bug#58168: string-lessp glitches and inconsistencies

[Mattias Engdegård]

We really want string< to be consistent with string= and itself since this is 
fundamental for string ordering in searching and sorting applications.
This means that for any pair of strings A and B, we should either have A<B, B<A 
or A=B.

Unfortunately:

  (let* ((a "ü")
         (b "\xfc"))
    (list (string= a b)
          (string< a b)
          (string< b a)))
=> (nil nil nil)

because string< considers the unibyte raw byte 0xFC and the multibyte char 
U+00FC to be the same, but string= thinks they are different.
We also distinguish raw bytes by multibyte-ness:

  (let* ((u "\x80")
         (m (string-to-multibyte u)))
    (list (string= u m)
          (string< u m)
          (string< m u)))
=> (nil t nil)

but this is a minor annoyance that we can live with: we strongly want string= 
to remain consistent with `equal` for strings.
So, what can be done? The current string< implementation uses the character 
order

 ASCII < ub raw 80..FF = mb U+0080..U+00FF < U+0100..10FFFF < mb raw 80..FF

in conflict with string= which unifies unibyte and multibyte ASCII but not raw 
bytes and Latin-1.
It suggests the following alternative collation orders:

A. ASCII < ub raw 80..FF < mb U+0080..10FFFF < mb raw 80..FF

which puts all non-ASCII multibyte chars after unibyte.

B. ASCII < ub raw 80..FF < mb raw 80..FF < mb U+0080..10FFFF

which inserts multibyte raw bytes after the unibyte ones, permitting any ub-ub 
and mb-mb comparisons to be made using memcmp, and a slow decoding loop only 
required for unibyte against non-ASCII multibyte strings.

C. ASCII < mb U+0080..10FFFF < mb raw 80..FF < ub raw 80..FF

which instead moves unibyte raw bytes to after the multibyte raw range. This 
has the same memcmp benefit as alternative B, but may be slightly faster for 
ub-mb comparisons since only unibyte 80..FF need to be remapped.

Any particular preference? Otherwise, I'll go with B or C, depending on what 
the resulting code looks like.