Prototype support for PEP 739 (continued)

[mgorny]

This is a continuation of work from #13945, see #13945 (comment). For a start, just rebased it against fresh git head.

[mgorny]

Okay, I think this is as far as I can go on my own.

Few final points:

• I've disabled it on macOS, since — at least judging by the existing code — we need to hack around frameworks, and the relevant data is not available in build-details.json. I don't have a macOS system to work on this properly. I think it's good enough to focus on getting it right for the more general POSIX case first.
• The remaining CI failures don't seem to be related to these changes.
• I've been able to successfully cross-compile NumPy with these changes.

[rgommers]

Nice to see it's ready for testing!

You may want to rewrite history already, since it's required in Meson development to have self-contained orthogonal commits before merging.

A few initial comments; I'll try to find time to test this asap!

[rgommers]

nit: should be a regular code comment (# ) it looks like. Same comment for 2 lines lower.

[mgorny]

I suppose it was meant as a docstring, but according to PEP 258, attribute docstrings go below the assignment, not above. While it was rejected, it seems that tools support this convention and I see it used e.g. in mesonbuild/cargo/manifest.py, so I'll just fix the location.

[rgommers]

When I look at the multiple if self.build_config ... else I wondered if it isn't better to combined build_config and info into the same structure in a single place. I bet you considered it though and decided against doing that?

[mgorny]

To be honest, I haven't given it much thought, but they certainly do trigger different code paths, and contain slightly different information. I don't think trying to combine them would make the code more readable.

[rgommers]

I've been able to successfully cross-compile NumPy with these changes.

Do you happen to have a reusable recipe for this?

[rgommers]

This seems to work as expected, here is a set of CI jobs that cross-compiles SciPy from x86-64 to aarch64 in the first job, and then uses an aarch64 runner in the second job to download the built wheel and run the test suite:

• scipy/scipy@main...rgommers:scipy:ci-cross
• passing CI log (modulo one test failure)

I think there was a script to generate the build-details.json file; couldn't find it so quickly so I hardcoded the paths for now. Overall this was pretty smooth though. I handwrote the json file by taking the example from PEP 739 and just modifying for the target interpreter, took only a minute or two to do that.

[mgorny]

I've been able to successfully cross-compile NumPy with these changes.

Do you happen to have a reusable recipe for this?

Not really reusable, no. I've used Gentoo's crossdev to get the needed dependencies, i.e.:

crossdev aarch64-unknown-linux-gnu
aarch64-unknown-linux-gnu-emerge -1v openblas python:3.14

As a result, I've been able to test on a build-details.json from a real Python installation, which amounted to:

`build-details.json`

{
  "schema_version": "1.0",
  "base_prefix": "/usr",
  "base_interpreter": "/usr/bin/python3.14",
  "platform": "linux-aarch64",
  "language": {
    "version": "3.14",
    "version_info": {
      "major": 3,
      "minor": 14,
      "micro": 0,
      "releaselevel": "beta",
      "serial": 3
    }
  },
  "implementation": {
    "name": "cpython",
    "cache_tag": "cpython-314",
    "version": {
      "major": 3,
      "minor": 14,
      "micro": 0,
      "releaselevel": "beta",
      "serial": 3
    },
    "hexversion": 51249331,
    "_multiarch": "aarch64-linux-gnu"
  },
  "abi": {
    "flags": [],
    "extension_suffix": ".cpython-314-aarch64-linux-gnu.so",
    "stable_abi_suffix": ".abi3.so"
  },
  "suffixes": {
    "source": [
      ".py"
    ],
    "bytecode": [
      ".pyc"
    ],
    "extensions": [
      ".cpython-314-x86_64-linux-gnu.so",
      ".abi3.so",
      ".so"
    ]
  },
  "libpython": {
    "dynamic": "/usr/lib/libpython3.14.so",
    "dynamic_stableabi": "/usr/lib/libpython3.so",
    "link_extensions": false
  },
  "c_api": {
    "headers": "/usr/include/python3.14",
    "pkgconfig_path": "/usr/lib/pkgconfig"
  }
}

Created a cross file:

meson --cross-file

[binaries]
c = 'aarch64-unknown-linux-gnu-gcc'
cpp = 'aarch64-unknown-linux-gnu-g++'
pkg-config = 'aarch64-unknown-linux-gnu-pkg-config'

[properties]
sys_root = '/usr/aarch64-unknown-linux-gnu'
pkg_config_libdir = '/usr/aarch64-unknown-linux-gnu/usr/lib/pkgconfig'
needs_exe_wrapper = true
longdouble_format = 'IEEE_QUAD_LE'

[host_machine]
system = 'linux'
cpu_family = 'aarch64'
cpu = 'aarch64'
endian = 'little'

Merged this branch with #14773, copied mesonbuild/modules/features from NumPy and run:

./meson.py setup --python.build-config /usr/aarch64-unknown-linux-gnu/usr/lib/python3.14/build-details.json --cross-file ~/tmp/meson-cross.txt ../numpy/build ../numpy

[rgommers]

I opened mesonbuild/meson-python#779 for the needed meson-python changes that go with this. With this PR and that one, I could remove all the hacks and cross-compile SciPy correctly.