Passing in dtype to array fails on NumPy master in pip-installed version

Reporting a bug

• I have tried using the latest released version of Numba (most recent is visible in the change log (https://github.com/numba/numba/blob/master/CHANGE_LOG).
• I have included below a minimal working reproducer (if you are unsure how to write one see http://matthewrocklin.com/blog/work/2018/02/28/minimal-bug-reports).

Please note that this doesn’t happen on the conda-installed version, only pip-installed

import numba
import numpy as np

dt1 = np.dtype("float64")
dt2 = np.dtype("float64")

dtr = np.result_type(dt1, dt2)

@numba.jit(
    nopython=True,
    nogil=True,
    locals={"data_curr": numba.np.numpy_support.from_dtype(dtr)},
)
def _dot_coo_coo(coords1, data1, coords2, data2):  # pragma: no cover
    return np.empty((0,), dtype=dtr)

coords1 = np.array([], dtype=np.int64).reshape((2, 0))
coords2 = coords1

data1 = np.array([], dtype=dt1)
data2 = data1

_dot_coo_coo(coords1, data1, coords2, data2)

Output

Traceback (most recent call last):
  File "sandbox.py", line 23, in <module>
    _dot_coo_coo(coords1, data1, coords2, data2)
TypeError: expected dtype object, got 'numpy.dtype[float64]'

Original issue by @TomAugspurger in pydata/sparse#383

Package    Version
---------- -------------------
llvmlite   0.33.0
numba      0.50.1
numpy      1.20.0.dev0+1ce5457
pip        20.1.1
scipy      1.5.2
setuptools 49.2.0
sparse     0.10.0
wheel      0.34.2

[stuartarchibald]

Note from Numba core dev meeting, preference is to fix up Numba’s use of the macro so as to not have to move NumPy base versions.

[eric-wieser]

This is probably because there is an ABI change in numpy 1.20, which means that numba must be recompiled. If you rebuild numpy master from source then numba from source, I expect the problem to go away.