A Guide to GHC's Extensions

last updated October 18 2018 for GHC 8.6.1

The latest official version of the Haskell standard was published in 2010. Since then, GHC, the most popular Haskell compiler, has grown almost 100 extensions which modify the language, sometimes quite drastically. Many of them have proven popular, to the point where it’s rare to find a library on Hackage that doesn’t use any extensions.

This is a problem for students of Haskell since most teaching materials only cover the base standard, leaving learners to fend for themselves in the extension jungle. This guide aims to make the journey a little easier. To that end, I sort GHC’s extensions into four tracks:

  • The basic track contains extensions that are generally simple and either commonly used or particularly useful for Haskell beginners. I also consider most of these extensions future-proof, in the sense that a future Haskell standard would probably include them without major changes.
  • The advanced track contains extensions that are useful, but either more complex or less popular than the ones in the basic track. This includes a bunch of ‘special interest’ extensions that are only relevant if you want to do particular things with Haskell.
  • The questionable track contains extensions that you probably shouldn’t use – either because they are actively unhelpful (badly designed or deprecated), or because they have failed to gain much adoption.
  • The miscellaneous track contains extensions that wouldn’t fit in anywhere else.

Each track is subdivided into topics that are mostly independent, so you can pick and choose according to your interests. Which extension belongs into which track is of course a matter of opinion, so take mine with a grain of salt.

For each extension, I provide the following information:

  • The GHC version which first introduced the extension. Note that some extensions have been changed considerably in later GHC versions, so there may be surprises lurking if you use extensions with older GHCs. When in doubt, consult the User’s Guide for your GHC version.
  • An estimate of how stable the extension is. GHC doesn’t provide formal stability guarantees for extensions, so this is just my outside view of the situation. Possible classifications are:
    • Stable: the extension is well-established and I don’t expect its design to change much anymore.
    • Mostly stable: The extension’s core design is stable, but details are somewhat likely to change.
    • Unstable: it is at least possible that the extension will be changed significantly in future GHC versions.
  • A link to the relevant section of the GHC User’s Guide. This is your primary reference, but it often contains a bunch of technical details that may obscure the ‘essence’ of an extension.
  • Links to external tutorials where available. You should take a look at these before diving into the User’s Guide.
  • A brief explanation of what the extension is all about. This is intended as a primer which allows you to determine if the extension is something worth knowing about. It’s not a full tutorial, and certainly not a reference.

Please let me know if I’ve made any mistakes. I don’t know many of these extensions too well myself, so there’s a good chance that some of the explanations are wrong. Also, if you have a suggestion for how to explain things more clearly (structure, choice of examples, etc.), I’d be grateful for that as well.

Of course, I’m not the first to cover the topic of GHC extensions. Some alternative takes:


1 Language Standards

There are essentially two versions of the Haskell standard: Haskell98 and Haskell2010. Haskell2010 is a conservative evolution of Haskell98 which just makes the language better, so there’s no reason to limit yourself to Haskell98. Consequently, this guide only covers extensions that aren’t in Haskell2010, and when I say “standard Haskell”, I mean Haskell2010.

Some day, we will hopefully see a new version of the standard that incorporates GHC’s more well-established extensions. Efforts in this direction are underway, but they’re moving so slowly that it’s hard to tell whether they’re moving at all. In the meantime, the pragmatic choice is to treat Haskell as an implementation-defined language and to use a conservative subset of GHC’s extensions.

2 Enabling GHC Extensions

It is good practice to specify exactly which extensions you are using. That way, if someone tries to compile your code with a different compiler (like GHCJS or Eta), the compiler can fail early if it doesn’t support all the extensions you rely on. There are two main ways to enable extensions:

  1. You can enable extensions in a specific source file by adding a LANGUAGE pragma at the top of the file:

    {-# LANGUAGE Extension1, Extension2 #-}
    module Mod where ...

    If the file is part of a Cabal project, you should also add the extensions to your Cabal file’s other-extensions field(s). The other-extensions field goes in a component block (library, executable, test-suite or benchmark) and should list all extensions used by the source files of that component.

  2. You can enable an extension globally in your project by adding it to a default-extensions field in your Cabal file:

    library lib
      default-extensions: Extension1, Extension2

    default-extensions works like other-extensions, but the listed extensions are enabled automatically for all source files belonging to the field’s component.

For some discussion on which style to prefer, see here, here and here.

3 Basic Track

3.1 Syntactic Niceties

3.1.1 TupleSections

Since 6.12 | Stable | User’s Guide | School of Haskell

Write (x,,) instead of \y z -> (x, y, z).

3.1.2 LambdaCase

Since 7.6.1 | Stable | User’s Guide | School of Haskell

Instead of

\x -> case x of
  1 -> True
  _ -> False


  1 -> True
  _ -> False

3.1.3 MultiWayIf

Since 7.6.1 | Stable | User’s Guide | School of Haskell

Instead of

if n == m
  then x
  else if n == k
    then y
    else z


if | n == m    -> x
   | n == k    -> y
   | otherwise -> z

3.1.4 BlockArguments

Since 8.6.1 | Unstable | User’s Guide

If you want to apply a function to a do block, you usually need to use the $ operator:

main = f $ do

This extension allows you to omit the $ and directly give the do block as an argument. This also works with lambdas, case statements and other blocks: (f \x -> y) is now equivalent to (f (\x -> y)).

3.1.5 TypeOperators

Since 6.8.1 | Stable | User’s Guide

In Haskell2010, expressions like (+) in a type are parsed as type variables, which is not very useful. This extension allows you to use operators as names for type constructors, type synonyms etc.:

type a + b = Either a b

Note that starting with GHC 8.6, using * as a type operator may lead to trouble. See StarIsType.

3.1.6 NumericUnderscores

Since 8.6.1 | Stable | User’s Guide

Write 100_000_001 instead of 100000001.

