DEP: Do not allow "abstract" dtype conversion/creation

These dtypes do not really make sense as instances. We can (somewhat)
reasonably define np.dtype(np.int64) as the default (machine endianess)
int64. (Arguably, it is unclear that `np.array(arr_of_>f8, dtype="f")`
should return arr_of_<f8, but that would be very noisy!)

However, `np.integer` as equivalent to long, is not well defined.
Similarly, `dtype=Decimal` may be neat to spell `dtype=object` when you
intend to put Decimal objects into the array. But it is misleading,
since there is no special meaning to it at this time.

The biggest issue with it, is that `arr.astype(np.floating)` looks
like it will let float32 or float128 pass, but it will force a
float64 output! Arguably downcasting is a bug in this case.

A related issue is `np.dtype("S")` and especially "S0". The dtype "S"
does make sense for most or all places where `dtype=...` can be
passed. However, it is conceptionally different from other dtypes, since
it will not end up being attached to the array (unlike "S2" which
would be). The dtype "S" really means the type number/DType class
of String, and not a specific dtype instance.

1 files changed, 11 insertions, 6 deletions

diff --git a/numpy/lib/_iotools.py b/numpy/lib/_iotools.py
index 251d2d2a7..48d130bac 100644
--- a/numpy/lib/_iotools.py
+++ b/numpy/lib/_iotools.py
@@ -504,18 +504,23 @@ class StringConverter:
     """
     #
     _mapper = [(nx.bool_, str2bool, False),
-               (nx.integer, int, -1)]
+               (nx.int_, int, -1),]
 
     # On 32-bit systems, we need to make sure that we explicitly include
-    # nx.int64 since ns.integer is nx.int32.
-    if nx.dtype(nx.integer).itemsize < nx.dtype(nx.int64).itemsize:
+    # nx.int64 since ns.int_ is nx.int32.
+    if nx.dtype(nx.int_).itemsize < nx.dtype(nx.int64).itemsize:
         _mapper.append((nx.int64, int, -1))
 
-    _mapper.extend([(nx.floating, float, nx.nan),
-                    (nx.complexfloating, complex, nx.nan + 0j),
+    _mapper.extend([(nx.float64, float, nx.nan),
+                    (nx.complex128, complex, nx.nan + 0j),
                     (nx.longdouble, nx.longdouble, nx.nan),
                     (nx.unicode_, asunicode, '???'),
-                    (nx.string_, asbytes, '???')])
+                    (nx.string_, asbytes, '???'),
+                    # If a non-default dtype is passed, fall back to generic
+                    # ones (should only be used for the converter)
+                    (nx.integer, int, -1),
+                    (nx.floating, float, nx.nan),
+                    (nx.complexfloating, complex, nx.nan + 0j),])
 
     (_defaulttype, _defaultfunc, _defaultfill) = zip(*_mapper)