10.1. Language options
As with all known Haskell systems, GHC implements some extensions to the standard Haskell language. They can all be enabled or disabled by command line flags or language pragmas. By default GHC understands the most recent Haskell version it supports, plus a handful of extensions.
Some of the Glasgow extensions serve to give you access to the underlying facilities with which we implement Haskell. Thus, you can get at the Raw Iron, if you are willing to write some nonportable code at a more primitive level. You need not be “stuck” on performance because of the implementation costs of Haskell’s “highlevel” features—you can always code “under” them. In an extreme case, you can write all your timecritical code in C, and then just glue it together with Haskell!
Before you get too carried away working at the lowest level (e.g., sloshing MutableByteArray#
s around your program), you may wish to check if there are libraries that provide a “Haskellised veneer” over the features you want. The separate libraries documentation describes all the libraries that come with GHC.
The language extensions control what variation of the language are permitted.
Language options can be controlled in two ways:
 Every language option can switched on by a commandline flag “
X...
” (e.g.XTemplateHaskell
), and switched off by the flag “XNo...
”; (e.g.XNoTemplateHaskell
).  Language options recognised by Cabal can also be enabled using the
LANGUAGE
pragma, thus{# LANGUAGE TemplateHaskell #}
(see LANGUAGE pragma).
GHC supports these language options:
Extension  Description 

Allow the user to write ambiguous types, and the type inference engine to infer them.  
Enable Applicative donotation desugaring  
Enable arrow notation extension  
Enable bang patterns.  
Enable support for binary literals.  
Allow  
Enable the CAPI calling convention.  
Enable constrained class methods.  
Enable a kind of constraints.  
Enable the C preprocessor.  
Enable detection of complete usersupplied kind signatures.  
Enable datatype promotion.  
Allow contexts on  
Enable default signatures.  
Enable deriving for any class.  
Enable deriving for the  
Enable deriving for the Foldable class. Implied by  
Enable deriving for the Functor class. Implied by  
Enable deriving for the Generic class.  
Enable deriving for the Lift class  
Enable deriving for the Traversable class. Implies  
Enables deriving strategies.  
Enable deriving instances  
Enable record field disambiguation. Implied by  
Allow definition of record types with identicallynamed fields.  
Allow empty case alternatives.  
Allow definition of empty  
Allow deriving instances of standard type classes for empty data types.  
Enable liberalised type synonyms.  
Enable explicit universal quantification. Implied by  
Enable using the keyword  
Use GHCi's extended default rules in a normal module.  
Enable flexible contexts.  
Enable flexible instances. Implies  
Enable foreign function interface.  
Enable functional dependencies. Implies  
Enable generalised algebraic data types. Implies  
Enable generalised algebraic data type syntax.  
Enable newtype deriving.  
Use the Haskell 2010 language variant.  
Use the Haskell 2010 language variant.  
Enable support for hexadecimal floating point literals.  
Enable Implicit Parameters.  
Don't implicitly  
 
Enable impredicative types. Implies  
Enable incoherent instances. Implies  
Enable instance signatures.  
Enable interruptible FFI.  
Enable kind signatures. Implied by  
Enable lambdacase expressions.  
Enable liberalised type synonyms.  
Allow  
Enable monad comprehensions.  
Enable monadfail desugaring.  
Enable do not generalise local bindings. Implied by  
Disable the monomorphism restriction.  
Enable multi parameter type classes. Implied by  
Enable multiway ifexpressions.  
Enable record puns.  
Enable named wildcards.  
Allow nested contexts to be at the same indentation level as its enclosing context.  
Enable support for negative literals.  
Enable support for  
Deprecated, does nothing. nullary (no parameter) type classes are now enabled using  
Enable support for 'fractional' integer literals.  
Enable support for numeric underscores.  
Enable overlapping instances.  
Enable overloaded labels.  
Enable overloaded lists.  
Enable overloaded string literals.  
Enable packagequalified imports.  
Enable parallel list comprehensions.  
Enable partial type signatures.  
Disable pattern guards. Implied by  
Enable pattern synonyms.  
Enable kind polymorphism. Implies  
Enable postfix operators.  
Allow  
Enable quasiquotation.  
Enable rank2 types. Synonym for  
Enable rankN types. Implied by  
Employ rebindable syntax. Implies  
Enable record wildcards. Implies  
Enable recursive do (mdo) notation.  
Enable role annotations.  
Enable the Safe Haskell Safe mode.  
Enable lexicallyscoped type variables.  
Enable standalone deriving.  
Allow the use of standalone kind signatures.  
Treat  
Enable static pointers.  
Make bindings in the current module strict by default.  
Enable default strict datatype fields.  
Enable Template Haskell.  
Enable quotation subset of Template Haskell.  
Disable support for traditional record syntax (as supported by Haskell 98)  
Enable generalised list comprehensions.  
Enable the Safe Haskell Trustworthy mode.  
Enable tuple sections.  
Enable type application syntax in terms and types.  
Enable type families. Implies  
Enable injective type families. Implies  
Deprecated. Enable kind polymorphism and datatype promotion.  
Enable type operators. Implies  
Enable type synonyms in instance heads. Implied by  
Enable unboxed sums.  
Enable the use of unboxed tuple syntax.  
Enable undecidable instances.  
Allow all superclass constraints, including those that may result in nontermination of the typechecker.  
Enable unicode syntax.  
Enable unlifted FFI types  
Enable unlifted newtypes.  
Enable Safe Haskell Unsafe mode.  
Enable view patterns. 
Although not recommended, the deprecated fglasgowexts
flag enables a large swath of the extensions supported by GHC at once.

fglasgowexts

The flag
fglasgowexts
is equivalent to enabling the following extensions:Enabling these options is the only effect of
fglasgowexts
. We are trying to move away from this portmanteau flag, and towards enabling features individually.

Haskell2010

Compile Haskell 2010 language variant. Enables the following language extensions:
EmptyDataDecls
ForeignFunctionInterface
PatternGuards
DoAndIfThenElse
RelaxedPolyRec

Haskell98

Compile using Haskell 98 language variant. Enables the following language extensions:
10.2. Unboxed types and primitive operations
GHC is built on a raft of primitive data types and operations; “primitive” in the sense that they cannot be defined in Haskell itself. While you really can use this stuff to write fast code, we generally find it a lot less painful, and more satisfying in the long run, to use higherlevel language features and libraries. With any luck, the code you write will be optimised to the efficient unboxed version in any case. And if it isn’t, we’d like to know about it.
All these primitive data types and operations are exported by the library GHC.Prim
, for which there is detailed online documentation <GHC.Prim.>. (This documentation is generated from the file compiler/prelude/primops.txt.pp
.)
If you want to mention any of the primitive data types or operations in your program, you must first import GHC.Prim
to bring them into scope. Many of them have names ending in #
, and to mention such names you need the MagicHash
extension.
The primops make extensive use of unboxed types and unboxed tuples, which we briefly summarise here.
10.2.1. Unboxed types
Most types in GHC are boxed, which means that values of that type are represented by a pointer to a heap object. The representation of a Haskell Int
, for example, is a twoword heap object. An unboxed type, however, is represented by the value itself, no pointers or heap allocation are involved.
Unboxed types correspond to the “raw machine” types you would use in C: Int#
(long int), Double#
(double), Addr#
(void *), etc. The primitive operations (PrimOps) on these types are what you might expect; e.g., (+#)
is addition on Int#
s, and is the machineaddition that we all know and love—usually one instruction.
Primitive (unboxed) types cannot be defined in Haskell, and are therefore built into the language and compiler. Primitive types are always unlifted; that is, a value of a primitive type cannot be bottom. (Note: a “boxed” type means that a value is represented by a pointer to a heap object; a “lifted” type means that terms of that type may be bottom. See the next paragraph for an example.) We use the convention (but it is only a convention) that primitive types, values, and operations have a #
suffix (see The magic hash). For some primitive types we have special syntax for literals, also described in the same section.
Primitive values are often represented by a simple bitpattern, such as Int#
, Float#
, Double#
. But this is not necessarily the case: a primitive value might be represented by a pointer to a heapallocated object. Examples include Array#
, the type of primitive arrays. Thus, Array#
is an unlifted, boxed type. A primitive array is heapallocated because it is too big a value to fit in a register, and would be too expensive to copy around; in a sense, it is accidental that it is represented by a pointer. If a pointer represents a primitive value, then it really does point to that value: no unevaluated thunks, no indirections. Nothing can be at the other end of the pointer than the primitive value. A numericallyintensive program using unboxed types can go a lot faster than its “standard” counterpart—we saw a threefold speedup on one example.
10.2.2. Unboxed type kinds
Because unboxed types are represented without the use of pointers, we cannot store them in use a polymorphic datatype at an unboxed type. For example, the Just
node of Just 42#
would have to be different from the Just
node of Just 42
; the former stores an integer directly, while the latter stores a pointer. GHC currently does not support this variety of Just
nodes (nor for any other datatype). Accordingly, the kind of an unboxed type is different from the kind of a boxed type.
The Haskell Report describes that *
(spelled Type
and imported from Data.Kind
in the GHC dialect of Haskell) is the kind of ordinary datatypes, such as Int
. Furthermore, type constructors can have kinds with arrows; for example, Maybe
has kind Type > Type
. Unboxed types have a kind that specifies their runtime representation. For example, the type Int#
has kind TYPE 'IntRep
and Double#
has kind TYPE 'DoubleRep
. These kinds say that the runtime representation of an Int#
is a machine integer, and the runtime representation of a Double#
is a machine doubleprecision floating point. In contrast, the kind Type
is actually just a synonym for TYPE
'LiftedRep
. More details of the TYPE
mechanisms appear in the section on runtime representation polymorphism.
Given that Int#
’s kind is not Type
, it then it follows that Maybe
Int#
is disallowed. Similarly, because type variables tend to be of kind Type
(for example, in (.) :: (b > c) > (a > b) > a > c
, all the type variables have kind Type
), polymorphism tends not to work over primitive types. Stepping back, this makes some sense, because a polymorphic function needs to manipulate the pointers to its data, and most primitive types are unboxed.
There are some restrictions on the use of primitive types:

You cannot define a newtype whose representation type (the argument type of the data constructor) is an unboxed type. Thus, this is illegal:
newtype A = MkA Int#
However, this restriction can be relaxed by enabling
UnliftedNewtypes
. The section on unlifted newtypes details the behavior of such types.  You cannot bind a variable with an unboxed type in a toplevel binding.
 You cannot bind a variable with an unboxed type in a recursive binding.

You may bind unboxed variables in a (nonrecursive, nontoplevel) pattern binding, but you must make any such patternmatch strict. (Failing to do so emits a warning
Wunbangedstrictpatterns
.) For example, rather than:data Foo = Foo Int Int# f x = let (Foo a b, w) = ..rhs.. in ..body..
you must write:
data Foo = Foo Int Int# f x = let !(Foo a b, w) = ..rhs.. in ..body..
since
b
has typeInt#
.
10.2.3. Unboxed tuples

UnboxedTuples

 Since

6.8.1
Unboxed tuples aren’t really exported by GHC.Exts
; they are a syntactic extension (UnboxedTuples
). An unboxed tuple looks like this:
(# e_1, ..., e_n #)
where e_1..e_n
are expressions of any type (primitive or nonprimitive). The type of an unboxed tuple looks the same.
Note that when unboxed tuples are enabled, (#
is a single lexeme, so for example when using operators like #
and #
you need to write ( # )
and ( # )
rather than (#)
and (#)
.
Unboxed tuples are used for functions that need to return multiple values, but they avoid the heap allocation normally associated with using fullyfledged tuples. When an unboxed tuple is returned, the components are put directly into registers or on the stack; the unboxed tuple itself does not have a composite representation. Many of the primitive operations listed in primops.txt.pp
return unboxed tuples. In particular, the IO
and ST
monads use unboxed tuples to avoid unnecessary allocation during sequences of operations.
There are some restrictions on the use of unboxed tuples:

The typical use of unboxed tuples is simply to return multiple values, binding those multiple results with a
case
expression, thus:f x y = (# x+1, y1 #) g x = case f x x of { (# a, b #) > a + b }
You can have an unboxed tuple in a pattern binding, thus
f x = let (# p,q #) = h x in ..body..
If the types of
p
andq
are not unboxed, the resulting binding is lazy like any other Haskell pattern binding. The above example desugars like this:f x = let t = case h x of { (# p,q #) > (p,q) } p = fst t q = snd t in ..body..
Indeed, the bindings can even be recursive.
10.2.4. Unboxed sums

UnboxedSums

 Since

8.2.1
Enable the use of unboxed sum syntax.
XUnboxedSums
enables new syntax for anonymous, unboxed sum types. The syntax for an unboxed sum type with N alternatives is
(# t_1  t_2  ...  t_N #)
where t_1
… t_N
are types (which can be unlifted, including unboxed tuples and sums).
Unboxed tuples can be used for multiarity alternatives. For example:
(# (# Int, String #)  Bool #)
The term level syntax is similar. Leading and preceding bars (
) indicate which alternative it is. Here are two terms of the type shown above:
(# (# 1, "foo" #)  #)  first alternative (#  True #)  second alternative
The pattern syntax reflects the term syntax:
case x of (# (# i, str #)  #) > ... (#  bool #) > ...
Unboxed sums are “unboxed” in the sense that, instead of allocating sums in the heap and representing values as pointers, unboxed sums are represented as their components, just like unboxed tuples. These “components” depend on alternatives of a sum type. Like unboxed tuples, unboxed sums are lazy in their lifted components.
The code generator tries to generate as compact layout as possible for each unboxed sum. In the best case, size of an unboxed sum is size of its biggest alternative plus one word (for a tag). The algorithm for generating the memory layout for a sum type works like this:
 All types are classified as one of these classes: 32bit word, 64bit word, 32bit float, 64bit float, pointer.
 For each alternative of the sum type, a layout that consists of these fields is generated. For example, if an alternative has
Int
,Float#
andString
fields, the layout will have an 32bit word, 32bit float and pointer fields. 
Layout fields are then overlapped so that the final layout will be as compact as possible. For example, suppose we have the unboxed sum:
(# (# Word32#, String, Float# #)  (# Float#, Float#, Maybe Int #) #)
The final layout will be something like
Int32, Float32, Float32, Word32, Pointer
The first
Int32
is for the tag. There are twoFloat32
fields because floating point types can’t overlap with other types, because of limitations of the code generator that we’re hoping to overcome in the future. The second alternative needs twoFloat32
fields: TheWord32
field is for theWord32#
in the first alternative. ThePointer
field is shared betweenString
andMaybe Int
values of the alternatives.As another example, this is the layout for the unboxed version of
Maybe a
type,(# (# #)  a #)
:Int32, Pointer
The
Pointer
field is not used when tag says that it’sNothing
. OtherwisePointer
points to the value inJust
. As mentioned above, this type is lazy in its lifted field. Therefore, the typedata Maybe' a = Maybe' (# (# #)  a #)
is precisely isomorphic to the type
Maybe a
, although its memory representation is different.In the degenerate case where all the alternatives have zero width, such as the
Bool
like(# (# #)  (# #) #)
, the unboxed sum layout only has anInt32
tag field (i.e., the whole thing is represented by an integer).
10.2.5. Unlifted Newtypes

UnliftedNewtypes

 Since

8.10.1
Enable the use of newtypes over types with nonlifted runtime representations.
GHC implements an UnliftedNewtypes
extension as specified in this GHC proposal. UnliftedNewtypes
relaxes the restrictions around what types can appear inside of a newtype
. For example, the type
newtype A = MkA Int#
is accepted when this extension is enabled. This creates a type A :: TYPE 'IntRep
and a data constructor MkA :: Int# > A
. Although the kind of A
is inferred by GHC, there is nothing visually distictive about this type that indicated that is it not of kind Type
like newtypes typically are. GADTSyntax can be used to provide a kind signature for additional clarity
newtype A :: TYPE 'IntRep where MkA :: Int# > A
The Coercible
machinery works with unlifted newtypes just like it does with lifted types. In either of the equivalent formulations of A
given above, users would additionally have access to a coercion between A
and Int#
.
As a consequence of the levitypolymorphic binder restriction, levitypolymorphic fields are disallowed in data constructors of data types declared using data
. However, since newtype
data constructor application is implemented as a coercion instead of as function application, this restriction does not apply to the field inside a newtype
data constructor. Thus, the type checker accepts
newtype Identity# :: forall (r :: RuntimeRep). TYPE r > TYPE r where MkIdentity# :: forall (r :: RuntimeRep) (a :: TYPE r). a > Identity# a
And with UnboxedSums enabled
newtype Maybe# :: forall (r :: RuntimeRep). TYPE r > TYPE (SumRep '[r, TupleRep '[]]) where MkMaybe# :: forall (r :: RuntimeRep) (a :: TYPE r). (# a  (# #) #) > Maybe# a
This extension also relaxes some of the restrictions around data family instances. In particular, UnliftedNewtypes
permits a newtype instance
to be given a return kind of TYPE r
, not just Type
. For example, the following newtype instance
declarations would be permitted:
class Foo a where data FooKey a :: TYPE 'IntRep class Bar (r :: RuntimeRep) where data BarType r :: TYPE r instance Foo Bool where newtype FooKey Bool = FooKeyBoolC Int# instance Bar 'WordRep where newtype BarType 'WordRep = BarTypeWordRepC Word#
It is worth noting that UnliftedNewtypes
is not required to give the data families themselves return kinds involving TYPE
, such as the FooKey
and BarType
examples above. The extension is only required for newtype instance
declarations, such as FooKeyBoolC
and BarTypeWorkRepC
above.
This extension impacts the determination of whether or not a newtype has a Complete UserSpecified Kind Signature (CUSK). The exact impact is specified the section on CUSKs.
10.3. Syntactic extensions
10.3.1. Unicode syntax

UnicodeSyntax

 Since

6.8.1
Enable the use of Unicode characters in place of their equivalent ASCII sequences.
The language extension UnicodeSyntax
enables Unicode characters to be used to stand for certain ASCII character sequences. The following alternatives are provided:
ASCII  Unicode alternative  Code point  Name 

 ∷  0x2237  PROPORTION 
 ⇒  0x21D2  RIGHTWARDS DOUBLE ARROW 
 →  0x2192  RIGHTWARDS ARROW 
 ←  0x2190  LEFTWARDS ARROW 
 ⤚  0x291a  RIGHTWARDS ARROWTAIL 
 ⤙  0x2919  LEFTWARDS ARROWTAIL 
 ⤜  0x291C  RIGHTWARDS DOUBLE ARROWTAIL 
 ⤛  0x291B  LEFTWARDS DOUBLE ARROWTAIL 
 ★  0x2605  BLACK STAR 
 ∀  0x2200  FOR ALL 
 ⦇  0x2987  Z NOTATION LEFT IMAGE BRACKET 
 ⦈  0x2988  Z NOTATION RIGHT IMAGE BRACKET 
 ⟦  0x27E6  MATHEMATICAL LEFT WHITE SQUARE BRACKET 
 ⟧  0x27E7  MATHEMATICAL RIGHT WHITE SQUARE BRACKET 
10.3.2. The magic hash

MagicHash

 Since

6.8.1
Enables the use of the hash character (
#
) as an identifier suffix.
The language extension MagicHash
allows #
as a postfix modifier to identifiers. Thus, x#
is a valid variable, and T#
is a valid type constructor or data constructor.
The hash sign does not change semantics at all. We tend to use variable names ending in “#” for unboxed values or types (e.g. Int#
), but there is no requirement to do so; they are just plain ordinary variables. Nor does the MagicHash
extension bring anything into scope. For example, to bring Int#
into scope you must import GHC.Prim
(see Unboxed types and primitive operations); the MagicHash
extension then allows you to refer to the Int#
that is now in scope. Note that with this option, the meaning of x#y = 0
is changed: it defines a function x#
taking a single argument y
; to define the operator #
, put a space: x # y = 0
.
The MagicHash
also enables some new forms of literals (see Unboxed types):

'x'#
has typeChar#

"foo"#
has typeAddr#

3#
has typeInt#
. In general, any Haskell integer lexeme followed by a#
is anInt#
literal, e.g.0x3A#
as well as32#
. 
3##
has typeWord#
. In general, any nonnegative Haskell integer lexeme followed by##
is aWord#
. 
3.2#
has typeFloat#
. 
3.2##
has typeDouble#
10.3.3. Negative literals

NegativeLiterals

 Since

7.8.1
Enable the use of unparenthesized negative numeric literals.
The literal 123
is, according to Haskell98 and Haskell 2010, desugared as negate (fromInteger 123)
. The language extension NegativeLiterals
means that it is instead desugared as fromInteger (123)
.
This can make a difference when the positive and negative range of a numeric data type don’t match up. For example, in 8bit arithmetic 128 is representable, but +128 is not. So negate (fromInteger 128)
will elicit an unexpected integerliteraloverflow message.
10.3.4. Fractional looking integer literals

NumDecimals

 Since

7.8.1
Allow the use of floatingpoint literal syntax for integral types.
Haskell 2010 and Haskell 98 define floating literals with the syntax 1.2e6
. These literals have the type Fractional a => a
.
The language extension NumDecimals
allows you to also use the floating literal syntax for instances of Integral
, and have values like (1.2e6 :: Num a => a)
10.3.5. Binary integer literals

BinaryLiterals

 Since

7.10.1
Allow the use of binary notation in integer literals.
Haskell 2010 and Haskell 98 allows for integer literals to be given in decimal, octal (prefixed by 0o
or 0O
), or hexadecimal notation (prefixed by 0x
or 0X
).
The language extension BinaryLiterals
adds support for expressing integer literals in binary notation with the prefix 0b
or 0B
. For instance, the binary integer literal 0b11001001
will be desugared into fromInteger 201
when BinaryLiterals
is enabled.
10.3.6. Hexadecimal floating point literals

HexFloatLiterals

 Since

8.4.1
Allow writing floating point literals using hexadecimal notation.
The hexadecimal notation for floating point literals is useful when you need to specify floating point constants precisely, as the literal notation corresponds closely to the underlying bitencoding of the number.
In this notation floating point numbers are written using hexadecimal digits, and so the digits are interpreted using base 16, rather then the usual 10. This means that digits left of the decimal point correspond to positive powers of 16, while the ones to the right correspond to negative ones.
You may also write an explicit exponent, which is similar to the exponent in decimal notation with the following differences:  the exponent begins with p
instead of e
 the exponent is written in base 10
(not 16)  the base of the exponent is 2
(not 16).
In terms of the underlying bit encoding, each hexadecimal digit corresponds to 4 bits, and you may think of the exponent as “moving” the floating point by one bit left (negative) or right (positive). Here are some examples:

0x0.1
is the same as1/16

0x0.01
is the same as1/256

0xF.FF
is the same as15 + 15/16 + 15/256

0x0.1p4
is the same as1

0x0.1p4
is the same as1/256

0x0.1p12
is the same as256
10.3.7. Numeric underscores

NumericUnderscores

 Since

8.6.1
Allow the use of underscores in numeric literals.
GHC allows for numeric literals to be given in decimal, octal, hexadecimal, binary, or float notation.
The language extension NumericUnderscores
adds support for expressing underscores in numeric literals. For instance, the numeric literal 1_000_000
will be parsed into 1000000
when NumericUnderscores
is enabled. That is, underscores in numeric literals are ignored when NumericUnderscores
is enabled. See also #14473.
For example:
 decimal million = 1_000_000 billion = 1_000_000_000 lightspeed = 299_792_458 version = 8_04_1 date = 2017_12_31  hexadecimal red_mask = 0xff_00_00 size1G = 0x3fff_ffff  binary bit8th = 0b01_0000_0000 packbits = 0b1_11_01_0000_0_111 bigbits = 0b1100_1011__1110_1111__0101_0011  float pi = 3.141_592_653_589_793 faraday = 96_485.332_89 avogadro = 6.022_140_857e+23  function isUnderMillion = (< 1_000_000) clip64M x  x > 0x3ff_ffff = 0x3ff_ffff  otherwise = x test8bit x = (0b01_0000_0000 .&. x) /= 0
About validity:
x0 = 1_000_000  valid x1 = 1__000000  valid x2 = 1000000_  invalid x3 = _1000000  invalid e0 = 0.0001  valid e1 = 0.000_1  valid e2 = 0_.0001  invalid e3 = _0.0001  invalid e4 = 0._0001  invalid e5 = 0.0001_  invalid f0 = 1e+23  valid f1 = 1_e+23  valid f2 = 1__e+23  valid f3 = 1e_+23  invalid g0 = 1e+23  valid g1 = 1e+_23  invalid g2 = 1e+23_  invalid h0 = 0xffff  valid h1 = 0xff_ff  valid h2 = 0x_ffff  valid h3 = 0x__ffff  valid h4 = _0xffff  invalid
10.3.8. Pattern guards

NoPatternGuards

 Implied by
 Since

6.8.1
Disable pattern guards.
10.3.9. View patterns

ViewPatterns

 Since

6.10.1
Allow use of view pattern syntax.
View patterns are enabled by the language extension ViewPatterns
. More information and examples of view patterns can be found on the Wiki page.
View patterns are somewhat like pattern guards that can be nested inside of other patterns. They are a convenient way of patternmatching against values of abstract types. For example, in a programming language implementation, we might represent the syntax of the types of the language as follows:
type Typ data TypView = Unit  Arrow Typ Typ view :: Typ > TypView  additional operations for constructing Typ's ...
The representation of Typ is held abstract, permitting implementations to use a fancy representation (e.g., hashconsing to manage sharing). Without view patterns, using this signature is a little inconvenient:
size :: Typ > Integer size t = case view t of Unit > 1 Arrow t1 t2 > size t1 + size t2
It is necessary to iterate the case, rather than using an equational function definition. And the situation is even worse when the matching against t
is buried deep inside another pattern.
View patterns permit calling the view function inside the pattern and matching against the result:
size (view > Unit) = 1 size (view > Arrow t1 t2) = size t1 + size t2
That is, we add a new form of pattern, written ⟨expression⟩ >
⟨pattern⟩ that means “apply the expression to whatever we’re trying to match against, and then match the result of that application against the pattern”. The expression can be any Haskell expression of function type, and view patterns can be used wherever patterns are used.
The semantics of a pattern (
⟨exp⟩ >
⟨pat⟩ )
are as follows:

Scoping: The variables bound by the view pattern are the variables bound by ⟨pat⟩.
Any variables in ⟨exp⟩ are bound occurrences, but variables bound “to the left” in a pattern are in scope. This feature permits, for example, one argument to a function to be used in the view of another argument. For example, the function
clunky
from Pattern guards can be written using view patterns as follows:clunky env (lookup env > Just val1) (lookup env > Just val2) = val1 + val2 ...other equations for clunky...
More precisely, the scoping rules are:

In a single pattern, variables bound by patterns to the left of a view pattern expression are in scope. For example:
example :: Maybe ((String > Integer,Integer), String) > Bool example (Just ((f,_), f > 4)) = True
Additionally, in function definitions, variables bound by matching earlier curried arguments may be used in view pattern expressions in later arguments:
example :: (String > Integer) > String > Bool example f (f > 4) = True
That is, the scoping is the same as it would be if the curried arguments were collected into a tuple.

In mutually recursive bindings, such as
let
,where
, or the top level, view patterns in one declaration may not mention variables bound by other declarations. That is, each declaration must be selfcontained. For example, the following program is not allowed:let {(x > y) = e1 ; (y > x) = e2 } in x
(For some amplification on this design choice see #4061.

