I wrote:
> I just noticed that the collation patch has modified char2wchar and
> wchar2char to accept a collation OID as argument ... but it hasn't done
> anything to make those arguments actually work. Since those functions
> depend on wcstombs and mbstowcs, which respond to LC_CTYPE and nothing
> else, this flat out does not work in non-default collations. What's
> more, there doesn't seem to be any such thing as wcstombs_l or
> mbstowcs_l (at least my Fedora box hasn't got them), so this can't be
> fixed within the available glibc API.
I poked around a bit and found out that wcstombs_l and mbstowcs_l *do*
exist in some BSD-derived systems (in particular, OS X has 'em; and
Windows too). Not clear why XBD didn't see fit to include them, because
so far as I can see there simply isn't any useful substitute.
What I think we need to do is:
* Redefine char2wchar and wchar2char as being like mbstowcs_l and
wcstombs_l (in particular, take pg_locale_t *not* a collation OID).
* Use mbstowcs_l and wcstombs_l where available.
* Where they're not, install the locale_t with uselocale(), do
mbstowcs or wcstombs, and revert to the former locale_t setting.
This is ugly as sin, and not thread-safe, but of course lots of
the backend is not thread-safe. A quick look at the glibc source
for uselocale() suggests that it won't be too horribly inefficient.
Comments, better ideas?
regards, tom lane