If it is always stored in a decomposed form, an XML library probably wouldn't want to use it, because it requires a composed form. And making the encoding an implementation detail makes it inefficient to use in situations where binary compatibility matters (serialization, for example).
I think the best solution is to store the string in the form it was originally recieved (decomposed or not), and instead provide composition functions or even iterator wrappers that compose on the fly. That would allow for composed strings to be used if needed (like in a XML library, but not imposing that requirement on all other users.
I don't think I can agree on that. If you do a lot of input/output, this might yield a better performance, but even in reading XML, you probably need to compare strings a lot, and if they are not normalised, this will really take a lot of processing. Correct me if I'm wrong, but a simple comparison of two non-normalized Unicode strings would take looking up the characters in the Unicode Character Database, decomposing every single character, gathering base characters and combining marks, and ordering the marks, then comparing them. And this must be done for every character. I don't have any numbers, of course, but I have this feeling it is going to be really really slow. Regards, Rogier