 Typing: If ⟨exp⟩ has type ⟨T1⟩
>
⟨T2⟩ and ⟨pat⟩ matches a ⟨T2⟩, then the whole view pattern matches a ⟨T1⟩. 
Matching: To the equations in Section 3.17.3 of the Haskell 98 Report, add the following:
case v of { (e > p) > e1 ; _ > e2 } = case (e v) of { p > e1 ; _ > e2 }
That is, to match a variable ⟨v⟩ against a pattern
(
⟨exp⟩>
⟨pat⟩)
, evaluate(
⟨exp⟩ ⟨v⟩)
and match the result against ⟨pat⟩. 
Efficiency: When the same view function is applied in multiple branches of a function definition or a case expression (e.g., in
size
above), GHC makes an attempt to collect these applications into a single nested case expression, so that the view function is only applied once. Pattern compilation in GHC follows the matrix algorithm described in Chapter 4 of The Implementation of Functional Programming Languages. When the top rows of the first column of a matrix are all view patterns with the “same” expression, these patterns are transformed into a single nested case. This includes, for example, adjacent view patterns that line up in a tuple, as inf ((view > A, p1), p2) = e1 f ((view > B, p3), p4) = e2
The current notion of when two view pattern expressions are “the same” is very restricted: it is not even full syntactic equality. However, it does include variables, literals, applications, and tuples; e.g., two instances of
view ("hi", "there")
will be collected. However, the current implementation does not compare up to alphaequivalence, so two instances of(x, view x > y)
will not be coalesced.
10.3.10. n+k patterns

NPlusKPatterns

 Implied by
 Since

6.12.1
Enable use of
n+k
patterns.
10.3.11. The recursive donotation

RecursiveDo

 Since

6.8.1
Allow the use of recursive
do
notation.
The donotation of Haskell 98 does not allow recursive bindings, that is, the variables bound in a doexpression are visible only in the textually following code block. Compare this to a letexpression, where bound variables are visible in the entire binding group.
It turns out that such recursive bindings do indeed make sense for a variety of monads, but not all. In particular, recursion in this sense requires a fixedpoint operator for the underlying monad, captured by the mfix
method of the MonadFix
class, defined in Control.Monad.Fix
as follows:
class Monad m => MonadFix m where mfix :: (a > m a) > m a
Haskell’s Maybe
, []
(list), ST
(both strict and lazy versions), IO
, and many other monads have MonadFix
instances. On the negative side, the continuation monad, with the signature (a > r) > r
, does not.
For monads that do belong to the MonadFix
class, GHC provides an extended version of the donotation that allows recursive bindings. The RecursiveDo
(language pragma: RecursiveDo
) provides the necessary syntactic support, introducing the keywords mdo
and rec
for higher and lower levels of the notation respectively. Unlike bindings in a do
expression, those introduced by mdo
and rec
are recursively defined, much like in an ordinary letexpression. Due to the new keyword mdo
, we also call this notation the mdonotation.
Here is a simple (albeit contrived) example:
{# LANGUAGE RecursiveDo #} justOnes = mdo { xs < Just (1:xs) ; return (map negate xs) }
or equivalently
{# LANGUAGE RecursiveDo #} justOnes = do { rec { xs < Just (1:xs) } ; return (map negate xs) }
As you can guess justOnes
will evaluate to Just [1,1,1,...
.
GHC’s implementation the mdonotation closely follows the original translation as described in the paper A recursive do for Haskell, which in turn is based on the work Value Recursion in Monadic Computations. Furthermore, GHC extends the syntax described in the former paper with a lower level syntax flagged by the rec
keyword, as we describe next.
10.3.11.1. Recursive binding groups
The extension RecursiveDo
also introduces a new keyword rec
, which wraps a mutuallyrecursive group of monadic statements inside a do
expression, producing a single statement. Similar to a let
statement inside a do
, variables bound in the rec
are visible throughout the rec
group, and below it. For example, compare
do { a < getChar do { a < getChar ; let { r1 = f a r2 ; rec { r1 < f a r2 ; ; r2 = g r1 } ; ; r2 < g r1 } ; return (r1 ++ r2) } ; return (r1 ++ r2) }
In both cases, r1
and r2
are available both throughout the let
or rec
block, and in the statements that follow it. The difference is that let
is nonmonadic, while rec
is monadic. (In Haskell let
is really letrec
, of course.)
The semantics of rec
is fairly straightforward. Whenever GHC finds a rec
group, it will compute its set of bound variables, and will introduce an appropriate call to the underlying monadic valuerecursion operator mfix
, belonging to the MonadFix
class. Here is an example:
rec { b < f a c ===> (b,c) < mfix (\ ~(b,c) > do { b < f a c ; c < f b a } ; c < f b a ; return (b,c) })
As usual, the metavariables b
, c
etc., can be arbitrary patterns. In general, the statement rec ss
is desugared to the statement
vs < mfix (\ ~vs > do { ss; return vs })
where vs
is a tuple of the variables bound by ss
.
Note in particular that the translation for a rec
block only involves wrapping a call to mfix
: it performs no other analysis on the bindings. The latter is the task for the mdo
notation, which is described next.
10.3.11.2. The mdo
notation
A rec
block tells the compiler where precisely the recursive knot should be tied. It turns out that the placement of the recursive knots can be rather delicate: in particular, we would like the knots to be wrapped around as minimal groups as possible. This process is known as segmentation, and is described in detail in Section 3.2 of A recursive do for Haskell. Segmentation improves polymorphism and reduces the size of the recursive knot. Most importantly, it avoids unnecessary interference caused by a fundamental issue with the socalled rightshrinking axiom for monadic recursion. In brief, most monads of interest (IO, strict state, etc.) do not have recursion operators that satisfy this axiom, and thus not performing segmentation can cause unnecessary interference, changing the termination behavior of the resulting translation. (Details can be found in Sections 3.1 and 7.2.2 of Value Recursion in Monadic Computations.)
The mdo
notation removes the burden of placing explicit rec
blocks in the code. Unlike an ordinary do
expression, in which variables bound by statements are only in scope for later statements, variables bound in an mdo
expression are in scope for all statements of the expression. The compiler then automatically identifies minimal mutually recursively dependent segments of statements, treating them as if the user had wrapped a rec
qualifier around them.
The definition is syntactic:

A generator ⟨g⟩ depends on a textually following generator ⟨g’⟩, if
 ⟨g’⟩ defines a variable that is used by ⟨g⟩, or
 ⟨g’⟩ textually appears between ⟨g⟩ and ⟨g’’⟩, where ⟨g⟩ depends on ⟨g’’⟩.
 A segment of a given
mdo
expression is a minimal sequence of generators such that no generator of the sequence depends on an outside generator. As a special case, although it is not a generator, the final expression in anmdo
expression is considered to form a segment by itself.
Segments in this sense are related to stronglyconnected components analysis, with the exception that bindings in a segment cannot be reordered and must be contiguous.
Here is an example mdo
expression, and its translation to rec
blocks:
mdo { a < getChar ===> do { a < getChar ; b < f a c ; rec { b < f a c ; c < f b a ; ; c < f b a } ; z < h a b ; z < h a b ; d < g d e ; rec { d < g d e ; e < g a z ; ; e < g a z } ; putChar c } ; putChar c }
Note that a given mdo
expression can cause the creation of multiple rec
blocks. If there are no recursive dependencies, mdo
will introduce no rec
blocks. In this latter case an mdo
expression is precisely the same as a do
expression, as one would expect.
In summary, given an mdo
expression, GHC first performs segmentation, introducing rec
blocks to wrap over minimal recursive groups. Then, each resulting rec
is desugared, using a call to Control.Monad.Fix.mfix
as described in the previous section. The original mdo
expression typechecks exactly when the desugared version would do so.
Here are some other important points in using the recursivedo notation:
 It is enabled with the extension
RecursiveDo
, or theLANGUAGE RecursiveDo
pragma. (The same extension enables bothmdo
notation, and the use ofrec
blocks insidedo
expressions.) 
rec
blocks can also be used insidemdo
expressions, which will be treated as a single statement. However, it is good style to either usemdo
orrec
blocks in a single expression.  If recursive bindings are required for a monad, then that monad must be declared an instance of the
MonadFix
class.  The following instances of
MonadFix
are automatically provided: List, Maybe, IO. Furthermore, theControl.Monad.ST
andControl.Monad.ST.Lazy
modules provide the instances of theMonadFix
class for Haskell’s internal state monad (strict and lazy, respectively).  Like
let
andwhere
bindings, name shadowing is not allowed within anmdo
expression or arec
block; that is, all the names bound in a singlerec
must be distinct. (GHC will complain if this is not the case.)
10.3.12. Applicative donotation

ApplicativeDo

 Since

8.0.1
Allow use of
Applicative
do
notation.
The language option ApplicativeDo
enables an alternative translation for the donotation, which uses the operators <$>
, <*>
, along with join
as far as possible. There are two main reasons for wanting to do this:
 We can use donotation with types that are an instance of
Applicative
andFunctor
, but notMonad
 In some monads, using the applicative operators is more efficient than monadic bind. For example, it may enable more parallelism.
Applicative donotation desugaring preserves the original semantics, provided that the Applicative
instance satisfies <*> = ap
and pure = return
(these are true of all the common monadic types). Thus, you can normally turn on ApplicativeDo
without fear of breaking your program. There is one pitfall to watch out for; see Things to watch out for.
There are no syntactic changes with ApplicativeDo
. The only way it shows up at the source level is that you can have a do
expression that doesn’t require a Monad
constraint. For example, in GHCi:
Prelude> :set XApplicativeDo Prelude> :t \m > do { x < m; return (not x) } \m > do { x < m; return (not x) } :: Functor f => f Bool > f Bool
This example only requires Functor
, because it is translated into (\x >
not x) <$> m
. A more complex example requires Applicative
,
Prelude> :t \m > do { x < m 'a'; y < m 'b'; return (x  y) } \m > do { x < m 'a'; y < m 'b'; return (x  y) } :: Applicative f => (Char > f Bool) > f Bool
Here GHC has translated the expression into
(\x y > x  y) <$> m 'a' <*> m 'b'
It is possible to see the actual translation by using ddumpds
, but be warned, the output is quite verbose.
Note that if the expression can’t be translated into uses of <$>
, <*>
only, then it will incur a Monad
constraint as usual. This happens when there is a dependency on a value produced by an earlier statement in the do
block:
Prelude> :t \m > do { x < m True; y < m x; return (x  y) } \m > do { x < m True; y < m x; return (x  y) } :: Monad m => (Bool > m Bool) > m Bool
Here, m x
depends on the value of x
produced by the first statement, so the expression cannot be translated using <*>
.
In general, the rule for when a do
statement incurs a Monad
constraint is as follows. If the doexpression has the following form:
do p1 < E1; ...; pn < En; return E
where none of the variables defined by p1...pn
are mentioned in E1...En
, and p1...pn
are all variables or lazy patterns, then the expression will only require Applicative
. Otherwise, the expression will require Monad
. The block may return a pure expression E
depending upon the results p1...pn
with either return
or pure
.
Note: the final statement must match one of these patterns exactly:
return E
return $ E
pure E
pure $ E
otherwise GHC cannot recognise it as a return
statement, and the transformation to use <$>
that we saw above does not apply. In particular, slight variations such as return . Just $ x
or let x
= e in return x
would not be recognised.
If the final statement is not of one of these forms, GHC falls back to standard do
desugaring, and the expression will require a Monad
constraint.
When the statements of a do
expression have dependencies between them, and ApplicativeDo
cannot infer an Applicative
type, it uses a heuristic algorithm to try to use <*>
as much as possible. This algorithm usually finds the best solution, but in rare complex cases it might miss an opportunity. There is an algorithm that finds the optimal solution, provided as an option:

foptimalapplicativedo

 Since

8.0.1
Enables an alternative algorithm for choosing where to use
<*>
in conjunction with theApplicativeDo
language extension. This algorithm always finds the optimal solution, but it is expensive:O(n^3)
, so this option can lead to long compile times when there are very largedo
expressions (over 100 statements). The defaultApplicativeDo
algorithm isO(n^2)
.
10.3.12.1. Strict patterns
A strict pattern match in a bind statement prevents ApplicativeDo
from transforming that statement to use Applicative
. This is because the transformation would change the semantics by making the expression lazier.
For example, this code will require a Monad
constraint:
> :t \m > do { (x:xs) < m; return x } \m > do { (x:xs) < m; return x } :: Monad m => m [b] > m b
but making the pattern match lazy allows it to have a Functor
constraint:
> :t \m > do { ~(x:xs) < m; return x } \m > do { ~(x:xs) < m; return x } :: Functor f => f [b] > f b
A “strict pattern match” is any pattern match that can fail. For example, ()
, (x:xs)
, !z
, and C x
are strict patterns, but x
and ~(1,2)
are not. For the purposes of ApplicativeDo
, a pattern match against a newtype
constructor is considered strict.
When there’s a strict pattern match in a sequence of statements, ApplicativeDo
places a >>=
between that statement and the one that follows it. The sequence may be transformed to use <*>
elsewhere, but the strict pattern match and the following statement will always be connected with >>=
, to retain the same strictness semantics as the standard donotation. If you don’t want this, simply put a ~
on the pattern match to make it lazy.
10.3.12.2. Things to watch out for
Your code should just work as before when ApplicativeDo
is enabled, provided you use conventional Applicative
instances. However, if you define a Functor
or Applicative
instance using donotation, then it will likely get turned into an infinite loop by GHC. For example, if you do this:
instance Functor MyType where fmap f m = do x < m; return (f x)
Then applicative desugaring will turn it into
instance Functor MyType where fmap f m = fmap (\x > f x) m
And the program will loop at runtime. Similarly, an Applicative
instance like this
instance Applicative MyType where pure = return x <*> y = do f < x; a < y; return (f a)
will result in an infinte loop when <*>
is called.
Just as you wouldn’t define a Monad
instance using the donotation, you shouldn’t define Functor
or Applicative
instance using donotation (when using ApplicativeDo
) either. The correct way to define these instances in terms of Monad
is to use the Monad
operations directly, e.g.
instance Functor MyType where fmap f m = m >>= return . f instance Applicative MyType where pure = return (<*>) = ap
10.3.13. Parallel List Comprehensions

ParallelListComp

 Since

6.8.1
Allow parallel list comprehension syntax.
Parallel list comprehensions are a natural extension to list comprehensions. List comprehensions can be thought of as a nice syntax for writing maps and filters. Parallel comprehensions extend this to include the zipWith
family.
A parallel list comprehension has multiple independent branches of qualifier lists, each separated by a 
symbol. For example, the following zips together two lists:
[ (x, y)  x < xs  y < ys ]
The behaviour of parallel list comprehensions follows that of zip, in that the resulting list will have the same length as the shortest branch.
We can define parallel list comprehensions by translation to regular comprehensions. Here’s the basic idea:
Given a parallel comprehension of the form:
[ e  p1 < e11, p2 < e12, ...  q1 < e21, q2 < e22, ... ... ]
This will be translated to:
[ e  ((p1,p2), (q1,q2), ...) < zipN [(p1,p2)  p1 < e11, p2 < e12, ...] [(q1,q2)  q1 < e21, q2 < e22, ...] ... ]
where zipN
is the appropriate zip for the given number of branches.
10.3.14. Generalised (SQLlike) List Comprehensions

TransformListComp

 Since

6.10.1
Allow use of generalised list (SQLlike) comprehension syntax. This introduces the
group
,by
, andusing
keywords.
Generalised list comprehensions are a further enhancement to the list comprehension syntactic sugar to allow operations such as sorting and grouping which are familiar from SQL. They are fully described in the paper Comprehensive comprehensions: comprehensions with “order by” and “group by”, except that the syntax we use differs slightly from the paper.
The extension is enabled with the extension TransformListComp
.
Here is an example:
employees = [ ("Simon", "MS", 80) , ("Erik", "MS", 100) , ("Phil", "Ed", 40) , ("Gordon", "Ed", 45) , ("Paul", "Yale", 60) ] output = [ (the dept, sum salary)  (name, dept, salary) < employees , then group by dept using groupWith , then sortWith by (sum salary) , then take 5 ]
In this example, the list output
would take on the value:
[("Yale", 60), ("Ed", 85), ("MS", 180)]
There are three new keywords: group
, by
, and using
. (The functions sortWith
and groupWith
are not keywords; they are ordinary functions that are exported by GHC.Exts
.)
There are five new forms of comprehension qualifier, all introduced by the (existing) keyword then
:

then f
This statement requires that f have the type forall a. [a] > [a] . You can see an example of its use in the motivating example, as this form is used to apply take 5 .

then f by e
This form is similar to the previous one, but allows you to create a function which will be passed as the first argument to f. As a consequence f must have the type
forall a. (a > t) > [a] > [a]
. As you can see from the type, this function lets f “project out” some information from the elements of the list it is transforming.An example is shown in the opening example, where
sortWith
is supplied with a function that lets it find out thesum salary
for any item in the list comprehension it transforms. 
then group by e using f
This is the most general of the groupingtype statements. In this form, f is required to have type
forall a. (a > t) > [a] > [[a]]
. As with thethen f by e
case above, the first argument is a function supplied to f by the compiler which lets it compute e on every element of the list being transformed. However, unlike the nongrouping case, f additionally partitions the list into a number of sublists: this means that at every point after this statement, binders occurring before it in the comprehension refer to lists of possible values, not single values. To help understand this, let’s look at an example: This works similarly to groupWith in GHC.Exts, but doesn't sort its input first groupRuns :: Eq b => (a > b) > [a] > [[a]] groupRuns f = groupBy (\x y > f x == f y) output = [ (the x, y)  x < ([1..3] ++ [1..2]) , y < [4..6] , then group by x using groupRuns ]
This results in the variable
output
taking on the value below:[(1, [4, 5, 6]), (2, [4, 5, 6]), (3, [4, 5, 6]), (1, [4, 5, 6]), (2, [4, 5, 6])]
Note that we have used the
the
function to change the type of x from a list to its original numeric type. The variable y, in contrast, is left unchanged from the list form introduced by the grouping. 
then group using f
With this form of the group statement, f is required to simply have the type
forall a. [a] > [[a]]
, which will be used to group up the comprehension so far directly. An example of this form is as follows:output = [ x  y < [1..5] , x < "hello" , then group using inits]
This will yield a list containing every prefix of the word “hello” written out 5 times:
["","h","he","hel","hell","hello","helloh","hellohe","hellohel","hellohell","hellohello","hellohelloh",...]
10.3.15. Monad comprehensions

MonadComprehensions

 Since

7.2.1
Enable list comprehension syntax for arbitrary monads.
Monad comprehensions generalise the list comprehension notation, including parallel comprehensions (Parallel List Comprehensions) and transform comprehensions (Generalised (SQLlike) List Comprehensions) to work for any monad.
Monad comprehensions support:

Bindings:
[ x + y  x < Just 1, y < Just 2 ]
Bindings are translated with the
(>>=)
andreturn
functions to the usual donotation:do x < Just 1 y < Just 2 return (x+y)

Guards:
[ x  x < [1..10], x <= 5 ]
Guards are translated with the
guard
function, which requires aMonadPlus
instance:do x < [1..10] guard (x <= 5) return x

Transform statements (as with
TransformListComp
):[ x+y  x < [1..10], y < [1..x], then take 2 ]
This translates to:
do (x,y) < take 2 (do x < [1..10] y < [1..x] return (x,y)) return (x+y)

Group statements (as with
TransformListComp
):[ x  x < [1,1,2,2,3], then group by x using GHC.Exts.groupWith ] [ x  x < [1,1,2,2,3], then group using myGroup ]

Parallel statements (as with
ParallelListComp
):[ (x+y)  x < [1..10]  y < [11..20] ]
Parallel statements are translated using the
mzip
function, which requires aMonadZip
instance defined in Control.Monad.Zip:do (x,y) < mzip (do x < [1..10] return x) (do y < [11..20] return y) return (x+y)
All these features are enabled by default if the MonadComprehensions
extension is enabled. The types and more detailed examples on how to use comprehensions are explained in the previous chapters Generalised (SQLlike) List Comprehensions and Parallel List Comprehensions. In general you just have to replace the type [a]
with the type Monad m => m a
for monad comprehensions.
Note
Even though most of these examples are using the list monad, monad comprehensions work for any monad. The base
package offers all necessary instances for lists, which make MonadComprehensions
backward compatible to builtin, transform and parallel list comprehensions.
More formally, the desugaring is as follows. We write D[ e  Q]
to mean the desugaring of the monad comprehension [ e  Q]
:
Expressions: e Declarations: d Lists of qualifiers: Q,R,S  Basic forms D[ e  ] = return e D[ e  p < e, Q ] = e >>= \p > D[ e  Q ] D[ e  e, Q ] = guard e >> \p > D[ e  Q ] D[ e  let d, Q ] = let d in D[ e  Q ]  Parallel comprehensions (iterate for multiple parallel branches) D[ e  (Q  R), S ] = mzip D[ Qv  Q ] D[ Rv  R ] >>= \(Qv,Rv) > D[ e  S ]  Transform comprehensions D[ e  Q then f, R ] = f D[ Qv  Q ] >>= \Qv > D[ e  R ] D[ e  Q then f by b, R ] = f (\Qv > b) D[ Qv  Q ] >>= \Qv > D[ e  R ] D[ e  Q then group using f, R ] = f D[ Qv  Q ] >>= \ys > case (fmap selQv1 ys, ..., fmap selQvn ys) of Qv > D[ e  R ] D[ e  Q then group by b using f, R ] = f (\Qv > b) D[ Qv  Q ] >>= \ys > case (fmap selQv1 ys, ..., fmap selQvn ys) of Qv > D[ e  R ] where Qv is the tuple of variables bound by Q (and used subsequently) selQvi is a selector mapping Qv to the ith component of Qv Operator Standard binding Expected type  return GHC.Base t1 > m t2 (>>=) GHC.Base m1 t1 > (t2 > m2 t3) > m3 t3 (>>) GHC.Base m1 t1 > m2 t2 > m3 t3 guard Control.Monad t1 > m t2 fmap GHC.Base forall a b. (a>b) > n a > n b mzip Control.Monad.Zip forall a b. m a > m b > m (a,b)
The comprehension should typecheck when its desugaring would typecheck, except that (as discussed in Generalised (SQLlike) List Comprehensions) in the “then f
” and “then group using f
” clauses, when the “by b
” qualifier is omitted, argument f
should have a polymorphic type. In particular, “then Data.List.sort
” and “then group using Data.List.group
” are insufficiently polymorphic.
Monad comprehensions support rebindable syntax (Rebindable syntax and the implicit Prelude import). Without rebindable syntax, the operators from the “standard binding” module are used; with rebindable syntax, the operators are looked up in the current lexical scope. For example, parallel comprehensions will be typechecked and desugared using whatever “mzip
” is in scope.
The rebindable operators must have the “Expected type” given in the table above. These types are surprisingly general. For example, you can use a bind operator with the type
(>>=) :: T x y a > (a > T y z b) > T x z b
In the case of transform comprehensions, notice that the groups are parameterised over some arbitrary type n
(provided it has an fmap
, as well as the comprehension being over an arbitrary monad.
10.3.16. New monadic failure desugaring mechanism

MonadFailDesugaring

 Since

8.0.1
Use the
MonadFail.fail
instead of the legacyMonad.fail
function when desugaring refutable patterns indo
blocks.
The XMonadFailDesugaring
extension switches the desugaring of do
blocks to use MonadFail.fail
instead of Monad.fail
.
This extension is enabled by default since GHC 8.6.1, under the MonadFail Proposal (MFP).
This extension is temporary, and will be deprecated in a future release.
10.3.17. Rebindable syntax and the implicit Prelude import

NoImplicitPrelude

 Since

6.8.1
Don’t import
Prelude
by default.
GHC normally imports Prelude.hi
files for you. If you’d rather it didn’t, then give it a XNoImplicitPrelude
option. The idea is that you can then import a Prelude of your own. (But don’t call it Prelude
; the Haskell module namespace is flat, and you must not conflict with any Prelude module.)

RebindableSyntax

 Implies
 Since

7.0.1
Enable rebinding of a variety of usuallybuiltin operations.
Suppose you are importing a Prelude of your own in order to define your own numeric class hierarchy. It completely defeats that purpose if the literal “1” means “Prelude.fromInteger 1
”, which is what the Haskell Report specifies. So the RebindableSyntax
extension causes the following pieces of builtin syntax to refer to whatever is in scope, not the Prelude versions:
 An integer literal
368
means “fromInteger (368::Integer)
”, rather than “Prelude.fromInteger (368::Integer)
”.  Fractional literals are handled in just the same way, except that the translation is
fromRational (3.68::Rational)
.  String literals are also handled the same way, except that the translation is
fromString ("368"::String)
.  The equality test in an overloaded numeric pattern uses whatever
(==)
is in scope.  The subtraction operation, and the greaterthanorequal test, in
n+k
patterns use whatever()
and(>=)
are in scope.  Negation (e.g. “
 (f x)
”) means “negate (f x)
”, both in numeric patterns, and expressions.  Conditionals (e.g. “
if
e1then
e2else
e3”) means “ifThenElse
e1 e2 e3”. Howevercase
expressions are unaffected.  “Do” notation is translated using whatever functions
(>>=)
,(>>)
, andfail
, are in scope (not the Prelude versions). List comprehensions,mdo
(The recursive donotation), and parallel array comprehensions, are unaffected.  Arrow notation (see Arrow notation) uses whatever
arr
,(>>>)
,first
,app
,()
andloop
functions are in scope. But unlike the other constructs, the types of these functions must match the Prelude types very closely. Details are in flux; if you want to use this, ask!  List notation, such as
[x,y]
or[m..n]
can also be treated via rebindable syntax if you useXOverloadedLists
; see Overloaded lists.  An overloaded label “
#foo
” means “fromLabel @"foo"
”, rather than “GHC.OverloadedLabels.fromLabel @"foo"
” (see Overloaded labels).
RebindableSyntax
implies NoImplicitPrelude
.
In all cases (apart from arrow notation), the static semantics should be that of the desugared form, even if that is a little unexpected. For example, the static semantics of the literal 368
is exactly that of fromInteger (368::Integer)
; it’s fine for fromInteger
to have any of the types:
fromInteger :: Integer > Integer fromInteger :: forall a. Foo a => Integer > a fromInteger :: Num a => a > Integer fromInteger :: Integer > Bool > Bool
Be warned: this is an experimental facility, with fewer checks than usual. Use dcorelint
to typecheck the desugared program. If Core Lint is happy you should be all right.
10.3.17.1. Things unaffected by RebindableSyntax
RebindableSyntax
does not apply to any code generated from a deriving
clause or declaration. To see why, consider the following code:
{# LANGUAGE RebindableSyntax, OverloadedStrings #} newtype Text = Text String fromString :: String > Text fromString = Text data Foo = Foo deriving Show
This will generate code to the effect of:
instance Show Foo where showsPrec _ Foo = showString "Foo"
But because RebindableSyntax
and OverloadedStrings
are enabled, the "Foo"
string literal would now be of type Text
, not String
, which showString
doesn’t accept! This causes the generated Show
instance to fail to typecheck. It’s hard to imagine any scenario where it would be desirable have RebindableSyntax
behavior within derived code, so GHC simply ignores RebindableSyntax
entirely when checking derived code.
10.3.18. Postfix operators

PostfixOperators

 Since

7.10.1
Allow the use of postfix operators
The PostfixOperators
extension enables a small extension to the syntax of left operator sections, which allows you to define postfix operators. The extension is this: the left section
(e !)
is equivalent (from the point of view of both type checking and execution) to the expression
((!) e)
(for any expression e
and operator (!)
. The strict Haskell 98 interpretation is that the section is equivalent to
(\y > (!) e y)
That is, the operator must be a function of two arguments. GHC allows it to take only one argument, and that in turn allows you to write the function postfix.
The extension does not extend to the lefthand side of function definitions; you must define such a function in prefix form.
10.3.19. Tuple sections

TupleSections

 Since

6.12
Allow the use of tuple section syntax
The TupleSections
extension enables partially applied tuple constructors. For example, the following program
(, True)
is considered to be an alternative notation for the more unwieldy alternative
\x > (x, True)
You can omit any combination of arguments to the tuple, as in the following
(, "I", , , "Love", , 1337)
which translates to
\a b c d > (a, "I", b, c, "Love", d, 1337)
If you have unboxed tuples enabled, tuple sections will also be available for them, like so
(# , True #)
Because there is no unboxed unit tuple, the following expression
(# #)
continues to stand for the unboxed singleton tuple data constructor.
10.3.20. Lambdacase

