The writeup suggests this implementation is optimized for the not-locked case: > uses an optimistic CAS (compare and swap) immediately, so that locking happens

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