BUG: fix bool no-op replace calls

[cpcloud]

closes #6332

[jreback]

if this is true, then isn't the result a noop? (regardless if its a bool or not?)

[cpcloud]

no. the values_to_mask must not be castable specifically to bool and arr.dtype must be be bool, that's what the arr.dtype == np.bool_ is there for.

[jreback]

and you don't want to have a BoolBlock handler instead of this?

[cpcloud]

actually yes you're right ... don't know why i didn't of this before but i can just override putmask in boolblock and check for the new dtype there

[cpcloud]

@jreback look okay to u?

[jreback]

why does this work?

[cpcloud]

For the same reason I mentioned in #6332. np.atleast_1d is a call to array plus some insurance that you get at least a 1D array. Using the SO example, that will create an array of one string. S dtypes cannot be safely cast to bool therefore I pass in the array of strings which will ultimately raise ValueError in mask_missing which will convert them to object dtype. Then mask_missing procedes to compare each non NA value with those in arr. Since no strings are equal to True or False, the mask is an array of Falses. Does that suffice?

[cpcloud]

dicts work too

In [18]: s
Out[18]:
0     True
1    False
2     True
dtype: bool

In [19]: s.replace({'asdf': 'asdb', True: 'a'})
Out[19]:
0        a
1    False
2        a
dtype: object

In [20]: s.replace({False: 'asdb', True: 'a'})
Out[20]:
0       a
1    asdb
2       a
dtype: object

If there's no test I'll add one

[cpcloud]

One thing that won't work as "expected" is df.replace({True: False, False: True}). The keys are iterated over in an arbitrary order and you'll end up with either all Trues or all Falses. I don't really think this is that common.

[jreback]

merge away

neat trick with atleast1 d