ENH: ufunc should warn when argument where is specified and out is default None.

[Game4Move78]

Proposed new feature or change:

My argument is that where makes sense unless out is default None, in which case it simply obscures the creation of a partially uninitialised array.

The casual numpy user accessing through the public API likely only wants garbage values if they have explicitly asked for them using empty_like. If they have decided not to provide out, it probably means they expect a partially transformed copy of the input array and have overlooked the note in the docstring.

Example:

User creates a function that takes log on values where log is defined .

def safe_log(val, out):
    return np.log(val, where=np.greater(val, 0), out=out)

User suddenly decides that they don't want to specify out every time, making safe_log unsafe.

def safe_log(val, out=None):
    return np.log(val, where=np.greater(val, 0), out=out)

[eric-wieser]

The difficulty in using empty_like is that you now have to guess the right return type rather than letting np.log pick it for you.

[Game4Move78]

@eric-wieser Thanks for your response. Is the difficulty with the return type of empty_like that it doesn't create scalars?