I wonder if the comparisons with Cranelift would still show the same build-time/run-time gap four years later.
From what I understand, Cranelift's current instruction selector is very close to copy-and-patch conceptually (it maps patterns of bytecode instructions to sequences of machine instructions), and the pass Cranelift spends the most time in, register allocation, still has to happen in copy-and-patch.
So if copy-and-patch still has a meaningful advantage over Cranelift, I'm very curious to know where it comes from.
From what I understand, Cranelift's current instruction selector is very close to copy-and-patch conceptually (it maps patterns of bytecode instructions to sequences of machine instructions), and the pass Cranelift spends the most time in, register allocation, still has to happen in copy-and-patch.
So if copy-and-patch still has a meaningful advantage over Cranelift, I'm very curious to know where it comes from.