The integer dtype used in array constructors need not be restricted to intp

[stuartarchibald]

This:

import numpy as np
from numba import njit

def foo():
    sz = np.int8(2)
    return np.zeros((sz, sz))

print(foo())
print(njit(foo)())

gives:

array([[ 0.,  0.],
       [ 0.,  0.]])
<traceback omitted>
LoweringError: Failed at nopython (nopython mode backend)
Operands must be the same type, got (i64, i8)

the root cause is a type mismatch in builder.mul as the arrlen const is fixed to types.intp:

numba/numba/targets/arrayobj.py

Lines 3042 to 3045 in b6b965a

	# compute array length
	arrlen = context.get_constant(types.intp, 1)
	for s in shapes:
	arrlen = builder.mul(arrlen, s)

and is probably fixable by obtaining the type(s) in the shape tuple here,

numba/numba/targets/arrayobj.py

Line 3091 in b6b965a

def _parse_shape(context, builder, ty, val):

and returning it as part of:

numba/numba/targets/arrayobj.py

Lines 3115 to 3122 in b6b965a

	def _parse_empty_args(context, builder, sig, args):
	"""
	Parse the arguments of a np.empty(), np.zeros() or np.ones() call.
	"""
	arrshapetype = sig.args[0]
	arrshape = args[0]
	arrtype = sig.return_type
	return arrtype, _parse_shape(context, builder, arrshapetype, arrshape)

and doing an appropriate cast to types.intp in the loop above that accumulates arrlen.

[stuartarchibald]

Closing, fixed by #4729