Why will performance go up when shifting from compilation targeting C to CPS-style compilation to assembly without compiler optimizations implemented? C compilers are crammed full of optimizations, so I would naively think that writing your own code generator is a lose in both complexity and performance.

Among other reasons, because when compiling to assembly you can invent your own, model-appropriate calling conventions.

interesting. as far as i could work out, that implementation is pretty direct, which means that tail calls will consume stack. is that right?

Yes, that's correct.