gmp, mpfr, Windows port

[Sisyphus]

Andreas Fabri wrote:
> 
> Hello,
> 
> I work in the CGAL project, which makes an open source
> library where the user can choose the type of arithmetic
> for coordinates of geometric objects. One option is
> the rational number type of GMP and we then use mpfr
> to compute enclosing double intervals for rational numbers,
> in order to do interval arithmetic.
> 
> CGAL is across platform library, which particularly means
> also supporting Windows compilers.  The proposed solution
> in this forum is to use mingw and g++ in order to
> create the dll, and to use the Microsoft lib tool
> to create the lib.  For mpfr we got the recipe to
> create the .a and to rename it to .lib
> Doing this  we encountered problems with missing
> memory allocation functions, and there  a suggestion
> on this mailing list is to also link libgcc.a from
> mingw, even if one uses Visual C++.   This compiles
> and links, but the executables using it often crash
> as soon as we use mpfr.
> 
> The first time that our testsuite worked without
> crashing, was with Brian Gladman's approach.
> 
> 
> Do you see a way to better integrate your work,
> because as it is now is really hard for your users.
> * mpfr is not up to date in the gmp distribution.
> * mpfr does not generate a dll (Brian has patches
>   for this).
> * I recently got a mail announcing a
>   release candidate for mpfr. Now I have to do
>   wait for Brian to redo his changes, unless the
>   mpfr folks incorporate his patchs in their code.
> 
> Yes this is work, but taking into account that
> Windows is a major platform  (if we like it or
> not), it seems worth the effort for the success
> of "arithmetic without limitations".
> 

I now build GMP as a static library with MinGW and MSYS because it 
builds flawlessly that way - but such a build does not seem to transfer 
across very well to MSVC (both 6.0 and 7.0). I keep getting "unresolved 
external symbols".

So I built GMP-4.1.4 as a dynamic lib using MinGW and MSYS.

./configure --disable-static --enable-shared
make

Running make check builds the test executables in the wrong place - so I 
transferred them to the correct location and re-ran make check. 
Eventually I got to run all of the tests - 2 of which failed. I 
submitted a bug report about all of this a few days ago. I suspect the 
bugs are in 'make check' rather than in the GMP library because, like I 
said, if I build as a static library, everything (including 'make 
check') proceeds flawlessly.

I then copy gmp.h into MinGW/include, VC6/include and VC7/include.
I then copy libgmp.dll.a into MinGW/lib/libgmp.a

And I copy 'libgmp-3.dll' to a folder that's in the system path.

Then, using VC7 I run:
lib /def:libgmp-3.dll-def /out:libgmp.lib
and copy the resultant 'libgmp.lib' into VC7/lib.

Then, using VC6 I run:
lib /def:libgmp-3.dll-def /out:libgmp.lib
and copy the resultant 'libgmp.lib' into VC6/lib.

Then, back in the MSYS shell, I build mpfr from mpfr-2.1.0-rc2:
./configure
make
make check

Then using VC7 in the mpfr build directory, I run:
lib /out:libmpfr.lib *.o
and copy the resultant 'libmpfr.lib' into VC7/lib

Then using VC6 in the mpfr build directory, I run:
lib /out:libmpfr.lib *.o
and copy the resultant 'libmpfr.lib' into VC6/lib

Then, with both VC6 and VC7 I can build mpfr apps with:
cl mpfr_test.c libmpfr.lib libgmp.lib libgcc.lib

('libgcc.lib', as you know, has been copied from MinGW's 'libgcc.a'.)

Seems to work ok for me - though one script hardly constitutes a 
thorough test :-)
(I hardly ever use the Microsoft compilers.)

I get some warnings along the way with VC7 - maybe there's something 
lurking beneath the surface just waiting to bight. Do you have a script 
that breaks the above procedure ? If you've got a script or 2 that would 
provide a thorough work out, I'd be interested to give them a run. (I 
run mostly perl scripts - don't have much in the way of C programs.)

With VC6 I don't think there were any warnings issued at any stage.

Admittedly it's a little kludgy, but it's fairly straightforward and the 
copying and 'lib' commands could be automated with a simple perl script 
or batch file.

Anyway - that's filled in *my* afternoon quite nicely .... hope there's 
something there that helps.

Cheers,
Rob