numpy calls interfere with external code in nopython=True functions

[cosama]

System

• conda python 3.7 with:
  • llvmlite 0.27.0-py37hd408876_0 --> 0.28.0-py37hd408876_0
  • numba 0.42.0-py37h962f231_0 --> 0.43.0-py37h962f231_0
• Tested on Fedora and Ubuntu 18.04

Minimal example

import numba
import numpy as np


@numba.jit(nopython=False)
def cross_issue(a=None):
    if a is None:
        a = 1
    np.sum([1, 2, 3])
    return a


print(cross_issue())

Expected result

• after the if statement a should be 1, however the function returns None.
• The behavior doesn't show up if the call to np.sum is removed. The weird behavior is also present with other functions (I originally discovered it with np.cross and saw it with np.mean too).
• I have honestly no idea how this even can be or how to debug it. Suggestions are welcome.

[cosama]

I forgot to mention, the above code works fine in numba 0.42.0 and 0.41.0.

[stuartarchibald]

Thanks for the report. What you are seeing is most likely due to this bug: #3879 which was fixed here: #3883, the recently released Numba 0.43.1 fixes this problem, perhaps try updating your Numba installation:

conda install numba=0.43.1

[cosama]

Yeah 0.43.1 fixed this. Sorry didn't find the other issue. I will close this then.

[stuartarchibald]

@cosama great, thanks for confirming.