[BUG] Auto-generation of "import_array" not working on recent Numpy

[da-woods]

Describe the bug

In #3524 I added auto-generation of np.import_array to help users out who'd forgotten to include it themselves. In that I relied on the NPY_NDARRAYOBJECT_H header include guard to identify "real" Numpy. In numpy/numpy@6dff0c2 they renamed their header include guard, thus breaking the fix.

See scipy/scipy#14732 and related issue #4394

Additional context

I'll submit a fix (and a more robust set of tests...) fairly shortly

[seberg]

Just to be sure, this does not affect any released versions of cython? Because if it does, I still think we should just modify NumPy back. Otherwise, there is probably indeed no point.

The SciPy issue uses newer features, but the important part seems only whether this automatic addition calling of import_array() is new.

[da-woods]

This affects the 3.0 pre-releases only (I don't know if you consider those "real" releases?).

In those we switched the Numpy .pxd files to use the Numpy API rather than looking directly into the PyArrayObject. It became obvious that was going to break a lot of user code so we added to automatic call of import_array in 3.0 alpha 2. The feature is not in the main 0.29.x releases

[seberg]

Ok, so the two features just happened to go hand-in-hand. Then I think we can indeed ignore this in NumPy: It is OK if alpha releases are broken :).

[charris]

It became obvious that was going to break a lot of user code

@seberg That might be a consideration. The guard check could be used to warn downstream users to fix things up, then at some point we could remove the NumPy fix.