3 ms·

A fair amount of compiler optimizations already work like this; I'm thinking in particular of LLVM IR passes.