Issue 22407: re.LOCALE is nonsensical for Unicode

Current implementation of re.LOCALE support for Unicode strings is nonsensical. It correctly works only on Latin1 locales (because Unicode string interpreted as Latin1 decoded bytes string. all characters outside UCS1 range considered as non-words), on other locales it got strange and useless results.

>>> import re, locale
>>> locale.setlocale(locale.LC_CTYPE, 'ru_RU.cp1251')
'ru_RU.cp1251'
>>> re.match(br'\w', 'µ'.encode('cp1251'), re.L)
<_sre.SRE_Match object; span=(0, 1), match=b'\xb5'>
>>> re.match(r'\w', 'µ', re.L)
<_sre.SRE_Match object; span=(0, 1), match='µ'>
>>> re.match(br'\w', 'ё'.encode('cp1251'), re.L)
<_sre.SRE_Match object; span=(0, 1), match=b'\xb8'>
>>> re.match(r'\w', 'ё', re.L)

Proposed patch fixes re.LOCALE support for Unicode strings. It uses the wide-character equivalents of C characters functions (towlower(), iswalpha(), etc).

The problem is that these functions are not exists in C89, they are introduced only in C99. Gcc understand them, we should check other compilers. However these functions are already used on FreeBSD and MacOS.

I don't think we should fix this in 2.x: some people may rely on the old behaviour, and it will be difficult for them to debug.
In 3.x, I simply propose we deprecate re.LOCALE for unicode strings and make it a no-op.

Yes, one of solution is to deprecate re.LOCALE for unicode strings and then 
make it incompatible with unicode strings. But I think it would be good to 
implement locale-aware matching.

Example.

>>> for a in 'Ii\u0130\u0131':
...     for b in 'Ii\u0130\u0131':
...         if a != b and re.match(a, b, re.I): print(a, '~', b)
... 
I ~ i
I ~ İ
i ~ I
i ~ İ
İ ~ I
İ ~ i

This is incorrect result in Turkish. Capital dotless "I" matches capital "İ" 
with dot above, and small dotless "ı" doesn't match anything.

Regex produces more relevant output, which includes matches for Turkish and 
English:

I ~ i
I ~ ı
i ~ I
i ~ İ
İ ~ i
ı ~ I

With locale tr_TR.utf8 (with the patch):

>>> for a in 'Ii\u0130\u0131':
...     for b in 'Ii\u0130\u0131':
...         if a != b and re.match(a, b, re.I|re.L): print(a, '~', b)
... 
I ~ ı
i ~ İ
İ ~ i
ı ~ I

This is correct result in Turkish.

Therefore there is a use case for this feature.

Ha, I always forget about the Turkish locale case...

Here is simple patch which just deprecate using of the re.LOCALE flag with str patterns. It also deprecates using of the re.LOCALE flag with the re.ASCII flag (with bytes patterns) and adds some re.LOCALE related tests.

If there are no objections I'll commit the re_deprecate_unicode_locale.patch patch. But it would be good if someone will review doc changes.

Looks like revision 561d1d0de518 was to fix this issue, but the NEWS entry has the wrong reference number

Indeed. Thank you Martin.

New changeset abc7fe393016 by Serhiy Storchaka in branch 'default':
Fixed issue number in Misc/NEWS for issue #22407.
https://hg.python.org/cpython/rev/abc7fe393016

messages: + msg228876
versions: - Python 2.7, Python 3.4