It turns out not so trivial to find the equivalent of "writable operand with 0 stride and usable dim > 0."
I thought I had a good handle on this but

• Reductions and accumulations can set an output stride = 0
• Fancy indexing uses an iterator that never really iterates, it is used to calculate the strides.

Some of that we can handle, but my guess is its either more complicated to detect (or needs a flag), or just doesn't belong in nditer but instead in the ufunc machinery in this case. Not sure though. Other pain points could be e.g. ufunc.at (but that is basically indexing).

EDIT: Ah well, dunno, have to think a bit more to recall how this all worked....

This may indeed be better done in the ufunc setup - the fact that reduce uses this suggests that for the iterator one should continue to allow maximum freedom.

Also, a different and I think better solution is to just make broadcasted arrays read-only; this is indeed why as_strided has a writable flag added to it in numpy 1.12 and broadcast_to returns read-only arrays. The idea was to eventually do the same for broadcast_arrays, though it is not so easy to see how to do that...

rebased to fix merge conflicts

I moved the check to the ufunc level, added the trivial test from the issue and modified memory_overlap tests to verify that if the error is triggered then ufunc(b, c, out=out) != ufunc(b, c). This should cover most of the common cases, without trying too hard to systematically catch them all.

I think I may have to think about it a while when I feel I have the time, but maybe Marten has a feel for it. First, I want to think a bit whether this approach is right (and if you were maybe right after all with the nditer approach). Second, I am a bit unsure whether a hard error is good.

I'm still not sure about this - I think I'd prefer to just not generate any writable broadcast arrays by default, and leaving any other usage up to the user, i.e., rather than prevent doing things that are potentially risky and fragile, make it such that it is a conscious choice to do so.

Closing. If any fix is needed it should be done in broadcast_arrays