Probably missing priority forwarding for inplace operators.

Probably indeed __array_priority__ handling. In previous numpy:

>>> a.__iadd__(sp.lil_matrix(a))
NotImplemented

current master:

>>> a.__iadd__(sp.lil_matrix(a))
Traceback (most recent call last):
  File "<ipython-input-3-c4acac29d15e>", line 1, in <module>
    a.__iadd__(sp.lil_matrix(a))
TypeError: ufunc 'add' output (typecode 'O') could not be coerced to provided output parameter (typecode 'l') according to the casting rule ''same_kind''

Hmm, we are still depending on ufuncs returning NotImplemented, that doesn't seem right. This seems like the sort of thing that should be handled at the ndarray level.

This fixes the inplace add

diff --git a/numpy/core/src/multiarray/number.c b/numpy/core/src/multiarray/number.c
index 0d667ac..2ee2c2d 100644
--- a/numpy/core/src/multiarray/number.c
+++ b/numpy/core/src/multiarray/number.c
@@ -628,6 +628,7 @@ array_bitwise_xor(PyArrayObject *m1, PyObject *m2)
 static PyObject *
 array_inplace_add(PyArrayObject *m1, PyObject *m2)
 {
+    BINOP_GIVE_UP_IF_NEEDED(m1, m2, nb_add, array_add);
     return PyArray_GenericInplaceBinaryFunction(m1, m2, n_ops.add);
 }

The result is that a is replaced by a matrix, python rules, but that is as before.

I seems that all the inplace operators suffer the same omission. Major test failure ;)

@juliantaylor What happens to the elision code if the inplace function returns NotImplemented?

EDIT: Looks like that is already checked before elision is attempted, so should not be a problem. However, the same check is now repeated twice in the path if elision is done, so might want to bypass that at some point.

Called it ;)

There's a hidden discussion here with some rationale

The result is that a is replaced by a matrix, python rules, but that is as before.

I think this was behaviour that was ruled as undesirable in that discussion. Evidently, it's too late for us to decide that.

This needs a doc fix as well as @charris's patch

See #9021.

@pv Would it be appropriate to raise a DeprecationWarning in this case while preserving past behavior? The return value really isn't consistent with the ndarray inplace behavior.