This cuts the average bytes/relative span from 3.5 to 3.2 on libcore, ultimately saving ~400kb of data.