LambdaCase

 Since

7.6.1
Allow the use of lambdacase syntax.
The LambdaCase
extension enables expressions of the form
\case { p1 > e1; ...; pN > eN }
which is equivalent to
\freshName > case freshName of { p1 > e1; ...; pN > eN }
Note that \case
starts a layout, so you can write
\case p1 > e1 ... pN > eN
10.3.21. Empty case alternatives

EmptyCase

 Since

7.8.1
Allow empty case expressions.
The EmptyCase
extension enables case expressions, or lambdacase expressions, that have no alternatives, thus:
case e of { }  No alternatives
or
\case { }  XLambdaCase is also required
This can be useful when you know that the expression being scrutinised has no nonbottom values. For example:
data Void f :: Void > Int f x = case x of { }
With dependentlytyped features it is more useful (see #2431). For example, consider these two candidate definitions of absurd
:
data a :~: b where Refl :: a :~: a absurd :: True :~: False > a absurd x = error "absurd"  (A) absurd x = case x of {}  (B)
We much prefer (B). Why? Because GHC can figure out that (True :~: False)
is an empty type. So (B) has no partiality and GHC is able to compile with Wincompletepatterns
and Werror
. On the other hand (A) looks dangerous, and GHC doesn’t check to make sure that, in fact, the function can never get called.
10.3.22. Multiway ifexpressions

MultiWayIf

 Since

7.6.1
Allow the use of multiway
if
syntax.
With MultiWayIf
extension GHC accepts conditional expressions with multiple branches:
if  guard1 > expr1  ...  guardN > exprN
which is roughly equivalent to
case () of _  guard1 > expr1 ... _  guardN > exprN
Multiway if expressions introduce a new layout context. So the example above is equivalent to:
if {  guard1 > expr1 ;  ... ;  guardN > exprN }
The following behaves as expected:
if  guard1 > if  guard2 > expr2  guard3 > expr3  guard4 > expr4
because layout translates it as
if {  guard1 > if {  guard2 > expr2 ;  guard3 > expr3 } ;  guard4 > expr4 }
Layout with multiway if works in the same way as other layout contexts, except that the semicolons between guards in a multiway if are optional. So it is not necessary to line up all the guards at the same column; this is consistent with the way guards work in function definitions and case expressions.
10.3.23. Local Fixity Declarations
A careful reading of the Haskell 98 Report reveals that fixity declarations (infix
, infixl
, and infixr
) are permitted to appear inside local bindings such those introduced by let
and where
. However, the Haskell Report does not specify the semantics of such bindings very precisely.
In GHC, a fixity declaration may accompany a local binding:
let f = ... infixr 3 `f` in ...
and the fixity declaration applies wherever the binding is in scope. For example, in a let
, it applies in the righthand sides of other let
bindings and the body of the let
C. Or, in recursive do
expressions (The recursive donotation), the local fixity declarations of a let
statement scope over other statements in the group, just as the bound name does.
Moreover, a local fixity declaration must accompany a local binding of that name: it is not possible to revise the fixity of name bound elsewhere, as in
let infixr 9 $ in ...
Because local fixity declarations are technically Haskell 98, no extension is necessary to enable them.
10.3.24. Import and export extensions
10.3.24.1. Hiding things the imported module doesn’t export
Technically in Haskell 2010 this is illegal:
module A( f ) where f = True module B where import A hiding( g )  A does not export g g = f
The import A hiding( g )
in module B
is technically an error (Haskell Report, 5.3.1) because A
does not export g
. However GHC allows it, in the interests of supporting backward compatibility; for example, a newer version of A
might export g
, and you want B
to work in either case.
The warning Wdodgyimports
, which is off by default but included with W
, warns if you hide something that the imported module does not export.
10.3.24.2. Packagequalified imports

PackageImports

 Since

6.10.1
Allow the use of packagequalified
import
syntax.
With the PackageImports
extension, GHC allows import declarations to be qualified by the package name that the module is intended to be imported from. For example:
import "network" Network.Socket
would import the module Network.Socket
from the package network
(any version). This may be used to disambiguate an import when the same module is available from multiple packages, or is present in both the current package being built and an external package.
The special package name this
can be used to refer to the current package being built.
Note
You probably don’t need to use this feature, it was added mainly so that we can build backwardscompatible versions of packages when APIs change. It can lead to fragile dependencies in the common case: modules occasionally move from one package to another, rendering any packagequalified imports broken. See also Thinning and renaming modules for an alternative way of disambiguating between module names.
10.3.24.3. Safe imports

Safe

 Since

7.2.1
Declare the Safe Haskell state of the current module.

Trustworthy

 Since

7.2.1
Declare the Safe Haskell state of the current module.

Unsafe

 Since

7.4.1
Declare the Safe Haskell state of the current module.
With the Safe
, Trustworthy
and Unsafe
language flags, GHC extends the import declaration syntax to take an optional safe
keyword after the import
keyword. This feature is part of the Safe Haskell GHC extension. For example:
import safe qualified Network.Socket as NS
would import the module Network.Socket
with compilation only succeeding if Network.Socket
can be safely imported. For a description of when a import is considered safe see Safe Haskell.
10.3.24.4. Explicit namespaces in import/export

ExplicitNamespaces

 Since

7.6.1
Enable use of explicit namespaces in module export lists.
In an import or export list, such as
module M( f, (++) ) where ... import N( f, (++) ) ...
the entities f
and (++)
are values. However, with type operators (Type operators) it becomes possible to declare (++)
as a type constructor. In that case, how would you export or import it?
The ExplicitNamespaces
extension allows you to prefix the name of a type constructor in an import or export list with “type
” to disambiguate this case, thus:
module M( f, type (++) ) where ... import N( f, type (++) ) ... module N( f, type (++) ) where data family a ++ b = L a  R b
The extension ExplicitNamespaces
is implied by TypeOperators
and (for some reason) by TypeFamilies
.
In addition, with PatternSynonyms
you can prefix the name of a data constructor in an import or export list with the keyword pattern
, to allow the import or export of a data constructor without its parent type constructor (see Import and export of pattern synonyms).
10.3.24.5. Writing qualified in postpositive position

ImportQualifiedPost

 Since

8.10.1
ImportQualifiedPost
allows the syntaximport M qualified
, that is, to annotate a module as qualified by writingqualified
after the module name.
To import a qualified module usually you must specify qualified
in prepositive position : import qualified M
. This often leads to a “hanging indent” (which is automatically inserted by some autoformatters and common in many code bases. For example:
import qualified A import B import C
The ImportQualifiedPost
extension allows qualified
to appear in postpositive position : import M qualified
. With this extension enabled, one can write:
import A qualified import B import C
It is an error if qualified
appears in both pre and postpositive positions.
The warning Wprepositivequalifiedsyntax
(off by default) reports on any occurrences of imports annotated qualified
using prepositive syntax.
10.3.25. More liberal syntax for function arguments

BlockArguments

 Since

8.6.1
Allow
do
expressions, lambda expressions, etc. to be directly used as a function argument.
In Haskell 2010, certain kinds of expressions can be used without parentheses as an argument to an operator, but not as an argument to a function. They include do
, lambda, if
, case
, and let
expressions. Some GHC extensions also define language constructs of this type: mdo
(The recursive donotation), \case
(Lambdacase), and proc
(Arrow notation).
The BlockArguments
extension allows these constructs to be directly used as a function argument. For example:
when (x > 0) do print x exitFailure
will be parsed as:
when (x > 0) (do print x exitFailure)
and
withForeignPtr fptr \ptr > c_memcpy buf ptr size
will be parsed as:
withForeignPtr fptr (\ptr > c_memcpy buf ptr size)
10.3.25.1. Changes to the grammar
The Haskell report defines the lexp
nonterminal thus (*
indicates a rule of interest)
lexp → \ apat1 … apatn > exp (lambda abstraction, n ≥ 1) *  let decls in exp (let expression) *  if exp [;] then exp [;] else exp (conditional) *  case exp of { alts } (case expression) *  do { stmts } (do expression) *  fexp fexp → [fexp] aexp (function application) aexp → qvar (variable)  gcon (general constructor)  literal  ( exp ) (parenthesized expression)  qcon { fbind1 … fbindn } (labeled construction)  aexp { fbind1 … fbindn } (labelled update)  …
The BlockArguments
extension moves these production rules under aexp
lexp → fexp fexp → [fexp] aexp (function application) aexp → qvar (variable)  gcon (general constructor)  literal  ( exp ) (parenthesized expression)  qcon { fbind1 … fbindn } (labeled construction)  aexp { fbind1 … fbindn } (labelled update)  \ apat1 … apatn > exp (lambda abstraction, n ≥ 1) *  let decls in exp (let expression) *  if exp [;] then exp [;] else exp (conditional) *  case exp of { alts } (case expression) *  do { stmts } (do expression) *  …
Now the lexp
nonterminal is redundant and can be dropped from the grammar.
Note that this change relies on an existing metarule to resolve ambiguities:
The grammar is ambiguous regarding the extent of lambda abstractions, let expressions, and conditionals. The ambiguity is resolved by the metarule that each of these constructs extends as far to the right as possible.
For example, f \a > a b
will be parsed as f (\a > a b)
, not as f
(\a > a) b
.
10.3.26. Summary of stolen syntax
Turning on an option that enables special syntax might cause working Haskell 98 code to fail to compile, perhaps because it uses a variable name which has become a reserved word. This section lists the syntax that is “stolen” by language extensions. We use notation and nonterminal names from the Haskell 98 lexical syntax (see the Haskell 98 Report). We only list syntax changes here that might affect existing working programs (i.e. “stolen” syntax). Many of these extensions will also enable new contextfree syntax, but in all cases programs written to use the new syntax would not be compilable without the option enabled.
There are two classes of special syntax:
 New reserved words and symbols: character sequences which are no longer available for use as identifiers in the program.
 Other special syntax: sequences of characters that have a different meaning when this particular option is turned on.
The following syntax is stolen:

forall

Stolen (in types) by default (see Lexical syntax).
forall
is a reserved keyword and never a type variable, in accordance with GHC Proposal #43. 
mdo

Stolen by:
RecursiveDo

foreign

Stolen by:
ForeignFunctionInterface

rec, proc, <, >, <<, >>, (, )

Stolen by:
Arrows

?varid

Stolen by:
ImplicitParams

[, [e, [p, [d, [t, [, [e

Stolen by:
QuasiQuotes
. Moreover, this introduces an ambiguity with list comprehension syntax. See the discussion on quasiquoting for details. 
$(, $$(, $varid, $$varid

Stolen by:
TemplateHaskell

[varid

Stolen by:
QuasiQuotes

⟨varid⟩, #⟨char⟩, #, ⟨string⟩, #, ⟨integer⟩, #, ⟨float⟩, #, ⟨float⟩, ##

Stolen by:
MagicHash

(#, #)

Stolen by:
UnboxedTuples

⟨varid⟩, !, ⟨varid⟩

Stolen by:
BangPatterns

pattern

Stolen by:
PatternSynonyms

static

Stolen by:
StaticPointers
10.4. Extensions to data types and type synonyms
10.4.1. Data types with no constructors

EmptyDataDecls

 Since

6.8.1
Allow definition of empty
data
types.
With the EmptyDataDecls
extension, GHC lets you declare a data type with no constructors.
You only need to enable this extension if the language you’re using is Haskell 98, in which a data type must have at least one constructor. Haskell 2010 relaxed this rule to allow data types with no constructors, and thus EmptyDataDecls
is enabled by default when the language is Haskell 2010.
For example:
data S  S :: Type data T a  T :: Type > Type
Syntactically, the declaration lacks the “= constrs” part. The type can be parameterised over types of any kind, but if the kind is not Type
then an explicit kind annotation must be used (see Explicitlykinded quantification).
Such data types have only one value, namely bottom. Nevertheless, they can be useful when defining “phantom types”.
In conjunction with the EmptyDataDeriving
extension, empty data declarations can also derive instances of standard type classes (see Deriving instances for empty data types).
10.4.2. Data type contexts

DatatypeContexts

 Since

7.0.1
Allow contexts on
data
types.
Haskell allows datatypes to be given contexts, e.g.
data Eq a => Set a = NilSet  ConsSet a (Set a)
give constructors with types:
NilSet :: Set a ConsSet :: Eq a => a > Set a > Set a
This is widely considered a misfeature, and is going to be removed from the language. In GHC, it is controlled by the deprecated extension DatatypeContexts
.
10.4.3. Infix type constructors, classes, and type variables
GHC allows type constructors, classes, and type variables to be operators, and to be written infix, very much like expressions. More specifically:
 A type constructor or class can be any nonreserved operator. Symbols used in types are always like capitalized identifiers; they are never variables. Note that this is different from the lexical syntax of data constructors, which are required to begin with a
:
. 
Data type and typesynonym declarations can be written infix, parenthesised if you want further arguments. E.g.
data a :*: b = Foo a b type a :+: b = Either a b class a :=: b where ... data (a :**: b) x = Baz a b x type (a :++: b) y = Either (a,b) y

Types, and class constraints, can be written infix. For example
x :: Int :*: Bool f :: (a :=: b) => a > b
 Backquotes work as for expressions, both for type constructors and type variables; e.g.
Int `Either` Bool
, orInt `a` Bool
. Similarly, parentheses work the same; e.g.(:*:) Int Bool
. 
Fixities may be declared for type constructors, or classes, just as for data constructors. However, one cannot distinguish between the two in a fixity declaration; a fixity declaration sets the fixity for a data constructor and the corresponding type constructor. For example:
infixl 7 T, :*:
sets the fixity for both type constructor
T
and data constructorT
, and similarly for:*:
.Int `a` Bool
.  The function arrow
>
isinfixr
with fixity 1.
10.4.4. Type operators

TypeOperators

 Implies
 Since

6.8.1
Allow the use and definition of types with operator names.
In types, an operator symbol like (+)
is normally treated as a type variable, just like a
. Thus in Haskell 98 you can say
type T (+) = ((+), (+))  Just like: type T a = (a,a) f :: T Int > Int f (x,y)= x
As you can see, using operators in this way is not very useful, and Haskell 98 does not even allow you to write them infix.
The language TypeOperators
changes this behaviour:
 Operator symbols become type constructors rather than type variables.

Operator symbols in types can be written infix, both in definitions and uses. For example:
data a + b = Plus a b type Foo = Int + Bool

There is now some potential ambiguity in import and export lists; for example if you write
import M( (+) )
do you mean the function(+)
or the type constructor(+)
? The default is the former, but withExplicitNamespaces
(which is implied byTypeOperators
) GHC allows you to specify the latter by preceding it with the keywordtype
, thus:import M( type (+) )
 The fixity of a type operator may be set using the usual fixity declarations but, as in Infix type constructors, classes, and type variables, the function and type constructor share a single fixity.
10.4.5. Liberalised type synonyms

LiberalTypeSynonyms

 Implies
 Since

6.8.1
Relax many of the Haskell 98 rules on type synonym definitions.
Type synonyms are like macros at the type level, but Haskell 98 imposes many rules on individual synonym declarations. With the LiberalTypeSynonyms
extension, GHC does validity checking on types only after expanding type synonyms. That means that GHC can be very much more liberal about type synonyms than Haskell 98.

You can write a
forall
(including overloading) in a type synonym, thus:type Discard a = forall b. Show b => a > b > (a, String) f :: Discard a f x y = (x, show y) g :: Discard Int > (Int,String)  A rank2 type g f = f 3 True

If you also use
UnboxedTuples
, you can write an unboxed tuple in a type synonym:type Pr = (# Int, Int #) h :: Int > Pr h x = (# x, x #)

You can apply a type synonym to a forall type:
type Foo a = a > a > Bool f :: Foo (forall b. b>b)
After expanding the synonym,
f
has the legal (in GHC) type:f :: (forall b. b>b) > (forall b. b>b) > Bool

You can apply a type synonym to a partially applied type synonym:
type Generic i o = forall x. i x > o x type Id x = x foo :: Generic Id []
After expanding the synonym,
foo
has the legal (in GHC) type:foo :: forall x. x > [x]
GHC currently does kind checking before expanding synonyms (though even that could be changed).
After expanding type synonyms, GHC does validity checking on types, looking for the following malformedness which isn’t detected simply by kind checking:
 Type constructor applied to a type involving foralls (if
ImpredicativeTypes
is off)  Partiallyapplied type synonym.
So, for example, this will be rejected:
type Pr = forall a. a h :: [Pr] h = ...
because GHC does not allow type constructors applied to forall types.
10.4.6. Existentially quantified data constructors

ExistentialQuantification

 Implies
 Since

6.8.1
Allow existentially quantified type variables in types.
The idea of using existential quantification in data type declarations was suggested by Perry, and implemented in Hope+ (Nigel Perry, The Implementation of Practical Functional Programming Languages, PhD Thesis, University of London, 1991). It was later formalised by Laufer and Odersky (Polymorphic type inference and abstract data types, TOPLAS, 16(5), pp. 14111430, 1994). It’s been in Lennart Augustsson’s hbc
Haskell compiler for several years, and proved very useful. Here’s the idea. Consider the declaration:
data Foo = forall a. MkFoo a (a > Bool)  Nil
The data type Foo
has two constructors with types:
MkFoo :: forall a. a > (a > Bool) > Foo Nil :: Foo
Notice that the type variable a
in the type of MkFoo
does not appear in the data type itself, which is plain Foo
. For example, the following expression is fine:
[MkFoo 3 even, MkFoo 'c' isUpper] :: [Foo]
Here, (MkFoo 3 even)
packages an integer with a function even
that maps an integer to Bool
; and MkFoo 'c'
isUpper
packages a character with a compatible function. These two things are each of type Foo
and can be put in a list.
What can we do with a value of type Foo
? In particular, what happens when we patternmatch on MkFoo
?
f (MkFoo val fn) = ???
Since all we know about val
and fn
is that they are compatible, the only (useful) thing we can do with them is to apply fn
to val
to get a boolean. For example:
f :: Foo > Bool f (MkFoo val fn) = fn val
What this allows us to do is to package heterogeneous values together with a bunch of functions that manipulate them, and then treat that collection of packages in a uniform manner. You can express quite a bit of objectorientedlike programming this way.
10.4.6.1. Why existential?
What has this to do with existential quantification? Simply that MkFoo
has the (nearly) isomorphic type
MkFoo :: (exists a . (a, a > Bool)) > Foo
But Haskell programmers can safely think of the ordinary universally quantified type given above, thereby avoiding adding a new existential quantification construct.
10.4.6.2. Existentials and type classes
An easy extension is to allow arbitrary contexts before the constructor. For example:
data Baz = forall a. Eq a => Baz1 a a  forall b. Show b => Baz2 b (b > b)
The two constructors have the types you’d expect:
Baz1 :: forall a. Eq a => a > a > Baz Baz2 :: forall b. Show b => b > (b > b) > Baz
But when pattern matching on Baz1
the matched values can be compared for equality, and when pattern matching on Baz2
the first matched value can be converted to a string (as well as applying the function to it). So this program is legal:
f :: Baz > String f (Baz1 p q)  p == q = "Yes"  otherwise = "No" f (Baz2 v fn) = show (fn v)
Operationally, in a dictionarypassing implementation, the constructors Baz1
and Baz2
must store the dictionaries for Eq
and Show
respectively, and extract it on pattern matching.
10.4.6.3. Record Constructors
GHC allows existentials to be used with records syntax as well. For example:
data Counter a = forall self. NewCounter { _this :: self , _inc :: self > self , _display :: self > IO () , tag :: a }
Here tag
is a public field, with a welltyped selector function tag :: Counter a > a
. The self
type is hidden from the outside; any attempt to apply _this
, _inc
or _display
as functions will raise a compiletime error. In other words, GHC defines a record selector function only for fields whose type does not mention the existentiallyquantified variables. (This example used an underscore in the fields for which record selectors will not be defined, but that is only programming style; GHC ignores them.)
To make use of these hidden fields, we need to create some helper functions:
inc :: Counter a > Counter a inc (NewCounter x i d t) = NewCounter { _this = i x, _inc = i, _display = d, tag = t } display :: Counter a > IO () display NewCounter{ _this = x, _display = d } = d x
Now we can define counters with different underlying implementations:
counterA :: Counter String counterA = NewCounter { _this = 0, _inc = (1+), _display = print, tag = "A" } counterB :: Counter String counterB = NewCounter { _this = "", _inc = ('#':), _display = putStrLn, tag = "B" } main = do display (inc counterA)  prints "1" display (inc (inc counterB))  prints "##"
Record update syntax is supported for existentials (and GADTs):
setTag :: Counter a > a > Counter a setTag obj t = obj{ tag = t }
The rule for record update is this:
the types of the updated fields may mention only the universallyquantified type variables of the data constructor. For GADTs, the field may mention only types that appear as a simple typevariable argument in the constructor’s result type.
For example:
data T a b where { T1 { f1::a, f2::b, f3::(b,c) } :: T a b }  c is existential upd1 t x = t { f1=x }  OK: upd1 :: T a b > a' > T a' b upd2 t x = t { f3=x }  BAD (f3's type mentions c, which is  existentially quantified) data G a b where { G1 { g1::a, g2::c } :: G a [c] } upd3 g x = g { g1=x }  OK: upd3 :: G a b > c > G c b upd4 g x = g { g2=x }  BAD (f2's type mentions c, which is not a simple  typevariable argument in G1's result type)
10.4.6.4. Restrictions
There are several restrictions on the ways in which existentiallyquantified constructors can be used.

When pattern matching, each pattern match introduces a new, distinct, type for each existential type variable. These types cannot be unified with any other type, nor can they escape from the scope of the pattern match. For example, these fragments are incorrect:
f1 (MkFoo a f) = a
Here, the type bound by
MkFoo
“escapes”, becausea
is the result off1
. One way to see why this is wrong is to ask what typef1
has:f1 :: Foo > a  Weird!
What is this “
a
” in the result type? Clearly we don’t mean this:f1 :: forall a. Foo > a  Wrong!
The original program is just plain wrong. Here’s another sort of error
f2 (Baz1 a b) (Baz1 p q) = a==q
It’s ok to say
a==b
orp==q
, buta==q
is wrong because it equates the two distinct types arising from the twoBaz1
constructors. 
You can’t patternmatch on an existentially quantified constructor in a
let
orwhere
group of bindings. So this is illegal:f3 x = a==b where { Baz1 a b = x }
Instead, use a
case
expression:f3 x = case x of Baz1 a b > a==b
In general, you can only patternmatch on an existentiallyquantified constructor in a
case
expression or in the patterns of a function definition. The reason for this restriction is really an implementation one. Typechecking binding groups is already a nightmare without existentials complicating the picture. Also an existential pattern binding at the top level of a module doesn’t make sense, because it’s not clear how to prevent the existentiallyquantified type “escaping”. So for now, there’s a simpletostate restriction. We’ll see how annoying it is. 
You can’t use existential quantification for
newtype
declarations. So this is illegal:newtype T = forall a. Ord a => MkT a
Reason: a value of type
T
must be represented as a pair of a dictionary forOrd t
and a value of typet
. That contradicts the idea thatnewtype
should have no concrete representation. You can get just the same efficiency and effect by usingdata
instead ofnewtype
. If there is no overloading involved, then there is more of a case for allowing an existentiallyquantifiednewtype
, because thedata
version does carry an implementation cost, but singlefield existentially quantified constructors aren’t much use. So the simple restriction (no existential stuff onnewtype
) stands, unless there are convincing reasons to change it. 
You can’t use
deriving
to define instances of a data type with existentially quantified data constructors. Reason: in most cases it would not make sense. For example:;data T = forall a. MkT [a] deriving( Eq )
To derive
Eq
in the standard way we would need to have equality between the single component of twoMkT
constructors:instance Eq T where (MkT a) == (MkT b) = ???
But
a
andb
have distinct types, and so can’t be compared. It’s just about possible to imagine examples in which the derived instance would make sense, but it seems altogether simpler simply to prohibit such declarations. Define your own instances!
10.4.7. Declaring data types with explicit constructor signatures

GADTSyntax

 Since

7.2.1
Allow the use of GADT syntax in data type definitions (but not GADTs themselves; for this see
GADTs
)
When the GADTSyntax
extension is enabled, GHC allows you to declare an algebraic data type by giving the type signatures of constructors explicitly. For example:
data Maybe a where Nothing :: Maybe a Just :: a > Maybe a
The form is called a “GADTstyle declaration” because Generalised Algebraic Data Types, described in Generalised Algebraic Data Types (GADTs), can only be declared using this form.
Notice that GADTstyle syntax generalises existential types (Existentially quantified data constructors). For example, these two declarations are equivalent:
data Foo = forall a. MkFoo a (a > Bool) data Foo' where { MKFoo :: a > (a>Bool) > Foo' }
Any data type that can be declared in standard Haskell 98 syntax can also be declared using GADTstyle syntax. The choice is largely stylistic, but GADTstyle declarations differ in one important respect: they treat class constraints on the data constructors differently. Specifically, if the constructor is given a typeclass context, that context is made available by pattern matching. For example:
data Set a where MkSet :: Eq a => [a] > Set a makeSet :: Eq a => [a] > Set a makeSet xs = MkSet (nub xs) insert :: a > Set a > Set a insert a (MkSet as)  a `elem` as = MkSet as  otherwise = MkSet (a:as)
A use of MkSet
as a constructor (e.g. in the definition of makeSet
) gives rise to a (Eq a)
constraint, as you would expect. The new feature is that patternmatching on MkSet
(as in the definition of insert
) makes available an (Eq a)
context. In implementation terms, the MkSet
constructor has a hidden field that stores the (Eq a)
dictionary that is passed to MkSet
; so when patternmatching that dictionary becomes available for the righthand side of the match. In the example, the equality dictionary is used to satisfy the equality constraint generated by the call to elem
, so that the type of insert
itself has no Eq
constraint.
For example, one possible application is to reify dictionaries:
data NumInst a where MkNumInst :: Num a => NumInst a intInst :: NumInst Int intInst = MkNumInst plus :: NumInst a > a > a > a plus MkNumInst p q = p + q
Here, a value of type NumInst a
is equivalent to an explicit (Num a)
dictionary.
All this applies to constructors declared using the syntax of Existentials and type classes. For example, the NumInst
data type above could equivalently be declared like this:
data NumInst a = Num a => MkNumInst (NumInst a)
Notice that, unlike the situation when declaring an existential, there is no forall
, because the Num
constrains the data type’s universally quantified type variable a
. A constructor may have both universal and existential type variables: for example, the following two declarations are equivalent:
data T1 a = forall b. (Num a, Eq b) => MkT1 a b data T2 a where MkT2 :: (Num a, Eq b) => a > b > T2 a
All this behaviour contrasts with Haskell 98’s peculiar treatment of contexts on a data type declaration (Section 4.2.1 of the Haskell 98 Report). In Haskell 98 the definition
data Eq a => Set' a = MkSet' [a]
gives MkSet'
the same type as MkSet
above. But instead of making available an (Eq a)
constraint, patternmatching on MkSet'
requires an (Eq a)
constraint! GHC faithfully implements this behaviour, odd though it is. But for GADTstyle declarations, GHC’s behaviour is much more useful, as well as much more intuitive.
The rest of this section gives further details about GADTstyle data type declarations.
 The result type of each data constructor must begin with the type constructor being defined. If the result type of all constructors has the form
T a1 ... an
, wherea1 ... an
are distinct type variables, then the data type is ordinary; otherwise is a generalised data type (Generalised Algebraic Data Types (GADTs)). 
As with other type signatures, you can give a single signature for several data constructors. In this example we give a single signature for
T1
andT2
:data T a where T1,T2 :: a > T a T3 :: T a

The type signature of each constructor is independent, and is implicitly universally quantified as usual. In particular, the type variable(s) in the “
data T a where
” header have no scope, and different constructors may have different universallyquantified type variables:data T a where  The 'a' has no scope T1,T2 :: b > T b  Means forall b. b > T b T3 :: T a  Means forall a. T a

A constructor signature may mention type class constraints, which can differ for different constructors. For example, this is fine:
data T a where T1 :: Eq b => b > b > T b T2 :: (Show c, Ix c) => c > [c] > T c
When pattern matching, these constraints are made available to discharge constraints in the body of the match. For example:
f :: T a > String f (T1 x y)  x==y = "yes"  otherwise = "no" f (T2 a b) = show a
Note that
f
is not overloaded; theEq
constraint arising from the use of==
is discharged by the pattern match onT1
and similarly theShow
constraint arising from the use ofshow
. 
Unlike a Haskell98style data type declaration, the type variable(s) in the “
data Set a where
” header have no scope. Indeed, one can write a kind signature instead:data Set :: Type > Type where ...
or even a mixture of the two:
data Bar a :: (Type > Type) > Type where ...
The type variables (if given) may be explicitly kinded, so we could also write the header for
Foo
like this:data Bar a (b :: Type > Type) where ...

