, SPREAD_GENERIC_INVOCATION);
+        return MethodHandles.invoke(spreadInvocation, arguments);
     }
 }

(Mind you, this is a reverse diff, but I'm too sleepy now to fix it; just swap + and -)

I think the problem was in the fact that the (now deprecated) MethodHandles.invoke() used to behave as "invokeExact" in the last August's builds (so that's why I had to genericize + spread explicitly), while now it maps to MethodHandle.invokeVarargs(), so in addition to it being deprecated, its behaviour was also changed in the past half a year and that broke things, as you have yourself experienced :-).

Should be good now.

BTW, Stephen Bannasch's MLVM build works completely okay - I abandoned the attempt to now build my own as it didn't work out quickly and I didn't want to waste a lot of time; his version + Java sources from Mercurial seem sufficient for debugging.

I'm still not back to the old 6 unit test errors, but this patch reduced them from 18 to 8; seems there are still two genuine new issues.
 
Attila.

> 
> Index: src/org/dynalang/dynalink/support/DynamicLinkerImpl.java
> ===================================================================
> --- src/org/dynalang/dynalink/support/DynamicLinkerImpl.java	(revision 232)
> +++ src/org/dynalang/dynalink/support/DynamicLinkerImpl.java	(working copy)
> @@ -115,8 +115,6 @@
>         final MethodType genericType = invocation.type().generic();
>         final MethodHandle genericizedInvocation =
>             MethodHandles.convertArguments(invocation, genericType);
> -        final MethodHandle spreadInvocation = MethodHandles.spreadArguments(
> -                genericizedInvocation, SPREAD_GENERIC_INVOCATION);
> -        return MethodHandles.invoke(spreadInvocation, arguments);
> +        return MethodHandles.invoke(genericizedInvocation, arguments);
>     }
> }
> 
> I committed a built version with this hack...no tests, etc yet for
> Duby's "dynamic" type, but that will come soon.
> 
> BTW, what's the current state of the art for emitting .java with an
> invokedynamic in it? Duby also has a .java backend, so I'll need to
> add indy support there as well (somehow).
> 
> - Charlie