Having the target on a 16 byte boundry is very important for performance reasons on a non-P4 (or even a P4/P4 Xeon if the target is not in the uOP cache). The reason is the IFetch buffers are loaded directly from tag lines, and thus, if they are not aligned, you are paying for fluff (i.e. bytes will be loaded that are never used).

As xMaster said, in an ideal world, 64 bytes would be even better, but could you imagine 63 bytes-o-NOPs between single ret functions? Only Microsoft... (I kid, actually their compilers arent that bad).