You can use strictness annotations, in the obvious places in the constructor type:
data Term a where Lit :: !Int > Term Int If :: Term Bool > !(Term a) > !(Term a) > Term a Pair :: Term a > Term b > Term (a,b)

You can use a
deriving
clause on a GADTstyle data type declaration. For example, these two declarations are equivalentdata Maybe1 a where { Nothing1 :: Maybe1 a ; Just1 :: a > Maybe1 a } deriving( Eq, Ord ) data Maybe2 a = Nothing2  Just2 a deriving( Eq, Ord )

The type signature may have quantified type variables that do not appear in the result type:
data Foo where MkFoo :: a > (a>Bool) > Foo Nil :: Foo
Here the type variable
a
does not appear in the result type of either constructor. Although it is universally quantified in the type of the constructor, such a type variable is often called “existential”. Indeed, the above declaration declares precisely the same type as thedata Foo
in Existentially quantified data constructors.The type may contain a class context too, of course:
data Showable where MkShowable :: Show a => a > Showable

You can use record syntax on a GADTstyle data type declaration:
data Person where Adult :: { name :: String, children :: [Person] } > Person Child :: Show a => { name :: !String, funny :: a } > Person
As usual, for every constructor that has a field
f
, the type of fieldf
must be the same (modulo alpha conversion). TheChild
constructor above shows that the signature may have a context, existentiallyquantified variables, and strictness annotations, just as in the nonrecord case. (NB: the “type” that follows the doublecolon is not really a type, because of the record syntax and strictness annotations. A “type” of this form can appear only in a constructor signature.)  Record updates are allowed with GADTstyle declarations, only fields that have the following property: the type of the field mentions no existential type variables.

As in the case of existentials declared using the Haskell98like record syntax (Record Constructors), recordselector functions are generated only for those fields that have welltyped selectors. Here is the example of that section, in GADTstyle syntax:
data Counter a where NewCounter :: { _this :: self , _inc :: self > self , _display :: self > IO () , tag :: a } > Counter a
As before, only one selector function is generated here, that for
tag
. Nevertheless, you can still use all the field names in pattern matching and record construction. 
In a GADTstyle data type declaration there is no obvious way to specify that a data constructor should be infix, which makes a difference if you derive
Show
for the type. (Data constructors declared infix are displayed infix by the derivedshow
.) So GHC implements the following design: a data constructor declared in a GADTstyle data type declaration is displayed infix byShow
iff (a) it is an operator symbol, (b) it has two arguments, (c) it has a programmersupplied fixity declaration. For exampleinfix 6 (::) data T a where (::) :: Int > Bool > T Int
10.4.8. Generalised Algebraic Data Types (GADTs)

GADTs

 Implies
 Since

6.8.1
Allow use of Generalised Algebraic Data Types (GADTs).
Generalised Algebraic Data Types generalise ordinary algebraic data types by allowing constructors to have richer return types. Here is an example:
data Term a where Lit :: Int > Term Int Succ :: Term Int > Term Int IsZero :: Term Int > Term Bool If :: Term Bool > Term a > Term a > Term a Pair :: Term a > Term b > Term (a,b)
Notice that the return type of the constructors is not always Term a
, as is the case with ordinary data types. This generality allows us to write a welltyped eval
function for these Terms
:
eval :: Term a > a eval (Lit i) = i eval (Succ t) = 1 + eval t eval (IsZero t) = eval t == 0 eval (If b e1 e2) = if eval b then eval e1 else eval e2 eval (Pair e1 e2) = (eval e1, eval e2)
The key point about GADTs is that pattern matching causes type refinement. For example, in the right hand side of the equation
eval :: Term a > a eval (Lit i) = ...
the type a
is refined to Int
. That’s the whole point! A precise specification of the type rules is beyond what this user manual aspires to, but the design closely follows that described in the paper Simple unificationbased type inference for GADTs, (ICFP 2006). The general principle is this: type refinement is only carried out based on usersupplied type annotations. So if no type signature is supplied for eval
, no type refinement happens, and lots of obscure error messages will occur. However, the refinement is quite general. For example, if we had:
eval :: Term a > a > a eval (Lit i) j = i+j
the pattern match causes the type a
to be refined to Int
(because of the type of the constructor Lit
), and that refinement also applies to the type of j
, and the result type of the case
expression. Hence the addition i+j
is legal.
These and many other examples are given in papers by Hongwei Xi, and Tim Sheard. There is a longer introduction on the wiki, and Ralf Hinze’s Fun with phantom types also has a number of examples. Note that papers may use different notation to that implemented in GHC.
The rest of this section outlines the extensions to GHC that support GADTs. The extension is enabled with GADTs
. The GADTs
extension also sets GADTSyntax
and MonoLocalBinds
.
 A GADT can only be declared using GADTstyle syntax (Declaring data types with explicit constructor signatures); the old Haskell 98 syntax for data declarations always declares an ordinary data type. The result type of each constructor must begin with the type constructor being defined, but for a GADT the arguments to the type constructor can be arbitrary monotypes. For example, in the
Term
data type above, the type of each constructor must end withTerm ty
, but thety
need not be a type variable (e.g. theLit
constructor).  It is permitted to declare an ordinary algebraic data type using GADTstyle syntax. What makes a GADT into a GADT is not the syntax, but rather the presence of data constructors whose result type is not just
T a b
.  You cannot use a
deriving
clause for a GADT; only for an ordinary data type. 
As mentioned in Declaring data types with explicit constructor signatures, record syntax is supported. For example:
data Term a where Lit :: { val :: Int } > Term Int Succ :: { num :: Term Int } > Term Int Pred :: { num :: Term Int } > Term Int IsZero :: { arg :: Term Int } > Term Bool Pair :: { arg1 :: Term a , arg2 :: Term b } > Term (a,b) If :: { cnd :: Term Bool , tru :: Term a , fls :: Term a } > Term a
However, for GADTs there is the following additional constraint: every constructor that has a field
f
must have the same result type (modulo alpha conversion) Hence, in the above example, we cannot merge thenum
andarg
fields above into a single name. Although their field types are bothTerm Int
, their selector functions actually have different types:num :: Term Int > Term Int arg :: Term Bool > Term Int

When patternmatching against data constructors drawn from a GADT, for example in a
case
expression, the following rules apply: The type of the scrutinee must be rigid.
 The type of the entire
case
expression must be rigid.  The type of any free variable mentioned in any of the
case
alternatives must be rigid.
A type is “rigid” if it is completely known to the compiler at its binding site. The easiest way to ensure that a variable a rigid type is to give it a type signature. For more precise details see Simple unificationbased type inference for GADTs. The criteria implemented by GHC are given in the Appendix.
10.5. Extensions to the record system
10.5.1. Traditional record syntax

NoTraditionalRecordSyntax

 Since

7.4.1
Disallow use of record syntax.
Traditional record syntax, such as C {f = x}
, is enabled by default. To disable it, you can use the NoTraditionalRecordSyntax
extension.
10.5.2. Record field disambiguation

DisambiguateRecordFields

 Since

6.8.1
Allow the compiler to automatically choose between identicallynamed record selectors based on type (if the choice is unambiguous).
In record construction and record pattern matching it is entirely unambiguous which field is referred to, even if there are two different data types in scope with a common field name. For example:
module M where data S = MkS { x :: Int, y :: Bool } module Foo where import M data T = MkT { x :: Int } ok1 (MkS { x = n }) = n+1  Unambiguous ok2 n = MkT { x = n+1 }  Unambiguous bad1 k = k { x = 3 }  Ambiguous bad2 k = x k  Ambiguous
Even though there are two x
’s in scope, it is clear that the x
in the pattern in the definition of ok1
can only mean the field x
from type S
. Similarly for the function ok2
. However, in the record update in bad1
and the record selection in bad2
it is not clear which of the two types is intended.
Haskell 98 regards all four as ambiguous, but with the DisambiguateRecordFields
extension, GHC will accept the former two. The rules are precisely the same as those for instance declarations in Haskell 98, where the method names on the lefthand side of the method bindings in an instance declaration refer unambiguously to the method of that class (provided they are in scope at all), even if there are other variables in scope with the same name. This reduces the clutter of qualified names when you import two records from different modules that use the same field name.
Some details:

Field disambiguation can be combined with punning (see Record puns). For example:
module Foo where import M x=True ok3 (MkS { x }) = x+1  Uses both disambiguation and punning

With
DisambiguateRecordFields
you can use unqualified field names even if the corresponding selector is only in scope qualified For example, assuming the same moduleM
as in our earlier example, this is legal:module Foo where import qualified M  Note qualified ok4 (M.MkS { x = n }) = n+1  Unambiguous
Since the constructor
MkS
is only in scope qualified, you must name itM.MkS
, but the fieldx
does not need to be qualified even thoughM.x
is in scope butx
is not (In effect, it is qualified by the constructor).
10.5.3. Duplicate record fields

DuplicateRecordFields

 Implies
 Since

8.0.1
Allow definition of record types with identicallynamed fields.
Going beyond DisambiguateRecordFields
(see Record field disambiguation), the DuplicateRecordFields
extension allows multiple datatypes to be declared using the same field names in a single module. For example, it allows this:
module M where data S = MkS { x :: Int } data T = MkT { x :: Bool }
Uses of fields that are always unambiguous because they mention the constructor, including construction and patternmatching, may freely use duplicated field names. For example, the following are permitted (just as with DisambiguateRecordFields
):
s = MkS { x = 3 } f (MkT { x = b }) = b
Field names used as selector functions or in record updates must be unambiguous, either because there is only one such field in scope, or because a type signature is supplied, as described in the following sections.
10.5.3.1. Selector functions
Fields may be used as selector functions only if they are unambiguous, so this is still not allowed if both S(x)
and T(x)
are in scope:
bad r = x r
An ambiguous selector may be disambiguated by the type being “pushed down” to the occurrence of the selector (see Type inference for more details on what “pushed down” means). For example, the following are permitted:
ok1 = x :: S > Int ok2 :: S > Int ok2 = x ok3 = k x  assuming we already have k :: (S > Int) > _
In addition, the datatype that is meant may be given as a type signature on the argument to the selector:
ok4 s = x (s :: S)
However, we do not infer the type of the argument to determine the datatype, or have any way of deferring the choice to the constraint solver. Thus the following is ambiguous:
bad :: S > Int bad s = x s
Even though a field label is duplicated in its defining module, it may be possible to use the selector unambiguously elsewhere. For example, another module could import S(x)
but not T(x)
, and then use x
unambiguously.
10.5.3.2. Record updates
In a record update such as e { x = 1 }
, if there are multiple x
fields in scope, then the type of the context must fix which record datatype is intended, or a type annotation must be supplied. Consider the following definitions:
data S = MkS { foo :: Int } data T = MkT { foo :: Int, bar :: Int } data U = MkU { bar :: Int, baz :: Int }
Without DuplicateRecordFields
, an update mentioning foo
will always be ambiguous if all these definitions were in scope. When the extension is enabled, there are several options for disambiguating updates:

Check for types that have all the fields being updated. For example:
f x = x { foo = 3, bar = 2 }
Here
f
must be updatingT
because neitherS
norU
have both fields. 
Use the type being pushed in to the record update, as in the following:
g1 :: T > T g1 x = x { foo = 3 } g2 x = x { foo = 3 } :: T g3 = k (x { foo = 3 })  assuming we already have k :: T > _

Use an explicit type signature on the record expression, as in:
h x = (x :: T) { foo = 3 }
The type of the expression being updated will not be inferred, and no constraintsolving will be performed, so the following will be rejected as ambiguous:
let x :: T x = blah in x { foo = 3 } \x > [x { foo = 3 }, blah :: T ] \ (x :: T) > x { foo = 3 }
10.5.3.3. Import and export of record fields
When DuplicateRecordFields
is enabled, an ambiguous field must be exported as part of its datatype, rather than at the top level. For example, the following is legal:
module M (S(x), T(..)) where data S = MkS { x :: Int } data T = MkT { x :: Bool }
However, this would not be permitted, because x
is ambiguous:
module M (x) where ...
Similar restrictions apply on import.
10.5.4. Record puns

NamedFieldPuns

 Since

6.10.1
Allow use of record puns.
Record puns are enabled by the language extension NamedFieldPuns
.
When using records, it is common to write a pattern that binds a variable with the same name as a record field, such as:
data C = C {a :: Int} f (C {a = a}) = a
Record punning permits the variable name to be elided, so one can simply write
f (C {a}) = a
to mean the same pattern as above. That is, in a record pattern, the pattern a
expands into the pattern a = a
for the same name a
.
Note that:

Record punning can also be used in an expression, writing, for example,
let a = 1 in C {a}
instead of
let a = 1 in C {a = a}
The expansion is purely syntactic, so the expanded righthand side expression refers to the nearest enclosing variable that is spelled the same as the field name.

Puns and other patterns can be mixed in the same record:
data C = C {a :: Int, b :: Int} f (C {a, b = 4}) = a
 Puns can be used wherever record patterns occur (e.g. in
let
bindings or at the toplevel). 
A pun on a qualified field name is expanded by stripping off the module qualifier. For example:
f (C {M.a}) = a
means
f (M.C {M.a = a}) = a
(This is useful if the field selector
a
for constructorM.C
is only in scope in qualified form.)
10.5.5. Record wildcards

RecordWildCards

 Implies
 Since

6.8.1
Allow the use of wildcards in record construction and pattern matching.
Record wildcards are enabled by the language extension RecordWildCards
. This exension implies DisambiguateRecordFields
.
For records with many fields, it can be tiresome to write out each field individually in a record pattern, as in
data C = C {a :: Int, b :: Int, c :: Int, d :: Int} f (C {a = 1, b = b, c = c, d = d}) = b + c + d
Record wildcard syntax permits a “..
” in a record pattern, where each elided field f
is replaced by the pattern f = f
. For example, the above pattern can be written as
f (C {a = 1, ..}) = b + c + d
More details:

Record wildcards in patterns can be mixed with other patterns, including puns (Record puns); for example, in a pattern
(C {a = 1, b, ..})
. Additionally, record wildcards can be used wherever record patterns occur, including inlet
bindings and at the toplevel. For example, the toplevel bindingC {a = 1, ..} = e
defines
b
,c
, andd
. 
Record wildcards can also be used in an expression, when constructing a record. For example,
let {a = 1; b = 2; c = 3; d = 4} in C {..}
in place of
let {a = 1; b = 2; c = 3; d = 4} in C {a=a, b=b, c=c, d=d}
The expansion is purely syntactic, so the record wildcard expression refers to the nearest enclosing variables that are spelled the same as the omitted field names.

For both pattern and expression wildcards, the “
..
” expands to the missing inscope record fields. Specifically the expansion of “C {..}
” includesf
if and only if:
f
is a record field of constructorC
.  The record field
f
is in scope somehow (either qualified or unqualified).
These rules restrict record wildcards to the situations in which the user could have written the expanded version. For example
module M where data R = R { a,b,c :: Int } module X where import M( R(R,a,c) ) f a b = R { .. }
The
R{..}
expands toR{a=a}
, omittingb
since the record field is not in scope, and omittingc
since the variablec
is not in scope (apart from the binding of the record selectorc
, of course). 

When record wildcards are use in record construction, a field
f
is initialised only iff
is in scope, and is not imported or bound at top level. For example,f
can be bound by an enclosing pattern match or let/wherebinding. For examplemodule M where import A( a ) data R = R { a,b,c,d :: Int } c = 3 :: Int f b = R { .. }  Expands to R { b = b, d = d } where d = b+1
Here,
a
is imported, andc
is bound at top level, so neither contribute to the expansion of the “..
”. The motivation here is that it should be easy for the reader to figure out what the “..
” expands to. 
Record wildcards cannot be used (a) in a record update construct, and (b) for data constructors that are not declared with record fields. For example:
f x = x { v=True, .. }  Illegal (a) data T = MkT Int Bool g = MkT { .. }  Illegal (b) h (MkT { .. }) = True  Illegal (b)
10.5.6. Record field selector polymorphism
The module GHC.Records defines the following:
class HasField (x :: k) r a  x r > a where getField :: r > a
A HasField x r a
constraint represents the fact that x
is a field of type a
belonging to a record type r
. The getField
method gives the record selector function.
This allows definitions that are polymorphic over record types with a specified field. For example, the following works with any record type that has a field name :: String
:
foo :: HasField "name" r String => r > String foo r = reverse (getField @"name" r)
HasField
is a magic builtin typeclass (similar to Coercible
, for example). It is given special treatment by the constraint solver (see Solving HasField constraints). Users may define their own instances of HasField
also (see Virtual record fields).
10.5.6.1. Solving HasField constraints
If the constraint solver encounters a constraint HasField x r a
where r
is a concrete datatype with a field x
in scope, it will automatically solve the constraint using the field selector as the dictionary, unifying a
with the type of the field if necessary. This happens irrespective of which extensions are enabled.
For example, if the following datatype is in scope
data Person = Person { name :: String }
the end result is rather like having an instance
instance HasField "name" Person String where getField = name
except that this instance is not actually generated anywhere, rather the constraint is solved directly by the constraint solver.
A field must be in scope for the corresponding HasField
constraint to be solved. This retains the existing representation hiding mechanism, whereby a module may choose not to export a field, preventing client modules from accessing or updating it directly.
Solving HasField
constraints depends on the field selector functions that are generated for each datatype definition:

If a record field does not have a selector function because its type would allow an existential variable to escape, the corresponding
HasField
constraint will not be solved. For example,{# LANGUAGE ExistentialQuantification #} data Exists t = forall x . MkExists { unExists :: t x }
does not give rise to a selector
unExists :: Exists t > t x
and we will not solveHasField "unExists" (Exists t) a
automatically. 
If a record field has a polymorphic type (and hence the selector function is higherrank), the corresponding
HasField
constraint will not be solved, because doing so would violate the functional dependency onHasField
and/or require impredicativity. For example,{# LANGUAGE RankNTypes #} data Higher = MkHigher { unHigher :: forall t . t > t }
gives rise to a selector
unHigher :: Higher > (forall t . t > t)
but does not lead to solution of the constraintHasField "unHigher" Higher a
. 
A record GADT may have a restricted type for a selector function, which may lead to additional unification when solving
HasField
constraints. For example,{# LANGUAGE GADTs #} data Gadt t where MkGadt :: { unGadt :: Maybe v } > Gadt [v]
gives rise to a selector
unGadt :: Gadt [v] > Maybe v
, so the solver will reduce the constraintHasField "unGadt" (Gadt t) b
by unifyingt ~ [v]
andb ~ Maybe v
for some fresh metavariablev
, rather as if we had an instanceinstance (t ~ [v], b ~ Maybe v) => HasField "unGadt" (Gadt t) b

If a record type has an oldfashioned datatype context, the
HasField
constraint will be reduced to solving the constraints from the context. For example,{# LANGUAGE DatatypeContexts #} data Eq a => Silly a = MkSilly { unSilly :: a }
gives rise to a selector
unSilly :: Eq a => Silly a > a
, so the solver will reduce the constraintHasField "unSilly" (Silly a) b
toEq a
(and unifya
withb
), rather as if we had an instanceinstance (Eq a, a ~ b) => HasField "unSilly" (Silly a) b
10.5.6.2. Virtual record fields
Users may define their own instances of HasField
, provided they do not conflict with the builtin constraint solving behaviour. This allows “virtual” record fields to be defined for datatypes that do not otherwise have them.
For example, this instance would make the name
field of Person
accessible using #fullname
as well:
instance HasField "fullname" Person String where getField = name
More substantially, an anonymous records library could provide HasField
instances for its anonymous records, and thus be compatible with the polymorphic record selectors introduced by this proposal. For example, something like this makes it possible to use getField
to access Record
values with the appropriate string in the typelevel list of fields:
data Record (xs :: [(k, Type)]) where Nil :: Record '[] Cons :: Proxy x > a > Record xs > Record ('(x, a) ': xs) instance HasField x (Record ('(x, a) ': xs)) a where getField (Cons _ v _) = v instance HasField x (Record xs) a => HasField x (Record ('(y, b) ': xs)) a where getField (Cons _ _ r) = getField @x r r :: Record '[ '("name", String) ] r = Cons Proxy "R" Nil) x = getField @"name" r
Since representations such as this can support field labels with kinds other than Symbol
, the HasField
class is polykinded (even though the builtin constraint solving works only at kind Symbol
). In particular, this allows users to declare scoped field labels such as in the following example:
data PersonFields = Name s :: Record '[ '(Name, String) ] s = Cons Proxy "S" Nil y = getField @Name s
In order to avoid conflicting with the builtin constraint solving, the following userdefined HasField
instances are prohibited (in addition to the usual rules, such as the prohibition on type families appearing in instance heads):

HasField _ r _
wherer
is a variable; 
HasField _ (T ...) _
ifT
is a data family (because it might have fields introduced later, using data instance declarations); 
HasField x (T ...) _
ifx
is a variable andT
has any fields at all (but this instance is permitted ifT
has no fields); 
HasField "foo" (T ...) _
ifT
has a fieldfoo
(but this instance is permitted if it does not).
If a field has a higherrank or existential type, the corresponding HasField
constraint will not be solved automatically (as described above), but in the interests of simplicity we do not permit users to define their own instances either. If a field is not in scope, the corresponding instance is still prohibited, to avoid conflicts in downstream modules.
10.6. Extensions to the “deriving” mechanism
Haskell 98 allows the programmer to add a deriving clause to a data type declaration, to generate a standard instance declaration for specified class. GHC extends this mechanism along several axes:
 The derivation mechanism can be used separately from the data type declaration, using the standalone deriving mechanism.
 In Haskell 98, the only derivable classes are
Eq
,Ord
,Enum
,Ix
,Bounded
,Read
, andShow
. Various language extensions extend this list. 
Besides the stock approach to deriving instances by generating all method definitions, GHC supports two additional deriving strategies, which can derive arbitrary classes:
 Generalised newtype deriving for newtypes and
 deriving any class using an empty instance declaration.
The user can optionally declare the desired deriving strategy, especially if the compiler chooses the wrong one by default.
10.6.1. Deriving instances for empty data types

EmptyDataDeriving

 Since

8.4.1
Allow deriving instances of standard type classes for empty data types.
One can write data types with no constructors using the EmptyDataDecls
flag (see Data types with no constructors), which is on by default in Haskell 2010. What is not on by default is the ability to derive type class instances for these types. This ability is enabled through use of the EmptyDataDeriving
flag. For instance, this lets one write:
data Empty deriving (Eq, Ord, Read, Show)
This would generate the following instances:
instance Eq Empty where _ == _ = True instance Ord Empty where compare _ _ = EQ instance Read Empty where readPrec = pfail instance Show Empty where showsPrec _ x = case x of {}
The EmptyDataDeriving
flag is only required to enable deriving of these four “standard” type classes (which are mentioned in the Haskell Report). Other extensions to the deriving
mechanism, which are explained below in greater detail, do not require EmptyDataDeriving
to be used in conjunction with empty data types. These include:

StandaloneDeriving
(see Standalone deriving declarations)  Type classes which require their own extensions to be enabled to be derived, such as
DeriveFunctor
(see Deriving instances of extra classes (Data, etc.)) 
DeriveAnyClass
(see Deriving any other class)
10.6.2. Inferred context for deriving clauses
The Haskell Report is vague about exactly when a deriving
clause is legal. For example:
data T0 f a = MkT0 a deriving( Eq ) data T1 f a = MkT1 (f a) deriving( Eq ) data T2 f a = MkT2 (f (f a)) deriving( Eq )
The natural generated Eq
code would result in these instance declarations:
instance Eq a => Eq (T0 f a) where ... instance Eq (f a) => Eq (T1 f a) where ... instance Eq (f (f a)) => Eq (T2 f a) where ...
The first of these is obviously fine. The second is still fine, although less obviously. The third is not Haskell 98, and risks losing termination of instances.
GHC takes a conservative position: it accepts the first two, but not the third. The rule is this: each constraint in the inferred instance context must consist only of type variables, with no repetitions.
This rule is applied regardless of flags. If you want a more exotic context, you can write it yourself, using the standalone deriving mechanism.
10.6.3. Standalone deriving declarations

StandaloneDeriving

 Since

6.8.1
Allow the use of standalone
deriving
declarations.
GHC allows standalone deriving
declarations, enabled by StandaloneDeriving
:
data Foo a = Bar a  Baz String deriving instance Eq a => Eq (Foo a)
The syntax is identical to that of an ordinary instance declaration apart from (a) the keyword deriving
, and (b) the absence of the where
part.
However, standalone deriving differs from a deriving
clause in a number of important ways:
 The standalone deriving declaration does not need to be in the same module as the data type declaration. (But be aware of the dangers of orphan instances (Orphan modules and instance declarations).

