Sebastiaan Visser Oktober 1st, 2013

fclabels 2.0

I’m excited to announce a brand new major version of fclabels. This Haskell library provides first class labels that can act as bidirectional record fields. The labels can be derived automatically using Template Haskell which means you don’t have to write any boilerplate yourself. The labels are implemented as lenses and are fully composable. Lenses can be used to get, set and modify parts of a data type in a consistent way.

In this post:

What’s new

This new version provides a major cleanup of the API and provides a lot more power while not compromising on the simplicity of the interface. We now support lenses for both monomorphic and polymorphic updates, we have total lenses, partial lenses and lenses that can fail with some error.

There are now several ways to derive the labels from Haskell data types, for both record and non-record types. We can derive labels in scope (the existing method), but we can now also generate them as expressions that can be named and typed manually. Data type declarations can also directly be translated into a variant with first class labels, without needing the underscore prefixes. Examples of the possible derivation methods can be found below.

The new version has full support for deriving labels for the fields in generalized algebraic data types (GADTs). It contains a sufficiently intelligent totality checker that looks at the indices of a GADT to decide whether to derive the labels as total or partial lenses.

A small set of predefined lenses is included for some of Haskell’s base types, like Maybe, Either, lists and tuples.

For an introduction into this package read the rest of this blog post or see the API documentation on Hackage or the source code on github.

Note that the library is heavily optimised for qualified imports and some names may occurs multiple times across the package. The module name should provide enough context.

Basic usage

Let’s say we have a simple record data type that represents an address. We can automatically derive first class labels for this type by prefixing the field names with an underscore and invoking the Template Haskell function mkLabel:

import Data.Label

data Address = Address
  { _city    :: String
  , _street  :: (String, Int)
  } deriving Show

mkLabel ''Address

Now we have two labels in scope named after the record fields:

city    :: Address :-> String
street  :: Address :-> (String, Int)

The labels are implemented as lenses, which means getter and modifier functions packed in one data type. The lens type above uses a function like operator (:->). We can use the get, modify, and set function to read or write values.

get    :: (f :-> a) -> f -> a
set    :: (f :-> a) -> a -> f -> f
modify :: (f :-> a) -> (a -> a) -> f -> f

Although not enforced by the type system, for predictable usage of the library we expect the following law to hold:

get lens (modify lens m f) == m (get lens f)

Because the Address data type contains only one constructor and doesn’t have any type variables, the derived labels are total and monomorphic. More lens types are available in this library as will be explained in the next section.

The labels can be used to get and update values from an Address:

rijksmuseum = Address
  "Amsterdam" ("Museumstraat", 1)

> get city rijksmuseum
"Amsterdam"

> modify city (map toUpper) rijksmuseum
Address "AMSTERDAM" ("Museumstraat", 1)

Labels can be composed with other labels to dig deeper into a data type. For example, we can access the individual components of the street tuple:

import Control.Category
import Prelude hiding ((.))
import qualified Data.Label.Base as L

> modify (L.snd . street) (+ 99)
Address "Amsterdam" ("Museumstraat", 100)

Composition in fclabels is done using the (.) function from Control.Category, a generalization of function composition. This way lens composition reads exactly like function composition.

This example explains the essence of the library, derivation of labels for our data types that are implemented as fully composable lenses. The lenses are first class and can be passed around and extended as we wish. Lenses allow both reading values from and writing values to a data type.

Different lens types

The package provides several different lens types. The two main distinctions we make is between lenses that allow only monomorphic updates and lenses that allow polymorphic updates, and between total lenses and partial lenses. All four combinations are possible.

As explained above, the lens in the Address example is a total monomorphic lens. But, for example, a lens that points to the value in the Left constructor of an Either will be of a different type: partial and polymorphic. It is partial because the accessor functions might fail in the case of a Right constructor, it is polymorphic because the type of the value can change on update.

All the lens types are built by specialization of a base lens type. The base lens type is an abstract polymorphic lens. Abstract means that the getters and setters are not just Haskell functions, but can run in a custom Category, allowing effects. For example, totality and partiality are considered effects.

