Note by Dale H. about implementation problems:
(Mailing list December 2015)
I think that the issue (from a performance perspective) is that you can't depend upon the value of the code point when doing collation — the main algorithm[5] is pretty much table based — In addition to the different sort orders based on characters there are even more arcane sort rules where characters at the end of a word can affect the sort order of the word (for more info see[4]).
It is worth looking at the Conformance section of the Unicode spec[1] as there are different levels of collation conformance .....
ICU conforms[2] to to UTS #10[3], the highest level of conformance ...
It looks like TwitterCLDR[6] uses the Main Algorithm[5] with tailoring[7]. They don't claim to be conformant to the Unicode Collation Algorithm[3], but they are covering a big chunk of the standard use cases ....