In most cases, you must supply an explicit context (in the example the context is
(Eq a)
), exactly as you would in an ordinary instance declaration. (In contrast, in aderiving
clause attached to a data type declaration, the context is inferred.)The exception to this rule is that the context of a standalone deriving declaration can infer its context when a single, extrawildcards constraint is used as the context, such as in:
deriving instance _ => Eq (Foo a)
This is essentially the same as if you had written
deriving Eq
after the declaration fordata Foo a
. Using this feature requires the use ofPartialTypeSignatures
(Partial Type Signatures). 
Unlike a
deriving
declaration attached to adata
declaration, the instance can be more specific than the data type (assuming you also useFlexibleInstances
, Relaxed rules for instance contexts). Consider for exampledata Foo a = Bar a  Baz String deriving instance Eq a => Eq (Foo [a]) deriving instance Eq a => Eq (Foo (Maybe a))
This will generate a derived instance for
(Foo [a])
and(Foo (Maybe a))
, but other types such as(Foo (Int,Bool))
will not be an instance ofEq
. 
Unlike a
deriving
declaration attached to adata
declaration, GHC does not restrict the form of the data type. Instead, GHC simply generates the appropriate boilerplate code for the specified class, and typechecks it. If there is a type error, it is your problem. (GHC will show you the offending code if it has a type error.)The merit of this is that you can derive instances for GADTs and other exotic data types, providing only that the boilerplate code does indeed typecheck. For example:
data T a where T1 :: T Int T2 :: T Bool deriving instance Show (T a)
In this example, you cannot say
... deriving( Show )
on the data type declaration forT
, becauseT
is a GADT, but you can generate the instance declaration using standalone deriving.The downside is that, if the boilerplate code fails to typecheck, you will get an error message about that code, which you did not write. Whereas, with a
deriving
clause the sideconditions are necessarily more conservative, but any error message may be more comprehensible. 
Under most circumstances, you cannot use standalone deriving to create an instance for a data type whose constructors are not all in scope. This is because the derived instance would generate code that uses the constructors behind the scenes, which would break abstraction.
The one exception to this rule is
DeriveAnyClass
, since deriving an instance viaDeriveAnyClass
simply generates an empty instance declaration, which does not require the use of any constructors. See the deriving any class section for more details.
In other ways, however, a standalone deriving obeys the same rules as ordinary deriving:
 A
deriving instance
declaration must obey the same rules concerning form and termination as ordinary instance declarations, controlled by the same flags; see Instance declarations. 
The standalone syntax is generalised for newtypes in exactly the same way that ordinary
deriving
clauses are generalised (Generalised derived instances for newtypes). For example:newtype Foo a = MkFoo (State Int a) deriving instance MonadState Int Foo
GHC always treats the last parameter of the instance (
Foo
in this example) as the type whose instance is being derived.
10.6.4. Deriving instances of extra classes (Data
, etc.)
Haskell 98 allows the programmer to add “deriving( Eq, Ord )
” to a data type declaration, to generate a standard instance declaration for classes specified in the deriving
clause. In Haskell 98, the only classes that may appear in the deriving
clause are the standard classes Eq
, Ord
, Enum
, Ix
, Bounded
, Read
, and Show
.
GHC extends this list with several more classes that may be automatically derived:
 With
DeriveGeneric
, you can derive instances of the classesGeneric
andGeneric1
, defined inGHC.Generics
. You can use these to define generic functions, as described in Generic programming.  With
DeriveFunctor
, you can derive instances of the classFunctor
, defined inGHC.Base
.  With
DeriveDataTypeable
, you can derive instances of the classData
, defined inData.Data
.  With
DeriveFoldable
, you can derive instances of the classFoldable
, defined inData.Foldable
.  With
DeriveTraversable
, you can derive instances of the classTraversable
, defined inData.Traversable
. Since theTraversable
instance dictates the instances ofFunctor
andFoldable
, you’ll probably want to derive them too, soDeriveTraversable
impliesDeriveFunctor
andDeriveFoldable
.  With
DeriveLift
, you can derive instances of the classLift
, defined in theLanguage.Haskell.TH.Syntax
module of thetemplatehaskell
package.
You can also use a standalone deriving declaration instead (see Standalone deriving declarations).
In each case the appropriate class must be in scope before it can be mentioned in the deriving
clause.
10.6.4.1. Deriving Functor
instances

DeriveFunctor

 Since

7.10.1
Allow automatic deriving of instances for the
Functor
typeclass.
With DeriveFunctor
, one can derive Functor
instances for data types of kind Type > Type
. For example, this declaration:
data Example a = Ex a Char (Example a) (Example Char) deriving Functor
would generate the following instance:
instance Functor Example where fmap f (Ex a1 a2 a3 a4) = Ex (f a1) a2 (fmap f a3) a4
The basic algorithm for DeriveFunctor
walks the arguments of each constructor of a data type, applying a mapping function depending on the type of each argument. If a plain type variable is found that is syntactically equivalent to the last type parameter of the data type (a
in the above example), then we apply the function f
directly to it. If a type is encountered that is not syntactically equivalent to the last type parameter but does mention the last type parameter somewhere in it, then a recursive call to fmap
is made. If a type is found which doesn’t mention the last type parameter at all, then it is left alone.
The second of those cases, in which a type is unequal to the type parameter but does contain the type parameter, can be surprisingly tricky. For example, the following example compiles:
newtype Right a = Right (Either Int a) deriving Functor
Modifying the code slightly, however, produces code which will not compile:
newtype Wrong a = Wrong (Either a Int) deriving Functor
The difference involves the placement of the last type parameter, a
. In the Right
case, a
occurs within the type Either Int a
, and moreover, it appears as the last type argument of Either
. In the Wrong
case, however, a
is not the last type argument to Either
; rather, Int
is.
This distinction is important because of the way DeriveFunctor
works. The derived Functor Right
instance would be:
instance Functor Right where fmap f (Right a) = Right (fmap f a)
Given a value of type Right a
, GHC must produce a value of type Right b
. Since the argument to the Right
constructor has type Either Int a
, the code recursively calls fmap
on it to produce a value of type Either Int b
, which is used in turn to construct a final value of type Right b
.
The generated code for the Functor Wrong
instance would look exactly the same, except with Wrong
replacing every occurrence of Right
. The problem is now that fmap
is being applied recursively to a value of type Either a Int
. This cannot possibly produce a value of type Either b Int
, as fmap
can only change the last type parameter! This causes the generated code to be illtyped.
As a general rule, if a data type has a derived Functor
instance and its last type parameter occurs on the righthand side of the data declaration, then either it must (1) occur bare (e.g., newtype Id a = Id a
), or (2) occur as the last argument of a type constructor (as in Right
above).
There are two exceptions to this rule:

Tuple types. When a nonunit tuple is used on the righthand side of a data declaration,
DeriveFunctor
treats it as a product of distinct types. In other words, the following code:newtype Triple a = Triple (a, Int, [a]) deriving Functor
Would result in a generated
Functor
instance like so:instance Functor Triple where fmap f (Triple a) = Triple (case a of (a1, a2, a3) > (f a1, a2, fmap f a3))
That is,
DeriveFunctor
patternmatches its way into tuples and maps over each type that constitutes the tuple. The generated code is reminiscent of what would be generated fromdata Triple a = Triple a Int [a]
, except with extra machinery to handle the tuple. 
Function types. The last type parameter can appear anywhere in a function type as long as it occurs in a covariant position. To illustrate what this means, consider the following three examples:
newtype CovFun1 a = CovFun1 (Int > a) deriving Functor newtype CovFun2 a = CovFun2 ((a > Int) > a) deriving Functor newtype CovFun3 a = CovFun3 (((Int > a) > Int) > a) deriving Functor
All three of these examples would compile without issue. On the other hand:
newtype ContraFun1 a = ContraFun1 (a > Int) deriving Functor newtype ContraFun2 a = ContraFun2 ((Int > a) > Int) deriving Functor newtype ContraFun3 a = ContraFun3 (((a > Int) > a) > Int) deriving Functor
While these examples look similar, none of them would successfully compile. This is because all occurrences of the last type parameter
a
occur in contravariant positions, not covariant ones.Intuitively, a covariant type is produced, and a contravariant type is consumed. Most types in Haskell are covariant, but the function type is special in that the lefthand side of a function arrow reverses variance. If a function type
a > b
appears in a covariant position (e.g.,CovFun1
above), thena
is in a contravariant position andb
is in a covariant position. Similarly, ifa > b
appears in a contravariant position (e.g.,CovFun2
above), thena
is ina
covariant position andb
is in a contravariant position.To see why a data type with a contravariant occurrence of its last type parameter cannot have a derived
Functor
instance, let’s suppose that aFunctor ContraFun1
instance exists. The implementation would look something like this:instance Functor ContraFun1 where fmap f (ContraFun g) = ContraFun (\x > _)
We have
f :: a > b
,g :: a > Int
, andx :: b
. Using these, we must somehow fill in the hole (denoted with an underscore) with a value of typeInt
. What are our options?We could try applying
g
tox
. This won’t work though, asg
expects an argument of typea
, andx :: b
. Even worse, we can’t turnx
into something of typea
, sincef
also needs an argument of typea
! In short, there’s no good way to make this work.On the other hand, a derived
Functor
instances for theCovFun
s are within the realm of possibility:instance Functor CovFun1 where fmap f (CovFun1 g) = CovFun1 (\x > f (g x)) instance Functor CovFun2 where fmap f (CovFun2 g) = CovFun2 (\h > f (g (\x > h (f x)))) instance Functor CovFun3 where fmap f (CovFun3 g) = CovFun3 (\h > f (g (\k > h (\x > f (k x)))))
There are some other scenarios in which a derived Functor
instance will fail to compile:
 A data type has no type parameters (e.g.,
data Nothing = Nothing
).  A data type’s last type variable is used in a
DatatypeContexts
constraint (e.g.,data Ord a => O a = O a
). 
A data type’s last type variable is used in an
ExistentialQuantification
constraint, or is refined in a GADT. For example,data T a b where T4 :: Ord b => b > T a b T5 :: b > T b b T6 :: T a (b,b) deriving instance Functor (T a)
would not compile successfully due to the way in which
b
is constrained.
When the last type parameter has a phantom role (see Roles), the derived Functor
instance will not be produced using the usual algorithm. Instead, the entire value will be coerced.
data Phantom a = Z  S (Phantom a) deriving Functor
will produce the following instance:
instance Functor Phantom where fmap _ = coerce
When a type has no constructors, the derived Functor
instance will simply force the (bottom) value of the argument using EmptyCase
.
data V a deriving Functor type role V nominal
will produce
 instance Functor V where

fmap _ z = case z of
10.6.4.2. Deriving Foldable
instances

DeriveFoldable

 Since

7.10.1
Allow automatic deriving of instances for the
Foldable
typeclass.
With DeriveFoldable
, one can derive Foldable
instances for data types of kind Type > Type
. For example, this declaration:
data Example a = Ex a Char (Example a) (Example Char) deriving Foldable
would generate the following instance:
instance Foldable Example where foldr f z (Ex a1 a2 a3 a4) = f a1 (foldr f z a3) foldMap f (Ex a1 a2 a3 a4) = mappend (f a1) (foldMap f a3)
The algorithm for DeriveFoldable
is adapted from the DeriveFunctor
algorithm, but it generates definitions for foldMap
, foldr
, and null
instead of fmap
. In addition, DeriveFoldable
filters out all constructor arguments on the RHS expression whose types do not mention the last type parameter, since those arguments do not need to be folded over.
When the type parameter has a phantom role (see Roles), DeriveFoldable
derives a trivial instance. For example, this declaration:
data Phantom a = Z  S (Phantom a)
will generate the following instance.
instance Foldable Phantom where foldMap _ _ = mempty
Similarly, when the type has no constructors, DeriveFoldable
will derive a trivial instance:
data V a deriving Foldable type role V nominal
will generate the following.
instance Foldable V where foldMap _ _ = mempty
Here are the differences between the generated code for Functor
and Foldable
:
#. When a bare type variable a
is encountered, DeriveFunctor
would generate f a
for an fmap
definition. DeriveFoldable
would generate f a z
for foldr
, f a
for foldMap
, and False
for null
.

When a type that is not syntactically equivalent to
a
, but which does containa
, is encountered,DeriveFunctor
recursively callsfmap
on it. Similarly,DeriveFoldable
would recursively callfoldr
andfoldMap
. Depending on the context,null
may recursively callnull
orall null
. For example, givendata F a = F (P a) data G a = G (P (a, Int)) data H a = H (P (Q a))
Foldable
deriving will producenull (F x) = null x null (G x) = null x null (H x) = all null x

DeriveFunctor
puts everything back together again at the end by invoking the constructor.DeriveFoldable
, however, builds up a value of some type. Forfoldr
, this is accomplished by chaining applications off
and recursivefoldr
calls on the state valuez
. ForfoldMap
, this happens by combining all values withmappend
. Fornull
, the values are usually combined with&&
. However, if any of the values is known to beFalse
, all the rest will be dropped. For example,data SnocList a = Nil  Snoc (SnocList a) a
will not produce
null (Snoc xs _) = null xs && False
(which would walk the whole list), but rather
null (Snoc _ _) = False
There are some other differences regarding what data types can have derived Foldable
instances:
 Data types containing function types on the righthand side cannot have derived
Foldable
instances. 
Foldable
instances can be derived for data types in which the last type parameter is existentially constrained or refined in a GADT. For example, this data type:data E a where E1 :: (a ~ Int) => a > E a E2 :: Int > E Int E3 :: (a ~ Int) => a > E Int E4 :: (a ~ Int) => Int > E a deriving instance Foldable E
would have the following generated
Foldable
instance:instance Foldable E where foldr f z (E1 e) = f e z foldr f z (E2 e) = z foldr f z (E3 e) = z foldr f z (E4 e) = z foldMap f (E1 e) = f e foldMap f (E2 e) = mempty foldMap f (E3 e) = mempty foldMap f (E4 e) = mempty
Notice how every constructor of
E
utilizes some sort of existential quantification, but only the argument ofE1
is actually “folded over”. This is because we make a deliberate choice to only fold over universally polymorphic types that are syntactically equivalent to the last type parameter. In particular:
 We don’t fold over the arguments of
E1
orE4
because even though(a ~ Int)
,Int
is not syntactically equivalent toa
.  We don’t fold over the argument of
E3
becausea
is not universally polymorphic. Thea
inE3
is (implicitly) existentially quantified, so it is not the same as the last type parameter ofE
.
10.6.4.3. Deriving Traversable
instances

DeriveTraversable

 Implies
 Since

7.10.1
Allow automatic deriving of instances for the
Traversable
typeclass.
With DeriveTraversable
, one can derive Traversable
instances for data types of kind Type > Type
. For example, this declaration:
data Example a = Ex a Char (Example a) (Example Char) deriving (Functor, Foldable, Traversable)
would generate the following Traversable
instance:
instance Traversable Example where traverse f (Ex a1 a2 a3 a4) = fmap (\b1 b3 > Ex b1 a2 b3 a4) (f a1) <*> traverse f a3
The algorithm for DeriveTraversable
is adapted from the DeriveFunctor
algorithm, but it generates a definition for traverse
instead of fmap
. In addition, DeriveTraversable
filters out all constructor arguments on the RHS expression whose types do not mention the last type parameter, since those arguments do not produce any effects in a traversal.
When the type parameter has a phantom role (see Roles), DeriveTraversable
coerces its argument. For example, this declaration:
data Phantom a = Z  S (Phantom a) deriving Traversable
will generate the following instance:
instance Traversable Phantom where traverse _ z = pure (coerce z)
When the type has no constructors, DeriveTraversable
will derive the laziest instance it can.
data V a deriving Traversable type role V nominal
will generate the following, using EmptyCase
:
instance Traversable V where traverse _ z = pure (case z of)
Here are the differences between the generated code in each extension:
 When a bare type variable
a
is encountered, bothDeriveFunctor
andDeriveTraversable
would generatef a
for anfmap
andtraverse
definition, respectively.  When a type that is not syntactically equivalent to
a
, but which does containa
, is encountered,DeriveFunctor
recursively callsfmap
on it. Similarly,DeriveTraversable
would recursively calltraverse
. 
DeriveFunctor
puts everything back together again at the end by invoking the constructor.DeriveTraversable
does something similar, but it works in anApplicative
context by chaining everything together with(<*>)
.
Unlike DeriveFunctor
, DeriveTraversable
cannot be used on data types containing a function type on the righthand side.
For a full specification of the algorithms used in DeriveFunctor
, DeriveFoldable
, and DeriveTraversable
, see this wiki page.
10.6.4.4. Deriving Data
instances

DeriveDataTypeable

 Since

6.8.1
Enable automatic deriving of instances for the
Data
typeclass
10.6.4.5. Deriving Typeable
instances
The class Typeable
is very special:

Typeable
is kindpolymorphic (see Kind polymorphism).  GHC has a custom solver for discharging constraints that involve class
Typeable
, and handwritten instances are forbidden. This ensures that the programmer cannot subvert the type system by writing bogus instances.  Derived instances of
Typeable
may be declared if theDeriveDataTypeable
extension is enabled, but they are ignored, and they may be reported as an error in a later version of the compiler. 
The rules for solving
Typeable
constraints are as follows:
A concrete type constructor applied to some types.
instance (Typeable t1, .., Typeable t_n) => Typeable (T t1 .. t_n)
This rule works for any concrete type constructor, including type constructors with polymorphic kinds. The only restriction is that if the type constructor has a polymorphic kind, then it has to be applied to all of its kinds parameters, and these kinds need to be concrete (i.e., they cannot mention kind variables).

A type variable applied to some types:
instance (Typeable f, Typeable t1, .., Typeable t_n) => Typeable (f t1 .. t_n)

A concrete type literal.:
instance Typeable 0  Type natural literals instance Typeable "Hello"  Typelevel symbols

10.6.4.6. Deriving Lift
instances

DeriveLift

 Since

8.0.1
Enable automatic deriving of instances for the
Lift
typeclass for Template Haskell.
The class Lift
, unlike other derivable classes, lives in templatehaskell
instead of base
. Having a data type be an instance of Lift
permits its values to be promoted to Template Haskell expressions (of type ExpQ
and TExpQ a
), which can then be spliced into Haskell source code.
Here is an example of how one can derive Lift
:
{# LANGUAGE DeriveLift #} module Bar where import Language.Haskell.TH.Syntax data Foo a = Foo a  a :^: a deriving Lift { instance (Lift a) => Lift (Foo a) where lift (Foo a) = [ Foo a ] lift ((:^:) u v) = [ (:^:) u v ] liftTyped (Foo a) = [ Foo a ] liftTyped ((:^:) u v) = [ (:^:) u v ] }  {# LANGUAGE TemplateHaskell #} module Baz where import Bar import Language.Haskell.TH.Lift foo :: Foo String foo = $(lift $ Foo "foo") fooExp :: Lift a => Foo a > Q Exp fooExp f = [ f ]
Note that the Lift
typeclass takes advantage of Levity polymorphism in order to support instances involving unboxed types. This means DeriveLift
also works for these types:
{# LANGUAGE DeriveLift, MagicHash #} module Unboxed where import GHC.Exts import Language.Haskell.TH.Syntax data IntHash = IntHash Int# deriving Lift { instance Lift IntHash where lift (IntHash i) = [ IntHash i ] liftTyped (IntHash i) = [ IntHash i ] }
10.6.5. Generalised derived instances for newtypes

GeneralisedNewtypeDeriving

GeneralizedNewtypeDeriving

 Since

6.8.1. British spelling since 8.6.1.
Enable GHC’s cunning generalised deriving mechanism for
newtype
s
When you define an abstract type using newtype
, you may want the new type to inherit some instances from its representation. In Haskell 98, you can inherit instances of Eq
, Ord
, Enum
and Bounded
by deriving them, but for any other classes you have to write an explicit instance declaration. For example, if you define
newtype Dollars = Dollars Int
and you want to use arithmetic on Dollars
, you have to explicitly define an instance of Num
:
instance Num Dollars where Dollars a + Dollars b = Dollars (a+b) ...
All the instance does is apply and remove the newtype
constructor. It is particularly galling that, since the constructor doesn’t appear at runtime, this instance declaration defines a dictionary which is wholly equivalent to the Int
dictionary, only slower!
DerivingVia
(see Deriving via) is a generalization of this idea.
10.6.5.1. Generalising the deriving clause
GHC now permits such instances to be derived instead, using the extension GeneralizedNewtypeDeriving
, so one can write
newtype Dollars = Dollars { getDollars :: Int } deriving (Eq,Show,Num)
and the implementation uses the same Num
dictionary for Dollars
as for Int
. In other words, GHC will generate something that resembles the following code
instance Num Int => Num Dollars
and then attempt to simplify the Num Int
context as much as possible. GHC knows that there is a Num Int
instance in scope, so it is able to discharge the Num Int
constraint, leaving the code that GHC actually generates
instance Num Dollars
One can think of this instance being implemented with the same code as the Num Int
instance, but with Dollars
and getDollars
added wherever necessary in order to make it typecheck. (In practice, GHC uses a somewhat different approach to code generation. See the A more precise specification section below for more details.)
We can also derive instances of constructor classes in a similar way. For example, suppose we have implemented state and failure monad transformers, such that
instance Monad m => Monad (State s m) instance Monad m => Monad (Failure m)
In Haskell 98, we can define a parsing monad by
type Parser tok m a = State [tok] (Failure m) a
which is automatically a monad thanks to the instance declarations above. With the extension, we can make the parser type abstract, without needing to write an instance of class Monad
, via
newtype Parser tok m a = Parser (State [tok] (Failure m) a) deriving Monad
In this case the derived instance declaration is of the form
instance Monad (State [tok] (Failure m)) => Monad (Parser tok m)
Notice that, since Monad
is a constructor class, the instance is a partial application of the newtype, not the entire left hand side. We can imagine that the type declaration is “etaconverted” to generate the context of the instance declaration.
We can even derive instances of multiparameter classes, provided the newtype is the last class parameter. In this case, a “partial application” of the class appears in the deriving
clause. For example, given the class
class StateMonad s m  m > s where ... instance Monad m => StateMonad s (State s m) where ...
then we can derive an instance of StateMonad
for Parser
by
newtype Parser tok m a = Parser (State [tok] (Failure m) a) deriving (Monad, StateMonad [tok])
The derived instance is obtained by completing the application of the class to the new type:
instance StateMonad [tok] (State [tok] (Failure m)) => StateMonad [tok] (Parser tok m)
As a result of this extension, all derived instances in newtype declarations are treated uniformly (and implemented just by reusing the dictionary for the representation type), except Show
and Read
, which really behave differently for the newtype and its representation.
Note
It is sometimes necessary to enable additional language extensions when deriving instances via GeneralizedNewtypeDeriving
. For instance, consider a simple class and instance using UnboxedTuples
syntax:
{# LANGUAGE UnboxedTuples #} module Lib where class AClass a where aMethod :: a > (# Int, a #) instance AClass Int where aMethod x = (# x, x #)
The following will fail with an “Illegal unboxed tuple” error, since the derived instance produced by the compiler makes use of unboxed tuple syntax,
{# LANGUAGE GeneralizedNewtypeDeriving #} import Lib newtype Int' = Int' Int deriving (AClass)
However, enabling the UnboxedTuples
extension allows the module to compile. Similar errors may occur with a variety of extensions, including:
10.6.5.2. A more precise specification
A derived instance is derived only for declarations of these forms (after expansion of any type synonyms)
newtype T v1..vn = MkT (t vk+1..vn) deriving (C t1..tj) newtype instance T s1..sk vk+1..vn = MkT (t vk+1..vn) deriving (C t1..tj)
where

v1..vn
are type variables, andt
,s1..sk
,t1..tj
are types.  The
(C t1..tj)
is a partial applications of the classC
, where the arity ofC
is exactlyj+1
. That is,C
lacks exactly one type argument. 
k
is chosen so thatC t1..tj (T v1...vk)
is wellkinded. (Or, in the case of adata instance
, so thatC t1..tj (T s1..sk)
is well kinded.)  The type
t
is an arbitrary type.  The type variables
vk+1...vn
do not occur in the typest
,s1..sk
, ort1..tj
. 
C
is notRead
,Show
,Typeable
, orData
. These classes should not “look through” the type or its constructor. You can still derive these classes for a newtype, but it happens in the usual way, not via this new mechanism. Confer with Default deriving strategy.  It is safe to coerce each of the methods of
C
. That is, the missing last argument toC
is not used at a nominal role in any of theC
’s methods. (See Roles.) 
C
is allowed to have associated type families, provided they meet the requirements laid out in the section on GND and associated types.
Then the derived instance declaration is of the form
instance C t1..tj t => C t1..tj (T v1...vk)
Note that if C
does not contain any class methods, the instance context is wholly unnecessary, and as such GHC will instead generate:
instance C t1..tj (T v1..vk)
As an example which does not work, consider
newtype NonMonad m s = NonMonad (State s m s) deriving Monad
Here we cannot derive the instance
instance Monad (State s m) => Monad (NonMonad m)
because the type variable s
occurs in State s m
, and so cannot be “etaconverted” away. It is a good thing that this deriving
clause is rejected, because NonMonad m
is not, in fact, a monad — for the same reason. Try defining >>=
with the correct type: you won’t be able to.
Notice also that the order of class parameters becomes important, since we can only derive instances for the last one. If the StateMonad
class above were instead defined as
class StateMonad m s  m > s where ...
then we would not have been able to derive an instance for the Parser
type above. We hypothesise that multiparameter classes usually have one “main” parameter for which deriving new instances is most interesting.
Lastly, all of this applies only for classes other than Read
, Show
, Typeable
, and Data
, for which the stock derivation applies (section 4.3.3. of the Haskell Report). (For the standard classes Eq
, Ord
, Ix
, and Bounded
it is immaterial whether the stock method is used or the one described here.)
10.6.5.3. Associated type families
GeneralizedNewtypeDeriving
also works for some type classes with associated type families. Here is an example:
class HasRing a where type Ring a newtype L1Norm a = L1Norm a deriving HasRing
The derived HasRing
instance would look like
instance HasRing (L1Norm a) where type Ring (L1Norm a) = Ring a
To be precise, if the class being derived is of the form
class C c_1 c_2 ... c_m where type T1 t1_1 t1_2 ... t1_n ... type Tk tk_1 tk_2 ... tk_p
and the newtype is of the form
newtype N n_1 n_2 ... n_q = MkN <reptype>
then you can derive a C c_1 c_2 ... c_(m1)
instance for N n_1 n_2 ... n_q
, provided that:

The type parameter
c_m
occurs once in each of the type variables ofT1
throughTk
. Imagine a class where this condition didn’t hold. For example:class Bad a b where type B a instance Bad Int a where type B Int = Char newtype Foo a = Foo a deriving (Bad Int)
For the derived
Bad Int
instance, GHC would need to generate something like this:instance Bad Int (Foo a) where type B Int = B ???
Now we’re stuck, since we have no way to refer to
a
on the righthand side of theB
family instance, so this instance doesn’t really make sense in aGeneralizedNewtypeDeriving
setting. 
C
does not have any associated data families (only type families). To see why data families are forbidden, imagine the following scenario:class Ex a where data D a instance Ex Int where data D Int = DInt Bool newtype Age = MkAge Int deriving Ex
For the derived
Ex
instance, GHC would need to generate something like this:instance Ex Age where data D Age = ???
But it is not clear what GHC would fill in for
???
, as each data family instance must generate fresh data constructors.
If both of these conditions are met, GHC will generate this instance:
instance C c_1 c_2 ... c_(m1) <reptype> => C c_1 c_2 ... c_(m1) (N n_1 n_2 ... n_q) where type T1 t1_1 t1_2 ... (N n_1 n_2 ... n_q) ... t1_n = T1 t1_1 t1_2 ... <reptype> ... t1_n ... type Tk tk_1 tk_2 ... (N n_1 n_2 ... n_q) ... tk_p = Tk tk_1 tk_2 ... <reptype> ... tk_p
Again, if C
contains no class methods, the instance context will be redundant, so GHC will instead generate instance C c_1 c_2 ... c_(m1) (N n_1 n_2 ... n_q)
.
Beware that in some cases, you may need to enable the UndecidableInstances
extension in order to use this feature. Here’s a pathological case that illustrates why this might happen:
class C a where type T a newtype Loop = MkLoop Loop deriving C
This will generate the derived instance:
instance C Loop where type T Loop = T Loop
Here, it is evident that attempting to use the type T Loop
will throw the typechecker into an infinite loop, as its definition recurses endlessly. In other cases, you might need to enable UndecidableInstances
even if the generated code won’t put the typechecker into a loop. For example:
instance C Int where type C Int = Int newtype MyInt = MyInt Int deriving C
This will generate the derived instance:
instance C MyInt where type T MyInt = T Int
Although typechecking T MyInt
will terminate, GHC’s termination checker isn’t sophisticated enough to determine this, so you’ll need to enable UndecidableInstances
in order to use this derived instance. If you do go down this route, make sure you can convince yourself that all of the type family instances you’re deriving will eventually terminate if used!
Note that DerivingVia
(see Deriving via) uses essentially the same specification to derive instances of associated type families as well (except that it uses the via
type instead of the underlying reptype
of a newtype).
10.6.6. Deriving any other class

DeriveAnyClass

 Since

7.10.1
Allow use of any typeclass in
deriving
clauses.
With DeriveAnyClass
you can derive any other class. The compiler will simply generate an instance declaration with no explicitlydefined methods. This is mostly useful in classes whose minimal set is empty, and especially when writing generic functions.
As an example, consider a simple prettyprinter class SPretty
, which outputs pretty strings:
{# LANGUAGE DefaultSignatures, DeriveAnyClass #} class SPretty a where sPpr :: a > String default sPpr :: Show a => a > String sPpr = show
If a user does not provide a manual implementation for sPpr
, then it will default to show
. Now we can leverage the DeriveAnyClass
extension to easily implement a SPretty
instance for a new data type:
data Foo = Foo deriving (Show, SPretty)
The above code is equivalent to:
data Foo = Foo deriving Show instance SPretty Foo
That is, an SPretty Foo
instance will be created with empty implementations for all methods. Since we are using DefaultSignatures
in this example, a default implementation of sPpr
is filled in automatically.
Note the following details
 In case you try to derive some class on a newtype, and
GeneralizedNewtypeDeriving
is also on,DeriveAnyClass
takes precedence. 
The instance context is determined by the type signatures of the derived class’s methods. For instance, if the class is:
class Foo a where bar :: a > String default bar :: Show a => a > String bar = show baz :: a > a > Bool default baz :: Ord a => a > a > Bool baz x y = compare x y == EQ
And you attempt to derive it using
DeriveAnyClass
:instance Eq a => Eq (Option a) where ... instance Ord a => Ord (Option a) where ... instance Show a => Show (Option a) where ... data Option a = None  Some a deriving Foo
Then the derived
Foo
instance will be:instance (Show a, Ord a) => Foo (Option a)
Since the default type signatures for
bar
andbaz
requireShow a
andOrd a
constraints, respectively.Constraints on the nondefault type signatures can play a role in inferring the instance context as well. For example, if you have this class:
class HigherEq f where (==#) :: f a > f a > Bool default (==#) :: Eq (f a) => f a > f a > Bool x ==# y = (x == y)
And you tried to derive an instance for it:
instance Eq a => Eq (Option a) where ... data Option a = None  Some a deriving HigherEq
Then it will fail with an error to the effect of:
No instance for (Eq a) arising from the 'deriving' clause of a data type declaration
That is because we require an
Eq (Option a)
instance from the default type signature for(==#)
, which in turn requires anEq a
instance, which we don’t have in scope. But if you tweak the definition ofHigherEq
slightly:class HigherEq f where (==#) :: Eq a => f a > f a > Bool default (==#) :: Eq (f a) => f a > f a > Bool x ==# y = (x == y)
Then it becomes possible to derive a
HigherEq Option
instance. Note that the only difference is that now the nondefault type signature for(==#)
brings in anEq a
constraint. Constraints from nondefault type signatures never appear in the derived instance context itself, but they can be used to discharge obligations that are demanded by the default type signatures. In the example above, the default type signature demanded anEq a
instance, and the nondefault signature was able to satisfy that request, so the derived instance is simply:instance HigherEq Option

