ENH: support shared libraries on Windows when explicitly enabled

[dnicolodi]

No description provided.

[rgommers]

This looks like a good approach, I like it. A couple of thoughts:

1. Internal shared libraries can be made to work today already in a package using meson-python, when explicitly installing the shared library without the package and using delvewheel (in an elaborate way) and install_rpath on other platforms, as demonstrated by BUG: Fix for scipy.special seterr, geterr, errstate scipy/scipy#20321. I think this PR is fine as is; I did have a concern that not having the option set shouldn't start raising an exception, so let's make sure that that remains the case.

2. As @QuLogic commented, it'd be good to elaborate on the best ways for a package to do this. I think there are two:

   • Using os.add_dll_directory, which should be reliable on Python >= 3.10
   • Preloading the DLL with ctypes.WinDLL - we've done this for a long time in NumPy and SciPy already for external shared libraries: https://github.com/scipy/scipy/blob/bb92c8014e21052e7dde67a76b28214dd1dcb94a/tools/openblas_support.py#L239-L274 (it should require fewer lines of code for an internal shared library that will always be present)

[dnicolodi]

With the documentation work done in #700 it is time to revisit this, I think. I think this PR is a good approach and the implementation is simple enough. However, I really don't like the option name I came up with. Does anyone have a better idea?

[rgommers]

A good descriptive name is not so easy. Ideally the name should make clear it's a boolean and allows meson-python to fold internal shared libraries targeted for install outside site-packages into the wheel on Windows. That's too much info though. Although shared-libs-win32 can probably be improved on. Perhaps something like allow-windows-internal-sharedlib? (I don't like win32 much, since in wheel tags it's specific to 32-bit Python).

[dnicolodi]

I've adjusted the option name and updated the documentation. Please have a look.

[rgommers]

The added docs and the option name look quite good to me, as do the code changes.

The docs CI job has a complaint still. And it seems like test coverage still needs to be added.

[dnicolodi]

Thanks for the review @rgommers I didn't add tests because I didn't want to deal with compilation with MSVC... I added them and indeed I have no idea why MSVC complains during linking (there is an issue with an invalid linker argument that I'll fix, but that does not seem to be the cause of the link failure). Any help would be appreciated

[dnicolodi]

All green!

[rgommers]

This looks good to go, nice work! Glad adding the test wasn't too much trouble.

One more docs suggestion inline.

One test suggestion here: the first commit and the FreeBSD patchelf addition show that RPATH handling works fine on FreeBSD now, so it may be nice to add freebsd to these lines in test_wheel.py:

@pytest.mark.skipif(sys.platform not in {'linux', 'darwin', 'sunos5'}, reason='Not supported on this platform')

Both of these are optional suggestions - happy for this to be merged as is if you prefer.

[dnicolodi]

One test suggestion here: the first commit and the FreeBSD patchelf addition show that RPATH handling works fine on FreeBSD now, so it may be nice to add freebsd to these lines in test_wheel.py:

Indeed. I'll do

[rgommers]

🎉 nice work!