The code block below lists the most important lenses from the library together with some convenient type synonyms. It might look like a lot at first, but you will notice a very simple and regular pattern:

-- Abstract polymorphic (base) lens:

  Lens cat (f -> g) (o -> i)

-- Abstract monomorphic lens:

  Lens cat f o

-- Total polymorphic lens and synonyms:

  Lens (->) (f -> g) (o -> i)
  Lens Total (f -> g) (o -> i)
  (f -> g) :-> (o -> i)

-- Total monomorphic lens and synonyms:

  Lens (->) f o
  Lens Total f o
  f :-> o

-- Partial polymorphic lens and synonyms:

  Lens (Kleisli Maybe) (f -> g) (o -> i)
  Lens Partial (f -> g) (o -> i)
  (f -> g) :~> (o -> i)

-- Partial monomorphic lens and synonyms:

  Lens (Kleisli Maybe) f o
  Lens Partial f o
  f :~> o

You get the pattern.

Because all the lenses in the package are specializations of the same base lens they can all be fully composed with each other.

Polymorphic and monomorphic lens types

The two function types in the signature for the polymorphic lenses indicate that by modifying the value in the record field from o -> i the data type changes from f -> g. This makes the lens allow polymorphic updates.

Lenses for the components of a tuple are nice examples of polymorphic lenses:

import Data.Label.Base (fst, snd)

fst :: Lens arr ((a, b) -> (o, b)) (a -> o)
snd :: Lens arr ((a, b) -> (a, o)) (b -> o)

In the case of fst modifying the value of type a to type o will update the entire tuple form (a, b) to (o, b). Something similar happens for snd and the second component of the tuple.

Monomorphic lenses simply specialise polymorphic lenses by having the same input and output types. A synonym from Data.Label.Mono can be used to simplify the type:

import qualified Data.Lens.Poly as Poly
import qualified Data.Lens.Mono as Mono

Poly.Lens cat (f -> f) (o -> o)

-- or simply:

Mono.Lens cat f o

Specializing contexts

The base lens is abstract which means that we can use it in different computational contexts. We can specialize the context to allow different effects.

By specializing the cat variable to normal function space (->) we get back total lenses. By specializing to Kleisli Maybe we end up with a partial lens that fails silently. By specializing to Kleisli (Either e) we end up with a partial lens that can fail with some error value. The library provides convenient names for these commonly used lens contexts:

type Total     = (->)
type Partial   = Kleisli Maybe
type Failing e = Kleisli (Either e)

Lens Total       (f -> g) (o -> i)
Lens Partial     (f -> g) (o -> i)
Lens (Failing e) (f -> g) (o -> i)

Specialized versions of the lens utilities are provided by the library in the modules Data.Label.Total, Data.Label.Partial and Data.Label.Failing.

For total lenses we provide the shortcut operator (:->) and for partial lenses we provide (:~>).

By default, labels derived using the Template Haskell functions that come with the library are abstract. They are not yet specialized to a concrete context. This is useful for composition. For example, we can freely lift total lenses into partial lenses and compose them with other partial lenses. Of course, conceptually we cannot lift partial lenses into total lenses. By using custom Arrow type classes (like ArrowZero and ArrowFail) we are actually able to enforce those constraints using the type system.

We can imagine this approach nicely scaling up to different effects. We could use Kleisli IO to build lenses for IORefs, Kleisli STM to build lenses for TVars, or some Database category to build lenses pointing to some record in an external database.

Isomorphisms

The library contains a type Iso which represents an isomorphism, which is a bidirectional function. Like lenses isomorphisms don’t use Haskell functions directly, but work in some custom Category.

data Iso cat i o = Iso (cat i o) (cat o i)

To clarify, specialized to normal function space this isomorphism would just be a pair of a -> b and b -> a.

Isomorphisms can be lifted into a lens using the iso function. After lifting they can be freely composed with other lenses. The library provides both a function for embedding monomorphic and polymorphic isomorphisms. Embedding an isomorphism into a lens will preserve the original context.

For example, we can create a partial lens from a Read/Show isomorphism:

import Safe (readMay)

readShow
  :: (Show a, Read a)
  => Iso Partial String a
readShow = Iso r s
  where r = Kleisli readMay
        s = Kleisli (Just . show)

asFloat :: Mono.Lens Partial (Int, String) Float
asFloat = iso readShow . L.snd

> Partial.modify asFloat (*4) (1, "2.1")
Just (1, "8.4")

> Partial.modify asFloat (*4) (1, "-")
Nothing

The readShow isomorphism is included in Data.Label.Base.