DeriveAnyClass
can be used with partially applied classes, such asdata T a = MKT a deriving( D Int )
which generates
instance D Int a => D Int (T a) where {}

DeriveAnyClass
can be used to fill in default instances for associated type families:{# LANGUAGE DeriveAnyClass, TypeFamilies #} class Sizable a where type Size a type Size a = Int data Bar = Bar deriving Sizable doubleBarSize :: Size Bar > Size Bar doubleBarSize s = 2*s
The
deriving( Sizable )
is equivalent to sayinginstance Sizeable Bar where {}
and then the normal rules for filling in associated types from the default will apply, making
Size Bar
equal toInt
.
10.6.7. Deriving strategies

DerivingStrategies

 Since

8.2.1
Allow multiple
deriving
, each optionally qualified with a strategy.
In most scenarios, every deriving
statement generates a typeclass instance in an unambiguous fashion. There is a corner case, however, where simultaneously enabling both the GeneralizedNewtypeDeriving
and DeriveAnyClass
extensions can make deriving become ambiguous. Consider the following example
{# LANGUAGE DeriveAnyClass, GeneralizedNewtypeDeriving #} newtype Foo = MkFoo Bar deriving C
One could either pick the DeriveAnyClass
approach to deriving C
or the GeneralizedNewtypeDeriving
approach to deriving C
, both of which would be equally as valid. GHC defaults to favoring DeriveAnyClass
in such a dispute, but this is not a satisfying solution, since that leaves users unable to use both language extensions in a single module.
To make this more robust, GHC has a notion of deriving strategies, which allow the user to explicitly request which approach to use when deriving an instance. To enable this feature, one must enable the DerivingStrategies
language extension. A deriving strategy can be specified in a deriving clause
newtype Foo = MkFoo Bar deriving newtype C
Or in a standalone deriving declaration
deriving anyclass instance C Foo
DerivingStrategies
also allows the use of multiple deriving clauses per data declaration so that a user can derive some instance with one deriving strategy and other instances with another deriving strategy. For example
newtype Baz = Baz Quux deriving (Eq, Ord) deriving stock (Read, Show) deriving newtype (Num, Floating) deriving anyclass C
Currently, the deriving strategies are:

stock
: Have GHC implement a “standard” instance for a data type, if possible (e.g.,Eq
,Ord
,Generic
,Data
,Functor
, etc.) 
anyclass
: UseDeriveAnyClass
(see Deriving any other class) 

newtype: Use
GeneralizedNewtypeDeriving


via
: UseDerivingVia
(see Deriving via)
10.6.7.1. Default deriving strategy
If an explicit deriving strategy is not given, multiple strategies may apply. In that case, GHC chooses the strategy as follows:

Stock type classes, i.e. those specified in the report and those enabled by language extensions, are derived using the
stock
strategy, with the following exception: For newtypes,
Eq
,Ord
,Ix
andBounded
are always derived using thenewtype
strategy, even withoutGeneralizedNewtypeDeriving
enabled. (There should be no observable difference to instances derived using the stock strategy.)  Also for newtypes,
Functor
,Foldable
andEnum
are derived using thenewtype
strategy ifGeneralizedNewtypeDeriving
is enabled and the derivation succeeds.
 For newtypes,

For other any type class:
 When
DeriveAnyClass
is enabled, useanyclass
.  When
GeneralizedNewtypeDeriving
is enabled and we are deriving for a newtype, then usenewtype
.
If both rules apply to a deriving clause, then
anyclass
is used and the user is warned about the ambiguity. The warning can be avoided by explicitly stating the desired deriving strategy.  When
10.6.8. Deriving via

DerivingVia

 Implies
 Since

8.6.1
This allows deriving
a class instance for a type by specifying another type of equal runtime representation (such that there exists a Coercible
instance between the two: see The Coercible constraint) that is already an instance of the that class.
DerivingVia
is indicated by the use of the via
deriving strategy. via
requires specifying another type (the via
type) to coerce
through. For example, this code:
{# LANGUAGE DerivingVia #} import Numeric newtype Hex a = Hex a instance (Integral a, Show a) => Show (Hex a) where show (Hex a) = "0x" ++ showHex a "" newtype Unicode = U Int deriving Show via (Hex Int)  >>> euroSign  0x20ac euroSign :: Unicode euroSign = U 0x20ac
Generates the following instance
instance Show Unicode where show :: Unicode > String show = Data.Coerce.coerce @(Hex Int > String) @(Unicode > String) show
This extension generalizes GeneralizedNewtypeDeriving
. To derive Num Unicode
with GND (deriving newtype Num
) it must reuse the Num Int
instance. With DerivingVia
, we can explicitly specify the representation type Int
:
newtype Unicode = U Int deriving Num via Int deriving Show via (Hex Int) euroSign :: Unicode euroSign = 0x20ac
Code duplication is common in instance declarations. A familiar pattern is lifting operations over an Applicative
functor. Instead of having catchall instances for f a
which overlap with all other such instances, like so:
instance (Applicative f, Semigroup a) => Semigroup (f a) .. instance (Applicative f, Monoid a) => Monoid (f a) ..
We can instead create a newtype App
(where App f a
and f a
are represented the same in memory) and use DerivingVia
to explicitly enable uses of this pattern:
{# LANGUAGE DerivingVia, DeriveFunctor, GeneralizedNewtypeDeriving #} import Control.Applicative newtype App f a = App (f a) deriving newtype (Functor, Applicative) instance (Applicative f, Semigroup a) => Semigroup (App f a) where (<>) = liftA2 (<>) instance (Applicative f, Monoid a) => Monoid (App f a) where mempty = pure mempty data Pair a = MkPair a a deriving stock Functor deriving (Semigroup, Monoid) via (App Pair a) instance Applicative Pair where pure a = MkPair a a MkPair f g <*> MkPair a b = MkPair (f a) (g b)
Note that the via
type does not have to be a newtype
. The only restriction is that it is coercible with the original data type. This means there can be arbitrary nesting of newtypes, as in the following example:
newtype Kleisli m a b = (a > m b) deriving (Semigroup, Monoid) via (a > App m b)
Here we make use of the Monoid ((>) a)
instance.
When used in combination with StandaloneDeriving
we swap the order for the instance we base our derivation on and the instance we define e.g.:
deriving via (a > App m b) instance Monoid (Kleisli m a b)
10.7. Pattern synonyms

PatternSynonyms

 Since

7.8.1
Allow the definition of pattern synonyms.
Pattern synonyms are enabled by the language extension PatternSynonyms
, which is required for defining them, but not for using them. More information and examples of pattern synonyms can be found on the Wiki page.
Pattern synonyms enable giving names to parametrized pattern schemes. They can also be thought of as abstract constructors that don’t have a bearing on data representation. For example, in a programming language implementation, we might represent types of the language as follows:
data Type = App String [Type]
Here are some examples of using said representation. Consider a few types of the Type
universe encoded like this:
App ">" [t1, t2]  t1 > t2 App "Int" []  Int App "Maybe" [App "Int" []]  Maybe Int
This representation is very generic in that no types are given special treatment. However, some functions might need to handle some known types specially, for example the following two functions collect all argument types of (nested) arrow types, and recognize the Int
type, respectively:
collectArgs :: Type > [Type] collectArgs (App ">" [t1, t2]) = t1 : collectArgs t2 collectArgs _ = [] isInt :: Type > Bool isInt (App "Int" []) = True isInt _ = False
Matching on App
directly is both hard to read and error prone to write. And the situation is even worse when the matching is nested:
isIntEndo :: Type > Bool isIntEndo (App ">" [App "Int" [], App "Int" []]) = True isIntEndo _ = False
Pattern synonyms permit abstracting from the representation to expose matchers that behave in a constructorlike manner with respect to pattern matching. We can create pattern synonyms for the known types we care about, without committing the representation to them (note that these don’t have to be defined in the same module as the Type
type):
pattern Arrow t1 t2 = App ">" [t1, t2] pattern Int = App "Int" [] pattern Maybe t = App "Maybe" [t]
Which enables us to rewrite our functions in a much cleaner style:
collectArgs :: Type > [Type] collectArgs (Arrow t1 t2) = t1 : collectArgs t2 collectArgs _ = [] isInt :: Type > Bool isInt Int = True isInt _ = False isIntEndo :: Type > Bool isIntEndo (Arrow Int Int) = True isIntEndo _ = False
In general there are three kinds of pattern synonyms. Unidirectional, bidirectional and explicitly bidirectional. The examples given so far are examples of bidirectional pattern synonyms. A bidirectional synonym behaves the same as an ordinary data constructor. We can use it in a pattern context to deconstruct values and in an expression context to construct values. For example, we can construct the value intEndo
using the pattern synonyms Arrow
and Int
as defined previously.
intEndo :: Type intEndo = Arrow Int Int
This example is equivalent to the much more complicated construction if we had directly used the Type
constructors.
intEndo :: Type intEndo = App ">" [App "Int" [], App "Int" []]
Unidirectional synonyms can only be used in a pattern context and are defined as follows:
pattern Head x < x:xs
In this case, Head
⟨x⟩ cannot be used in expressions, only patterns, since it wouldn’t specify a value for the ⟨xs⟩ on the righthand side. However, we can define an explicitly bidirectional pattern synonym by separately specifying how to construct and deconstruct a type. The syntax for doing this is as follows:
pattern HeadC x < x:xs where HeadC x = [x]
We can then use HeadC
in both expression and pattern contexts. In a pattern context it will match the head of any list with length at least one. In an expression context it will construct a singleton list.
Explicitly bidirectional pattern synonyms offer greater flexibility than implicitly bidirectional ones in terms of the syntax that is permitted. For instance, the following is not a legal implicitly bidirectional pattern synonym:
pattern StrictJust a = Just !a
This is illegal because the use of BangPatterns
on the righthand sides prevents it from being a well formed expression. However, constructing a strict pattern synonym is quite possible with an explicitly bidirectional pattern synonym:
pattern StrictJust a < Just !a where StrictJust !a = Just a
Constructing an explicitly bidirectional pattern synonym also:
 can create different data constructors from the underlying data type, not just the one appearing in the pattern match;
 can call any functions or conditional logic, especially validation, of course providing it constructs a result of the right type;
 can use guards on the lhs of the
=
;  can have multiple equations.
For example:
data PosNeg = Pos Int  Neg Int pattern Smarter{ nonneg } < Pos nonneg where Smarter x = if x >= 0 then (Pos x) else (Neg x)
Or using guards:
pattern Smarter{ nonneg } < Pos nonneg where Smarter x  x >= 0 = (Pos x)  otherwise = (Neg x)
There is an extensive Haskell folk art of smart constructors, essentially functions that wrap validation around a constructor, and avoid exposing its representation. The downside is that the underlying constructor can’t be used as a matcher. Pattern synonyms can be used as genuinely smart constructors, for both validation and matching.
The table below summarises where each kind of pattern synonym can be used.
Context  Unidirectional  Bidirectional  Explicitly Bidirectional 

Pattern  Yes  Yes  Yes 
Expression  No  Yes (Inferred)  Yes (Explicit) 
10.7.1. Record Pattern Synonyms
It is also possible to define pattern synonyms which behave just like record constructors. The syntax for doing this is as follows:
pattern Point :: Int > Int > (Int, Int) pattern Point{x, y} = (x, y)
The idea is that we can then use Point
just as if we had defined a new datatype MyPoint
with two fields x
and y
.
data MyPoint = Point { x :: Int, y :: Int }
Whilst a normal pattern synonym can be used in two ways, there are then seven ways in which to use Point
. Precisely the ways in which a normal record constructor can be used.
Usage  Example 

As a constructor 

As a constructor with record syntax 

In a pattern context 

In a pattern context with record syntax 

In a pattern context with field puns 

In a record update 

Using record selectors 

For a unidirectional record pattern synonym we define record selectors but do not allow record updates or construction.
The syntax and semantics of pattern synonyms are elaborated in the following subsections. There are also lots more details in the paper.
See the Wiki page for more details.
10.7.2. Syntax and scoping of pattern synonyms
A pattern synonym declaration can be either unidirectional, bidirectional or explicitly bidirectional. The syntax for unidirectional pattern synonyms is:
pattern pat_lhs < pat
the syntax for bidirectional pattern synonyms is:
pattern pat_lhs = pat
and the syntax for explicitly bidirectional pattern synonyms is:
pattern pat_lhs < pat where pat_lhs = expr  lhs restricted, see below
We can define either prefix, infix or record pattern synonyms by modifying the form of pat_lhs
. The syntax for these is as follows:
Prefix 

Infix 

Record 

The pat_lhs
for explicitly bidirectional construction cannot use Record syntax. (Because the rhs expr might be constructing different data constructors.) It can use guards with multiple equations.
Pattern synonym declarations can only occur in the top level of a module. In particular, they are not allowed as local definitions.
The variables in the lefthand side of the definition are bound by the pattern on the righthand side. For bidirectional pattern synonyms, all the variables of the righthand side must also occur on the lefthand side; also, wildcard patterns and view patterns are not allowed. For unidirectional and explicitly bidirectional pattern synonyms, there is no restriction on the righthand side pattern.
Pattern synonyms cannot be defined recursively.
COMPLETE pragmas can be specified in order to tell the pattern match exhaustiveness checker that a set of pattern synonyms is complete.
10.7.3. Import and export of pattern synonyms
The name of the pattern synonym is in the same namespace as proper data constructors. Like normal data constructors, pattern synonyms can be imported and exported through association with a type constructor or independently.
To export them on their own, in an export or import specification, you must prefix pattern names with the pattern
keyword, e.g.:
module Example (pattern Zero) where data MyNum = MkNum Int pattern Zero :: MyNum pattern Zero = MkNum 0
Without the pattern
prefix, Zero
would be interpreted as a type constructor in the export list.
You may also use the pattern
keyword in an import/export specification to import or export an ordinary data constructor. For example:
import Data.Maybe( pattern Just )
would bring into scope the data constructor Just
from the Maybe
type, without also bringing the type constructor Maybe
into scope.
To bundle a pattern synonym with a type constructor, we list the pattern synonym in the export list of a module which exports the type constructor. For example, to bundle Zero
with MyNum
we could write the following:
module Example ( MyNum(Zero) ) where
If a module was then to import MyNum
from Example
, it would also import the pattern synonym Zero
.
It is also possible to use the special token ..
in an export list to mean all currently bundled constructors. For example, we could write:
module Example ( MyNum(.., Zero) ) where
in which case, Example
would export the type constructor MyNum
with the data constructor MkNum
and also the pattern synonym Zero
.
Bundled pattern synonyms are type checked to ensure that they are of the same type as the type constructor which they are bundled with. A pattern synonym P
can not be bundled with a type constructor T
if P
’s type is visibly incompatible with T
.
A module which imports MyNum(..)
from Example
and then reexports MyNum(..)
will also export any pattern synonyms bundled with MyNum
in Example
. A more complete specification can be found on the wiki.
10.7.4. Typing of pattern synonyms
Given a pattern synonym definition of the form
pattern P var1 var2 ... varN < pat
it is assigned a pattern type of the form
pattern P :: CReq => CProv => t1 > t2 > ... > tN > t
where ⟨CReq⟩ and ⟨CProv⟩ are type contexts, and ⟨t1⟩, ⟨t2⟩, …, ⟨tN⟩ and ⟨t⟩ are types. Notice the unusual form of the type, with two contexts ⟨CReq⟩ and ⟨CProv⟩:
 ⟨CReq⟩ are the constraints required to match the pattern.
 ⟨CProv⟩ are the constraints made available (provided) by a successful pattern match.
For example, consider
data T a where MkT :: (Show b) => a > b > T a f1 :: (Num a, Eq a) => T a > String f1 (MkT 42 x) = show x pattern ExNumPat :: (Num a, Eq a) => (Show b) => b > T a pattern ExNumPat x = MkT 42 x f2 :: (Eq a, Num a) => T a > String f2 (ExNumPat x) = show x
Here f1
does not use pattern synonyms. To match against the numeric pattern 42
requires the caller to satisfy the constraints (Num a, Eq a)
, so they appear in f1
’s type. The call to show
generates a (Show b)
constraint, where b
is an existentially type variable bound by the pattern match on MkT
. But the same pattern match also provides the constraint (Show b)
(see MkT
’s type), and so all is well.
Exactly the same reasoning applies to ExNumPat
: matching against ExNumPat
requires the constraints (Num a, Eq a)
, and provides the constraint (Show b)
.
Note also the following points
 In the common case where
CProv
is empty, (i.e.,()
), it can be omitted altogether in the above pattern type signature forP
. 
However, if
CProv
is nonempty, whileCReq
is, the above pattern type signature forP
must be specified asP :: () => CProv => t1 > t2 > .. > tN > t
 The GHCi
:info
command shows pattern types in this format. 
You may specify an explicit pattern signature, as we did for
ExNumPat
above, to specify the type of a pattern, just as you can for a function. As usual, the type signature can be less polymorphic than the inferred type. For example Inferred type would be 'a > [a]' pattern SinglePair :: (a, a) > [(a, a)] pattern SinglePair x = [x]
Just like signatures on valuelevel bindings, pattern synonym signatures can apply to more than one pattern. For instance,
pattern Left', Right' :: a > Either a a pattern Left' x = Left x pattern Right' x = Right x

The rules for lexicallyscoped type variables (see Lexically scoped type variables) apply to patternsynonym signatures. As those rules specify, only the type variables from an explicit, syntacticallyvisible outer
forall
(the universals) scope over the definition of the pattern synonym; the existentials, bound by the inner forall, do not. For exampledata T a where MkT :: Bool > b > (b>Int) > a > T a pattern P :: forall a. forall b. b > (b>Int) > a > T a pattern P x y v < MkT True x y (v::a)
Here the universal type variable
a
scopes over the definition ofP
, but the existentialb
does not. (c.f. discussion on #14998.) 
For a bidirectional pattern synonym, a use of the pattern synonym as an expression has the type
(CReq, CProv) => t1 > t2 > ... > tN > t
So in the previous example, when used in an expression,
ExNumPat
has typeExNumPat :: (Num a, Eq a, Show b) => b > T t
Notice that this is a tiny bit more restrictive than the expression
MkT 42 x
which would not require(Eq a)
. 
Consider these two pattern synonyms:
data S a where S1 :: Bool > S Bool pattern P1 :: Bool > Maybe Bool pattern P1 b = Just b pattern P2 :: () => (b ~ Bool) => Bool > S b pattern P2 b = S1 b f :: Maybe a > String f (P1 x) = "no no no"  Typeincorrect g :: S a > String g (P2 b) = "yes yes yes"  Fine
Pattern
P1
can only match against a value of typeMaybe Bool
, so functionf
is rejected because the type signature isMaybe a
. (To see this, imagine expanding the pattern synonym.)On the other hand, function
g
works fine, because matching againstP2
(which wraps the GADTS
) provides the local equality(a~Bool)
. If you were to give an explicit pattern signatureP2 :: Bool > S Bool
, thenP2
would become less polymorphic, and would behave exactly likeP1
so thatg
would then be rejected.In short, if you want GADTlike behaviour for pattern synonyms, then (unlike concrete data constructors like
S1
) you must write its type with explicit provided equalities. For a concrete data constructor likeS1
you can write its type signature as eitherS1 :: Bool > S Bool
orS1 :: (b~Bool) => Bool > S b
; the two are equivalent. Not so for pattern synonyms: the two forms are different, in order to distinguish the two cases above. (See #9953 for discussion of this choice.)
10.7.5. Matching of pattern synonyms
A pattern synonym occurrence in a pattern is evaluated by first matching against the pattern synonym itself, and then on the argument patterns.
More precisely, the semantics of pattern matching is given in Section 3.17 of the Haskell 2010 report. To the informal semantics in Section 3.17.2 we add this extra rule:

If the pattern is a constructor pattern
(P p1 ... pn)
, whereP
is a pattern synonym defined byP x1 ... xn = p
orP x1 ... xn < p
, then: Match the value
v
againstp
. If this match fails or diverges, so does the whole (pattern synonym) match. Otherwise the match againstp
must bind the variablesx1 ... xn
; let them be bound to valuesv1 ... vn
.  Match
v1
againstp1
,v2
againstp2
and so on. If any of these matches fail or diverge, so does the whole match.  If all the matches against the
pi
succeed, the match succeeds, binding the variables bound by thepi
. (Thexi
are not bound; they remain local to the pattern synonym declaration.)
 Match the value
For example, in the following program, f
and f'
are equivalent:
pattern Pair x y < [x, y] f (Pair True True) = True f _ = False f' [x, y]  True < x, True < y = True f' _ = False
Note that the strictness of f
differs from that of g
defined below:
g [True, True] = True g _ = False *Main> f (False:undefined) *** Exception: Prelude.undefined *Main> g (False:undefined) False
10.8. Class and instances declarations
10.8.1. Class declarations
This section, and the next one, documents GHC’s typeclass extensions. There’s lots of background in the paper Type classes: exploring the design space (Simon Peyton Jones, Mark Jones, Erik Meijer).
10.8.1.1. Multiparameter type classes

MultiParamTypeClasses

 Implies
 Since

6.8.1
Allow the definition of typeclasses with more than one parameter.
Multiparameter type classes are permitted, with extension MultiParamTypeClasses
. For example:
class Collection c a where union :: c a > c a > c a ...etc.
10.8.1.2. The superclasses of a class declaration

FlexibleContexts

 Since

6.8.1
Allow the use of complex constraints in class declaration contexts.
In Haskell 98 the context of a class declaration (which introduces superclasses) must be simple; that is, each predicate must consist of a class applied to type variables. The extension FlexibleContexts
(The context of a type signature) lifts this restriction, so that the only restriction on the context in a class declaration is that the class hierarchy must be acyclic. So these class declarations are OK:
class Functor (m k) => FiniteMap m k where ... class (Monad m, Monad (t m)) => Transform t m where lift :: m a > (t m) a
As in Haskell 98, the class hierarchy must be acyclic. However, the definition of “acyclic” involves only the superclass relationships. For example, this is okay:
class C a where op :: D b => a > b > b class C a => D a where ...
Here, C
is a superclass of D
, but it’s OK for a class operation op
of C
to mention D
. (It would not be OK for D
to be a superclass of C
.)
With the extension that adds a kind of constraints, you can write more exotic superclass definitions. The superclass cycle check is even more liberal in these case. For example, this is OK:
class A cls c where meth :: cls c => c > c class A B c => B c where
A superclass context for a class C
is allowed if, after expanding type synonyms to their righthandsides, and uses of classes (other than C
) to their superclasses, C
does not occur syntactically in the context.
10.8.1.3. Constrained class method types

ConstrainedClassMethods

 Since

6.8.1
Allows the definition of further constraints on individual class methods.
Haskell 98 prohibits class method types to mention constraints on the class type variable, thus:
class Seq s a where fromList :: [a] > s a elem :: Eq a => a > s a > Bool
The type of elem
is illegal in Haskell 98, because it contains the constraint Eq a
, which constrains only the class type variable (in this case a
). this case a
). More precisely, a constraint in a class method signature is rejected if

The constraint mentions at least one type variable. So this is allowed:
class C a where op1 :: HasCallStack => a > a op2 :: (?x::Int) => Int > a

All of the type variables mentioned are bound by the class declaration, and none is locally quantified. Examples:
class C a where op3 :: Eq a => a > a  Rejected: constrains class variable only op4 :: D b => a > b  Accepted: constrains a locallyquantified variable `b` op5 :: D (a,b) => a > b  Accepted: constrains a locallyquantified variable `b`
GHC lifts this restriction with language extension ConstrainedClassMethods
. The restriction is a pretty stupid one in the first place, so ConstrainedClassMethods
is implied by MultiParamTypeClasses
.
10.8.1.4. Default method signatures

DefaultSignatures

 Since

7.2.1
Allows the definition of default method signatures in class definitions.
Haskell 98 allows you to define a default implementation when declaring a class:
class Enum a where enum :: [a] enum = []
The type of the enum
method is [a]
, and this is also the type of the default method. You can lift this restriction and give another type to the default method using the extension DefaultSignatures
. For instance, if you have written a generic implementation of enumeration in a class GEnum
with method genum
in terms of GHC.Generics
, you can specify a default method that uses that generic implementation:
class Enum a where enum :: [a] default enum :: (Generic a, GEnum (Rep a)) => [a] enum = map to genum
We reuse the keyword default
to signal that a signature applies to the default method only; when defining instances of the Enum
class, the original type [a]
of enum
still applies. When giving an empty instance, however, the default implementation (map to genum)
is filledin, and typechecked with the type (Generic a, GEnum (Rep a)) => [a]
.
The type signature for a default method of a type class must take on the same form as the corresponding main method’s type signature. Otherwise, the typechecker will reject that class’s definition. By “take on the same form”, we mean that the default type signature should differ from the main type signature only in their contexts. Therefore, if you have a method bar
:
class Foo a where bar :: forall b. C => a > b > b
Then a default method for bar
must take on the form:
default bar :: forall b. C' => a > b > b
C
is allowed to be different from C'
, but the righthand sides of the type signatures must coincide. We require this because when you declare an empty instance for a class that uses DefaultSignatures
, GHC implicitly fills in the default implementation like this:
instance Foo Int where bar = default_bar @Int
Where @Int
utilizes visible type application (Visible type application) to instantiate the b
in default bar :: forall b. C' => a > b > b
. In order for this type application to work, the default type signature for bar
must have the same type variable order as the nondefault signature! But there is no obligation for C
and C'
to be the same (see, for instance, the Enum
example above, which relies on this).
To further explain this example, the righthand side of the default type signature for bar
must be something that is alphaequivalent to forall b. a > b > b
(where a
is bound by the class itself, and is thus free in the methods’ type signatures). So this would also be an acceptable default type signature:
default bar :: forall x. C' => a > x > x
But not this (since the free variable a
is in the wrong place):
default bar :: forall b. C' => b > a > b
Nor this, since we can’t match the type variable b
with the concrete type Int
:
default bar :: C' => a > Int > Int
That last one deserves a special mention, however, since a > Int > Int
is a straightforward instantiation of forall b. a > b > b
. You can still write such a default type signature, but you now must use type equalities to do so:
default bar :: forall b. (C', b ~ Int) => a > b > b
We use default signatures to simplify generic programming in GHC (Generic programming).
10.8.1.5. Nullary type classes

NullaryTypeClasses

 Since

7.8.1
Allows the use definition of type classes with no parameters. This extension has been replaced by
MultiParamTypeClasses
.
Nullary (no parameter) type classes are enabled with MultiParamTypeClasses
; historically, they were enabled with the (now deprecated) NullaryTypeClasses
. Since there are no available parameters, there can be at most one instance of a nullary class. A nullary type class might be used to document some assumption in a type signature (such as reliance on the Riemann hypothesis) or add some globally configurable settings in a program. For example,
class RiemannHypothesis where assumeRH :: a > a  Deterministic version of the Miller test  correctness depends on the generalised Riemann hypothesis isPrime :: RiemannHypothesis => Integer > Bool isPrime n = assumeRH (...)
The type signature of isPrime
informs users that its correctness depends on an unproven conjecture. If the function is used, the user has to acknowledge the dependence with:
instance RiemannHypothesis where assumeRH = id
10.8.2. Functional dependencies

FunctionalDependencies

 Implies
 Since

6.8.1
Allow use of functional dependencies in class declarations.
Functional dependencies are implemented as described by Mark Jones in [Jones2000].
Functional dependencies are introduced by a vertical bar in the syntax of a class declaration; e.g.
class (Monad m) => MonadState s m  m > s where ... class Foo a b c  a b > c where ...
More documentation can be found in the Haskell Wiki.

