Review request for JDK-8007060

[Hannes Wallnoefer]

I did some octane testing with this patch today and didn't note any 
performance degradations. I also noticed that only one octane test uses 
non-builtin methods on primitives at all, and that is earley-boyer.

Based on this I uploaded a slightly changed patch to 
http://cr.openjdk.java.net/~hannesw/8007060/:

1) Adding the wrapFilter is now a one-way, i.e. once added it stays 
there and also handles object this-arguments (it did that already 
anyway). This makes NonStrictFunctionGuard simpler because we only have 
to check whether this-argument is ScriptObject (we install an ordinary 
FunctionGuard when the wrap filter is used).

2) I moved primitive wrapping from ScriptRuntime.apply() to 
ScriptFunction.invoke(). It really belongs there and also makes related 
code a bit nicer.

Hannes

Am 2013-01-30 14:20, schrieb Hannes Wallnoefer:
> Please review http://cr.openjdk.java.net/~hannesw/8007060/
>
> This removes the NashornInvocationTarget class whose purpose was to 
> track primitiveness of callsite targets and replaces it in two ways.
>
> 1) For functions invoked via ScriptRuntime.apply() (this includes 
> ordinary megamorphic callsites) we simply check function strictness 
> and this-argument primitiveness at runtime. We should have done that 
> anyway, I think there were several uses of this method that forgot to 
> wrap primitives when they should have.
>
> 2) For ordinary monomorphic callsites we check function strictness and 
> this-argument primitiveness at link time. If this is primitive we 
> install a wrapper methodhandle as argument filter. We then use a new 
> NonStrictFunctionGuard that in addition to the FunctionGuard check 
> tests whether primitiveness of this-argument has changed by doing one 
> simple instanceof check. (This should be very rare, as invocation of 
> script functions on JS primitives should generally be rare.
>
> I've tested this with test, test262parallel, and octane. Haven't 
> tested node.jar yet. Haven't done any in-depth performance tests but I 
> think there should be no notable regression.
>
> thanks,
> Hannes