Patch ldflags and libs in GCC_compile under windows

[lucianopaz]

Closes #944

On windows systems, the dll or lib libraries don't follow gcc's library naming convention, which means that their file name doesn't have the lib prefix (e.g. libmkl_core.dll). For this reason, gcc's library search fails to find these dynamic link libraries or static libs. An undesired consequence was that aesara (and old theano) was not able to set the correct blas__ldflags to link with mkl, even if the library had been correctly installed.

This PR introduces two GCC_compiler static methods that patch the libs or ldflags that are used during compilation. The static methods specifically target dll and lib files. They search for the library files in the provided library directories, and if a dll or lib is found, the ldflag is set to the absolute path that points to the dll or lib file. As a sidenote, dlls are preferred over libs.

As a consequence, the normal unix-like ldflags will be patched to work in windows as they do in unix. Thanks to this, aesara is now able to set the correct ldflags in windows to link to mkl installations.

[lucianopaz]

Does anyone know how to fix the unreachable code error from mypy? That code only runs on windows, and mypy doesn’t complain when I run it on windows.

[lucianopaz]

It looks like mypy is failing because of this bug

[brandonwillard]

It looks like mypy is failing because of this bug

It looks more like this one: python/mypy#5678. (Not the generators/yield parts, but the way Mypy is handing sys.platform.) Actually, they appear to both be fundamentally the same problem.

In general, it looks like Mypy is taking the value of sys.platform too literally.

Anyway, you can add a # type: ignore there.

[lucianopaz]

The failing test (tests/sparse/sandbox/test_sp.py::TestSP::test_maxpool) passes locally. Maybe the test is flaky?

[brandonwillard]

The failing test (tests/sparse/sandbox/test_sp.py::TestSP::test_maxpool) passes locally. Maybe the test is flaky?

Yeah, that looks like a bad, unseeded test.

[codecov]

Codecov Report

Merging #947 (0dc06ef) into main (b60cf72) will increase coverage by 0.06%.
The diff coverage is 97.22%.

@@            Coverage Diff             @@
##             main     #947      +/-   ##
==========================================
+ Coverage   78.92%   78.98%   +0.06%     
==========================================
  Files         152      152              
  Lines       47701    47795      +94     
  Branches    10862    10888      +26     
==========================================
+ Hits        37649    37753     +104     
+ Misses       7550     7542       -8     
+ Partials     2502     2500       -2     

Impacted Files	Coverage Δ
aesara/link/c/cmodule.py	52.40% <97.22%> (+1.33%)	⬆️
aesara/tensor/elemwise.py	88.17% <0.00%> (-0.27%)	⬇️
aesara/compile/function/types.py	79.91% <0.00%> (-0.14%)	⬇️
aesara/scalar/basic.py	79.05% <0.00%> (ø)
aesara/scan/op.py	84.84% <0.00%> (+0.24%)	⬆️
aesara/tensor/elemwise_cgen.py	95.74% <0.00%> (+0.29%)	⬆️
aesara/printing.py	51.66% <0.00%> (+3.07%)	⬆️

[hectormz]

Looking forward to this PR! Was just scratching my head about this on Windows. Thanks @lucianopaz

[lucianopaz]

Codecov seems to be linking the old commit instead of the new one