Jones2000

“Type Classes with Functional Dependencies”, Mark P. Jones, In Proceedings of the 9th European Symposium on Programming, ESOP 2000, Berlin, Germany, March 2000, SpringerVerlag LNCS 1782, .
10.8.2.1. Rules for functional dependencies
In a class declaration, all of the class type variables must be reachable (in the sense mentioned in The context of a type signature) from the free variables of each method type. For example:
class Coll s a where empty :: s insert :: s > a > s
is not OK, because the type of empty
doesn’t mention a
. Functional dependencies can make the type variable reachable:
class Coll s a  s > a where empty :: s insert :: s > a > s
Alternatively Coll
might be rewritten
class Coll s a where empty :: s a insert :: s a > a > s a
which makes the connection between the type of a collection of a
’s (namely (s a)
) and the element type a
. Occasionally this really doesn’t work, in which case you can split the class like this:
class CollE s where empty :: s class CollE s => Coll s a where insert :: s > a > s
10.8.2.2. Background on functional dependencies
The following description of the motivation and use of functional dependencies is taken from the Hugs user manual, reproduced here (with minor changes) by kind permission of Mark Jones.
Consider the following class, intended as part of a library for collection types:
class Collects e ce where empty :: ce insert :: e > ce > ce member :: e > ce > Bool
The type variable e
used here represents the element type, while ce
is the type of the container itself. Within this framework, we might want to define instances of this class for lists or characteristic functions (both of which can be used to represent collections of any equality type), bit sets (which can be used to represent collections of characters), or hash tables (which can be used to represent any collection whose elements have a hash function). Omitting standard implementation details, this would lead to the following declarations:
instance Eq e => Collects e [e] where ... instance Eq e => Collects e (e > Bool) where ... instance Collects Char BitSet where ... instance (Hashable e, Collects a ce) => Collects e (Array Int ce) where ...
All this looks quite promising; we have a class and a range of interesting implementations. Unfortunately, there are some serious problems with the class declaration. First, the empty function has an ambiguous type:
empty :: Collects e ce => ce
By “ambiguous” we mean that there is a type variable e
that appears on the left of the =>
symbol, but not on the right. The problem with this is that, according to the theoretical foundations of Haskell overloading, we cannot guarantee a welldefined semantics for any term with an ambiguous type.
We can sidestep this specific problem by removing the empty member from the class declaration. However, although the remaining members, insert and member, do not have ambiguous types, we still run into problems when we try to use them. For example, consider the following two functions:
f x y = insert x . insert y g = f True 'a'
for which GHC infers the following types:
f :: (Collects a c, Collects b c) => a > b > c > c g :: (Collects Bool c, Collects Char c) => c > c
Notice that the type for f
allows the two parameters x
and y
to be assigned different types, even though it attempts to insert each of the two values, one after the other, into the same collection. If we’re trying to model collections that contain only one type of value, then this is clearly an inaccurate type. Worse still, the definition for g is accepted, without causing a type error. As a result, the error in this code will not be flagged at the point where it appears. Instead, it will show up only when we try to use g
, which might even be in a different module.
10.8.2.2.1. An attempt to use constructor classes
Faced with the problems described above, some Haskell programmers might be tempted to use something like the following version of the class declaration:
class Collects e c where empty :: c e insert :: e > c e > c e member :: e > c e > Bool
The key difference here is that we abstract over the type constructor c
that is used to form the collection type c e
, and not over that collection type itself, represented by ce
in the original class declaration. This avoids the immediate problems that we mentioned above: empty has type Collects e c => c e
, which is not ambiguous.
The function f
from the previous section has a more accurate type:
f :: (Collects e c) => e > e > c e > c e
The function g
from the previous section is now rejected with a type error as we would hope because the type of f
does not allow the two arguments to have different types. This, then, is an example of a multiple parameter class that does actually work quite well in practice, without ambiguity problems. There is, however, a catch. This version of the Collects
class is nowhere near as general as the original class seemed to be: only one of the four instances for Collects
given above can be used with this version of Collects because only one of them—the instance for lists—has a collection type that can be written in the form c
e
, for some type constructor c
, and element type e
.
10.8.2.2.2. Adding functional dependencies
To get a more useful version of the Collects
class, GHC provides a mechanism that allows programmers to specify dependencies between the parameters of a multiple parameter class (For readers with an interest in theoretical foundations and previous work: The use of dependency information can be seen both as a generalisation of the proposal for “parametric type classes” that was put forward by Chen, Hudak, and Odersky, or as a special case of Mark Jones’s later framework for “improvement” of qualified types. The underlying ideas are also discussed in a more theoretical and abstract setting in a manuscript [Jones1999], where they are identified as one point in a general design space for systems of implicit parameterisation). To start with an abstract example, consider a declaration such as:
class C a b where ...

Jones1999

“Exploring the Design Space for Typebased Implicit Parameterization”, Mark P. Jones, Oregon Graduate Institute of Science & Technology, Technical Report, July 1999.
which tells us simply that C
can be thought of as a binary relation on types (or type constructors, depending on the kinds of a
and b
). Extra clauses can be included in the definition of classes to add information about dependencies between parameters, as in the following examples:
class D a b  a > b where ... class E a b  a > b, b > a where ...
The notation a > b
used here between the 
and where
symbols — not to be confused with a function type — indicates that the a
parameter uniquely determines the b
parameter, and might be read as “a
determines b
.” Thus D
is not just a relation, but actually a (partial) function. Similarly, from the two dependencies that are included in the definition of E
, we can see that E
represents a (partial) onetoone mapping between types.
More generally, dependencies take the form x1 ... xn > y1 ... ym
, where x1
, …, xn
, and y1
, …, yn
are type variables with n>0 and m>=0, meaning that the y
parameters are uniquely determined by the x
parameters. Spaces can be used as separators if more than one variable appears on any single side of a dependency, as in t > a b
. Note that a class may be annotated with multiple dependencies using commas as separators, as in the definition of E
above. Some dependencies that we can write in this notation are redundant, and will be rejected because they don’t serve any useful purpose, and may instead indicate an error in the program. Examples of dependencies like this include a > a
, a > a a
, a >
, etc. There can also be some redundancy if multiple dependencies are given, as in a>b
, b>c
, a>c
, and in which some subset implies the remaining dependencies. Examples like this are not treated as errors. Note that dependencies appear only in class declarations, and not in any other part of the language. In particular, the syntax for instance declarations, class constraints, and types is completely unchanged.
By including dependencies in a class declaration, we provide a mechanism for the programmer to specify each multiple parameter class more precisely. The compiler, on the other hand, is responsible for ensuring that the set of instances that are in scope at any given point in the program is consistent with any declared dependencies. For example, the following pair of instance declarations cannot appear together in the same scope because they violate the dependency for D
, even though either one on its own would be acceptable:
instance D Bool Int where ... instance D Bool Char where ...
Note also that the following declaration is not allowed, even by itself:
instance D [a] b where ...
The problem here is that this instance would allow one particular choice of [a]
to be associated with more than one choice for b
, which contradicts the dependency specified in the definition of D
. More generally, this means that, in any instance of the form:
instance D t s where ...
for some particular types t
and s
, the only variables that can appear in s
are the ones that appear in t
, and hence, if the type t
is known, then s
will be uniquely determined.
The benefit of including dependency information is that it allows us to define more general multiple parameter classes, without ambiguity problems, and with the benefit of more accurate types. To illustrate this, we return to the collection class example, and annotate the original definition of Collects
with a simple dependency:
class Collects e ce  ce > e where empty :: ce insert :: e > ce > ce member :: e > ce > Bool
The dependency ce > e
here specifies that the type e
of elements is uniquely determined by the type of the collection ce
. Note that both parameters of Collects are of kind Type
; there are no constructor classes here. Note too that all of the instances of Collects
that we gave earlier can be used together with this new definition.
What about the ambiguity problems that we encountered with the original definition? The empty function still has type Collects e ce => ce
, but it is no longer necessary to regard that as an ambiguous type: Although the variable e
does not appear on the right of the =>
symbol, the dependency for class Collects
tells us that it is uniquely determined by ce
, which does appear on the right of the =>
symbol. Hence the context in which empty is used can still give enough information to determine types for both ce
and e
, without ambiguity. More generally, we need only regard a type as ambiguous if it contains a variable on the left of the =>
that is not uniquely determined (either directly or indirectly) by the variables on the right.
Dependencies also help to produce more accurate types for user defined functions, and hence to provide earlier detection of errors, and less cluttered types for programmers to work with. Recall the previous definition for a function f
:
f x y = insert x y = insert x . insert y
for which we originally obtained a type:
f :: (Collects a c, Collects b c) => a > b > c > c
Given the dependency information that we have for Collects
, however, we can deduce that a
and b
must be equal because they both appear as the second parameter in a Collects
constraint with the same first parameter c
. Hence we can infer a shorter and more accurate type for f
:
f :: (Collects a c) => a > a > c > c
In a similar way, the earlier definition of g
will now be flagged as a type error.
Although we have given only a few examples here, it should be clear that the addition of dependency information can help to make multiple parameter classes more useful in practice, avoiding ambiguity problems, and allowing more general sets of instance declarations.
10.8.3. Instance declarations
An instance declaration has the form
instance ( assertion1, ..., assertionn) => class type1 ... typem where ...
The part before the “=>
” is the context, while the part after the “=>
” is the head of the instance declaration.
10.8.3.1. Instance resolution
When GHC tries to resolve, say, the constraint C Int Bool
, it tries to match every instance declaration against the constraint, by instantiating the head of the instance declaration. Consider these declarations:
instance context1 => C Int a where ...  (A) instance context2 => C a Bool where ...  (B)
GHC’s default behaviour is that exactly one instance must match the constraint it is trying to resolve. For example, the constraint C Int Bool
matches instances (A) and (B), and hence would be rejected; while C Int Char
matches only (A) and hence (A) is chosen.
Notice that
 When matching, GHC takes no account of the context of the instance declaration (
context1
etc).  It is fine for there to be a potential of overlap (by including both declarations (A) and (B), say); an error is only reported if a particular constraint matches more than one.
See also Overlapping instances for flags that loosen the instance resolution rules.
10.8.3.2. Relaxed rules for the instance head

TypeSynonymInstances

 Since

6.8.1
Allow definition of type class instances for type synonyms.

FlexibleInstances

 Implies
 Since

6.8.1
Allow definition of type class instances with arbitrary nested types in the instance head.
In Haskell 98 the head of an instance declaration must be of the form C (T a1 ... an)
, where C
is the class, T
is a data type constructor, and the a1 ... an
are distinct type variables. In the case of multiparameter type classes, this rule applies to each parameter of the instance head (Arguably it should be okay if just one has this form and the others are type variables, but that’s the rules at the moment).
GHC relaxes this rule in two ways:

With the
TypeSynonymInstances
extension, instance heads may use type synonyms. As always, using a type synonym is just shorthand for writing the RHS of the type synonym definition. For example:type Point a = (a,a) instance C (Point a) where ...
is legal. The instance declaration is equivalent to
instance C (a,a) where ...
As always, type synonyms must be fully applied. You cannot, for example, write:
instance Monad Point where ...

The
FlexibleInstances
extension allows the head of the instance declaration to mention arbitrary nested types. For example, this becomes a legal instance declarationinstance C (Maybe Int) where ...
See also the rules on overlap.
The
FlexibleInstances
extension impliesTypeSynonymInstances
.
However, the instance declaration must still conform to the rules for instance termination: see Instance termination rules.
10.8.3.3. Relaxed rules for instance contexts
In Haskell 98, the class constraints in the context of the instance declaration must be of the form C a
where a
is a type variable that occurs in the head.
The FlexibleContexts
extension relaxes this rule, as well as relaxing the corresponding rule for type signatures (see The context of a type signature). Specifically, FlexibleContexts
, allows (wellkinded) class constraints of form (C t1 ... tn)
in the context of an instance declaration.
Notice that the extension does not affect equality constraints in an instance context; they are permitted by TypeFamilies
or GADTs
.
However, the instance declaration must still conform to the rules for instance termination: see Instance termination rules.
10.8.3.4. Instance termination rules

UndecidableInstances

 Since

6.8.1
Permit definition of instances which may lead to typechecker nontermination.
Regardless of FlexibleInstances
and FlexibleContexts
, instance declarations must conform to some rules that ensure that instance resolution will terminate. The restrictions can be lifted with UndecidableInstances
(see Undecidable instances).
The rules are these:

The Paterson Conditions: for each class constraint
(C t1 ... tn)
in the context No type variable has more occurrences in the constraint than in the head
 The constraint has fewer constructors and variables (taken together and counting repetitions) than the head
 The constraint mentions no type functions. A type function application can in principle expand to a type of arbitrary size, and so are rejected out of hand
 The Coverage Condition. For each functional dependency, ⟨tvs⟩_{left}
>
⟨tvs⟩_{right}, of the class, every type variable in S(⟨tvs⟩_{right}) must appear in S(⟨tvs⟩_{left}), where S is the substitution mapping each type variable in the class declaration to the corresponding type in the instance head.
These restrictions ensure that instance resolution terminates: each reduction step makes the problem smaller by at least one constructor. You can find lots of background material about the reason for these restrictions in the paper Understanding functional dependencies via Constraint Handling Rules.
For example, these are okay:
instance C Int [a]  Multiple parameters instance Eq (S [a])  Structured type in head  Repeated type variable in head instance C4 a a => C4 [a] [a] instance Stateful (ST s) (MutVar s)  Head can consist of type variables only instance C a instance (Eq a, Show b) => C2 a b  Nontype variables in context instance Show (s a) => Show (Sized s a) instance C2 Int a => C3 Bool [a] instance C2 Int a => C3 [a] b
But these are not:
 Context assertion no smaller than head instance C a => C a where ...  (C b b) has more occurrences of b than the head instance C b b => Foo [b] where ...
The same restrictions apply to instances generated by deriving
clauses. Thus the following is accepted:
data MinHeap h a = H a (h a) deriving (Show)
because the derived instance
instance (Show a, Show (h a)) => Show (MinHeap h a)
conforms to the above rules.
A useful idiom permitted by the above rules is as follows. If one allows overlapping instance declarations then it’s quite convenient to have a “default instance” declaration that applies if something more specific does not:
instance C a where op = ...  Default
10.8.3.5. Undecidable instances
Sometimes even the termination rules of Instance termination rules are too onerous. So GHC allows you to experiment with more liberal rules: if you use the experimental extension UndecidableInstances
, both the Paterson Conditions and the Coverage Condition (described in Instance termination rules) are lifted. Termination is still ensured by having a fixeddepth recursion stack. If you exceed the stack depth you get a sort of backtrace, and the opportunity to increase the stack depth with freductiondepth=⟨n⟩
. However, if you should exceed the default reduction depth limit, it is probably best just to disable depth checking, with freductiondepth=0
. The exact depth your program requires depends on minutiae of your code, and it may change between minor GHC releases. The safest bet for released code – if you’re sure that it should compile in finite time – is just to disable the check.
For example, sometimes you might want to use the following to get the effect of a “class synonym”:
class (C1 a, C2 a, C3 a) => C a where { } instance (C1 a, C2 a, C3 a) => C a where { }
This allows you to write shorter signatures:
f :: C a => ...
instead of
f :: (C1 a, C2 a, C3 a) => ...
The restrictions on functional dependencies (Functional dependencies) are particularly troublesome. It is tempting to introduce type variables in the context that do not appear in the head, something that is excluded by the normal rules. For example:
class HasConverter a b  a > b where convert :: a > b data Foo a = MkFoo a instance (HasConverter a b,Show b) => Show (Foo a) where show (MkFoo value) = show (convert value)
This is dangerous territory, however. Here, for example, is a program that would make the typechecker loop:
class D a class F a b  a>b instance F [a] [[a]] instance (D c, F a c) => D [a]  'c' is not mentioned in the head
Similarly, it can be tempting to lift the coverage condition:
class Mul a b c  a b > c where (.*.) :: a > b > c instance Mul Int Int Int where (.*.) = (*) instance Mul Int Float Float where x .*. y = fromIntegral x * y instance Mul a b c => Mul a [b] [c] where x .*. v = map (x.*.) v
The third instance declaration does not obey the coverage condition; and indeed the (somewhat strange) definition:
f = \ b x y > if b then x .*. [y] else y
makes instance inference go into a loop, because it requires the constraint (Mul a [b] b)
.
The UndecidableInstances
extension is also used to lift some of the restrictions imposed on type family instances. See Decidability of type synonym instances.
10.8.3.6. Overlapping instances

OverlappingInstances

Deprecated extension to weaken checks intended to ensure instance resolution termination.

IncoherentInstances

 Since

6.8.1
Deprecated extension to weaken checks intended to ensure instance resolution termination.
In general, as discussed in Instance resolution, GHC requires that it be unambiguous which instance declaration should be used to resolve a typeclass constraint. GHC also provides a way to loosen the instance resolution, by allowing more than one instance to match, provided there is a most specific one. Moreover, it can be loosened further, by allowing more than one instance to match irrespective of whether there is a most specific one. This section gives the details.
To control the choice of instance, it is possible to specify the overlap behavior for individual instances with a pragma, written immediately after the instance
keyword. The pragma may be one of: {# OVERLAPPING #}
, {# OVERLAPPABLE #}
, {# OVERLAPS #}
, or {# INCOHERENT #}
.
The matching behaviour is also influenced by two modulelevel language extension flags: OverlappingInstances
and IncoherentInstances
. These extensions are now deprecated (since GHC 7.10) in favour of the finegrained perinstance pragmas.
A more precise specification is as follows. The willingness to be overlapped or incoherent is a property of the instance declaration itself, controlled as follows:
 An instance is incoherent if: it has an
INCOHERENT
pragma; or if the instance has no pragma and it appears in a module compiled withIncoherentInstances
.  An instance is overlappable if: it has an
OVERLAPPABLE
orOVERLAPS
pragma; or if the instance has no pragma and it appears in a module compiled withOverlappingInstances
; or if the instance is incoherent.  An instance is overlapping if: it has an
OVERLAPPING
orOVERLAPS
pragma; or if the instance has no pragma and it appears in a module compiled withOverlappingInstances
; or if the instance is incoherent.
Now suppose that, in some client module, we are searching for an instance of the target constraint (C ty1 .. tyn)
. The search works like this:
 Find all instances \(I\) that match the target constraint; that is, the target constraint is a substitution instance of \(I\). These instance declarations are the candidates.
 If no candidates remain, the search failes

Eliminate any candidate \(IX\) for which there is another candidate \(IY\) such that both of the following hold:
 \(IY\) is strictly more specific than \(IX\). That is, \(IY\) is a substitution instance of \(IX\) but not vice versa.
 Either \(IX\) is overlappable, or \(IY\) is overlapping. (This “either/or” design, rather than a “both/and” design, allow a client to deliberately override an instance from a library, without requiring a change to the library.)
 If all the remaining candidates are incoherent, the search suceeds, returning an arbitrary surviving candidate.
 If more than one nonincoherent candidate remains, the search fails.
 Otherwise there is exactly one nonincoherent candidate; call it the “prime candidate”.
 Now find all instances, or inscope given constraints, that unify with the target constraint, but do not match it. Such noncandidate instances might match when the target constraint is further instantiated. If all of them are incoherent toplevel instances, the search succeeds, returning the prime candidate. Otherwise the search fails.
Notice that these rules are not influenced by flag settings in the client module, where the instances are used. These rules make it possible for a library author to design a library that relies on overlapping instances without the client having to know.
Errors are reported lazily (when attempting to solve a constraint), rather than eagerly (when the instances themselves are defined). Consider, for example
instance C Int b where .. instance C a Bool where ..
These potentially overlap, but GHC will not complain about the instance declarations themselves, regardless of flag settings. If we later try to solve the constraint (C Int Char)
then only the first instance matches, and all is well. Similarly with (C Bool Bool)
. But if we try to solve (C Int Bool)
, both instances match and an error is reported.
As a more substantial example of the rules in action, consider
instance {# OVERLAPPABLE #} context1 => C Int b where ...  (A) instance {# OVERLAPPABLE #} context2 => C a Bool where ...  (B) instance {# OVERLAPPABLE #} context3 => C a [b] where ...  (C) instance {# OVERLAPPING #} context4 => C Int [Int] where ...  (D)
Now suppose that the type inference engine needs to solve the constraint C Int [Int]
. This constraint matches instances (A), (C) and (D), but the last is more specific, and hence is chosen.
If (D) did not exist then (A) and (C) would still be matched, but neither is most specific. In that case, the program would be rejected, unless IncoherentInstances
is enabled, in which case it would be accepted and (A) or (C) would be chosen arbitrarily.
An instance declaration is more specific than another iff the head of former is a substitution instance of the latter. For example (D) is “more specific” than (C) because you can get from (C) to (D) by substituting a := Int
.
The final bullet (about unifiying instances) makes GHC conservative about committing to an overlapping instance. For example:
f :: [b] > [b] f x = ...
Suppose that from the RHS of f
we get the constraint C b [b]
. But GHC does not commit to instance (C), because in a particular call of f
, b
might be instantiated to Int
, in which case instance (D) would be more specific still. So GHC rejects the program.
If, however, you enable the extension IncoherentInstances
when compiling the module that contains (D), GHC will instead pick (C), without complaining about the problem of subsequent instantiations.
Notice that we gave a type signature to f
, so GHC had to check that f
has the specified type. Suppose instead we do not give a type signature, asking GHC to infer it instead. In this case, GHC will refrain from simplifying the constraint C Int [b]
(for the same reason as before) but, rather than rejecting the program, it will infer the type
f :: C b [b] => [b] > [b]
That postpones the question of which instance to pick to the call site for f
by which time more is known about the type b
. You can write this type signature yourself if you use the FlexibleContexts
extension.
Exactly the same situation can arise in instance declarations themselves. Suppose we have
class Foo a where f :: a > a instance Foo [b] where f x = ...
and, as before, the constraint C Int [b]
arises from f
’s right hand side. GHC will reject the instance, complaining as before that it does not know how to resolve the constraint C Int [b]
, because it matches more than one instance declaration. The solution is to postpone the choice by adding the constraint to the context of the instance declaration, thus:
instance C Int [b] => Foo [b] where f x = ...
(You need FlexibleInstances
to do this.)
In the unification check in the final bullet, GHC also uses the “inscope given constraints”. Consider for example
instance C a Int g :: forall b c. C b Int => blah g = ...needs (C c Int)...
Here GHC will not solve the constraint (C c Int)
from the toplevel instance, because a particular call of g
might instantiate both b
and c
to the same type, which would allow the constraint to be solved in a different way. This latter restriction is principally to make the constraintsolver complete. (Interested folk can read Note [Instance and Given overlap]
in TcInteract
.) It is easy to avoid: in a type signature avoid a constraint that matches a toplevel instance. The flag Wsimplifiableclassconstraints
warns about such signatures.
Warning
Overlapping instances must be used with care. They can give rise to incoherence (i.e. different instance choices are made in different parts of the program) even without IncoherentInstances
. Consider:
{# LANGUAGE OverlappingInstances #} module Help where class MyShow a where myshow :: a > String instance MyShow a => MyShow [a] where myshow xs = concatMap myshow xs showHelp :: MyShow a => [a] > String showHelp xs = myshow xs {# LANGUAGE FlexibleInstances, OverlappingInstances #} module Main where import Help data T = MkT instance MyShow T where myshow x = "Used generic instance" instance MyShow [T] where myshow xs = "Used more specific instance" main = do { print (myshow [MkT]); print (showHelp [MkT]) }
In function showHelp
GHC sees no overlapping instances, and so uses the MyShow [a]
instance without complaint. In the call to myshow
in main
, GHC resolves the MyShow [T]
constraint using the overlapping instance declaration in module Main
. As a result, the program prints
"Used more specific instance" "Used generic instance"
(An alternative possible behaviour, not currently implemented, would be to reject module Help
on the grounds that a later instance declaration might overlap the local one.)
10.8.3.7. Instance signatures: type signatures in instance declarations

InstanceSigs

 Since

7.6.1
Allow type signatures for members in instance definitions.
In Haskell, you can’t write a type signature in an instance declaration, but it is sometimes convenient to do so, and the language extension InstanceSigs
allows you to do so. For example:
data T a = MkT a a instance Eq a => Eq (T a) where (==) :: T a > T a > Bool  The signature (==) (MkT x1 x2) (MkTy y1 y2) = x1==y1 && x2==y2
Some details

The type signature in the instance declaration must be more polymorphic than (or the same as) the one in the class declaration, instantiated with the instance type. For example, this is fine:
instance Eq a => Eq (T a) where (==) :: forall b. b > b > Bool (==) x y = True
Here the signature in the instance declaration is more polymorphic than that required by the instantiated class method.
 The code for the method in the instance declaration is typechecked against the type signature supplied in the instance declaration, as you would expect. So if the instance signature is more polymorphic than required, the code must be too.

