Byteswapped complex256 on Debian PPC randomly incorrect subtraction (Trac #2174)

[numpy-gitbot]

Original ticket http://projects.scipy.org/numpy/ticket/2174 on 2012-06-20 by atmention:matthew-brett, assigned to unknown.

PPC 32-bit on Debian Wheezy, numpy 1.6.2 and current numpy trunk:

    In [2]: np.__version__
    Out[2]: '1.7.0.dev-651ef74'

import numpy as np

arr = np.arange(10, dtype=np.complex256)
bs_arr = arr.byteswap().newbyteorder('S')
print arr
print bs_arr
print arr - bs_arr
print arr - bs_arr
print arr - bs_arr

The result of subtraction is wrong most (~90%) of the time; with the byteswapped array acting as if it is all zeros.

(np-devel)[mb312atmention:joshlegacy ~/tmp]$ python funny_bs.py
[ 0.0+0.0j  1.0+0.0j  2.0+0.0j  3.0+0.0j  4.0+0.0j  5.0+0.0j  6.0+0.0j
 7.0+0.0j  8.0+0.0j  9.0+0.0j]
[ 0.0+0.0j  1.0+0.0j  2.0+0.0j  3.0+0.0j  4.0+0.0j  5.0+0.0j  6.0+0.0j
 7.0+0.0j  8.0+0.0j  9.0+0.0j]
[ 0.0+0.0j  1.0+0.0j  2.0+0.0j  3.0+0.0j  4.0+0.0j  5.0+0.0j  6.0+0.0j
 7.0+0.0j  8.0+0.0j  9.0+0.0j]
[ 0.0+0.0j  1.0+0.0j  2.0+0.0j  3.0+0.0j  4.0+0.0j  5.0+0.0j  6.0+0.0j
 7.0+0.0j  8.0+0.0j  9.0+0.0j]
[ 0.0+0.0j  1.0+0.0j  2.0+0.0j  3.0+0.0j  4.0+0.0j  5.0+0.0j  6.0+0.0j
 7.0+0.0j  8.0+0.0j  9.0+0.0j]

About 10% of the time the arrays appear to behave correctly:

(np-devel)[mb312atmention:joshlegacy ~/tmp]$ python funny_bs.py
[ 0.0+0.0j  1.0+0.0j  2.0+0.0j  3.0+0.0j  4.0+0.0j  5.0+0.0j  6.0+0.0j
 7.0+0.0j  8.0+0.0j  9.0+0.0j]
[ 0.0+0.0j  1.0+0.0j  2.0+0.0j  3.0+0.0j  4.0+0.0j  5.0+0.0j  6.0+0.0j
 7.0+0.0j  8.0+0.0j  9.0+0.0j]
[ 0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j
 0.0+0.0j  0.0+0.0j  0.0+0.0j]
[ 0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j
 0.0+0.0j  0.0+0.0j  0.0+0.0j]
[ 0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j
 0.0+0.0j  0.0+0.0j  0.0+0.0j]

More rarely there is a mix of incorrect and correct. Notice below that the first pass of arr - bs_arr is incorrect, and the second and third repeat of exactly the same statement on the same variables are correct.

(np-devel)[mb312atmention:joshlegacy ~/tmp]$ python funny_bs.py 
[ 0.0+0.0j  1.0+0.0j  2.0+0.0j  3.0+0.0j  4.0+0.0j  5.0+0.0j  6.0+0.0j
  7.0+0.0j  8.0+0.0j  9.0+0.0j]
[ 0.0+0.0j  1.0+0.0j  2.0+0.0j  3.0+0.0j  4.0+0.0j  5.0+0.0j  6.0+0.0j
  7.0+0.0j  8.0+0.0j  9.0+0.0j]
[ 0.0+0.0j  1.0+0.0j  2.0+0.0j  3.0+0.0j  4.0+0.0j  5.0+0.0j  6.0+0.0j
  7.0+0.0j  8.0+0.0j  9.0+0.0j]
[ 0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j
  0.0+0.0j  0.0+0.0j  0.0+0.0j]
[ 0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j  0.0+0.0j
  0.0+0.0j  0.0+0.0j  0.0+0.0j]

Float128 and other numpy dtypes appear to be correct.

I'm happy to give access to this machine if someone needs to debug.