Add ability to generate universal binaries for macOS?

[virtuald]

As far as I can tell, meson does not currently build fat binaries. I've found that I can do it like this (this is building a python extension, but presumably it works for any shared module?):

wpiutil_module_x86_64 = py.extension_module(
  '_wpiutil_x86_64',
  c_args: ['-arch', 'x86_64'],
  cpp_args: ['-arch', 'x86_64'],
  link_args: ['-arch', 'x86_64'],
  sources: [wpiutil_sources],
  dependencies: [wpiutil_deps],
  install: false,
)
wpiutil_module_arm64 = py.extension_module(
  '_wpiutil_arm64',
  c_args: ['-arch', 'arm64'],
  cpp_args: ['-arch', 'arm64'],
  link_args: ['-arch', 'arm64'],
  sources: [wpiutil_sources],
  dependencies: [wpiutil_deps],
  install: false,
)
wpiutil_module = custom_target(
  command: [find_program('lipo'), '-create', '@INPUT0@', '@INPUT1@', '-output', '@OUTPUT0@'],
  install: true,
  install_dir: sw_py.get_install_dir(pure: false) / 'wpiutil',
  input: [wpiutil_module_x86_64, wpiutil_module_arm64],
  output: ['_wpiutil' + sw_py.get_variable('EXT_SUFFIX')],
)

While this is straightforward it is annoying to have to make this conditional on the platform. Other sources I've seen have recommended creating two different meson build directories and using a cross file per arch -- but that seems way more annoying.

My thought is maybe shared_module et al could get a 'build_universal' flag, or maybe a global setting could be set that would cause any targets to just transparently generate the above rules? This certainly feels like a common use case that would simplify building fat binaries on macos.

[eli-schwartz]

It cannot be done reliably without either two build directories or else per project opt-in, because there are projects which do code generation based on the host_machine (whether that is generating C sources or generating headers doesn't matter, but both are things that happen), so you'd have to know on a per project basis that none of that is happening.

And if you do know that, then if memory serves correctly, universal binary advocates say you should pass both -arch flags at the same time, and the Apple compiler will do two compilation passes and fuse the results itself, without the need for lipo.

I'm not certain what the answer to this general topic is. :(

[virtuald]

... fml, I didn't realize you could just set arch twice. That certainly seems easy enough.

The remaining challenge for me would be to communicate back to meson-python that fat binaries are being generated so that it sets the wheel tag correctly, but that's a different problem.

[eli-schwartz]

@rgommers I seem to recall you've talked about universal binary fusing before?

[rgommers]

We've never prioritized support for universal2 for meson-python, because universal2 is a really poor idea and packages should publish thin wheels for x86-64 and arm64, at which point pip will never pick the universal2 wheel (and hence why spend CI time on building them for a package?).

It's easy to create a universal2 wheel though from thin wheels if you really need one, with delocate-fuse:

pip install delocate
delocate-fuse $amd64_wheel $arm64_wheel -w .

I'm not opposed to implementing support in meson-python of course, but it's not a priority. In meson there is nothing to do I'd think.