Skip to content

How can I write lenses without depending on lens?

tom-bop edited this page Jun 14, 2018 · 12 revisions

Part of the magic of lens (referring to the package) is that you can write lenses (referring to the concept) without actually depending on lens.

This page provides a guide for package maintainers who'd like to add lens-compatible lenses to their library incurring minimal dependencies. It covers all of the basic lens types: Lens, Prism, Iso, Traversal, Fold, Getter, Setter, Review, Action, and Monadic Fold.

The various guides follow, ordered by needed dependencies.

A note on inlining. Generally, all lens-like functions get their best performance when aggressively INLINE'd. GHC can infer inlining within a module, but the INLINE pragma is recommended for all lens-like functions so that they'll be inlined in other modules as well. (<edwardk> INLINE all the things // lens really really needs to inline)

Lens

No dependencies.

The simplest interface is Lens. This is useful for exposing record fields or pieces of internal structure of your data types. The Lens type is

type Lens s t a b = Functor f => (a -> f b) -> s -> f t

which is very similar to Data.Traversable.traverse but can be built for traversals which do not necessarily traverse over slots denoted in the type signature. You write Lenses in much the same way you write Traversable instances, but using only the Functor interface

data State = State !Word8 !Word8

-- _word1 :: Lens' State Word8
_word1 :: Functor f => (Word8 -> f Word8) -> State -> f State
_word1 inj (State wa wb) = flip State wb <$> inj wa
{-# INLINE _word1 #-}

Generally, you must apply the inj :: a -> f b function to the subcomponent of your type you wish to focus on with your Lens then use fmap to reconstruct your type inside of the Functor.

Traversal

No dependencies.

If you'd like to provide sequential access to several like-typed subcomponents of your type, build a Traversal. The technique is identical to Lens except generalized to use the greater power afforded by Applicative. The Traversal type is

type Traversal s t a b = Applicative f => (a -> f b) -> s -> f t

So, we now know that the inj function gives us an Applicative and thus we can use it repeatedly before reassembling our type.

data State = State !Word8 !Word8

-- traverseWords :: Traversal' State Word8
traverseWords :: Applicative f => (Word8 -> f Word8) -> State -> f State
traverseWords inj (State wa wb) = State <$> inj wa <*> inj wb
{-# INLINE traverseWords #-}

Like Lens, this is similar to writing a Traversable instance (unsurprisingly) but even closer since we can use (<*>).

Getter

(Technically) depends on contravariant which includes tagged, transformers, transformers-compat, mtl, but...

Getters are not usually useful to define as a package maintainer. Users of lens can build getters on the fly by using to :: (s -> a) -> IndexPreservingGetter s a if you provide a function from the larger type to the subcomponent.

Fold

Depends on contravariant which includes tagged, transformers, transformers-compat, mtl.

Folds are a summary of a Traversal. They arise when after applying inj to your summary value it's no longer possible to reconstruct your original type. Yet, they still have a type which ends up returning a Functor "containing" your reconstructed type.

type Fold s a = 
    (Contravariant f, Applicative f) => (a -> f a) -> s -> f s

To write a Fold you use a trick afforded by the Contravariant and Applicative instances together---they force the parameter of f to be phantom by letting you write a strange internal function

import Data.Functor.Contravariant (contramap)

coerce :: (Contravariant f, Applicative f) => f a -> f b
coerce = contramap (const ()) . fmap (const ())
{-# INLINE coerce #-}

Then we use this to ignore our usual obligation to rebuild the total structure.

data State = State !Word8 !Word8

-- mixing :: Fold State Word8
mixing :: 
  (Contravariant f, Applicative f) => (Word8 -> f Word8) -> State -> f State
mixing inj (State wa wb) = coerce . inj $ wa `xor` wb
{-# INLINE mixing #-}

Iso

Depends on profunctors which includes tagged, comonad, transformers, containers, deepseq, array, and nats.

An Iso is an isomorphism between two types. It's useful whenever you have two representations of the same type, each perhaps exposing a different access pattern or aspect, but you have two functions to :: A -> B and fro :: B -> A which are mutual inverses.

type Iso s t a b =
  (Profunctor p, Functor f) => p a (f b) -> p s (f t)

Isos are easy to construct using only your two functions to and fro and

dimap :: Profunctor p => (s -> a) -> (b -> t) -> p a b -> p s t

For instance, the running example State is isomorphic to a pair of Word8s.

data State = State !Word8 !Word8

-- asAPair :: Iso' State (Word8, Word8)
asAPair :: (Functor f, Profunctor p) =>
           p (Word8, Word8) (f (Word8, Word8)) -> p State (f State)
asAPair = dimap (\(State wa wb) -> (wa, wb)) (fmap $ uncurry State)
{-# INLINE asAPair #-}

Note that the use of fmap on the second function is necessary for lifting your isomorphism witness up into the Functor.

Prism

Depends on profunctors which includes tagged, comonad, transformers, containers, deepseq, array, and nats.

Prisms usually target one branch of a sum type. In order to qualify, they must be able to "maybe" access the branch that they're focused on and be able to review their target type to reconstruct the whole.

They have the type

type Prism s t a b = (Choice p, Applicative f) => p a (f b) -> p s (f t)

and we'll construct a new example using a sum type. Prisms are built like Isos, but since our isomorphisms is partial in one direction, the Choice constraint is used to give that flexibility.

data Doing a = Done a | Doing (Doing a)

-- using right' from package profunctors
-- 
--   class Profunctor p => Choice p where
--     right' :: p a b -> p (Either c a) (Either c b)
--     ...

-- _Done :: Prism' (Doing a) a
_Done :: (Choice p, Applicative f) => p a (f a) -> p (Doing a) (f (Doing a))
_Done = dimap to fro . right' where
  to (Done a)    = Right a
  to x           = Left x
  fro (Left it)  = pure it
  fro (Right fa) = Done <$> fa
{-# INLINE _Done #-}

Here, the only degrees of freedom are in the Righty branches of to and from. In fact, this pattern is completely captured by the lens function prism usually used to construct Prisms. It can be easily included as a helper function in your library if you're defining many Prisms.

prism :: 
  (Choice p, Applicative f) => 
  (b -> t) -> (s -> Either t a) -> p a (f b) -> p s (f t)
prism bt seta = dimap seta (either pure (fmap bt)) . right'
{-# INLINE prism #-}

Todo:

  1. generalize this to the 4-parameter s t a b formulations.

  2. finish out the other types

Clone this wiki locally