Reshaped 0-size arrays have non-zero strides

[takagi]

Describe the issue:

Since NumPy 1.23, the strides of empty (0 size) arrays are set to zero as changed in #21477. However, once such arrays are reshaped, the results have non-zero strides computed in a C-contiguous manner.

>>> import numpy as np
>>> a = np.ones((2, 0, 3))
>>> a.strides
(0, 0, 0)
>>> b = a.reshape((3, 0, 2))
>>> b.strides
(16, 16, 8)  # also should be (0, 0, 0)

NumPy/Python version information:

1.23.0 3.10.0 (default, Oct 15 2021, 11:40:42) [GCC 7.5.0]

[eric-wieser]

As a reminder of why this matters; I think we concluded that it was undefined behavior for an array to contain a pointer to unallocated memory (which I think b[1:] does in your example); while in practice this doesn't seem to matter today, a future GCC optimization would be free to break things in arbitrary and possibly silent ways.

[seberg]

I would mostly be worried about crashes if some system thinks that e.g. an overflow in pointer arithmetic should send an abort signal?

@takagi is this important for you to be fixed ASAP for the 1.23.x series, and preferably 1.23.1?

[takagi]

I have no cases this problem immediately makes harm. (I'm in CuPy development, and found this case in its test)