It's not "overhead from sandboxing" that's a problem, it's the fact that translating from bytecode to bytecode to bytecode will cause it to lose some ... succinctness at each step, giving the final machine code generator a harder time coming up with efficient code for it.