BUG: Fix in-place permutation

[bashtage]

Alter test to check if arrays are the same to avoid in-place modification of some array-like objects

closes #11975

[bashtage]

It is hard to write the ideal test that would fail on master but pass in this PR since it requires a non-trivial array-like object that is False when tested with is equality after the np.asarray but shares memory. RangeIndex from pandas has these attributed. I wrote a test with a recarray but it passes on master as well.

Are there any classes like this in NumPy? Or are there other tests that define any that I could borrow?

[mattip]

class N(np.ndarray):
    pass
a = np.arange(10).view(N)
b = np.asarray(a)
b is a  # False
np.may_share_memory(a, b)  # True

[bashtage]

Thanks. works perfectly.

[charris]

Might try

In [1]: class Foo(object):
   ...:     a = np.arange(5)
   ...:     def __array__(self):
   ...:         return self.a
   ...:     

In [2]: a = Foo()

In [3]: a.__array__()
Out[3]: array([0, 1, 2, 3, 4])

In [4]: np.random.permutation(a)
Out[4]: array([0, 1, 3, 4, 2])

In [5]: a.__array__()
Out[5]: array([0, 1, 3, 4, 2])

As a test that currently fails. I suspect that is closer to the original.

EDIT: The failing pandas case is from an object that subclasses Index, which in turn has

    def __array__(self, dtype=None):
        """ the array interface, return my values """
        return self._data.view(np.ndarray)

so it is the use of __array__ that caused the original problem.

[charris]

I'm not convinced that we can solve this problem in all generality without just making a copy, but we can make a start :)

[charris]

Thanks @bashtage.