purescript-prelude
purescript-prelude copied to clipboard
Separate exact and inexact division
This would be a pretty seismic breaking change, so I would not be at all surprised if we end up deciding to close this issue on this basis, but it comes up semi-frequently and I think it is at least worth considering and having a discussion to link to.
At the moment, we have just one division operator, div
aka /
, which is provided by the EuclideanRing class. This class also provides a mod
operator:
class EuclideanRing a where
div :: a -> a -> a
mod :: a -> a -> a
The mod
operator is required to be compatible with /
in the sense that a = (a / b) * b + a `mod` b
. This means that for types that support exact division, such as Rational
and Number
, we are required to set mod _ _ = 0
. In a theoretical sense this is fine: every field is a Euclidean ring, and all of the laws are satisfied. Practically, however, this can be a bit of a footgun:
- PureScript beginners often expect
Prelude.mod
to behave the same way as the%
operator in JS and other languages, so that e.g.5.0 `mod` 2.0 = 1.0
and8.0 `mod` 3.0 = 2.0
. In my view, it is almost always a mistake to usemod
with a type that is aField
. - Functions like
lcm
andgcd
which make use ofmod
internally are likely to produce unexpected results when used with types that haveField
instances. For example,lcm = const
andgcd = flip const
for any type that has aField
instance. It's probably almost always an error to uselcm
orgcd
with types that haveField
instances.
If something is probably almost always an error, it suggests that we should be leaning on the compiler to let us know when we are doing it. My proposal is that we define a new division operator to separate inexact/integer division from exact division, as follows:
- Add
exactDiv :: a -> a -> a
toField
, with the requirement thatexactDiv a b = a * recip b
(provided thatb
is nonzero). TheexactDiv
function could be a member ofField
or it could be a separate function with aField
constraint, but I think the former would be nice to allow people to supply more efficient implementations; division of floating point numbers is usually implemented in hardware, for example. - Weaken the superclass constraints on
Field
from(EuclideanRing a, DivisionRing a)
to(CommutativeRing a, DivisionRing a)
- Change the operator alias
/
to refer toexactDiv
- Add a new operator alias
//
for EuclideanRing'sdiv
, which we could describe as "integer division".
After making these changes, we are free to separate the behaviour of mod
from that of /
. We can choose to either remove EuclideanRing
instances for types like Number
and Rational
entirely, so that using mod
on them is a type error, or we can choose to provide instances with a more Euclidean-ring-like behaviour. For example, we would be free to define mod
on Number
so that it does behave (more or less) the same way as JS's %
; in this case we'd have to define //
such that it always returns a integer, which is probably appropriate since we would be describing //
as "integer division". For example we might have:
-
4.5 // 1.25 = 3.0
- since1.25 * 3.0 = 3.75
-
4.5 `mod` 1.25 = 0.75
matching JS's%
which expresses that 1.25 goes into 4.5 three times, with 0.75 left over. Doing this would also allow lcm
and gcd
to behave more sensibly; with this instance, lcm
and gcd
would work the same way on integral Numbers as they would on Int. For fractional Numbers it'd still be a bit weird because of floating point inaccuracy, but for rationals we'd be able to have eg. lcm (1%2) (1%3) = 1%1
.
This approach is partially inspired by what Python does: /
is for exact division and //
is for integer division. https://docs.python.org/3.1/tutorial/introduction.html#numbers