3.2 Overloaded Literals

3.2.1 OverloadedStrings

Since 6.8.1 | Stable | User’s Guide | School of Haskell | 24 Days

Overload string literals, similar to numeric literals. This means that "a string" has type IsString a => a, and you can define your own instances of IsString. For example, a library for regular expressions might define an instance IsString Regex that parses regular expressions, enabling you to write

r :: Regex
r = "^x$"

3.2.2 OverloadedLists

Since 7.8.1 | Mostly stable | User’s Guide

Overload list literals, similar to numeric literals (and string literals with OverloadedStrings). This means that [1, 2] has type (IsList l, Num (Item l)) => l. (Item is an associated type of the IsList class; see TypeFamilies.) You can then define instances of IsList for list-like types like vectors and sets and write

s :: Set
s = [1, 2, 3]

3.3 Patterns

3.3.1 ViewPatterns

Since 6.10.1 | Stable | User’s Guide | School of Haskell | 24 Days

A view pattern lets you apply a function as part of a pattern, then match against the result of that function. So instead of

toNonEmpty :: Set a -> Maybe (NonEmpty a)
toNonEmpty s = case toList s of
  []     -> Nothing
  (x:xs) -> x :| xs


toNonEmpty :: Set a -> Maybe (NonEmpty a)
toNonEmpty (toList -> []    ) = Nothing
toNonEmpty (toList -> (x:xs)) = x :| xs

3.3.2 PatternSynonyms

Since 7.8.1 | Mostly stable | User’s Guide | School of Haskell | 24 Days

Pattern synonyms allow you to define additional patterns for a type. For example, you can define a pattern that matches on the first two elements of a list:

pattern Head2 x y <- (x:y:_)

firstTwo :: [a] -> Maybe (a, a)
firstTwo (Head2 x y) = Just (x, y)
firstTwo _           = Nothing

Pattern synonyms are useful when you want to hide the representation of a datatype. For example, the containers package defines a type Seq representing finite lists. It is implemented as a special sort of tree, but the implementation is not exposed. Instead, the package defines pattern synonyms like Empty and :<| which allow you to match on a Seq as if it were a list:

head :: Seq a -> Maybe a
head Empty     = Nothing
head (x :<| _) = Just x

3.4 Type System

3.4.1 ExplicitForAll

Since 6.12.1 | Stable | User’s Guide | School of Haskell

The type signature f :: a -> b -> b means that for all types a and b, we have a function from a and b to b. For example, f can be used as a function Int -> Bool -> Bool, where a := Int and b := Bool. ExplicitForAll allows us to make this “for all” explicit1 by writing

f :: forall a b. a -> b -> b

This is exactly the same type signature as above. Of course, that’s not very useful by itself, but the explicit forall becomes relevant in combination with the next three extensions.

3.4.2 TypeApplications

Since 8.0.1 | Stable | User’s Guide | Kwang Yul Seo

A polymorphic function like

id :: forall a. a -> a

is really a function with two arguments: One type, a, and one value of type a. TypeApplications allows you to give the type argument explicitly:

f :: Int -> Int
f x = id @Int x

This is particularly useful when combined with GHCi’s :t command, which you can use to view specialised type signatures:

> :t foldr
foldr :: Foldable t => (a -> b -> b) -> b -> t a -> b
> :t foldr @[] @Int @Bool
foldr @[] @Int @Bool :: (Int -> Bool -> Bool) -> Bool -> [Int] -> Bool

3.4.3 ScopedTypeVariables

Since 6.8.1 | Stable | User’s Guide | 24 Days | School of Haskell

Consider the following slightly-contrived-for-demonstration-purposes implementation of the minimum function:

minimum :: forall a. Ord a => [a] -> Maybe a
minimum [] = Nothing
minimum xs = Just $ foldl1' go xs
    go :: a -> a -> a
    go = min

GHC doesn’t accept this (even with ExplicitForAll) because from its point of view, the a in minimum’s type signature and the a in go’s type signature are different variables that just happen to share the same name. Therefore, we don’t have a constraint Ord a in go, and so we can’t use min (a method of Ord).

With ScopedTypeVariables, our definition of minimum is accepted: Both as are now considered the same type variable, and the Ord a constraint from minimum ‘trickles down’ to go.

Caveat: This only works because we quantify over a explicitly with forall a. If we omit the forall a, we get the same error as before.

3.4.4 RankNTypes

Since 6.8.1 | Stable | User’s Guide | School of Haskell | 24 Days | Gregor Riegler | Chris Done

The ST monad can be used to safely run stateful computations. To run an ST computation, you use the function runST, which has type

runST :: forall a. (forall s. ST s a) -> a

This function has one type argument, a, and one value argument of type forall s. ST s a. RankNTypes is required to allow this argument type, because it contains a forall – without RankNTypes, the only place where a forall may appear is at the beginning of a type signature.

The meaning of this argument type is that the argument must be polymorphic in s, meaning that the argument must have type ST s a for any s and some a. Contrast this with the different type signature

runST :: forall a s. ST s a -> a

This allows us to pass m :: ST Int Bool as an argument to runST, setting a := Bool and s := Int. With the correct signature of runST, on the other hand, m cannot be used as an argument because it would have to be of type ST s Bool for any s.

RankNTypes is arguably an advanced extension, but it is required for ST, which is in base.

3.4.5 LiberalTypeSynonyms

Since 6.8.1 | Stable | User’s Guide | School of Haskell

Standard Haskell places a lot of restrictions on type synonyms; for example, you can’t use forall in a type synonym. LiberalTypeSynonyms lifts most of these restrictions, which can occasionally come in handy.

3.5 Records

3.5.1 NamedFieldPuns

Since 6.10.1 | Stable | User’s Guide | Wikibook

Suppose we have a record storing some configuration:

data Configuration = Configuration
    { localHost :: String
    , localPort :: Int
    , ...

