Since they're all pretty simple (except the spill code..) merge in a bunch of performance improvements from the FrostByte branch,
for an about 15% speedup w/g++-4.3, and abour 10% or so with g++-4.2. The more aggressive
stuff on my TODO list will probably have to wait till FrostByte -42, though.

- Tweak inlining setup to work better with g++-4.3
- Tweak the generated VP dispatch/decode loop to have smaller & faster code, and to have
less unneeded specializations
- Avoid some can-fit checks when boxing things we know can't be smallint'd
- Streamline/simplify the function call code somewhat.
- Specialized < op for comparisons with numberic upper bound
- Do the string value -> string object conversion directly when legal
- Make the string [] accessor share substring data, avoding some allocations
- Remove a double-lookup in hashtable due to ReadOnly check in JSObject::put
- Do tearoffs deferred only