Segmentation fault for np.empty_like(a).astype(bool) for an object array a

[Horstage]

Using Python 3.8.9 on linux with numpy 1.21.0 I get an Segmentation fault when typecasting an object array to bool which was created by np.empty_like.

Minimal working example:

import numpy as np

class Foo:
    pass

a = np.array([Foo(), Foo()], dtype=Foo)
b = np.empty_like(a)
b.astype(bool)

To replicate environment use docker container python:3.8-slim and pip install numpy.
e.g.

docker run -it --entrypoint=/bin/bash python:3.8-slim
pip install numpy==1.21.0

Workaround for those who still have to work with this:

b = np.copy(a)
b[:] = None
b.astype(bool)

[cbodenst]

Also reproducible with
np.empty_like([None]).astype(bool)

[seberg]

Ugly :/. I am wondering if 1.20 also does random stuff here, but for some reason doesn't run into the crash. Anyway, the fix is easy, the problem why this slipped past is that empty_like may well be the only reasonable way to create a half-initialized object array (an object array filled with NULL rather than None – we consider NULL and None to be equivalent).