To extract only some data out of a Configuration, we can use record pattern matching syntax:

localAddr :: Configuration -> String
localAddr Configuration { localHost = localHost, localPort = localPort }
    = localHost ++ ":" ++ show localPort

Obviously, writing field = field gets old quickly. NamedFieldPuns reduces the noise:

localAddr :: Configuration -> String
localAddr Configuration { localHost , localPort }
    = localHost ++ ":" ++ show localPort

3.5.2 RecordWildCards

Since 6.8.1 | Stable | User’s Guide | 24 Days | Kwang Yul Seo

This extension takes the principle behind NamedFieldPuns one step further: We can now write (continuing the example):

localAddr :: Configuration -> String
localAddr Configuration { .. }
    = localHost ++ ":" ++ show localPort

The { .. } is equivalent to writing one field pun for every field of Configuration.

3.6 Classes

3.6.1 FlexibleInstances

Since 6.8.1 | Stable | User’s Guide

Haskell2010 doesn’t allow you to write instances like

instance C (Maybe Int) where ...

because Maybe must only be applied to type variables. FlexibleInstances lifts this restriction. Additionally, it allows you to declare instances for type synonyms.

3.6.2 FlexibleContexts

Since 6.8.1 | Stable | User’s Guide

Haskell2010 places some restrictions on the superclass constraints that can appear in a class declaration. FlexibleContexts lifts these restrictions.

3.6.3 DeriveFunctor, DeriveFoldable, DeriveTraversable

Since 7.10.1 | Stable | User’s Guide | 24 Days

This adds Functor, Foldable and Traversable to the list of classes that GHC can derive for you automatically. So, if you ever need to define your own list type, you get almost every relevant function for free.

3.6.4 GeneralizedNewtypeDeriving

Since 6.8.1 | Stable | User’s Guide | 24 Days

In Haskell land, we like to introduce lots of newtypes to prevent errors:

newtype Dollars = Dollars Int

Yet that also forces us to write lots of boilerplate instances:

instance Num Dollars where
  Dollars x + Dollars y = Dollars (x + y)
  Dollars x * Dollars y = Dollars (x * y)

All we do here is wrap and unwrap the Dollar constructor. With GeneralizedNewtypeDeriving, GHC can write these instances for us:

newtype Dollars = Dollars Int
  deriving (Num)

and this work for almost every class, not just the built-in Eq, Ord etc.

There is an unfortunate interaction between GeneralizedNewtypeDeriving and type families; see RoleAnnotations. However, that doesn’t need to concern you most of the time.

3.6.5 InstanceSigs

Since 7.6.1 | Stable | User’s Guide

This extension allows you to give type signatures in instances, which can be useful documentation:

data Result a = Success a | Failure

instance Functor Result where
  fmap :: (a -> b) -> Result a -> Result b
  fmap f (Success x) = Success $ f x
  fmap f Failure     = Failure

Without InstanceSigs, the signature for fmap would be illegal.

3.6.6 ConstrainedClassMethods

Since 6.8.1 | Stable | User’s Guide

Haskell2010 doesn’t allow you to have additional constraints on class methods, so the following is disallowed (since f, a class method, has an Eq constraint):

class Foo a where
  f :: Eq a => a -> a

ConstrainedClassMethods lifts this (quoting the User’s Guide) “pretty stupid” restriction.

3.6.7 MultiParamTypeClasses

Since 6.8.1 | Stable | User’s Guide | 24 Days | Wikibook | Dennis Gosnell

In standard Haskell, classes can only apply to a single type (here a):

class Show a where ...

MultiParamTypeClasses lifts this restriction, allowing us to write classes that apply to multiple types. For example, the popular mtl package defines a class

class MonadReader r m where ...

which denotes monads m that have access to a value of type r. (That’s not quite the whole story – MonadReader needs an additional functional dependency.)

Type families provide a better alternative to MultiParamTypeClasses in many cases, but the latter are older and still used by many popular packages.

3.6.8 FunctionalDependencies

Since 6.8.1 | Stable | User’s Guide | 24 Days | Wikibook | Dennis Gosnell

With MultiParamTypeClasses, you’ll soon encounter situations where some of a class’s type parameters already determine another type parameter. For example, with MonadReader, if m is Reader Int, then r can only be Int, and this applies generally: For every m, there is at most one valid r.

Functional dependencies let you express precisely this fact:

class MonadReader r m | m -> r where ...

The m -> r is a functional dependency which tells GHC that for any m, there is at most one r with instance MonadReader m r (and GHC then won’t allow us to declare multiple such instances). This is useful because it allows GHC to infer r from m. Without the functional dependency, you would have to give a lot more type annotations because GHC would frequently fail to infer which r you mean.

3.6.9 DeriveGeneric

Since 7.2.1 | Mostly stable | User’s Guide | 24 Days | Mark Karpov | Danny Gratzer | Haskell Wiki

DeriveGeneric enables deriving of the Generic class, which provides support for the popular GHC.Generics flavour of generic programming. ‘Generic’ means that you can define functions which work for a variety of datatypes, without having to write separate code for each datatype.

The linked tutorials discuss how to write a library using generic programming, which is an advanced topic. However, to use these libraries (for example Aeson) you just need to put a deriving Generic next to your datatypes, which is why this extension is in the basic track.

3.7 Typed Holes

3.7.1 Typed Holes and Type Wildcards

Since 7.8.1 | Mostly stable | User’s Guide | Chris Barrett | Haskell Wiki

Basic typed holes don’t require an extension and are enabled by default. They can occur in terms and in types.

In terms, you can write an underscore, called a hole, and GHC will tell you what type it expects you to fill in instead of the hole. For example, if you write

id :: a -> a
id x = _

then GHC will throw an error saying that it expects something of type a and that there is, conveniently, an x :: a in scope.

In type signatures, you can write an underscore, called a type wildcard, and GHC will tell you what type it inferred. Continuing the example:

id :: a -> _
id x = x

In this case, GHC will throw an error telling you that the _ must stand for a. Of course, typed holes and type wildcards are much more interesting with more complex functions and types.

