bug#18051: [Emacs-diffs] trunk r117726: Add string collation.

[Eli Zaretskii]

> Date: Wed, 27 Aug 2014 12:08:52 -0700
> From: Paul Eggert <eggert@cs.ucla.edu>
> CC: dmantipov@yandex.ru, 18051@debbugs.gnu.org
> 
> First, the current algorithm looks only at LC_COLLATE, but the usual 
> approach is to default LC_COLLATE to LANG if LC_COLLATE isn't set, and 
> to have LC_ALL override LC_COLLATE.  Shouldn't Emacs take a similar 
> approach, for compatibility?

I think we agreed to have a variable that holds the non-default locale
as a Lisp string.  LANG and LC_COLLATE will then be used internally by
newlocale and/or wcscoll_t, as users expect.  I don't think it's
appropriate for a primitive to take arguments from environment
variables, certainly not those on process-environment.  If some Lisp
application would want to do that, let them.

> More generally, it strikes me that string-collate-lessp will be quite 
> slow due to the overhead of looking up the locale environment string and 
> creating and destroying a locale for each string comparison.

The lookup will no longer be relevant, when we switch to a variable.

As for creating and destroying the locale, I guess you are right.

> Instead, shouldn't Emacs should have a locale object that the Emacs
> Lisp programmer can create, an object that encapsulates the low
> level locale_t object, and which can be passed as an optional
> argument to string-collate-lessp?

That's what Guile does.  But it will complicate using these functions
in sorting routines.  Perhaps binding a variable to the object will
do.

Alternatively, a simple one-slot cache internal to string_collate will
probably remove most of the overhead.  (You will see that
w32_compare_strings already employs a similar cache.)