Views using Applicative

We have seen how to compose lenses horizontally using the (.) operator from Control.Category. We can also compose lenses vertically using an instance of the Applicative type class. Vertical composition gives us a small DSL for building views. The following example might clarify this idea:

asTup3 :: Address :-> (String, String, Int)
asTup3 = Mono.point $
  (,,) <$> L.fst3 >- city
       <*> L.snd3 >- L.fst . street
       <*> L.trd3 >- L.snd . street

Here we create a view on the Address data type as a 3-tuple.

We connect the lenses for 3-tuples to the lenses for the Address data type using the (>-) operator, which can be read as points-to. Now we use the Applicative instance on the connected lenses together with the output constructor (,,) to create a view on the Address type. Because we cannot directly make Lens an instance of Applicative (the types don’t allow this) we use an more generic type Point underneath. We can turn a Point back into a proper lens using the point function.

Using the view we can now use 3-tuples to manipulate addresses.

> let addr = Address "-" ("Neude", 1)
> set (L.fst3 . asTup3) "Utrecht" addr
Address "Utrecht" ("Neude", 11)

For multi-constructor types we can use the Alternative instance for Point to get views on all the constructors. For example, to swap the sides of an Either:

swapE :: Either a b :~> Either b a
swapE = Poly.point $
      Left  <$> L.left  >- L.right
  <|> Right <$> L.right >- L.left

Deriving lenses in scope

There are three basic ways to derive labels for Haskell data types in fclabels. The first method, which we’ve already seen with the Address example, is to derive labels in scope for all the record fields prefixed with an underscore. We can use the function mkLabel, or mkLabels for this.

A more general version mkLabelsWith can be used to configure the deriving process. We can decide to generate type signature or not, we can create concrete type signatures or keep the lens abstract. We can decide whether to let partial lenses fail silently using ArrowZero or preserve the error with ArrowFail. We can provide a custom renaming function so we can use an different naming strategy than just stripping underscores.

Lenses as expressions

A second derivation method is to derive labels as expressions in a n-tuple, using the getLabel function. This allows us to pattern match on the tuple and provide custom names and types for the lenses. This is especially useful for deriving lenses for existing non-record data types. Most of the lenses in Data.Label.Base are derived this way. For example, the lenses for the Either type:

left  :: (Either a b -> Either o b) :~> (a -> o)
right :: (Either a b -> Either a o) :~> (b -> o)

(left, right) = $(getLabel ''Either)

Because of the abstract nature of the generated lenses and the top level pattern match, it might be required to use NoMonomorphismRestriction in some cases.

Direct derivation

The third option is to derive labels for a record by directly wrapping the declaration with the fclabels function. The wrapped record definition will be brought into scope together with the derived labels. The labels will be named exactly as in the record definition, the original fields will be stripped from the data type:

fclabels [d|
  data Pt = Pt
    { ptX :: Double
    , ptY :: Double
    } deriving Show
  |]

> modify ptX (+ 2.0) (Pt 1.0 2.0)
Record 3.0 2.0

There are several advantages to this approach: there is no need for the underscore prefixes, there is no need to explicitly hide the original labels from being exported from the module, and the record fields will not show up in the Show instance.

Multiple data types are allowed within one quotation passed to fclabels and all non-type declaration will be brought in scope untouched.

Lenses for GADTs

We can now also derive labels for GATDs, which wasn’t possible in the previous version. The interesting aspect of (some) GADTs is that even when the type might have multiple constructors, the record fields might still be total. This can happen when the type indices of the GADT are restrictive enough. For example:

data Gadt a where
  P :: { _fa :: a, _fb :: b } -> Gadt (a, b)
  Q :: { _bi :: (Bool, Int) } -> Gadt (Bool, Int)
  R :: { _ls :: [Int]       } -> Gadt [Int]

mkLabel ''Gadt

This will bring the partial labels fa, fb and bi in scope together with the total label for ls. A simple totaility checker will try to figure out if label types overlap or not.

Note that this feature has not been proven correct for all combination of indices yet, but for now it seems to cover a great deal of cases.


Install using Cabal from Hackage

Source code on github