Closed
Description
Original ticket http://projects.scipy.org/numpy/ticket/1299 on 2009-11-19 by trac user crobin, assigned to unknown.
I think there is a bug with __setitem__ when using advanced selection on a structured array with objects when dimension is at least 2.
x = numpy.rec.fromarrays([[0,1],['a','b']], formats='i8,O')[None]
print x
x[x.nonzero()] = x.ravel()[:1]
print x
This gives:
[[(0, 'a') (1, 'b')]]
[[(0, 'a') (0, 'b')]] ## WRONG
The very last 'b' should be an 'a'.
As a bonus, the reference count of the string ‘b’ decreases, even though it still remains in the array x. So the interpreter will crash if you repeat the above:
while True:
x[x.nonzero()] = x.ravel()[:1]
...leads to:
Fatal Python error: deletion of interned string failed
zsh: IOT instruction (core dumped) python2.6
Note that this bug does not seem to occur when x is dimension 1, or when basic selection is used. The offset of the object in the dtype does not seem to matter.