unfortunately i can't just use cranelift's regalloc2 because i'm doing copy&patch, and the register windowing restriction kind of nixes that. plus requiring exact register windows at each basic block for stitching regions together at later tier-ups, which i think it would also handle poorly
XFEDI
ActivityPubRemote profile
Posts
rewrote my register allocator and the code it outputs is twice as slow 👍
pretty sure im at the point i have to write a proper SSA IR backend for my JIT 😢