Cross target use of target-less @overload leads to lowering errors.

[stuartarchibald]

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/main/CHANGE_LOG).
• I have included a self contained code sample to reproduce the problem.
  i.e. it's possible to run as 'python bug.py'.

This incorrect use of @overload is actually a simplified version of the issue in #8529. Note that the @overload has no target= kwarg, as such the CUDA target cannot find it, but also as the target is merely implied as CPU to preserve existing behaviours, this manifests as a lowering error as there's no bar function implementation in the target context for CUDA.

import numpy as np
from numba.cuda import jit
from numba.extending import overload

def bar():
    pass


@overload(bar)
def ol_bar(x):
    def impl(x):
        return 12
    return impl


@jit
def call_bar(arr):
    arr[0] = bar(3)

tmp = np.zeros((1,))
call_bar[1, 1](tmp)
print(tmp)

The error message ends with:

NotImplementedError: No definition for lowering <... Closure>(int64,) -> Literal[int](12)

which is not ideal.

If the example is updated so that the overload is declared as cpu only via @overload(bar, target='cpu'), then the error message is more helpful:

Function resolution cannot find any matches for function '<function bar at 0x<snip>>` for the current target: '<class 'numba.core.target_extension.CUDA'>'.