Re: Normalising is REALLY slow

[Chris Warburton]

A thought occurs: the only reason we're merging overlapping reps
together is to ensure that each name gets the smallest equivalent when
finalised.

We *could* use hash tables directly: this would avoid the need to
finalise, but would make such merges tricky. There might be a way to
quickly merge two hash tables such that we can a) identify the smallest
name for each replacement and b) update all of the redundancies to point
at that name, but at the moment it looks like maintaining lists and
finalising is the simpler way.

Still, we don't actually need to merge the reps; all we need to do is
ensure that each rep contains the smallest name for its contents;
i.e. it's OK for multiple overlapping reps to exist, as long as their
overlap includes the smallest element.

This should let us avoid comparing many of the contents of the reps;
instead, for each rep we just have to check if its smallest value
(constant time) appears in the old-of any other rep; if so, add the
new-of the latter to the former.