Comment by Joker_vD

So, binary translation at instruction-level granularity? I imagine widening the scope to basic blocks would help with quality of the generated code, but I suppose 2/3 of native speed is quite good already.