[Numpy-discussion] Updates on using recent mingw for numpy/scipy

[David Cournapeau]

Hi there,

During pycon.fr sprints, I took some time to look more into building
numpy/scipy wheels on windows with recent mingw (gcc 4.x series).

tl;dr: While I made some progress, there remains some hard-to-track issues.
I will prepare a vagrant setup so that other people can work on this as
well so that I don't remain the bottleneck.

details:

 - all the tests were done on 32 bits. While 64 bits would add some
challenges, I think most mingw-specific issues are independent of 32 vs 64
bits.
 - works mean building + running numpy/scipy.test() without 'bad' failures
 - compiling numpy without any blas/lapack, but completely static (no
dependency on mingw at runtime) works with both mingw-w64 and mingw.
 - compiling numpy with a custom-built blas/lapack 3.4.2 gives me a working
numpy with mingw, but mingw-w64 gives a very weird failure: importing numpy
once segfaults at some location, twice at a difference one, three times at
yet another one, and then never segault anymore (unless I build a new one
or reboot windows). I suspect some weird runtime interactions.
 - compiling scipy with mingw + custom-built blas/lapack works, except that
it crashes when exciting the process (with some C runtime-related errors).
 - compiling both numpy and scipy without statically linking gfortran
runtime does work.

I am hoping that those errors are caused by some invalid link order. As
controlling those precisely with distutils is horrible, I have not been
able to pursue this any further.

cheers,
David
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/numpy-discussion/attachments/20131105/cd110759/attachment.html>