@becker33 @scheibelp I wonder if we want to do this globally for all Python packages. The headers are almost always installed in prefix/lib/pythonX.Y/site-packages, not in prefix/include.

@adamjstewart would it make more sense to use find_headers with that directory name, or this method using find_all_headers?

The problem is we need to recursively search both prefix.lib and prefix.lib64.

@adamjstewart : Any news about this? Was it fixed globally?

This also relates to #14513. This hasn't yet been fixed globally but I think it should be relatively easy to do after the recent Python refactor I did. I'll try to submit a PR sometime this week to do that.

The headers are almost always installed in prefix/lib/pythonX.Y/site-packages, not in prefix/include.

This is actually a bad habit for Python packages. NumPy has multiple headers and library directories where it hides static libraries. SciPy uses all of them, here is what it needs to do:

# NumPy include directory - needed in all submodules
incdir_numpy = run_command(py3,
  [
    '-c',
    'import os; os.chdir(".."); import numpy; print(numpy.get_include())'
  ],
  check: true
).stdout().strip()

inc_np = include_directories(incdir_numpy)

incdir_f2py = incdir_numpy / '..' / '..' / 'f2py' / 'src'
inc_f2py = include_directories(incdir_f2py)
fortranobject_c = incdir_f2py / 'fortranobject.c'

cc = meson.get_compiler('c')
npymath_path = incdir_numpy / '..' / 'lib'
npymath_lib = cc.find_library('npymath', dirs: npymath_path)
npyrandom_path = incdir_numpy / '..' / '..' / 'random' / 'lib'
npyrandom_lib = cc.find_library('npyrandom', dirs: npyrandom_path)

I'd say that the best solution would be for NumPy to install into the correct/standard library and include directories, and/or add proper pkg-config support. Also build systems can (and will) support NumPy as a custom dependency. Spack support for custom within-sitepackages locations doesn't make as much sense to me, because if another package relies on that it will not work anywhere else (PyPI, conda-forge, etc.) so it only makes sense for end user code that's only used with Spack. And end users probably cargo cult solutions from either the NumPy docs or from the build files of other packages.

@rgommers you may be interested in #28527 which should hopefully solve this issue for all Python packages, not just numpy.