Re: Remaining dependency on setlocale()

On Thu, 2025-07-24 at 11:10 -0700, Jeff Davis wrote:
> The main problem is with strerror_r()...

Postgres messages, like "division by zero" are translated just fine
without LC_CTYPE; gettext() only needs LC_MESSAGES and the server
encoding. So these are fine.

We use strerror_r() to translate the system errno into a readable
message, like "No such file or directory", i.e. the %m replacements.
That needs LC_CTYPE set (just for the encoding, not the
language/region) as well as LC_MESSAGES (for the language/region).

When using a locale provider other than libc, it's unfortunate to
require LC_CTYPE to be set for just this one single purpose. The locale
itself, e.g. the "en_US" part, is not used at all; only the encoding
part of the setting is relevant. And there is no value other than "C"
that works on all platforms. It's fairly confusing to explain why the
LC_CTYPE setting is required for the builtin or ICU providers at all.
Also, while it's far from the biggest challenge when it comes to
multithreading, it does cause thread-safety headaches on platforms
without uselocale().

Perhaps we could get the ASCII message and run it through gettext()?
That would be extra work for translators, but perhaps not a lot, given
that it's a small and static set of messages in practice. That would
also have the benefit that either NLS is enabled or not -- right now,
since the translation happens in two different ways you can end up with
partially-translated messages. It would also result in consistent
translations across platforms.

Regards,
    Jeff Davis