[Numpy-discussion] Ill-defined in-place operations (#1085)

[Pauli Virtanen]

Mixing views and mutating (eg. in-place) operations can cause surprising 
and ill-defined results. Consider 
http://projects.scipy.org/numpy/ticket/1085:

>>> import numpy as np
>>> x = np.array([[1,2], [3,4]])
>>> x
array([[1, 2],
       [3, 4]])
>>> x += x.T
>>> x
array([[2, 5],
       [8, 8]])
>>> y = np.array([[1,2], [3,4]], order='F')
>>> y
array([[1, 2],
       [3, 4]])
>>> y += y.T
>>> y
array([[2, 7],
       [5, 8]])

The result depends on the order in which the elements happen to lie in 
the memory. Predicting the outcome is nearly impossible. (Also, I think 
Numpy tries to optimize the order of the loops, making it even more 
difficult?)

This is a sort of a pitfall. Should Numpy issue a warning every time a 
mutating operation is performed on an array, with input data that is a 
view on the same array?

Some alternatives:

  a) Raise warning for all arrays, even 1-D.

  b) Raise warning only for multidimensional arrays. (Assume the user can
     predict what happens in 1-D.)

  c) Don't raise any warnings, just educate users in the documentation.

This is not the first time this has been discussed on this ML, but the 
last discussion did not (IIRC) lead to any conclusions. (But I think it 
was established that detecting when views really overlap is a problem 
with no cheap solution. So walking up ->base pointers or comparing ->data 
pointers seemed nearly the only feasible way.)

-- 
Pauli Virtanen