In this particular case, however, simply making memcpy() handle overlapping moves correctly would not break anything. Well, I suppose there's a theoretical possibility that someone is counting on the old behavior in the backwards-overlapping case, but that would be bizarre; surely this is the kind of code that should get broken, if any of it even exists.

If memcpy() had been fixed 30 years ago to do overlapping moves correctly, as it could and should have been, that would have been the end of it; we would not be having this conversation.