Re: Remaining dependency on setlocale()

On Thu, 2025-11-20 at 16:58 -0800, Jeff Davis wrote:
> On Wed, 2025-11-12 at 19:59 +0100, Peter Eisentraut wrote:
> > Many of these issues are pre-existing, but I just figured it has
> > reached 
> > a point where we need to do something about it.
> 
> I tried to simplify things in this patch series, assuming that we
> have
> some tolerance for small behavior changes.
> 
> 0001: No behavior change here, same patch as before. Uncontroversial
> simplification, so I plan to commit this soon.

Committed.

New series attached, which I tried to put in an order that would be
reasonable for commit.

0001-0004: Pure refactoring patches. I intend to commit a couple of
these soon.

0005: No behavioral change, and not much change at all. Computes the
"max_chr" for regexes (a performance optimization for low codepoints)
more consistently and simply based on the encoding.

0006: fixes longstanding ltree bug due to inconsistency between the
database locale and the global LC_CTYPE setting when using a non-libc
provider. The end result is also cleaner: use the database locale
consistently, like tsearch. I don't intend to backport this, unless
someone thinks it should be, but it should come with a release note to
reindex ltree indexes if using a non-libc provider.

0007: remove the char_tolower() API completely. We'd lose a pattern
matching optimization for single-byte encodings with libc and a non-C
locale, but it's a significant simplification. We could go even further
and change this to use casefolding rather than lower(), but that seems
like a separate change.

0008: Multibyte-aware extraction of pattern prefixes. The previous code
gave up on any byte that it didn't understand, which made prefixes
unnecessarily short. This patch is also cleaner.

0009: Changes fuzzystrmatch to use pg_ascii_toupper(). Most functions
in the extension are unaffected, but soundex() can be affected, and I'm
not sure what exactly it's supposed to do with non-ASCII.

0010: For downcase_identifier(), use a new provider-specific
pg_strfold_ident() method. The ICU version of this method is a work-in-
progress, because right now it depends on libc. I suppose it should
decode to UTF-32, then go through u_tolower(), then re-encode -- but
can the re-encoding fail? In any case, it would be a behavior change
for identifier casefolding with ICU and a single-byte encoding, which
is probably OK but the risk is non-zero.

0011: POC patch to introduce lc_collate GUC. It would only affect
extensions, PLs, libraries, or other non-core code that happens to call
strcoll() or strxfrm(). This would address Daniel's complaint, but it's
more flexible. And by being a GUC, it's clear that we shouldn't depend
on it for any stored data. We can do something similar for LC_CTYPE
after we eliminate dependencies in core code.

Regards,
	Jeff Davis