`irfftn` corrupts input when dimensions >=3

[rainwoodman]

Originally posted as numpy/numpy#10895

Seems to be introduced recently. It doesn't affect fftn, ifftn or rfft, irfft.

Here is the reproducer.

import numpy
from numpy.testing import assert_allclose

def showbug(nd):
    size = [4] * nd
    x = numpy.random.uniform(size=size)
    y = numpy.fft.rfftn(x)
    y1 = y.copy()

    numpy.fft.irfftn(y1)

    assert_allclose(y1, y)

showbug(3)

Traceback (most recent call last):
  File "fftbug.py", line 14, in <module>
    showbug(3)
  File "fftbug.py", line 12, in showbug
    assert_allclose(y1, y)
  File "/home/yfeng1/anaconda3/install/lib/python3.6/site-packages/numpy/testing/utils.py", line 1395, in assert_allclose
    verbose=verbose, header=header, equal_nan=equal_nan)
  File "/home/yfeng1/anaconda3/install/lib/python3.6/site-packages/numpy/testing/utils.py", line 778, in assert_array_compare
    raise AssertionError(msg)
AssertionError: 
Not equal to tolerance rtol=1e-07, atol=0

(mismatch 100.0%)
 x: array([[[ 1.612765+0.j      , -0.069468+0.278309j, -1.196650+0.j      ],
        [ 1.794408+0.j      ,  0.084526+0.187052j,  0.642339+0.j      ],
        [ 1.536380+0.j      , -0.811544-0.006539j,  0.736002+0.j      ],...
 y: array([[[  3.227387e+01+0.j      ,  -6.852522e-01-0.432786j,
          -2.638166e+00+0.j      ],
        [ -9.017300e-01+2.401103j,   1.732144e+00+0.885346j,...

[rainwoodman]

Seems to be related to the logic in the _datacopied() function returned True when it is not copied. Perhaps better to pass along the information on the stack explicitly rather than detecting it.

[rainwoodman]

Thanks -- when will this go to a release?

[oleksandr-pavlyk]

I will modify the conda-forge/mkl_fft-feedstock later today, which will trigger a new build shortly thereafter.