bug#58168: string-lessp glitches and inconsistencies

[Mattias Engdegård]

2 okt. 2022 kl. 07.36 skrev Eli Zaretskii <eliz@gnu.org>:

>> Comparison between objects is not only useful when someone cares about their 
>> order, as in presenting a sorted list to the user. Often what is important 
>> is an ability to impose an order, preferably total, for use in building and 
>> searching data structures. I came across this bug when implementing a string 
>> set.
> 
> Always converting to multibyte handles this case, doesn't it?

I don't think it does -- string= treats raw bytes in unibyte and multibyte 
strings as distinct; converting to multibyte does not preserve (in)equality.

>> Actually I was talking about multibyte-multibyte comparisons.
> 
> Then why did you mention raw bytes? their multibyte representation
> presents no performance problems

In a way they do -- the way raw bytes are represented (they start with C0 or 
C1) causes memcmp to sort them between U+007F and U+0080. If we accept that 
then comparisons are fast since memcmp will compare many character per 
data-dependent branch. The current code requires several data-dependent 
branches for each character.

While we could probably bring down the comparison cost slightly by clever 
hand-coding, it's unlikely to be even nearly as fast as a memcmp and much 
messier. Since users are unlikely to care much about the ordering between raw 
bytes and something else (as long as there is an order), it would be a cheap 
way to improve performance while at the same time fixing the string< / string= 
mismatch.

> You can compare under the assumption that a unibyte string is
> pure-ASCII until you bump into the first non-ASCII one.  If that
> happens, abandon the comparison, convert the unibyte string to its
> multibyte representation, and compare again.

I don't quite see how that would improve performance but may be missing 
something.