[cfe-dev] Libc++ Windows preliminary test results

[Howard Hinnant]

On Sep 26, 2011, at 11:56 AM, Ruben Van Boxem wrote:

> Hi everyone,
> 
> I have finally gotten libc++ in a linkable/testable state for Windows (MinGW). This means testit can be run, and useful results are obtained.
> 
> Currently I'm linking GCC's libcupc++ to provide the __cxa_* functions and therefore I have to link with -Wl,--allow-multiple-definition. Nonetheless, there are a lot of tests passing, so I guess this will do for now. If you can say which tests can be fixed easily or how, I can provide patches.
> 
> ****************************************************
> Results for /home/Ruben/libc++/test:
> using clang version 3.0 ()
> Target: x86_64-w64-mingw32
> Thread model: posix
> with -std=c++0x -stdlib=libc++ -I/home/Ruben/libc++/include 
> ----------------------------------------------------
> sections without tests   : 1
> sections with failures   : 192
> sections without failures: 871
>                        +   ----
> total number of sections : 1064
> ----------------------------------------------------
> number of tests failed   : 621
> number of tests passed   : 3703
>                        +   ----
> total number of tests    : 4324
> ****************************************************
> Summary: 16.77% of the tests fail. I think this is a pretty good baseline, as I only fixed the compile errors for the library itself, and provided naive alternatives to any missing functions. Compared to Mac OS X's ~2% failures, it sucks, but well, libc++ wasn't written with Windows in mind :)
> The most interesting and fixable results:
>  - locale_t objects are created using the UNIX locale names, Windows uses different names, resulting in failures across the board here. I believe these strings are implementation-defined anyways, so there should either be a global locale-name-defining header (which defines macros for all the locale names used) and use these macros in the tests themselves.
>  - stdint macros seem absent, this may be a Clang stdint.h problem, or a mingw-w64 stdint.h problem, or both. I'll check that in detail.
>  - <atomic> and exception_ptr stuff is broken, it unimplemented.
>  - wchar_t is assumed to be 4 bytes wide. This is an unportable assumption. This results in warnings when assigning wchar_t's, but also with L'\x40003' wchar_t's where the hex value is out of range.
>  - <uchar.h>/<cuchar> is assumed to be present, which it is not on Windows. It is, though, somewhere in Kyrgyzstan: http://en.wikipedia.org/wiki/Uchar
>  - swprintf prototype is not the same for Windows (I did redefine it in support/win32/support.h header, maybe an inline overload would be better)
>  - clang chokes on decltype of vswprintf, thinking the __attribute__((__dllimport__)) is the function name.
>  - failures in complex, may be platform-related as well.
>  - setenv is not available on Windows, putenv is (and seems adequate for the tests).
>  - lots of io related errors in numerics/rand. Will need to investigate where those come from... Probably Microsoft using a different string form than expected.
>  - aligned storage tests, probably due to sizeof(long) != 8.
>  - lots of passes (ie not only failures)! Yay for me! And for Howard, cause he's probably the main reason for most of the passes :)
> 
> I attached the test log in case anyone is also interested in following up on anything. I'll run the test suite through a GCC-compiled libc++ too, just to make sure Clang isn't messing up all too much...

Thanks Ruben.  I've reviewed your test log and your summary above seems on the mark to me.

Howard