integers(x, x) can generate (x-1) if x is an int64

[dycw]

Hi,

Just stumbled upon this by accident. Using hypothesis 3.68.0 and numpy 1.15.2:

from hypothesis import given
from hypothesis.strategies import data, integers
from hypothesis.extra.numpy import from_dtype, integer_dtypes

@given(data=data(), dtype=integer_dtypes())
def test_numpy_integers(data, dtype):
    x = data.draw(from_dtype(dtype))
    assert data.draw(integers(x, x)) == x

For example, with the int64 type we can fail with x = 9007199254740993, with integers() generating x - 1. I've not seen examples of anything other than x - 1 being generated.

[Zalathar]

I had a quick look into this, and was surprised to find that floor(x) != x for some values of x that are numpy.int64.

Since we call floor and ceil on the arguments to integers(), this causes strange things to happen.

I haven't had a chance to figure out what's actually going on here.

[Zac-HD]

The return type of numpy.ceil is float64, just like math.ceil in Python 2. Apparently there is also an intermediate cast to float in the __ceil__ method of integer dtype values - for example x = 9007199254740993 = 2 ** 53 + 1, and np.ceil(x) == 9007199254740992.0. This is actually also the Python2 behaviour of math.ceil (which we work around in Hypothesis), so it's certainly not unprecedented!

I've added a comment to the upstream issue, but for us the best option is probably to use the Python 2 implementation on all versions for now.

[dycw]

Thanks @Zac-HD, so efficient!