A pattern that presents itself the more often the more type safety is being introduced via newtype
is to project a value (or several values) to a newtype
wrapper, do some operations, and then retract the projection. An ubiquitous example is that of Sum
and Product
monoids:
λ x + y = getSum $ Sum x `mappend` Sum y
λ 1 + 2
3
I imagine a collection of functions like withSum
, withSum2
, and so on, may be automagically rolled out for each newtype
. Or maybe a parametrized Identity
may be created, for use with ApplicativeDo
. Or maybe there are some other approaches that I could not think of.
I wonder if there is some prior art or theory around this.
P.S. I am unhappy with coerce
, for two reasons:
safety I thought it is not very safe. After being pointed that it is actually safe, I tried a few things and I could not do anything harmful, because it requires a type annotation when there is a possibility of ambiguity. For example:
λ newtype F = F Int deriving Show
λ newtype G = G Int deriving Show
λ coerce . (mappend (1 :: Sum Int)) . coerce $ F 1 :: G
G 2
λ coerce . (mappend (1 :: Product Int)) . coerce $ F 1 :: G
G 1
λ coerce . (mappend 1) . coerce $ F 1 :: G
...
• Couldn't match representation of type ‘a0’ with that of ‘Int’
arising from a use of ‘coerce’
...
But I would still not welcome coerce
, because it is far too easy to strip a safety label and
shoot someone, once the reaching for it becomes habitual. Imagine that, in a cryptographic
application, there are two values: x :: Prime Int
and x' :: Sum Int
. I would much rather
type getPrime
and getSum
every time I use them, than coerce
everything and have one day
made a catastrophic mistake.
usefulness coerce
does not bring much to the table regarding a shorthand for
certain operations. The leading example of my post, that I repeat here:
λ getSum $ Sum 1 `mappend` Sum 2
3
— Turns into something along the lines of this spiked monster:
λ coerce $ mappend @(Sum Integer) (coerce 1) (coerce 2) :: Integer
3
— Which is hardly of any benfit.
Your "spiked monster" example is better handled by putting the summands into a list and using the ala
function available here, which has type:
ala :: (Coercible a b, Coercible a' b')
=> (a -> b)
-> ((a -> b) -> c -> b')
-> c
-> a'
where
a
is the unwrapped base type.b
is the newtype that wraps a
.a -> b
is the newtype constructor.((a -> b) -> c -> b')
is a function that, knowing how to wrap values of the base type a
, knows how to process a value of type c
(almost always a container of a
s) and return a wrapped result b'
. In practice this function is almost always foldMap
.a'
the unwrapped final result. The unwrapping is handled by ala
itself.in your case, it would be something like:
ala Sum foldMap [1,2::Integer]
"ala" functions can be implemented through means other than coerce
, for example using generics to handle the unwrapping, or even lenses.
coerce
from Data.Coerce can be pretty great for this sort of thing. You can use it to convert between different types with the same representation (like between a type and a newtype wrapper, or vice versa). For example:
λ coerce (3 :: Int) :: Sum Int
Sum {getSum = 3}
it :: Sum Int
λ coerce (3 :: Sum Int) :: Int
3
it :: Int
It was developed to solve the problem that it is cost-free to e.g. convert an Int
into a Sum Int
by applying Sum
, but it isn't necessarily cost-free to e.g convert a [Int]
to a [Sum Int]
by applying map Sum
. The compiler might be able to optimise away the traversal of the list spine from map
or it might not, but we know that the same structure in memory can serve as either a [Int]
or a [Sum Int]
, because the list structure doesn't depend on any properties of the elements and the element types have identical representation between those two cases. coerce
(plus the role system) allows us to make use of this fact to convert between the two in a way that is guaranteed not to do any runtime work, but still have the compiler check that it's safe to do so:
λ coerce [1, 2, 3 :: Int] :: [Sum Int]
[Sum {getSum = 1},Sum {getSum = 2},Sum {getSum = 3}]
it :: [Sum Int]
Something that wasn't at all obvious to me at first is that coerce
is not limited to coercing "structures"! Because all it's doing is allowing us to substitute types (including parts of compound types) when the representations are identical, it works just as well to coerce code:
λ addInt = (+) @ Int
addInt :: Int -> Int -> Int
λ let addSum :: Sum Int -> Sum Int -> Sum Int
| addSum = coerce addInt
|
addSum :: Sum Int -> Sum Int -> Sum Int
λ addSum (Sum 3) (Sum 19)
Sum {getSum = 22}
it :: Sum Int
(In the above example I had to define a monotype version of +
because coerce
is so generic the type system otherwise doesn't know which version of +
I'm asking to coerce to Sum Int -> Sum Int -> Sum Int
; I could instead have given an inline type signature on the argument to coerce
, but that looks less tidy. Often in real usage the context is sufficient to determine the "source" and "target" types of the coerce
)
I once wrote a library that provided a few different ways of paramterising types via newtypes, and provided similar APIs with each scheme. The modules implementing the APIs were full of type signatures and foo' = coerce foo
style definitions; it felt really nice that I was barely doing any work other than stating the types that I wanted.
Your example (using mappend
on Sum
to implement addition, without having to explicitly convert back and forth) could look like:
λ let (+) :: Int -> Int -> Int
| (+) = coerce (mappend @ (Sum Int))
|
(+) :: Int -> Int -> Int
λ 3 + 8
11
it :: Int
If you love us? You can donate to us via Paypal or buy me a coffee so we can maintain and grow! Thank you!
Donate Us With