Compiling multidimensional arrays

[Charles Turner <chturne at gmail dot com>]

I've spent today trying to get multidimensional array
subscripting/generalised setting to work. It works from the REPL, but
not when compiling. That is, I've fixed the various ArrayGet/ArraySet
reflectors, ApplyToArgs, SetArray and various compilation helpers, but
I'm currently having fun trying to fix the code in CompileArrays.

Currently, CompileArrays only handles 1d arrays correctly. I want to
make it handle nd arrays. It seems fairly straightforward: just emit
N-1 aaload instructions, followed by the appropriate *aload
instruction to retrieve the value. I'm still swimming around in
CodeAttr getting myself familiarised with the code generator, but I
noticed something very weird in CodeAttr#emitNewArray.

The branching order is an instanceof ObjectType test, and if that
fails, an instanceof ArrayType test. What? ArrayType is a subclass of
ObjectType! How can that branch be taken? Because of that, it seems
Kawa will never emit the multianewarray instruction even if it
correctly handled md arrays in the first place. In r1378, Per made the
following comment:

(emitNewArray):  Use anewarray even if element type is ArrayType.

I assume that's because md arrays don't work anyway, so it makes sense
to use the faster anewarray when creating arrays of one dimension, but
why keep the old instanceof ArrayType branch around and not even
comment why it's there? Am I missing something?

I'm going to take a crack at implementing my above scheme after some
rest. I'll special case the 1d array to use anewarray of course. I'm
also special casing 1d arrays in the reflectors and friends because I
measured a 10x time speedup as opposed to the way you get at arbitrary
arrays using java.lang.reflect.Array#get and a loop. Not sure if it's
worth special casing 2d arrays as well, how much do people use them? I
often find myself simulating them on a vector anyway :/