3.7.2 NamedWildCards

Since 7.10.1 | Mostly stable | User’s Guide | School of Haskell

This extension allows you to add an arbitrary name to a type wildcard, so you can write

id :: _typeofid
id x = x

GHC will use the name in the error message. You can also use the same named wildcard multiple times in a signature, in which case GHC assumes that all its occurrences refer to the same type:

id :: _a -> _a
id x = x

Without this extension, _a would be interpreted as a regular type variable.

3.7.3 PartialTypeSignatures

Since 7.10.1 | Mostly stable | User’s Guide | School of Haskell

When you use a type wildcard in a signature, you usually get an (informative) error. This extension turns these errors into warnings so you can use partial type signatures in your regular code. This lets you omit the boring parts of a complex type signature.

4 Advanced Track

4.1 Type System

4.1.1 ExistentialQuantification

Since 6.8.1 | Stable | User’s Guide | 24 Days | Wikibook | Arnaud Bailly | Jonathan Fischoff

An existentially quantified type allows you to ‘forget’ the type of a value, typically remembering only that the type belonged to some class. For example, we can define a type for any value that can be shown:

data Showable = forall a. Show a => Showable a

intShowable :: Int -> Showable
intShowable n = Showable n

stringShowable :: String -> Showable
stringShowable s = Showable s

showShowable :: Showable -> String
showShowable (Showable x) = show x

As demonstrated by intShowable and stringShowable, a Showable can contain a value of any type, as long as that type implements the Show class. showShowable demonstrates that we can use this fact to show a Showable. Note that we can’t do anything else with a Showable: the existential type ‘forgets’ everything about the value contained in it except that it can be shown.

Existential types look like an obvious solution to various problems, but it often turns out that these problems are better solved in other ways. Jonathan Fischoff’s article lays out some alternatives to existential quantification.

4.1.2 GADTSyntax

Since 7.2.1 | Stable | User’s Guide | School of Haskell

Constructors can be understood as special functions whose return type is the datatype to which they belong. For example, the Maybe type has constructors

Just :: a -> Maybe a
Nothing :: Maybe a

GADTSyntax provides an alternative syntax for datatype declarations which reflects this understanding:

data Maybe a where
  Just    :: a -> Maybe a
  Nothing :: Maybe a

GADTSyntax lets you define exactly those datatypes which you can define normally, so there’s not much reason to enable this extension, but we need it for GADTs.

4.1.3 GADTs

Since 6.8.1 | Stable | User’s Guide | Wikibook | Haskell Wiki | Matt Parsons

When you write a regular datatype in GADTSyntax, the return type of all constructors must be the datatype applied to its type variables. In our example from above, the return type of every constructor is (and must be) exactly Maybe a:

data Maybe a where
  Just    :: a -> Maybe a
  Nothing ::      Maybe a

GADTs lift this restriction, so different constructors can instantiate the datatype’s type variables differently:

data RestrictedMaybe a where
  JustInt    :: Int    -> RestrictedMaybe Int
  JustString :: String -> RestrictedMaybe String
  Nothing    ::           RestrictedMaybe a

A RestrictedMaybe can contain either an Int or a String, but no other types. This is not particularly useful, but the same principle can be applied to define a wide variety of interesting types.

4.1.4 TypeFamilies

Since 6.8.1 | Stable | User’s Guide | 24 Days | mchaver | Haskell Wiki | Matt Parsons | Kwang Yul Seo

Type families are the cornerstone of contemporary type-level programming in Haskell, so if you want to write programmes with fancy types (or use certain libraries), you should probably learn about them. The basics are not so hard, though there are a fair few corner cases. A short description couldn’t really do the concept justice, so I just recommend the above tutorials. Matt Parsons in particular gives a concise overview of type-level programming which also covers several other extensions.

4.1.5 TypeFamilyDependencies

Since 8.0.1 | Mostly stable | User’s Guide | Jan Stolarek | Paper (pdf)

Type family dependencies (aka injective type families) are the equivalent of functional dependencies for type families.

4.1.6 AllowAmbiguousTypes

Since 7.8.1 | Mostly stable | User’s Guide

GHC has an ambiguity check which raises an error if a function looks like it could never be called without leading to an ambiguous constraint. However, the check is incomplete, meaning that it sometimes rejects functions which can, in fact, be called. If you’re doing fancy type-level programming, you may therefore need to disable the check with AllowAmbiguousTypes (but you probably don’t).

Despite the scary name, this extension does not compromise type safety. If a function call leads to an ambiguous constraint, this is still an error. The only difference is that the error is reported at the function call, not at the function definition.

4.2 Kinds

4.2.1 KindSignatures

Since 6.8.1 | Stable | User’s Guide

This extension allows you specify the kinds of type variables (wherever type variables may occur). This can be used for documentation, and it’s sometimes necessary when working with fancy kinds. An unnecessarily verbose example:

class Functor (f :: * -> *) where
  fmap :: forall (a :: *) (b :: *). (a -> b) -> f a -> f b

4.2.2 ConstraintKinds

Since 7.4.1 | Stable | User’s Guide | Kwang Yul Seo | Wolfgang Jeltsch

A constraint is anything that can appear to the left of the => arrow in a type signature (mostly type class constraints like (Ord a, Eq b)). Constraints are type-like things which have a special kind, Constraint. ConstraintKinds makes this kind available, so you can define, for example, constraint synonyms:

type Serializable a = (Show a, Read a)

roundtrip :: Serializable a => a -> a
roundtrip = read . show

4.2.3 DataKinds

Since 7.4.1 | Mostly stable | User’s Guide | Matt Parsons | Kwang Yul Seo | Christian Marie

With DataKinds enabled, when you define a data type like

data Nat = Zero | Succ Nat

you also get a kind 'Nat corresponding to the type Nat, as well as types 'Zero and 'Succ corresponding to the constructors Zero and Succ. Hence

-- 'Nat is a kind
'Zero :: 'Nat
'Succ :: 'Nat -> 'Nat

Together with type families, this allows you to compute, for example, natural numbers at the type level, much like you would at the value level.

