[NumPy-Tickets] [NumPy] #1640: structured array incorrectly gives a broadcast error

[NumPy Trac]

#1640: structured array incorrectly gives a broadcast error
------------------------+---------------------------------------------------
 Reporter:  m-paradox   |       Owner:  somebody
     Type:  defect      |      Status:  new     
 Priority:  normal      |   Milestone:  2.0.0   
Component:  numpy.core  |     Version:  devel   
 Keywords:  structured  |  
------------------------+---------------------------------------------------
Changes (by m-paradox):

  * keywords:  => structured
  * component:  Other => numpy.core


Comment:

 It looks like the problem is caused by two features conflicting:
  * dtypes with fields are compared one field at a time, and ANDed together
  * types with a shape have their shape appended to the array's shape
 Here's what the code is effectively doing to get the failure, in function
 _void_compare of arrayobject.c
 {{{
 >>> p = np.rec.recarray((2,),dtype=[('f0','f4',3),('f1','f4',1)])
 >>> q = p.copy()
 # in _void_compare:
 >>> a = p.f0             # line 836
 >>> b = q.f0             # line 841
 >>> temp = a==b          # line 847
 >>> temp.shape      # (problem here) _void_compare assumes this has the
 same
 (2, 3)              # shape as p and q, but it doesn't
 >>> res = temp           # line 855
 >>> a = p.f1             # line 836
 >>> b = q.f1             # line 841
 >>> temp = a==b          # line 847
 >>> temp.shape
 (2,)                # this is the shape that _void_compare expects
 >>> temp2 = res & temp   # line 858
 Traceback (most recent call last):
   File "<stdin>", line 1, in <module>
 ValueError: shape mismatch: objects cannot be broadcast to a single shape

 }}}

-- 
Ticket URL: <http://projects.scipy.org/numpy/ticket/1640#comment:1>
NumPy <http://projects.scipy.org/numpy>
My example project