That describes the 'token' use case, right? While MMX operands let the clib devs play with 16-byte/dword/word units, we are principally looking at very short strings. As soon as you do a 16 byte compare w/delimiting the null byte, your optimization is lost.
I think we are of one mind on this, sniping aside. I started with an svn cp today from asf subversion, and chose to focus on only the svn_cstring_ (excluding svn_string and svn_stringbuf ops), but there is room if we give the nod and should treat them as 3 seperate groupings. First commit inbound in the morning with lots of room for optimization.