Setting ma fill_value of object and long int dtypes doesn't work if the fill_value has already been used

[eric-wieser]

This relates to this comment

# without reading `fill_value`
>>> a = np.ma.array([None])
>>> a.fill_value = object()
>>> a.fill_value
<object object at 0x000000000066A0E0>

# with reading `fill_value`
>>> a = np.ma.array([None])
>>> a.fill_value
'?'
>>> a.fill_value = object()
>>> a.fill_value
'<'

This fails because after reading fill_value for the first time, _fill_value.dtype is S1, which becomes fixed for the lifetime of the object.

In general, the problem is that a.fill_value.dtype is not a.dtype:

>>> np.ma.array(1, np.int8).fill_value.dtype
dtype('int32')  # maybe defensible, handy if the array is upcast later
>>> np.ma.array(1, np.int64).fill_value.dtype
dtype('int32')  # unacceptable loss of range (maybe windows only)!
>>> np.ma.array(1, object).fill_value.dtype
dtype('S1')  # this issue

It's tricky to fix this without breaking code that assumes np.ma.array(1, np.int8).fill_value = 999999. Do we care about that code?