4.2.4 PolyKinds

Since 7.4.1 | Mostly stable | User’s Guide

This extension enables kind-polymorphic types like the type-level identity function (which of course requires TypeFamilies):

type family Id (a :: k) :: k where
  Id a = a

We can now apply Id to types of different kinds:

type IdInt   = Id Int    -- kind *
type IdMonad = Id Monad  -- kind (* -> *)

4.3 Empty Types

4.3.1 EmptyCase

Since 7.8.1 | Stable | User’s Guide

Haskell2010 allows you to define empty data types, which have no constructors. The only value of such a type is undefined. EmptyCase allows you to match on such values with a case statement that has zero alternatives, corresponding to the zero constructors:

data Void

absurd :: Void -> a
absurd v = case v of {}

4.3.2 EmptyDataDeriving

Since 8.4.1 | Mostly stable | User’s Guide

Haskell2010 does not allow you to derive the four standard type classes Eq, Ord, Read and Show for empty types. This is a simple oversight, which this extension fixes.

4.4 Classes

4.4.1 StandaloneDeriving

Since 6.8.1 | Stable | User’s Guide

deriving clauses must usually be attached to a datatype declaration. This extension allows you to derive classes ‘after the fact’, even in a different module. Standalone deriving clauses are also a little more liberal than attached deriving clauses; for example, they support some GADTs.

data Result a = Success a | Failure

deriving instance Eq a => Eq (Result a)

4.4.2 DefaultSignatures

Since 7.2.1 | Stable | User’s Guide | Mark Karpov | Danny Gratzer | Haskell Wiki

When declaring a class, you can give default implementations for some or all of the class methods:

class Eq a where
  (==) :: a -> a -> Bool
  a == b = not (a /= b)

  (/=) :: a -> a -> Bool
  a /= b = not (a == b)

With DefaultSignatures, you can give type signatures for default methods. These can be more specific than the implemented method’s signature, so you can provide a default implementation only if the type in question has instances of other classes. For example, one could define

class Eq a where
  (==) :: a -> a -> Bool
  default (==) :: Ord a => a -> a -> Bool
  a == b = compare a b == EQ

instance Eq Int

instance Ord Int where
  compare a b = ...

DefaultSignatures is useful primarily for generic programming with GHC.Generics.

4.4.3 DeriveAnyClass

Since 7.10.1 | Stable | User’s Guide

This extensions allows you to write a deriving clause for any class, and GHC will simply generate an empty instance declaration (except for those classes which it knows how to derive). This makes sense if there are default implementations for all of the class’s methods, which happens frequently when using GHC.Generics. For example, we may define a pretty-printing class which falls back to Show (using DefaultSignatures):

class Pretty a where
  prettyPrint :: a -> String
  default prettyPrint :: Show a => a -> String
  prettyPrint = show

data Stuff = Stuff
  deriving (Show, Pretty)

4.4.4 DerivingStrategies

Since 8.2.1 | Mostly stable | User’s Guide | Ryan Scott

With both DeriveAnyClass and GeneralizedNewtypeDeriving enabled, it is unclear how to process the following declaration:

newtype N = N Int
  deriving Num

GHC could either use the DeriveAnyClass strategy and create an empty instance declaration, or it could use GeneralizedNewtypeDeriving. DerivingStrategies allows you to specify which strategy you want:

newtype N = N Int
  deriving newtype Num

4.4.5 DerivingVia

Since 8.6.1 | Unstable | User’s Guide | Paper (pdf)

With this extension, you can write

newtype N = N Int
  deriving Semigroup via (Product Int)

Product Int (from Data.Monoid) is another newtype for Int whose Semigroup instance uses multiplication. With the deriving via clause, our type N ‘inherits’ this Semigroup instance. Another possible choice would be deriving Semigroup via (Sum Int), in which case we would get a Semigroup instance that uses addition.

A deriving via clause requires that the type via which we are deriving (here Product Int) and the type whose instance we are deriving (N) have the same runtime representation (a machine integer).

4.4.6 QuantifiedConstraints

Since 8.6.1 | Unstable | User’s Guide | Ryan Scott | Paper (pdf)

This extension allows you to universally quantify constraints (i.e. to say that a constraint C x must be fulfilled for every x). Using this feature, the class of monad transformers, MonadTrans, could be rewritten like this:

class (forall m. Monad m => Monad (t m)) => MonadTrans t where

The forall-quantified constraint expresses our expectation that applying the transformer t to any monad m should again result in a monad.

4.4.7 UndecidableInstances

Since 6.8.1 | Stable | User’s Guide | Dennis Gosnell

GHC places some restrictions on instance declarations to ensure that it can always resolve instances in finite time. These restrictions are incomplete, meaning that they disallow some instance declarations that are perfectly fine. UndecidableInstances can therefore be used to disable GHC’s checks. You should not have to do this often.

4.4.8 UndecidableSuperClasses

Since 8.0.1 | Stable | User’s Guide | Edward Kmett (video)

GHC usually does not allow a class to be a superclass of itself, to ensure that typeclass resolution terminates. It can occasionally be useful to disable this check.

4.4.9 RoleAnnotations

Since 7.8.1 | Mostly stable | User’s Guide | Richard Eisenberg

GHC assigns every type variable (in datatypes, classes, etc.) a role. If a type variable has the wrong role, GeneralizedNewtypeDeriving can be used to break type safety. GHC’s heuristics for role assignment mostly do the right thing, but it can be necessary to help them out by specifying roles explicitly. This extension allows you to do that.

Technically, you should think about roles whenever you write a type variable, but I’m not sure anyone actually does that.

4.5 Records

4.5.1 DisambiguateRecordFields

Since 6.8.1 | Stable | User’s Guide

Say you have two records with the same field name,

module M where

data UserPrefs = UserPrefs
    { user :: Int
    , prefs :: String


module N where

data Profile = Profile
    { user :: Int
    , profileData :: String

Haskell2010 will only let you use the user record selector qualified, even where it’s entirely obvious which of the two record you mean. For example, the following is not accepted (with M imported):

updateProfile :: (Int -> Int) -> Profile -> Profile
updateProfile f Profile { user = oldUser, profile = profile }
    = Profile { user = f oldUser, profile = profile}

With DisambiguateRecordFields, GHC accepts this definition (if and only if UserPrefs and Profile are defined in different modules). However, this only works because both occurrences of user are under a Profile constructor.

