Division broken when dtype specified

In a head build using python 3:

>>> np.__version__
'1.8.0.dev-069e9b0'
>>> np.divide(.4, 1.0, dtype=np.float)
0.0
>>> np.divide(.4, 1.0)
0.40000000000000002

[Dobatymo]

there are still problems with division: https://stackoverflow.com/questions/45278112/numpy-integer-division-sometimes-yields-wrong-results-when-casted

>>> np.true_divide(200, 100, dtype=np.float64)
-0.56000000000000005
>>> np.true_divide(200, 100, dtype=np.float32)
2.0

python 3, numpy 1.13.1

[njsmith]

Yeah, the lack of a coherent theory about what dtype means bothers me too. Defining it to mean “this is the output type I want” would make some sense, and I think actually work if our type resolution code would try to pick the “best” intermediate type instead of casting int64 to float16, but (a) that’s not how our type resolution code works, and (b) I feel like there are cases where dtype is used to indicate intermediate precision or similar.