NumPy seems to attempt to iterate instead of calling __array__.

Consider:

class ArrayLike(object):
  def __init__(self, array):
    self.array = array

  def __len__(self):
    return len(self.array)
  
  def __iter__(self):
    print('calling __iter__')
    return iter(self.array)
  
  def __getitem__(self, index):
    print('calling __getitem__ with index={}'.format(index))
    return self.array[index]

  def __array__(self, dtype=None):
    print('calling __array__')
    return np.asarray(self.array, dtype=dtype)

>>> a = ArrayLike(np.arange(3))

>>> np.array(a)
calling __array__
array([0, 1, 2])

>>> list(a)
calling __iter__
[0, 1, 2]

>>> np.array([a])
calling __array__
calling __iter__
calling __iter__
array([[0, 1, 2]])

So actually, it looks like NumPy calls __array__, and then __iter__ twice!

Is there a good reason for this behavior? If not, maybe we can fix it? Either way, a pointer to the relevant place in the codebase would be appreciated.

This causes trouble for xarray users (pydata/xarray#1247) because iterating over xarray objects creates new xarray objects, which takes ~100 us each. That adds up fast for large arrays.