BLD: adopting the meson build backend

[neutrinoceros]

I would like to propose migrating from setuptools to meson-python as the build backend, in order to align the package with modern practices from numpy, scipy, matplotlib, to name a few.
I'm sure @rgommers would like the idea, I'm mostly wondering how disruptive it could or should be, really. The one thing I don't know if we can port from setup.py is the versioneer hook, though to be perfectly honest, I don't see a lot of value in having a dynamic version number, especially in this package that is currently in maintenance mode.

[rgommers]

Works for me. bottleneck wasn't top of my list for such a migration, but it should be easy to do and be an upgrade. Build time isn't really a concern here, but good logging will still be nice.

versioneer isn't too useful imho, and half-broken anyway (I believe it still gives version 0.0.0 or some such thing in CI). versioneer is a mess, all major users in the NumPy ecosystem have dropped it. I'd just copy the version handling setup from NumPy.

[rdbisme]

Hello guys,

I'm not a huge fan of meson. If anything, I'd probably migrate to something like https://github.com/scikit-build/scikit-build. But since you're all more familiar with meson in other projects, feel free to proceed and thanks for raising this.

[rgommers]

scikit-build is legacy and should be avoided, but scikit-build-core (= the scikit-build rewrite) would also be a significant upgrade from setuptools. I'm less familiar with it than meson-python, but I'd be +1 on that too.

[neutrinoceros]

I'm neutral myself. I proposed meson because I have some experience with it and know that Ralph is involved in meson-python, while I personally have no experience with scikit-build-core, but that's also a good reason for me to try it: I think it's the last notable build backend I never worked with.
Anyway I mostly want to remove setuptools, to be honest.

[rdbisme]

You're more active and have more experience with meson, so go ahead with that :)

[neutrinoceros]

I got started on this but quickly realized there was a significant hurdle in my way: templating.
Basically I don't know how approach re-implementing/replacing bn_template.py in meson. @rgommers, any idea where to start ?

[neutrinoceros]

actually I think I got it. I should be able to rewire this module through meson's generator

[rgommers]

@neutrinoceros generator is right. It's very similar to NumPy's .c.src template processing, so you should be able to copy that almost literally, see:

• https://github.com/numpy/numpy/blob/b6e8ad83299fc6f87107492df0039dadfb5b9b98/numpy/meson.build#L308-L312
• https://github.com/numpy/numpy/blob/main/numpy/_build_utils/process_src_template.py

The most useful thing to copy from process_src_template.py is the argparse CLI code and directory handling, and the shebang line (#!/usr/bin/env python3 - the exact form matters). You'll want to be writing the generated .c files to the build directory, not in-place.