4.5.2 DuplicateRecordFields

Since 8.0.1 | Stable | User’s Guide

Extending DisambiguateRecordFields, DuplicateRecordFields allows multiple records in the same module to share a field name.

Moreover, GHC can now sometimes disambiguate uses of record fields based on type information. (DisambiguateRecordFields, in contrast, only looks at whether record fields occur under a constructor.) For example, the following is accepted:

updateProfile :: (Int -> Int) -> Profile -> Profile
updateProfile f p = p { user = f $ user (p :: Profile) }

Exactly when GHC can disambiguate record fields is a little tricky – for instance, updateProfile is not accepted without the type annotation on p.

4.5.3 OverloadedLabels

Since 8.0.1 | Unstable | User’s Guide

With this extension, the new syntactic form #foo is desugared to fromLabel @foo, with fromLabel a class method of

class IsLabel (x :: Symbol) a where
  fromLabel :: a

(Symbol is a type-level string.) GHC by itself defines no instances of IsLabel. As I understand it, this mechanism was intended to be used for properly overloaded record fields, but it seems like there is no satisfying solution to this problem yet.

4.6 Strictness

4.6.1 BangPatterns

Since 6.8.1 | Stable | User’s Guide | 24 Days | School of Haskell

With this extension, you can prefix a pattern with a bang (!):

f :: Int -> Bool
f !n = True

When something is matched against the bang pattern, it will be evaluated to weak head normal form before the body of f is evaluated. For example, consider the following programme:

fibonacci :: Int -> Int
fibonacci 0 = 0
fibonacci 1 = 1
fibonacci n = fibonacci (n - 1) + fibonacci (n - 2)

main :: IO ()
main = print $ f (fibonacci 100)

Running main will take a while because it has to (very inefficiently) compute fibonacci 100 before f can return its result. Without the bang pattern, the programme terminates almost instantly because we don’t actually need the result of fibonacci 100.

4.6.2 Strict, StrictData

Since 8.0.1 | Mostly stable | User’s Guide | Johan Tibell

StrictData makes all fields of datatypes strict by default. You can then use ~ to make some fields lazy again. For example, with StrictData active, the declaration

data D = D Int ~Double

is equivalent to the usual

data D = D !Int Double

Strict works like StrictData, but in addition to fields of datatypes, it also makes most other things (patterns, let/where bindings, …) strict by default.

4.7 Do Notation

4.7.1 ApplicativeDo

Since 8.0.1 | Mostly stable | User’s Guide | Paper (pdf)

A definition like

f = do
  x <- mx
  y <- my
  return $ g x y

is equivalent to f = g <$> mx <*> my, which uses only Applicative combinators. Despite this, Haskell2010 will require a Monad constraint – and use the monadic combinators – as soon as you use do syntax. ApplicativeDo changes this, allowing you to use do notation with applicatives that aren’t also monads. Moreover, for some monads, the applicative combinators are more efficient than the monadic ones, in which case ApplicativeDo may improve performance.

4.7.2 RecursiveDo

Since 6.8.1 | Stable | User’s Guide | 24 Days | Roman Cheplyaka | Will Fancher

Some monads (lists, Maybe, ST, IO, …) support a notion of cyclic computation, in which a data structure is built by using parts of it that have already been built. This is captured in the MonadFix class. RecursiveDo adds some syntactic sugar so you can more easily write recursive monadic computations.

4.8 Literals

4.8.1 NegativeLiterals

Since 7.8.1 | Stable | User’s Guide

The literal -1 is usually desugared to negate (fromInteger 1). With NegativeLiterals, it is instead desugared to fromInteger (-1). This should not make a difference most of the time.

4.8.2 NumDecimals

Since 7.8.1 | Stable | User’s Guide

Floating-point literals like 2.1e6 usually have type Floating a => a. With NumDecimals, floating point literals which denote integers, like 2e6, have type Num a => a instead, so you can use scientific notation for integers.

4.8.3 BinaryLiterals

Since 7.10.1 | Stable | User’s Guide

This extension allows you to write integer literals in binary: 0b10 desugars to fromInteger 2.

4.8.4 HexFloatLiterals

Since 8.4.1 | Mostly stable | User’s Guide

This extension allows you to write floating point literals in hexadecimal, which corresponds closely to the underlying binary representation.

4.9 Template Haskell

4.9.1 TemplateHaskell, TemplateHaskellQuotes

Since 6.0/8.01 | Mostly stable | User’s Guide | 24 Days | Mark Karpov | Matt Parsons | Haskell Wiki

Template Haskell is a major extension that allows you to generate code at compile time. Basically, you write Haskell programmes that generate Haskell code (i.e. abstract syntax trees). GHC then runs your programmes at compile time and combines the generated code with your regular code. This is very powerful because you can generate whatever code you want, but it’s also a bit of a sledgehammer with a bunch of rough edges.

TemplateHaskellQuotes enables a safer, but also considerably less useful subset of the TemplateHaskell functionality.

4.9.2 QuasiQuotes

Since 6.10.1 | Stable | User’s Guide | Edsko de Vries

Quasi-quotes are an extension to Template Haskell which allow you to embed other languages into Haskell. For example:

query :: Query
query = [sql| select * from users |]

sql is, in effect, a function which parses an SQL query from a user-provided string at compile time, returning the Haskell type Query.

4.9.3 DeriveLift

Since 7.2.1 | Stable | User’s Guide

Template Haskell makes use of a Lift class to convert expressions into abstract syntax trees. This extension allows you to have your Lift instances derived automatically.

4.10 Low-Level Hacking

4.10.1 MagicHash

Since 6.8.1 | Stable | User’s Guide

GHC likes to give its primitives names that end in a hash (#). Haskell2010 disallows the hash in identifiers, so you have to enable MagicHash to be able to refer to these primitives.

