boost::regex match but not std::regex

[llvmbot]

Bugzilla Link	40052
Version	6.0
OS	FreeBSD
Attachments	Testcase
Reporter	LLVM Bugzilla Contributor
CC	@mclow

Extended Description

I got the attached program. It has a global locale set (de_DE.UTF-8).

I think it might be a bug in libc++ because
on Windows(MSVC 2013 & MSVC 2017) and on Linux (gcc 8.2 + libstdc++) this regex (from std) matches with the global locale from boost. Also the regex from boost matches (replace std::regex by boost::regex).

This bug triggers only (also on my box and only on freebsd with clang and libc++) when i use boost::locale. With std::locale() it matches.

I already submitted this bug to FreeBSD and to boost.org.

For reference, here are the links
https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=233994

Boost.locale says:
[quote]
Boost.Regex and Boost.Locale aren't related, the locale generated by Boost.Locale is "C" locale with addons unrelated to Boost.Regex
[/quote]
boostorg/locale#35

FreeBSD says, it is a bug in boost.locale.

As both of my direct upstream bugtrackers seem to "dislike" this bug, i report it to clang/libc++ directly.

[llvmbot]

assigned to @mclow

[llvmbot]

Expected Output of the Testcase:
All ok

Got:
Bug triggered

[mclow]

My first thought was that you had a "high ascii" character in the test case, and that was getting treated differently by the locale. That appears not to be the case.

[llvmbot]

Updated testcase with facets

[llvmbot]

I also tried to set the facets one by one. It just triggers at

• all_characters
• collation_facet
• all_categories

I also updated the testcase to show how i did it.

[llvmbot]

The bug also goes away if i remove the icase flag from the regex.

[mclow]

Is it specific to the "de_DE.UTF-8" locale, or does it happen with others?
I'm thinking of other UTF-8 locales, like "en_US.UTF-8" or "fr_FR.UTF-8"

[mclow]

Check all the characters for tolower
Here's a program that calls translate_nocase for all the possible character values.
Can you run that and email me the output, please?

On my (Mac OS) machine, the values C0 -> DE are translated.

[mclow]

Whoops. I sent this to the wrong place. This should have been sent to https://reviews.llvm.org/D55746 instead. It may be related, but that's not for sure yet.

[llvmbot]

[expanded testcase that run 192 locales (1 segfaults if i do it) x 3 backends x 12 facets x icase onhttps://user-images.githubusercontent.com/60944935/143758461-cba7d36c-a8ae-4849-88cd-758f0286b165.gz)
I ran 13824 testcases.
192 locales (1 segfaults if i do it) x 3 backends x 12 facets x icase on/off.

One locale segfaults, but it is always the facets "all_characters , collation_facet, all_categories". C and POSIX work always ok.

[llvmbot]

result csv that shows the testresults of the locale combinations

[llvmbot]

Pivot Analysis of the result
The Pivot analysis of the data shows, its BUG always in "all_characters , collation_facet, all_categories". C and POSIX work always ok. Every other locale is affected. Even en_US.UTF-8.

[Lord-Kamina]

I have made a new issue over at the boost.locale repository for what, I am assuming, is this same issue, but on mac, with AppleClang:

boostorg/locale#249

Using wandbox, you will see the issue does not seem to affect GCC and libstdc++