The new framewalker stuff is all done now. For interpreted frames you need to write all kinds of fiddly little accessors so the garbage collector can find the method pointer, the local variables, the monitors and the expression stack, and any other objects that may be lying around in there, but for compiled frames it’s simple: at any point at which the stack could be walked you just emit a map which says “in a stack frame with such and such a PC, slots 1, 2, 4, 5 and 8 contain pointers to objects”. The tricky bit is the PC: I don’t have access to the real one, so I had to fake one up, but it’s all working now — and surviving garbage collections — which is pretty cool! The garbage collector interface was the single biggest thing I was worried about, so it’s nice to have it under my belt, with all the old hacks removed.
Since finishing the framewalking stuff I’ve also implemented VM calls, which are the places where compiled code drops into C to do things too complicated to want to write in assembly. Making Shark fail gracefully when it hits unknown bytecodes was an amazing idea, as it shifted the focus from the simple grind of implementing bytecodes to the really critical — and interesting! — things. Doing it this way around means I can get all the infrastructure solid, then spend a week or so churning out the remaining ninety or so bytecodes.
In other news, with nearly 150 methods compiled, my simple testcase is now seven times faster with Shark than without