Ufunc where parameter causes array to be silently read out of bounds

[sergiocallegari]

When the where argument of a ufunc has a shape that exceeds that of the input argument of the ufunc, arrays are silently read out of bounds. For instance

np.log([10], where=[True, False])
Out[58]: array([  2.30258509,  31.5       ])

Is this the intended behavior?

[pv]

Are you sure it's a read out of bounds? It can also be just uninitialized memory (because where=False for the element, no value is set to it and it is not zeroed).

[eric-wieser]

A pretty simple test is to reverse the where array:

>>> np.log([10], where=[False, True])
array([  5.03403489e+223,   2.30258509e+000])

An out of bounds error shouldn't mess up for entry 0, so this looks like it's the intended behaviour @pv describes.

[callegar]

OK, now I understand. Since the where is broadcast with the argument, the output get the length of the wider. The output array is allocated but not initialized. Because the function is not computed where "where" is false, some buckets in the output remain uninitialized.

Makes sense, sorry for the noise. Wonder if the documentation could be enriched a little.

[eric-wieser]

Wonder if the documentation could be enriched a little.

You're right, the descriptions don't suggest that at all right now. The relevant docstrings are in numpy\core\code_generators\ufunc_docstrings.py and doc\source\reference\ufuncs.rst

What you're seeing is the interaction between out == None and where != True - the docs for where only describe the case when out != None.

Feel free to submit a patch with improved wording, or even just file an issue suggesting the new docstring.