4.10.2 UnboxedTuples

Since 6.8.1 | Stable | User’s Guide | Michael Snoyman

An unboxed tuple, written (# Int, Bool #), allows you to return multiple values from a function without the overhead of a regular tuple (heap allocation, pointer dereferencing, etc.). When you return an unboxed tuple, its contents will be passed directly via registers or the stack.

4.10.3 UnboxedSums

Since 8.2.1 | Unstable | User’s Guide | Ömer Sinan Ağacan

An unboxed sum, written (# Int | Bool #), is like an anonymous sum type with multiple alternatives. Unlike regular sum types, GHC will try to represent unboxed sums as compactly as possible.

4.11 Safe Haskell

4.11.1 Safe, Trustworthy, Unsafe

Since 7.2.1/7.2.1/7.4.1 | Mostly stable | User’s Guide | Kristen Kozak (video with notes by Joe Nelson) | Edward Z. Yang | Paper (pdf)

Haskell has various features, like unsafePerformIO, which can be (mis)used to circumvent the type system, module abstraction and other desirable properties. Safe Haskell allows you to control the use of these features to some degree, but I don’t think it ever really caught on.

4.12 Miscellaneous

4.12.1 CPP

Since forever | Stable | User’s Guide | Aelve

The C preprocessor is (unfortunately) the most common and easiest way to do simple compile-time metaprogramming. A typical use case is supporting different GHC versions, which sometimes requires slightly different code.

CPP is not technically an extension, but it is enabled with a {-# LANGUAGE CPP #-} pragma (or a GHC flag).

4.12.2 NoImplicitPrelude

Since 6.8.1 | Stable | User’s Guide

The module Prelude is usually imported implicitly in every module you write. NoImplicitPrelude disables this. This is useful if you want to use one of many alternative preludes developed by the community, whose exported names frequently clash with those from Prelude.

4.12.3 RebindableSyntax

Since 7.0.1 | Mostly stable | User’s Guide | 24 Days

Certain syntactic forms desugar to ordinary Haskell functions. For example, do syntax is desugared into applications of the >>= and return combinators; literals N are desugared into fromInteger N; etc. Usually, this desugaring uses fixed functions, so the fromInteger is really Prelude.fromInteger. With RebindableSyntax, the desugaring uses whatever functions are in scope, so you can define your own fromInteger and GHC will use that.

4.12.4 UnicodeSyntax

Since 6.8.1 | Stable | User’s Guide | Haskell Wiki

This extension allows you to use Unicode symbols for certain keywords, like instead of forall. (Even without UnicodeSyntax, you can use Unicode for identifiers.)

4.12.5 NoMonomorphismRestriction

Since 6.8.1 | Stable | User’s Guide | Neil Mitchell | Stack Overflow | Haskell Wiki

The monomorphism restriction is a highly technical feature of Haskell’s type system which sometimes makes GHC infer unexpected types for definitions without an explicit type signature. Since GHC 7.8.1, it is on by default in compiled modules, but off by default in GHCi. NoMonomorphismRestriction allows you to switch it off even in compiled modules, but there are subtle disadvantages to this.

4.12.6 PostfixOperators

Since 7.10.1 | Stable | User’s Guide | School of Haskell

The operator section (56 !) is usually equivalent to \y -> (!) 56 y. With PostfixOperators, it is instead treated as (!) 56, so you can define and use postfix operators (though only with ugly parentheses):

(!) :: Int -> Int
(!) n = product [1..n]

test :: Int
test = (56 !)

4.12.7 PackageImports

Since 6.10.1 | Stable | User’s Guide

With this extension, an import can specify which package to import from:

import "network" Network.Socket

This can be used if two packages you depend on define the same module, though that’s very rare in practice.

5 Questionable Track

5.1 Comprehensions

The extensions in this section aren’t particularly bad; they just have very few users as far as I can tell, and I don’t give them much of a chance to get in a future Haskell standard (except maybe ParallelListComp).

5.1.1 ParallelListComp

Since 6.8.1 | Stable | User’s Guide | 24 Days

This extension extends list comprehension syntax so you can write

[ x * y | x <- xs | y <- ys ]

instead of

[ x * y | (x, y) <- zip xs ys ]

5.1.2 TransformListComp

Since 6.10.1 | Stable | User’s Guide | 24 Days

This extension adds SQL-inspired constructs like group by and using to the list comprehension syntax. I’ve never seen this used in the wild.

5.1.3 MonadComprehensions

Since 7.2.1 | Stable | User’s Guide | 24 Days

As it turns out, list comprehension syntax can be sensibly used with any monad, not just lists. MonadComprehensions enables exactly that and includes the features from ParallelListComp and TransformListComp. I would argue, however, that the comprehension syntax loses much of its intuitive appeal when used with other monads.

5.2 Disabling Standard Features

5.2.1 NoTraditionalRecordSyntax

Since 7.4.1 | Stable | User’s Guide

This disables the usual record syntax, like C { f = x }.

5.2.2 NoPatternGuards

Since 6.8.1 | Stable | User’s Guide

This disables pattern guards, which are a Haskell2010 feature similar to ViewPatterns.

5.3 Miscellaneous

5.3.1 Arrows

Since 6.8.1 | Stable | User’s Guide | 24 Days

Arrows are a generalisation of monads. The Arrows extension provides a notation for these constructs, similar to do notation for monads. Outside of the Opaleye database library and certain ‘arrowised’ functional reactive programming libraries, arrows have not caught on in the community.

5.3.2 StaticPointers

Since 7.10.1 | Mostly stable | User’s Guide | 24 Days

Static Pointers allow you to get a reference to a closure (i.e. a computation, possibly with associated data) which can be serialised and deserialised. The intended use case is distributed programming, where computations can be turned into static pointers and then sent back and forth between nodes. However, as far as I know, the only user of this extension is Cloud Haskell.

