Lift spill slots to imaginary registers

[mysterymath]

The greedy register allocator is heuristic and imperfect; despite our best efforts, it may still spill values to the stack slot when it could have spilled them to imaginary registers. We should add a pass to lift spill slots to registers after register allocation; this should, in concert with copy propagation, reduce the amount of copying between static stack and imaginary registers.

[gandhi56]

Is there a test that reflects this regression?

[johnwbyrd]

I wouldn't call this a regression. This is an opportunity for improved quality of codegen.

[mysterymath]

I'm just straight up working on a more modern register allocator to replace the one in LLVM. I'm leaving these bugs open for now, since they're the next best option if that falls through. But if it pans out, ill be able to close a half dozen or so bugs of this kind, and remove a lot of ugliness in the backend to boot.