Bug 749392

Created attachment 530396 [details]
patch from upstream to lookup an ecoding for building wininst correctly

Description of problem:
The files:
/usr/lib64/python2.7/distutils/command/wininst-6.0.exe
/usr/lib64/python2.7/distutils/command/wininst-7.1.exe
/usr/lib64/python2.7/distutils/command/wininst-8.0.exe
/usr/lib64/python2.7/distutils/command/wininst-9.0-amd64.exe
/usr/lib64/python2.7/distutils/command/wininst-9.0.exe

are included in the python-devel package but they aren't being compiled from source.  At today's FPC meeting, the FPC decided that Windows binaries should be compiled from source as well.  We have a mingw-32 cross-compiler which will let us compile the win32 executables.  We'll have to wait for the mingw-64 stack in order to compile the amd64.exe (and yes, FPC decided we should just not have the 64 bit exe until then).

As background, there's a good deal of hinkiness about these binaries, what they do, and how they achieve it.  I'll summarise a discussion I had with epienbro of the MinGW SIG.  Someone with a Windows box can test further and if our assumptions seem off we can revisit this issue with more data.

According to epienbro we should be able to cross-compile the win32 .exe's with mingw32.  There are some caveats but they probably also apply to upstream in some ways so we're only making the problem slightly worse.

First off, it appears that code compiled with one compiler on windows may or may not work with code compiled using a different compiler.  This includes, not just MinGW but also different versions of Visual C++.  So the whole idea behind wininst.exe may be flawed unless it's entirely statically linked.

Second, eventually we may be able to statically link everything but MSVCRT ( the windows equivalent of libc) which MinGW does not ship.  If everything else is statically linked, this should significantly reduce the problems as MinGW is intended to build binaries which can work with the version of MSVCRT that is found on the user's system.  If it doesn't work, it's a bug that we can have the MinGW folks look into fixing.

Thirdly, with dynamic linking, even upstreams wininst's might still be a problem for distributors as the wininst that I build on my Windows system may detect that I'm building on a system that needs MSVCRT 9.0 but my user's systems may only have MSVCRT 6.0 so the wininst would not necessarily be portable between our systems.  (They may be, but that's where the compiler differences come in again).

Fourthly, in this particular case, we're using LoadLibrary() (roughly equivalent to dlopen in Linux) to load libpython.  That should shield us from some of the undesired effects of being built with two different compilers.

With all that said, what we need to do is rm the .exe's in %prep.  As part of the build, running somethig similar to this to build:

i686-pc-mingw32-gcc -c extract.c -o extract.o
i686-pc-mingw32-gcc -c install.c -o install.o
i686-pc-mingw32-windres install.rc install_rc.o
i686-pc-mingw32-gcc extract.o install.o install_rc.o -o wininst.exe /usr/i686-pc-mingw32/sys-root/mingw/lib/libz.a -luuid -lole32 -lgdi32 -lcomctl32

We'll need to BuildRequire: mingw32-zlib-static mingw32-gcc

Given the above notes about different VC++ versions and mingw, we can try building this and simply linking the various *.exe's together.  I believe that mingw also has the ability to specify a specific version of the CRT to pretend to link against (they're stubs in mingw; the user's windows machine will provide the actual version) but I don't know that that will help us here.


Additional info:

One reason that I don't feel too horribly about doing this is that wininst appears to be broken on Linux right now.  I had to apply a patch from upstream to get it to work: http://bugs.python.org/issue13151#msg145473

I'll add that as an attachment