By avoiding redundant alignment+alloc checks (startVector does prep) and virtual calls (by making the class final) in inner loops.