5.3.3 ImplicitParams

Since 6.8.1 | Stable | User’s Guide

This extension allows you to have function parameters which are automatically propagated to calling functions. They have never caught on.

5.3.4 ExtendedDefaultRules

Since 6.8.1 | Mostly stable | User’s Guide | Kwang Yul Seo

The type of

show (sum [1..100])

is ambiguous: sum [1..100] has type (Num a, Enum a) => a and show has type Show a => a -> String – so which a do we pick? Int, Integer and Double would all be candidates. For convenience, Haskell has a defaulting mechanism which kicks in here: a is defaulted to Integer based on its constraints. (With -Wtype-defaults, GHC warns you about this.) ExtendedDefaultRules applies this principle to more classes than standard Haskell. This is intended for GHCi, where ExtendedDefaultRules is enabled by default. In normal code, you usually want less defaulting, not more.

5.3.5 DeriveDataTypeable

Since 6.8.1 | Stable | User’s Guide | Chris Done

This extension allows you to derive instances of the Data class automatically. This class supports “Scrap your Boilerplate”-style generic programming, which seems to have been mostly obsoleted by GHC.Generics-style generic programming.

Recent GHCs derive instances of Typeable without you even writing a deriving clause, so you don’t need DeriveDataTypeable for that anymore.

5.3.6 NPlusKPatterns

Since 6.12.1 | Stable | User’s Guide

n+k patterns are a misfeature of Haskell98 that has been removed for Haskell2010, so this extension is effectively deprecated.

5.3.7 ImpredicativeTypes

Since 6.10.1 | Unstable | User’s Guide

Impredicative types have been effectively unsupported for a while. The User’s Guide says that “GHC has extremely flaky support for impredicative polymorphism”, which is probably a euphemism.

5.4 Deprecated Extensions

5.4.1 DatatypeContexts

Since 7.0.1 | Stable | User’s Guide

Haskell2010 allows you to put a context on datatype declarations:

data Ord a => Set a = ...

However, due to some weird design choices, this doesn’t work like you’d think and is pretty much useless.

5.4.2 NullaryTypeClasses

Since 7.8.1 | Stable | User’s Guide

A special case of MultiParamTypeClasses.

5.4.3 OverlappingInstances, IncoherentInstances

Since 6.8.1 | Mostly stable | User’s Guide

Consider the following situation:

class Foo a where
  f :: a -> a

instance Foo a where
  f = id

instance Foo [a] where
  f = reverse

The two instances overlap, meaning that if we look, say, for an instance Foo [Int], both instances match. This is disallowed by default. With OverlappingInstances, it is allowed and the more specific instance (the second) will be chosen. With IncoherentInstances, GHC doesn’t complain even if there is no single most specific instance; it will choose an arbitrary one.

OverlappingInstances and IncoherentInstances are deprecated since GHC 7.10.1 because the same functionality is now provided by the OVERLAPPING, OVERLAPPABLE, OVERLAPS and INCOHERENT pragmas (also explained in the linked section of the User’s Guide). These are preferable because they apply to a single instance, whereas the extensions disable the overlap checks for an entire module.

5.4.4 TypeInType

Since 8.0.1 | Unstable | User’s Guide

In GHC versions prior to 8.6.1, TypeInType merges the type and kind languages of GHC, meaning that types and kinds become the same thing. This implies that the type * has kind *; hence the extension’s name. It also means that everything we can do at the type level (type synonyms, type families, etc.), we can also do at the kind level.

Since GHC 8.6.1, TypeInType is a deprecated alias of PolyKinds, DataKinds and KindSignatures. Its functionality has been integrated into these other extensions.

5.4.5 Rank2Types

Since 6.8.1 | Stable | User’s Guide

A more restricted form of RankNTypes.

6 Miscellaneous Track

6.1 Extensions That Only Make Sense With Other Extensions

6.1.1 ExplicitNamespaces

Since 7.6.1 | Mostly stable | User’s Guide

With TypeOperators, you can define the type (++). If you want to export this type while the function (++) is also in scope, you need to be able to distinguish the two in the module export list. ExplicitNamespaces allows you to do so, by writing

module M ((++), type (++)) where ...

The same works when importing from M. ExplicitNamespaces also works with pattern synonyms, which can be prefixed with pattern. Both TypeOperators and PatternSynonyms imply ExplicitNamespaces, so you don’t need to enable it manually.

6.1.2 MonoLocalBinds

Since 6.12.1 | Stable | User’s Guide | Simon Peyton-Jones

Let and where bindings without an explicit type signature are usually generalised as much as possible. With MonoLocalBinds, they are generalised a little less. This leads to more predictable type inference when using type families or GADTs, so these two extensions imply MonoLocalBinds.

6.2 Temporary Extensions

6.2.1 MonadFailDesugaring

Since 8.0.1 | Unstable | User’s Guide

When a pattern match fails in a do block, the Monad class’s fail method is called with an error message. For many monads, fail is simply error, which introduces implicit partiality, which is no good.

As part of the MonadFail proposal, do syntax will in future use a subclass of Monad, MonadFail, when a do block contains a pattern that may fail. MonadFailDesugaring enables this behaviour. The extension is temporary because it will eventually become standard. MonadFailDesugaring is turned on by default since GHC 8.6.1.

6.2.2 StarIsType

Since 8.6.1 | Unstable | User’s Guide

In standard Haskell, * is the kind of all types (i.e. Int :: *, etc). This is an unfortunate choice of syntax because, among other things, it conflicts with TypeOperators: Int * Bool should probably be the type operator (*) applied to Int and Bool, rather than Int applied to * and Bool.

With NoStarIsType, * becomes a regular type operator with no special meaning. (To refer to the kind of all types, use Data.Kind.Type instead.) As of GHC 8.6.1, StarIsType is enabled by default, but this default will change in the future. Note that even with StarIsType, GHC 8.6.1 removes some special parsing rules for *, breaking backwards compatibility. See the migration plan.

  1. Don’t let anyone tell you that Haskellers are bad at naming.↩︎