Can you link to where the example comes from? (edit: updated top post with working link)

It is in the documentation, under the Iterating Over Arrays section - Reduction Iteration example for summation of all elements in an array
Here is the link https://docs.scipy.org/doc/numpy-1.13.0/reference/arrays.nditer.html

Looks like it fails for the same reason that the following does:

x = np.lib.stride_tricks.as_strided(0, shape=(4,), strides=(0,))
x += np.arange(len(x))
x
# array([3, 3, 3, 3]), expecting array([6, 6, 6, 6])

Strangely, using np.add(y, x, out=y) works correctly

Looks like a regression in 1.13 - 1.12 and 1.11 work correctly.

I suspect this is due to the overlap-detection added in 1.13 (#8043, @pv)

It seems the out= case was deliberately made an exception

Thanks for chiming in

Looking at these lines from @pv's PR, it seems that this change was deliberate, and the out argument behavior is a lingering bug (I can no longer reproduce this), and the documentation is out of date

Yeah, I somewhat guess it is a change in behaviour that was in some sense anticipated. Is this actually creating problems for you? The nditer API, is a bit too close to the C-API for its own good/nice usage sometimes....

Closing, this was never a bug/regression, but a deliberate change in our API to protect the vast majority of users (of course some like me in my early days had code that did use this).