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

[Howard Hinnant]

Committed revision 140384.

Howard

On Sep 23, 2011, at 9:22 AM, Ruben Van Boxem wrote:

> 2011/9/23 Ruben Van Boxem <vanboxem.ruben at gmail.com>
> 2011/9/22 Ruben Van Boxem <vanboxem.ruben at gmail.com>
> 2011/9/22 Ruben Van Boxem <vanboxem.ruben at gmail.com>
> 2011/9/22 Eli Friedman <eli.friedman at gmail.com>
> On Thu, Sep 22, 2011 at 12:35 PM, Ruben Van Boxem
> <vanboxem.ruben at gmail.com> wrote:
> > 2011/9/22 Howard Hinnant <hhinnant at apple.com>
> >>
> >> Patch committed revision 140328.
> >
> > Thanks!
> >
> > Now that I've got your attention, on to more... involved... matters :-)
> >
> > In order to support libc++'s current implementation using
> > cat[gets|open|close] and nl_types.h (which I ignored for now), I would need
> > to pull in some external code, namely mingw's libcatgets and a supporting
> > iconv library. These can all go under support/win32 and leave all other
> > platforms unchanged of course, and leave the rest of libc++ as coherent as
> > possible on all platforms. I understand Howard has an acceptable aversion to
> > including too much foreign code in libc++, and hope you can see the need for
> > this here.
> >
> > There are some issues though:
> >  - libcatgets is licensed under GPLv2. If I can go through with this, I can
> > ask/plead/beg the original author to allow a license change or ask him to
> > make an exception for libc++. I know of no existing alternative.
> >  - for the iconv part, GNU libiconv is of course out of the question. There
> > are two projects of interest: APR-iconv (Apache License 2.0) and win-iconv
> > (BSD 2-clause license), the latter being tiny but primitive. The first is a
> > truly first class implementation as far as I can see, and with minor
> > refactoring could be included in the libc++ Windows build.
> >  - I am not an Open Source license lawyer, and do not know what would be
> > acceptable for libc++ in this regard.
> >
> > I want to ask what's "best", before starting with the implementation. The
> > only real alternative is writing all this (ie catgets with supporting iconv
> > code) from scratch, which I don't see myself doing in the near to distant
> > future. Another way is to completely replace libc++ parts that use/rely on
> > these API's, but this would make libc++ a lot more assymetric across
> > platforms. In theory, these support files could help other platforms where
> > these API's are missing.
> >
> > Please let me know what you think. Thanks!
> 
> MSVC doesn't provide a non-trivial implementation of
> std::messages::do_open etc., so I don't see why it matters if libc++
> does not provide one on Windows.
> 
> Hmm, glancing through libstdc++'s headers, it doesn't seem to do much more :(. Then the question becomes (if the authors/copyright holders of the support code in question do not want to have it used in win32 libc++), should I just mimic current behavior of the other "Big 2" (MSVC/GCC) in this regard? What would I be missing out on then? Is it really only n3290's 22.4.7?
> 
> Implementing stubs was easy, so I just went ahead with it. I can always improve on it later if I want. Attached is another incremental patch.
> 
> Notes:
>  - inclue/__bit_reference: Win32 <yvals.h> has a line: "#define _C2 1" which obviously messes up the templates in this file. I added n extra underscore to work around this, and did the same to _C1 for consistency.
>  - include/locale: include my support header for vasprintf, make the std::message functions stubs for Windows, mirroring other C++ Standard Library implementations on Windows.
>  - include/support/win32/support.h: add two missing functions and implement them naively in function of not-missing functions.
>  - src/locale.cpp: exclude langinfo.h inclusion for Windows. More work is needed here later to provide an alternative.
>  - src/support/win32/suppport.cpp: implementation.
> 
> Please comment on this if something looks weird or wrong.
> 
> Clang failed me (http://llvm.org/bugs/show_bug.cgi?id=10989) so, I'll need to resort to GCC and its unhelpful error messages in the meantime.
> 
> Ruben
> 
> This time with patch. Note that the 32% is a rough estimate, mostly meant for my own encouragement.
> 
> PS: thanks for the quick patching for the assembler error.
> 
> I have reached 100%, that is, linking Clang object files fails horribly, and GCC can't inline var_args functions, but every source file builds with GCC and Clang apart from these non-libc++ related issues.
> 
> The patch explained:
>  - include/__bit_reference: Win32 <yvals.h> has a line: "#define _C2 1" which obviously messes up the templates in this file. I added an extra underscore to work around this, and did the same to _C1 for consistency.
>  - include/locale: include my support header for vasprintf, make the std::message functions stubs for Windows, mirroring other C++ Standard Library implementations on Windows.
>  - include/support/win32/support.h: add two missing functions, and define a swprintf for MINGX32. Technically, VS2003 also has a bad declaration, but as it doesn't support a lot more that isn't really relevant for libc++.
>  - include/suppport/win32/locale.h:
>    - add missing defines for prefixed versions of functions.
>    - Redo newlocale to accept a third parameter. Functionality should be ok, as the third parameter is always 0 in libc++.
>    - Implement is(w)blank naively, which are missing from msvcrt. 
>    - define LC_ALL_MASK
>    - remove the nl_types enum definition, as it is now unused.
>  - src/locale.cpp: exclude langinfo.h inclusion for Windows, instead including <locale.h>. msvcrt does not provide a C99 compliant lconv implementation, so for Windows, some int_* prefixed struct members are not present. I used the non-prefixed ones instead. Technically incorrect, but should work OK in most cases.
>  - src/string.cpp: include Win32 support header.
>  - src/thread.cpp: omit inclusion of sys/sysctl.h on Windows.
>  - src/support/win32/suppport.cpp:
>    - fix vasprintf.
>    - implement the two new functions naively in function of not-missing functions. These are slow and wasteful, but should work correctly at least.
> 
> I think the changes to core libc++ are minor, considering everything. Please comment if you think otherwise. On the other hand, please commit when you're OK with it.
> 
> This leads to the next problem(s):
>  - libc++ dll needs exports a la __cdeclspec(dllexport)/__cdeclspec(dllimport). See LLVM/Clang DLL discussions. Maybe using the same idea is possible.
>  - GCC (4.6) cannot inline vararg functions, thus is cannot fully compile libc++. On the other hand, Clang fails linking the dll due to a bunch of undefined references, among which cxxabi stuff, and C std library stuff. I'll investigate.
>  - MSVC has trouble parsing (I think) "inline namespace":
> [  4%] Building CXX object lib/CMakeFiles/cxx.dir/__/src/algorithm.cpp.obj
> algorithm.cpp
> M:\Development\Source\libc++\include\__config(14) : warning C4068: unknown pragma
> M:\Development\Source\libc++\include\cstddef(46) : warning C4068: unknown pragma
> M:\Development\Source\libc++\include\cstddef(50) : error C2143: syntax error : missing ';' before 'using'
> M:\Development\Source\libc++\include\cstddef(50) : error C4430: missing type specifier - int assumed. Note: C++ does not support default-int
> M:\Development\Source\libc++\include\cstddef(93) : error C2143: syntax error : missing ';' before 'namespace'
> M:\Development\Source\libc++\include\cstddef(93) : error C4430: missing type specifier - int assumed. Note: C++ does not support default-int
> 
> Ruben
> 
> <windows.patch>