Only update peak costs on the first parse pass

On the second parse pass we can reuse the initial data. This
saves ~200ms of parse time in my test case, and brings the time
down to ~1.6s from 1.8s.