1d array, column of fortran array, with F and C contiguity, raises error for not being C-contiguous.

Thank you for this great and useful library 😃

I have noticed that functions like the following foo do not compile:

from numba import njit
import numpy as np


@njit('f8[::1](f8[::1])')
def foo2(b):
    return b

@njit('f8[::1](f8[::1,:])')
def foo(a):
    return foo2(a[:, 0])

The following error is raised:

TypingError: Invalid use of type(CPUDispatcher(<function foo2 at 0x000001759D4C7550>)) with parameters (array(float64, 1d, F))
Known signatures:
 * (array(float64, 1d, C),) -> array(float64, 1d, C)
During: resolving callee type: type(CPUDispatcher(<function foo2 at 0x000001759D4C7550>))

It indicates that the input argument given to function foo2 is not C-contiguous.

However, by checking the flags of fortran-arrays columns, we can see that they present both contiguities, e.g.:

arr = np.array([[1.0, 2.0], [3.0, 4.0]], order='F')
print(f'{arr[:, 0].flags}\n{arr[:, 1].flags}')

C_CONTIGUOUS : True
F_CONTIGUOUS : True
OWNDATA : False
WRITEABLE : True
ALIGNED : True
WRITEBACKIFCOPY : False
UPDATEIFCOPY : False

C_CONTIGUOUS : True
F_CONTIGUOUS : True
OWNDATA : False
WRITEABLE : True
ALIGNED : True
WRITEBACKIFCOPY : False
UPDATEIFCOPY : False

I think this is related to #5967. There a similiar problem appeared when using empty arrays with F-order.

Numba version: 0.56.3 (latest as of today).

[stuartarchibald]

That’s really helpful, thanks again!, and my bad for not realising that I could have used that kind of type signature.

No problem. Typically signatures are not needed as type inference just “works it out”, but those options above present the methods of describing just about anything within the type system if needed.

[javierttgg]

That’s really helpful, thanks again!, and my bad for not realising that I could have used that kind of type signature.