I think there is an opportunity to speed up some unicode normalisations significantly.

In 3.9 at least, the normalisation appears to be dependent on the length of the string:

    >>> setup="from unicodedata import normalize; s = 'reverse'"
    >>> t1 = Timer('normalize("NFKC", s)', setup=setup)
    >>> setup="from unicodedata import normalize; s = 'reverse'*1000"
    >>> t2 = Timer('normalize("NFKC", s)', setup=setup)
    >>> 
    >>> min(t1.repeat(repeat=7))
    0.04854234401136637
    >>> min(t2.repeat(repeat=7))
    9.98313440399943

But ASCII strings are always in normalised form, for all four normalisation forms. In CPython, with PEP 393 (Flexible String Representation), it should be a constant-time operation to detect whether a string is pure ASCII, and avoid scanning the string or attempting the normalisation.