The function linalg.cross has a current bug. With the current development version, we get

>>> u = [1, 2, 3]
>>> v = [4, 5, 6]
>>> np.linalg.cross(u,v)
...
AttributeError: 'list' object has no attribute 'shape'

It's missing an asanyarray or asarray before asking for the shape. Which is preferrable?

After either change is made, we get:

>>> u = [1, 2, 3]
>>> v = [4, 5, 6]
>>> np.linalg.cross(u,v)
array([-3,  6, -3])

I also addded a test to check that lists are converted properly to arrays.

There is another issue happening, namely an IndexError can still occur, rather than an AxisError. Compare the behavoir of linalg.cross and cross.

>>> u = np.asarray([1, 2, 3])
>>> v = np.asarray([4, 5, 6])
>>> np.linalg.cross(u, v, axis=-2)
...
IndexError: tuple index out of range

>>> u = np.asarray([1, 2, 3])
>>> v = np.asarray([4, 5, 6])
>>> np.cross(u, v, axis=-2)
...
AxisError: axisa: axis -2 is out of bounds for array of dimension 1

I'm not sure if we want to call this a bug, or a feature, that different error messages are thrown. I had to deal with this while creating linalg.cross2d and noticed the issue. The issue comes by not using normalize_axes_index in linalg.cross before trying to access the shape. Fixing it would result in calling normalize_axis_index twice (as the check is done in cross).

I can addreses the inconsistency of error messages in this PR, or a separate, or just call it a feature. Either way, a user is being told they have an indexing issue, though I think the AxisError message is more informative.

@charris this one would be nice to backport for 2.0.0