Fix regression #7355: cannot set items in structured array data types

[sklam]

This is a minimal fix for #7355 and does not replace #7359. We need a minimal fix for 0.54.1

Edit, adding details:

This is the easiest fix because:

• making the base Array.copy() return a covariant type is the reason of the regression for nestedarray.
• The new array subclass feature can override copy() to specialize accordingly.
• Requiring copy() is less invasive and array subclass support is new and experimental.

This change will require array subclass to override copy().

[esc]

Probably Fixes #7426 too.

[stuartarchibald]

Thanks for the patch. I've tested it with this... on HEAD of mainline with the diff to the tests that this patch introduces, the corresponding test fails. Later this PR is checked out and the same runs fine:

$ git rev-parse HEAD
ed6ae6d3c2a615d329604c96e034efc876cf8230

$ git diff
diff --git a/numba/tests/test_record_dtype.py b/numba/tests/test_record_dtype.py
index 8f1137f..699017e 100644
--- a/numba/tests/test_record_dtype.py
+++ b/numba/tests/test_record_dtype.py
@@ -315,6 +315,11 @@ def set_field4(rec):
     return rec
 
 
+def set_field_slice(arr):
+    arr['k'][:] = 0.0
+    return arr
+
+
 recordtype = np.dtype([('a', np.float64),
                        ('b', np.int16),
                        ('c', np.complex64),
@@ -865,6 +870,10 @@ class TestRecordDtype(unittest.TestCase):
         cfunc = self.get_cfunc(pyfunc, (nbrecord,))
         self.assertEqual(cfunc(rec), pyfunc(rec))
 
+        pyfunc = set_field_slice
+        cfunc = self.get_cfunc(pyfunc, (nbrecord,))
+        np.testing.assert_array_equal(cfunc(rec), pyfunc(rec))
+
     def test_structure_dtype_with_titles(self):
         # the following is the definition of int4 vector type from pyopencl
         vecint4 = np.dtype([(('x', 's0'), 'i4'), (('y', 's1'), 'i4'),

$ ./runtests.py numba.tests.test_record_dtype.TestRecordDtype.test_record_two_arrays
E
======================================================================
ERROR: test_record_two_arrays (numba.tests.test_record_dtype.TestRecordDtype)
Tests that comparison of NestedArrays by key is working correctly. If
----------------------------------------------------------------------
Traceback (most recent call last):
<snip>
numba.core.errors.TypingError: No implementation of function Function(<built-in function setitem>) found for signature:
 
 >>> setitem(nestedarray(int32, (10, 20)), slice<a:b>, float64)
 
There are 16 candidate implementations:
  - Of which 14 did not match due to:
  Overload of function 'setitem': File: <numerous>: Line N/A.
    With argument(s): '(nestedarray(int32, (10, 20)), slice<a:b>, float64)':
   No match.
  - Of which 2 did not match due to:
  Overload in function 'SetItemBuffer.generic': File: numba/core/typing/arraydecl.py: Line 171.
    With argument(s): '(nestedarray(int32, (10, 20)), slice<a:b>, float64)':
   Rejected as the implementation raised a specific error:
     TypeError: __init__() got an unexpected keyword argument 'ndim'

$ #git stash/checkout this PR etc.

$ git rev-parse HEAD
1c3c10a528fb3e5ca0071886af49cc700a811577

$ ./runtests.py numba.tests.test_record_dtype.TestRecordDtype.test_record_two_arrays
.
----------------------------------------------------------------------
Ran 1 test in 0.584s

OK

I've also manually tested this PR against all the examples in #7355 and #7426, and those reported here: https://numba.discourse.group/t/883/7

[stuartarchibald]

Note: This is as 0.53 https://github.com/numba/numba/blob/release0.53/numba/core/types/npytypes.py#L441-L451