I do too care. It's just that I'm not intelligent/experienced enough to understand what you're doing in a way that would allow me to help. I have very little experience with pl design, and even less with compiler/vm implementation (read, none). However, I'm very interested in how it turns out in case I could learn something and especially because I want to use it when you're finished.
So please, keep working on it, and updating us about the changes. And it's not like there's anything else going on around here for you to distract us from anyway.
Well, this is my first time hacking together a VM anyway ^^. And arc2c is the first time I've hacked on a compiler, and I didn't even start arc2c.
Here are some interesting bits and pieces about VM's that I got via google trawl:
http://www.sidhe.org/~dan/blog/ - Squawks of the Parrot, a blog made by one of the first developers of Parrot VM. SNAP does things very differently from Parrot, partly because of its shared-nothing constraints, partly because this is a lisplike, and mostly because the Parrot VM is just something ^^.
http://blog.mozilla.com/dmandelin/2008/06/03/squirrelfish/ - about the squirrelfish VM. This is the main reason why I decided to use bytecode instead of AST traversal, and why I made a bit of an effort to reduce dependence on the stack (by introducing the -local-push and -clos-push variants of some bytecodes). It also has a bit about direct threading, which is portably implementable only on GNU compilers, and is the fastest portable implementation of bytecode interpreters (faster methods require assembly coding, or at least (in the case of dynamic inlining) some knowledge of what type of code can be copied without modification on what type of processor).
http://research.sun.com/self/papers/papers.html - a bunch of papers about Self, a smalltalk-like language and its implementation. I hear tell it's the most powerful VM on the planet, and its results are what is fueling Sun's recent JVM optimizations. The page itself does not contain the papers, but you can search for their titles and have a good chance of finding them online (or at least, when citeseer is up ^^)
And some more discussion:
Dynamic inlining is a sort of "poor man's JIT". Instead of generating a bytecode sequence as a vector of pointers-to-labels (as in direct threading), we directly copy the bytecode's implementation into the stream.
JIT is still the fastest way to implement an interpreter.
