Caterpillar

So what about the efficiency of using micro-instructions as undo instructions? Let's get back to our example, the expression 1 2 + which translates to the instructions push 1, push 2, add. We can undo the add with pop, push 1, push 2, the push 2 with a pop, and the push 1 with another pop.

So pop, push 1, push 2, pop, pop. Five micro-instructions to undo three regular ones; pretty much the same as before. This does not seem any more efficient. But remember, there are far fewer micro-instructions than regular instructions. So their encoding can be more compact. Those five wouldn't take up a lot of bytes.

Is this the best approach? I don't know. But it seems like the best bet for an initial attempt, for the implementation simplicity alone.