They separate the problem into two phases: the first is only worried about correctness, then the second deals with optimization while (eventually) preserving correctness.