The writeup suggests this implementation is optimized for the not-locked case:

> uses an optimistic CAS (compare and swap) immediately, so that locking happens quickly when there's no contention