Float to integer overflow doesn't match behaviour
Brought up in #408, we don't handle float conversion overflow correctly between x86 host and ARM.
We need a unit test to ensure correct behaviour here.