[cfe-dev] [PATCH] Libc++ Windows fixes

Ruben Van Boxem vanboxem.ruben at gmail.com
Thu Sep 29 04:08:24 PDT 2011


2011/9/29 Ruben Van Boxem <vanboxem.ruben at gmail.com>

> 2011/9/29 Ruben Van Boxem <vanboxem.ruben at gmail.com>
>
>> 2011/9/29 Howard Hinnant <hhinnant at apple.com>
>>
>>>
>>> On Sep 28, 2011, at 6:17 PM, David Chisnall wrote:
>>>
>>> > On 28 Sep 2011, at 22:45, Howard Hinnant wrote:
>>> >
>>> >> I think FreeBSD should define _LLIBCPP_STABLE_APPLE_ABI too.  David?
>>>  And maybe with David's latest patch to FreeBSD, should also #define
>>> _LIBCPP_GET_C_LOCALE 0.
>>> >
>>> > FreeBSD should now be mirroring the Apple implementation for anything
>>> xlocale related.  If there are any discrepancies, they are bugs and I'll fix
>>> them in libc, rather than hack around them in libc++.
>>> >
>>> > I'm not entirely sure what defining _LLIBCPP_STABLE_APPLE_ABI means -
>>> perhaps it could be replaced with a more descriptive name?
>>>
>>> See revision 140734, and let me know what needs fixing (hopefully
>>> nothing).
>>>
>>
>> This improves the behavior on Windows (uses the C locale instead of
>> returning NULL like for Linux). Attached patch also removes unreachable code
>> (_LIBCPP_HAS_DEFAULTRUNELOCALE is defined for both __APPLE__ and
>> __FreeBSD__, so the additional checks in locale.cpp are never reached).
>>
>> I'm now also wondering about _LIBCPP_WCTYPE_IS_MASK. I've been looking at
>> MSDN, but can't figure out if it matches. I did find a warning that if a
>> plain "char" is passed to Windows is* functions, unexpected results are
>> obtained (
>> http://msdn.microsoft.com/en-us/library/4yc6feha%28v=vs.80%29.aspx). It
>> seems that char_type could be char in the common case, and this could yield
>> unreliable behavior?
>>
>> Ruben
>>
>
> Never mind this patch. It's broken. Is there any reason not to call plain
> "toupper" in do_toupper directly? Or at least "toupper_(c,__cloc())"?
>
> Ruben
>

Okay, brace yourself...

I've tried to remove locale.cpp's dependency on glibc's toupper and tolower
transformation tables, which are only available in glibc. I have coded a
fallback do_tolower and do_toupper using basic ASCII subtract and add tricks
if glibc or the DefaultRuneLocale is not available. I know this is slower,
but it works at least. It uses islower_l and isupper_l with __cloc to check.
The change in behavior should be:
Mac: none
FreeBSD: none
Linux/glibc: none
other platforms (including Windows): no __classic_upper/lower_table
functions declared, defined, or used.
I also declared a symbol present in msvcrt.dll that represents the mask
table and added that case to classic_table.

I would prefer having a (global) static "C" locale object that can just be
given through __cloc, although I don't know if this is safe, but as long as
__cloc is only used in _l functions (which I think it is), this approach of
a persistent (global, not function-local) C locale object should work (is
initialized before any threads could call __cloc at the same time) and fix
the FIXME in __cloc.

I also tried to fix the ctype<wchar_t>::do_is to keep track of the result,
and now should work if multiple flags were requested.

Comments welcome!

Ruben
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20110929/c2ec60b5/attachment.html>


More information about the cfe-dev mailing list