One stylistic reason for wanting to write a type signature is simple documentation. Another is that you may want to bring scoped type variables into scope. For example:
class C a where foo :: b > a > (a, [b]) instance C a => C (T a) where foo :: forall b. b > T a > (T a, [b]) foo x (T y) = (T y, xs) where xs :: [b] xs = [x,x,x]
Provided that you also specify
ScopedTypeVariables
(Lexically scoped type variables), theforall b
scopes over the definition offoo
, and in particular over the type signature forxs
.
10.8.4. Overloaded string literals

OverloadedStrings

 Since

6.8.1
Enable overloaded string literals (e.g. string literals desugared via the
IsString
class).
GHC supports overloaded string literals. Normally a string literal has type String
, but with overloaded string literals enabled (with OverloadedStrings
) a string literal has type (IsString a) => a
.
This means that the usual string syntax can be used, e.g., for ByteString
, Text
, and other variations of string like types. String literals behave very much like integer literals, i.e., they can be used in both expressions and patterns. If used in a pattern the literal will be replaced by an equality test, in the same way as an integer literal is.
The class IsString
is defined as:
class IsString a where fromString :: String > a
The only predefined instance is the obvious one to make strings work as usual:
instance IsString [Char] where fromString cs = cs
The class IsString
is not in scope by default. If you want to mention it explicitly (for example, to give an instance declaration for it), you can import it from module Data.String
.
Haskell’s defaulting mechanism (Haskell Report, Section 4.3.4) is extended to cover string literals, when OverloadedStrings
is specified. Specifically:
 Each type in a
default
declaration must be an instance ofNum
or ofIsString
.  If no
default
declaration is given, then it is just as if the module contained the declarationdefault( Integer, Double, String)
.  The standard defaulting rule is extended thus: defaulting applies when all the unresolved constraints involve standard classes or
IsString
; and at least one is a numeric class orIsString
.
So, for example, the expression length "foo"
will give rise to an ambiguous use of IsString a0
which, because of the above rules, will default to String
.
A small example:
module Main where import Data.String( IsString(..) ) newtype MyString = MyString String deriving (Eq, Show) instance IsString MyString where fromString = MyString greet :: MyString > MyString greet "hello" = "world" greet other = other main = do print $ greet "hello" print $ greet "fool"
Note that deriving Eq
is necessary for the pattern matching to work since it gets translated into an equality comparison.
10.8.5. Overloaded labels

OverloadedLabels

 Since

8.0.1
Enable use of the
#foo
overloaded label syntax.
GHC supports overloaded labels, a form of identifier whose interpretation may depend both on its type and on its literal text. When the OverloadedLabels
extension is enabled, an overloaded label can be written with a prefix hash, for example #foo
. The type of this expression is IsLabel "foo" a => a
.
The class IsLabel
is defined as:
class IsLabel (x :: Symbol) a where fromLabel :: a
This is rather similar to the class IsString
(see Overloaded string literals), but with an additional type parameter that makes the text of the label available as a typelevel string (see TypeLevel Literals). Note that fromLabel
had an extra Proxy# x
argument in GHC 8.0, but this was removed in GHC 8.2 as a type application (see Visible type application) can be used instead.
There are no predefined instances of this class. It is not in scope by default, but can be brought into scope by importing GHC.OverloadedLabels. Unlike IsString
, there are no special defaulting rules for IsLabel
.
During typechecking, GHC will replace an occurrence of an overloaded label like #foo
with fromLabel @"foo"
. This will have some type alpha
and require the solution of a class constraint IsLabel "foo" alpha
.
The intention is for IsLabel
to be used to support overloaded record fields and perhaps anonymous records. Thus, it may be given instances for base datatypes (in particular (>)
) in the future.
If RebindableSyntax
is enabled, overloaded labels will be desugared using whatever fromLabel
function is in scope, rather than always using GHC.OverloadedLabels.fromLabel
.
When writing an overloaded label, there must be no space between the hash sign and the following identifier. The MagicHash
extension makes use of postfix hash signs; if OverloadedLabels
and MagicHash
are both enabled then x#y
means x# y
, but if only OverloadedLabels
is enabled then it means x #y
. The UnboxedTuples
extension makes (#
a single lexeme, so when UnboxedTuples
is enabled you must write a space between an opening parenthesis and an overloaded label. To avoid confusion, you are strongly encouraged to put a space before the hash when using OverloadedLabels
.
When using OverloadedLabels
(or other extensions that make use of hash signs) in a .hsc
file (see Writing Haskell interfaces to C code: hsc2hs), the hash signs must be doubled (write ##foo
instead of #foo
) to avoid them being treated as hsc2hs
directives.
Here is an extension of the record access example in TypeLevel Literals showing how an overloaded label can be used as a record selector:
{# LANGUAGE DataKinds, KindSignatures, MultiParamTypeClasses, FunctionalDependencies, FlexibleInstances, OverloadedLabels, ScopedTypeVariables #} import GHC.OverloadedLabels (IsLabel(..)) import GHC.TypeLits (Symbol) data Label (l :: Symbol) = Get class Has a l b  a l > b where from :: a > Label l > b data Point = Point Int Int deriving Show instance Has Point "x" Int where from (Point x _) _ = x instance Has Point "y" Int where from (Point _ y) _ = y instance Has a l b => IsLabel l (a > b) where fromLabel x = from x (Get :: Label l) example = #x (Point 1 2)
10.8.6. Overloaded lists

OverloadedLists

 Since

7.8.1
Enable overloaded list syntax (e.g. desugaring of lists via the
IsList
class).
GHC supports overloading of the list notation. Let us recap the notation for constructing lists. In Haskell, the list notation can be used in the following seven ways:
[]  Empty list [x]  x : [] [x,y,z]  x : y : z : [] [x .. ]  enumFrom x [x,y ..]  enumFromThen x y [x .. y]  enumFromTo x y [x,y .. z]  enumFromThenTo x y z
When the OverloadedLists
extension is turned on, the aforementioned seven notations are desugared as follows:
[]  fromListN 0 [] [x]  fromListN 1 (x : []) [x,y,z]  fromListN 3 (x : y : z : []) [x .. ]  fromList (enumFrom x) [x,y ..]  fromList (enumFromThen x y) [x .. y]  fromList (enumFromTo x y) [x,y .. z]  fromList (enumFromThenTo x y z)
This extension allows programmers to use the list notation for construction of structures like: Set
, Map
, IntMap
, Vector
, Text
and Array
. The following code listing gives a few examples:
['0' .. '9'] :: Set Char [1 .. 10] :: Vector Int [("default",0), (k1,v1)] :: Map String Int ['a' .. 'z'] :: Text
List patterns are also overloaded. When the OverloadedLists
extension is turned on, these definitions are desugared as follows
f [] = ...  f (toList > []) = ... g [x,y,z] = ...  g (toList > [x,y,z]) = ...
(Here we are using viewpattern syntax for the translation, see View patterns.)
10.8.6.1. The IsList
class
In the above desugarings, the functions toList
, fromList
and fromListN
are all methods of the IsList
class, which is itself exported from the GHC.Exts
module. The type class is defined as follows:
class IsList l where type Item l fromList :: [Item l] > l toList :: l > [Item l] fromListN :: Int > [Item l] > l fromListN _ = fromList
The IsList
class and its methods are intended to be used in conjunction with the OverloadedLists
extension.
 The type function
Item
returns the type of items of the structurel
.  The function
fromList
constructs the structurel
from the given list ofItem l
.  The function
fromListN
takes the input list’s length as a hint. Its behaviour should be equivalent tofromList
. The hint can be used for more efficient construction of the structurel
compared tofromList
. If the given hint is not equal to the input list’s length the behaviour offromListN
is not specified.  The function
toList
should be the inverse offromList
.
It is perfectly fine to declare new instances of IsList
, so that list notation becomes useful for completely new data types. Here are several example instances:
instance IsList [a] where type Item [a] = a fromList = id toList = id instance (Ord a) => IsList (Set a) where type Item (Set a) = a fromList = Set.fromList toList = Set.toList instance (Ord k) => IsList (Map k v) where type Item (Map k v) = (k,v) fromList = Map.fromList toList = Map.toList instance IsList (IntMap v) where type Item (IntMap v) = (Int,v) fromList = IntMap.fromList toList = IntMap.toList instance IsList Text where type Item Text = Char fromList = Text.pack toList = Text.unpack instance IsList (Vector a) where type Item (Vector a) = a fromList = Vector.fromList fromListN = Vector.fromListN toList = Vector.toList
10.8.6.2. Rebindable syntax
When desugaring list notation with OverloadedLists
GHC uses the fromList
(etc) methods from module GHC.Exts
. You do not need to import GHC.Exts
for this to happen.
However if you use RebindableSyntax
, then GHC instead uses whatever is in scope with the names of toList
, fromList
and fromListN
. That is, these functions are rebindable; c.f. Rebindable syntax and the implicit Prelude import.
10.8.6.3. Defaulting
Currently, the IsList
class is not accompanied with defaulting rules. Although feasible, not much thought has gone into how to specify the meaning of the default declarations like:
default ([a])
10.8.6.4. Speculation about the future
The current implementation of the OverloadedLists
extension can be improved by handling the lists that are only populated with literals in a special way. More specifically, the compiler could allocate such lists statically using a compact representation and allow IsList
instances to take advantage of the compact representation. Equipped with this capability the OverloadedLists
extension will be in a good position to subsume the OverloadedStrings
extension (currently, as a special case, string literals benefit from statically allocated compact representation).
10.8.7. Undecidable (or recursive) superclasses

UndecidableSuperClasses

 Since

8.0.1
Allow all superclass constraints, including those that may result in nontermination of the typechecker.
The language extension UndecidableSuperClasses
allows much more flexible constraints in superclasses.
A class cannot generally have itself as a superclass. So this is illegal
class C a => D a where ... class D a => C a where ...
GHC implements this test conservatively when type functions, or type variables, are involved. For example
type family F a :: Constraint class F a => C a where ...
GHC will complain about this, because you might later add
type instance F Int = C Int
and now we’d be in a superclass loop. Here’s an example involving a type variable
class f (C f) => C f class c => Id c
If we expanded the superclasses of C Id
we’d get first Id (C Id)
and thence C Id
again.
But superclass constraints like these are sometimes useful, and the conservative check is annoying where no actual recursion is involved.
Moreover genuninelyrecursive superclasses are sometimes useful. Here’s a reallife example (#10318)
class (Frac (Frac a) ~ Frac a, Fractional (Frac a), IntegralDomain (Frac a)) => IntegralDomain a where type Frac a :: Type
Here the superclass cycle does terminate but it’s not entirely straightforward to see that it does.
With the language extension UndecidableSuperClasses
GHC lifts all restrictions on superclass constraints. If there really is a loop, GHC will only expand it to finite depth.
10.9. Type families

TypeFamilies

 Implies
 Since

6.8.1
Allow use and definition of indexed type and data families.
Indexed type families form an extension to facilitate typelevel programming. Type families are a generalisation of associated data types [AssocDataTypes2005] and associated type synonyms [AssocTypeSyn2005] Type families themselves are described in Schrijvers 2008 [TypeFamilies2008]. Type families essentially provide typeindexed data types and named functions on types, which are useful for generic programming and highly parameterised library interfaces as well as interfaces with enhanced static information, much like dependent types. They might also be regarded as an alternative to functional dependencies, but provide a more functional style of typelevel programming than the relational style of functional dependencies.
Indexed type families, or type families for short, are type constructors that represent sets of types. Set members are denoted by supplying the type family constructor with type parameters, which are called type indices. The difference between vanilla parametrised type constructors and family constructors is much like between parametrically polymorphic functions and (adhoc polymorphic) methods of type classes. Parametric polymorphic functions behave the same at all type instances, whereas class methods can change their behaviour in dependence on the class type parameters. Similarly, vanilla type constructors imply the same data representation for all type instances, but family constructors can have varying representation types for varying type indices.
Indexed type families come in three flavours: data families, open type synonym families, and closed type synonym families. They are the indexed family variants of algebraic data types and type synonyms, respectively. The instances of data families can be data types and newtypes.
Type families are enabled by the language extension TypeFamilies
. Additional information on the use of type families in GHC is available on the Haskell wiki page on type families.

AssocDataTypes2005

“Associated Types with Class”, M. Chakravarty, G. Keller, S. Peyton Jones, and S. Marlow. In Proceedings of “The 32nd Annual ACM SIGPLANSIGACT Symposium on Principles of Programming Languages (POPL’05)”, pages 113, ACM Press, 2005.

AssocTypeSyn2005

“Type Associated Type Synonyms”. M. Chakravarty, G. Keller, and S. Peyton Jones. In Proceedings of “The Tenth ACM SIGPLAN International Conference on Functional Programming”, ACM Press, pages 241253, 2005.

TypeFamilies2008

“Type Checking with Open Type Functions”, T. Schrijvers, S. PeytonJones, M. Chakravarty, and M. Sulzmann, in Proceedings of “ICFP 2008: The 13th ACM SIGPLAN International Conference on Functional Programming”, ACM Press, pages 5162, 2008.
10.9.1. Data families
Data families appear in two flavours: (1) they can be defined on the toplevel or (2) they can appear inside type classes (in which case they are known as associated types). The former is the more general variant, as it lacks the requirement for the typeindexes to coincide with the class parameters. However, the latter can lead to more clearly structured code and compiler warnings if some type instances were  possibly accidentally  omitted. In the following, we always discuss the general toplevel form first and then cover the additional constraints placed on associated types.
10.9.1.1. Data family declarations
Indexed data families are introduced by a signature, such as
data family GMap k :: Type > Type
The special family
distinguishes family from standard data declarations. The result kind annotation is optional and, as usual, defaults to Type
if omitted. An example is
data family Array e
Named arguments can also be given explicit kind signatures if needed. Just as with GADT declarations named arguments are entirely optional, so that we can declare Array
alternatively with
data family Array :: Type > Type
Unlike with ordinary data definitions, the result kind of a data family does not need to be Type
. It can alternatively be:

Of the form
TYPE r
for somer
(see Levity polymorphism). For example:data family DF1 :: TYPE IntRep data family DF2 (r :: RuntimeRep) :: TYPE r data family DF3 :: Type > TYPE WordRep

A bare kind variable (with
PolyKinds
enabled). For example:data family DF4 :: k data family DF5 (a :: k) :: k data family DF6 :: (k > Type) > k
Data instances’ kinds must end in Type
, however. This restriction is slightly relaxed when the UnliftedNewtypes
extension is enabled, as it permits a newtype instance
’s kind to end in TYPE r
for some r
.
10.9.1.2. Data instance declarations
Instance declarations of data and newtype families are very similar to standard data and newtype declarations. The only two differences are that the keyword data
or newtype
is followed by instance
and that some or all of the type arguments can be nonvariable types, but may not contain forall types or type synonym families. However, data families are generally allowed in type parameters, and type synonyms are allowed as long as they are fully applied and expand to a type that is itself admissible  exactly as this is required for occurrences of type synonyms in class instance parameters. For example, the Either
instance for GMap
is
data instance GMap (Either a b) v = GMapEither (GMap a v) (GMap b v)
In this example, the declaration has only one variant. In general, it can be any number.
When ExplicitForAll
is enabled, type and kind variables used on the left hand side can be explicitly bound. For example:
data instance forall a (b :: Proxy a). F (Proxy b) = FProxy Bool
When an explicit forall
is present, all type and kind variables mentioned which are not already in scope must be bound by the forall
:
data instance forall (a :: k). F a = FOtherwise – rejected: k not in scope data instance forall k (a :: k). F a = FOtherwise – accepted
When the flag Wunusedtypepatterns
is enabled, type variables that are mentioned in the patterns on the left hand side, but not used on the right hand side are reported. Variables that occur multiple times on the left hand side are also considered used. To suppress the warnings, unused variables should be either replaced or prefixed with underscores. Type variables starting with an underscore (_x
) are otherwise treated as ordinary type variables.
This resembles the wildcards that can be used in Partial Type Signatures. However, there are some differences. No error messages reporting the inferred types are generated, nor does the extension PartialTypeSignatures
have any effect.
A type or kind variable explicitly bound using ExplicitForAll
but not used on the left hand side will generate an error, not a warning.
Data and newtype instance declarations are only permitted when an appropriate family declaration is in scope  just as a class instance declaration requires the class declaration to be visible. Moreover, each instance declaration has to conform to the kind determined by its family declaration. This implies that the number of parameters of an instance declaration matches the arity determined by the kind of the family.
A data family instance declaration can use the full expressiveness of ordinary data
or newtype
declarations:

Although, a data family is introduced with the keyword “
data
”, a data family instance can use eitherdata
ornewtype
. For example:data family T a data instance T Int = T1 Int  T2 Bool newtype instance T Char = TC Bool

A
data instance
can use GADT syntax for the data constructors, and indeed can define a GADT. For example:data family G a b data instance G [a] b where G1 :: c > G [Int] b G2 :: G [a] Bool
 You can use a
deriving
clause on adata instance
ornewtype instance
declaration.
Even if data families are defined as toplevel declarations, functions that perform different computations for different family instances may still need to be defined as methods of type classes. In particular, the following is not possible:
data family T a data instance T Int = A data instance T Char = B foo :: T a > Int foo A = 1 foo B = 2
Instead, you would have to write foo
as a class operation, thus:
class Foo a where foo :: T a > Int instance Foo Int where foo A = 1 instance Foo Char where foo B = 2
Given the functionality provided by GADTs (Generalised Algebraic Data Types), it might seem as if a definition, such as the above, should be feasible. However, type families  in contrast to GADTs  are open; i.e., new instances can always be added, possibly in other modules. Supporting pattern matching across different data instances would require a form of extensible case construct.
10.9.1.3. Overlap of data instances
The instance declarations of a data family used in a single program may not overlap at all, independent of whether they are associated or not. In contrast to type class instances, this is not only a matter of consistency, but one of type safety.
10.9.2. Synonym families
Type families appear in three flavours: (1) they can be defined as open families on the toplevel, (2) they can be defined as closed families on the toplevel, or (3) they can appear inside type classes (in which case they are known as associated type synonyms). Toplevel families are more general, as they lack the requirement for the typeindexes to coincide with the class parameters. However, associated type synonyms can lead to more clearly structured code and compiler warnings if some type instances were  possibly accidentally  omitted. In the following, we always discuss the general toplevel forms first and then cover the additional constraints placed on associated types. Note that closed associated type synonyms do not exist.
10.9.2.1. Type family declarations
Open indexed type families are introduced by a signature, such as
type family Elem c :: Type
The special family
distinguishes family from standard type declarations. The result kind annotation is optional and, as usual, defaults to Type
if omitted. An example is
type family Elem c
Parameters can also be given explicit kind signatures if needed. We call the number of parameters in a type family declaration, the family’s arity, and all applications of a type family must be fully saturated with respect to that arity. This requirement is unlike ordinary type synonyms and it implies that the kind of a type family is not sufficient to determine a family’s arity, and hence in general, also insufficient to determine whether a type family application is well formed. As an example, consider the following declaration:
type family F a b :: Type > Type  F's arity is 2,  although its overall kind is Type > Type > Type > Type
Given this declaration the following are examples of wellformed and malformed types:
F Char [Int]  OK! Kind: Type > Type F Char [Int] Bool  OK! Kind: Type F IO Bool  WRONG: kind mismatch in the first argument F Bool  WRONG: unsaturated application
The result kind annotation is optional and defaults to Type
(like argument kinds) if omitted. Polykinded type families can be declared using a parameter in the kind annotation:
type family F a :: k
In this case the kind parameter k
is actually an implicit parameter of the type family.
At definition site, the arity determines what inputs can be matched on:
data PT (a :: Type) type family F1 :: k > Type type instance F1 = PT  OK, 'k' can be matched on. type family F0 :: forall k. k > Type type instance F0 = PT  Error:  • Expected kind ‘forall k. k > Type’,  but ‘PT’ has kind ‘Type > Type’  • In the type ‘PT’  In the type instance declaration for ‘F0’
Both F1
and F0
have kind forall k. k > Type
, but their arity differs.
At use sites, the arity determines if the definition can be used in a higherrank scenario:
type HRK (f :: forall k. k > Type) = (f Int, f Maybe, f True) type H1 = HRK F0  OK type H2 = HRK F1  Error:  • Expected kind ‘forall k. k > Type’,  but ‘F1’ has kind ‘k0 > Type’  • In the first argument of ‘HRK’, namely ‘F1’  In the type ‘HRK F1’  In the type declaration for ‘H2’
This is a consequence of the requirement that all applications of a type family must be fully saturated with respect to their arity.
10.9.2.2. Type instance declarations
Instance declarations of type families are very similar to standard type synonym declarations. The only two differences are that the keyword type
is followed by instance
and that some or all of the type arguments can be nonvariable types, but may not contain forall types or type synonym families. However, data families are generally allowed, and type synonyms are allowed as long as they are fully applied and expand to a type that is admissible  these are the exact same requirements as for data instances. For example, the [e]
instance for Elem
is
type instance Elem [e] = e
Type arguments can be replaced with underscores (_
) if the names of the arguments don’t matter. This is the same as writing type variables with unique names. Unused type arguments can be replaced or prefixed with underscores to avoid warnings when the Wunusedtypepatterns
flag is enabled. The same rules apply as for Data instance declarations.
Also in the same way as Data instance declarations, when ExplicitForAll
is enabled, type and kind variables can be explicilty bound in a type instance declaration.
Type family instance declarations are only legitimate when an appropriate family declaration is in scope  just like class instances require the class declaration to be visible. Moreover, each instance declaration has to conform to the kind determined by its family declaration, and the number of type parameters in an instance declaration must match the number of type parameters in the family declaration. Finally, the righthand side of a type instance must be a monotype (i.e., it may not include foralls) and after the expansion of all saturated vanilla type synonyms, no synonyms, except family synonyms may remain.
10.9.2.3. Closed type families
A type family can also be declared with a where
clause, defining the full set of equations for that family. For example:
type family F a where F Int = Double F Bool = Char F a = String
A closed type family’s equations are tried in order, from top to bottom, when simplifying a type family application. In this example, we declare an instance for F
such that F Int
simplifies to Double
, F Bool
simplifies to Char
, and for any other type a
that is known not to be Int
or Bool
, F a
simplifies to String
. Note that GHC must be sure that a
cannot unify with Int
or Bool
in that last case; if a programmer specifies just F a
in their code, GHC will not be able to simplify the type. After all, a
might later be instantiated with Int
.
A closed type family’s equations have the same restrictions and extensions as the equations for open type family instances. For instance, when ExplicitForAll
is enabled, type or kind variables used on the left hand side of an equation can be explicitly bound, such as in:
type family R a where forall t a. R (t a) = [a] forall a. R a = a
A closed type family may be declared with no equations. Such closed type families are opaque typelevel definitions that will never reduce, are not necessarily injective (unlike empty data types), and cannot be given any instances. This is different from omitting the equations of a closed type family in a hsboot
file, which uses the syntax where ..
, as in that case there may or may not be equations given in the hs
file.
10.9.2.4. Type family examples
Here are some examples of admissible and illegal type instances:
type family F a :: Type type instance F [Int] = Int  OK! type instance F String = Char  OK! type instance F (F a) = a  WRONG: type parameter mentions a type family type instance F (forall a. (a, b)) = b  WRONG: a forall type appears in a type parameter type instance F Float = forall a.a  WRONG: righthand side may not be a forall type type family H a where  OK! H Int = Int H Bool = Bool H a = String type instance H Char = Char  WRONG: cannot have instances of closed family type family K a where  OK! type family G a b :: Type > Type type instance G Int = (,)  WRONG: must be two type parameters type instance G Int Char Float = Double  WRONG: must be two type parameters
10.9.2.5. Compatibility and apartness of type family equations
There must be some restrictions on the equations of type families, lest we define an ambiguous rewrite system. So, equations of open type families are restricted to be compatible. Two type patterns are compatible if
 all corresponding types and implicit kinds in the patterns are apart, or
 the two patterns unify producing a substitution, and the righthand sides are equal under that substitution.
Two types are considered apart if, for all possible substitutions, the types cannot reduce to a common reduct.
The first clause of “compatible” is the more straightforward one. It says that the patterns of two distinct type family instances cannot overlap. For example, the following is disallowed:
type instance F Int = Bool type instance F Int = Char
The second clause is a little more interesting. It says that two overlapping type family instances are allowed if the righthand sides coincide in the region of overlap. Some examples help here:
type instance F (a, Int) = [a] type instance F (Int, b) = [b]  overlap permitted type instance G (a, Int) = [a] type instance G (Char, a) = [a]  ILLEGAL overlap, as [Char] /= [Int]
Note that this compatibility condition is independent of whether the type family is associated or not, and it is not only a matter of consistency, but one of type safety.
For a polykinded type family, the kinds are checked for apartness just like types. For example, the following is accepted:
type family J a :: k type instance J Int = Bool type instance J Int = Maybe
These instances are compatible because they differ in their implicit kind parameter; the first uses Type
while the second uses Type > Type
.
The definition for “compatible” uses a notion of “apart”, whose definition in turn relies on type family reduction. This condition of “apartness”, as stated, is impossible to check, so we use this conservative approximation: two types are considered to be apart when the two types cannot be unified, even by a potentially infinite unifier. Allowing the unifier to be infinite disallows the following pair of instances:
type instance H x x = Int type instance H [x] x = Bool
The type patterns in this pair equal if x
is replaced by an infinite nesting of lists. Rejecting instances such as these is necessary for type soundness.
Compatibility also affects closed type families. When simplifying an application of a closed type family, GHC will select an equation only when it is sure that no incompatible previous equation will ever apply. Here are some examples:
type family F a where F Int = Bool F a = Char type family G a where G Int = Int G a = a
In the definition for F
, the two equations are incompatible – their patterns are not apart, and yet their righthand sides do not coincide. Thus, before GHC selects the second equation, it must be sure that the first can never apply. So, the type F a
does not simplify; only a type such as F Double
will simplify to Char
. In G
, on the other hand, the two equations are compatible. Thus, GHC can ignore the first equation when looking at the second. So, G a
will simplify to a
.
Incompatibilities between closed type family equations can be displayed in :info
when fprintaxiomincomps
is enabled.
However see Type, class and other declarations for the overlap rules in GHCi.
10.9.2.6. Decidability of type synonym instances

UndecidableInstances

Relax restrictions on the decidability of type synonym family instances.
In order to guarantee that type inference in the presence of type families decidable, we need to place a number of additional restrictions on the formation of type instance declarations (c.f., Definition 5 (Relaxed Conditions) of “Type Checking with Open Type Functions”). Instance declarations have the general form
type instance F t1 .. tn = t
where we require that for every type family application (G s1 .. sm)
in t
,

s1 .. sm
do not contain any type family constructors,  the total number of symbols (data type constructors and type variables) in
s1 .. sm
is strictly smaller than int1 .. tn
, and  for every type variable
a
,a
occurs ins1 .. sm
at most as often as int1 .. tn
.
These restrictions are easily verified and ensure termination of type inference. However, they are not sufficient to guarantee completeness of type inference in the presence of, so called, ‘’loopy equalities’’, such as a ~ [F a]
, where a recursive occurrence of a type variable is underneath a family application and data constructor application  see the above mentioned paper for details.
If the option UndecidableInstances
is passed to the compiler, the above restrictions are not enforced and it is on the programmer to ensure termination of the normalisation of type families during type inference.
10.9.3. Wildcards on the LHS of data and type family instances
When the name of a type argument of a data or type instance declaration doesn’t matter, it can be replaced with an underscore (_
). This is the same as writing a type variable with a unique name.
data family F a b :: Type data instance F Int _ = Int  Equivalent to data instance F Int b = Int type family T a :: Type type instance T (a,_) = a  Equivalent to type instance T (a,b) = a
This use of underscore for wildcard in a type pattern is exactly like pattern matching in the term language, but is rather different to the use of a underscore in a partial type signature (see Type Wildcards).
A type variable beginning with an underscore is not treated specially in a type or data instance declaration. For example:
data instance F Bool _a = _a > Int  Equivalent to data instance F Bool a = a > Int
Contrast this with the special treatment of named wildcards in type signatures (Named Wildcards).
10.9.4. Associated data and type families
A data or type synonym family can be declared as part of a type class, thus:
class GMapKey k where data GMap k :: Type > Type ... class Collects ce where type Elem ce :: Type ...
When doing so, we (optionally) may drop the “family
” keyword.
The type parameters must all be type variables, of course, and some (but not necessarily all) of then can be the class parameters. Each class parameter may only be used at most once per associated type, but some may be omitted and they may be in an order other than in the class head. Hence, the following contrived example is admissible:
class C a b c where type T c a x :: Type
Here c
and a
are class parameters, but the type is also indexed on a third parameter x
.
10.9.4.1. Associated instances
When an associated data or type synonym family instance is declared within a type class instance, we (optionally) may drop the instance
keyword in the family instance:
instance (GMapKey a, GMapKey b) => GMapKey (Either a b) where data GMap (Either a b) v = GMapEither (GMap a v) (GMap b v) ... instance Eq (Elem [e]) => Collects [e] where type Elem [e] = e ...
The data or type family instance for an associated type must follow the rule that the type indexes corresponding to class parameters must have precisely the same as type given in the instance head. For example:
class Collects ce where type Elem ce :: Type instance Eq (Elem [e]) => Collects [e] where  Choose one of the following alternatives: type Elem [e] = e  OK type Elem [x] = x  BAD; '[x]' is different to '[e]' from head type Elem x = x  BAD; 'x' is different to '[e]' type Elem [Maybe x] = x  BAD: '[Maybe x]' is different to '[e]'
Note the following points:
 An instance for an associated family can only appear as part of an instance declarations of the class in which the family was declared, just as with the equations of the methods of a class.

The type variables on the right hand side of the type family equation must, as usual, be explicitly bound by the left hand side. This restriction is relaxed for kind variables, however, as the right hand side is allowed to mention kind variables that are implicitly bound. For example, these are legitimate:
data family Nat :: k > k > Type  k is implicitly bound by an invisible kind pattern newtype instance Nat :: (k > Type) > (k > Type) > Type where Nat :: (forall xx. f xx > g xx) > Nat f g class Funct f where type Codomain f :: Type instance Funct ('KProxy :: KProxy o) where  o is implicitly bound by the kind signature  of the LHS type pattern ('KProxy) type Codomain 'KProxy = NatTr (Proxy :: o > Type)
 The instance for an associated type can be omitted in class instances. In that case, unless there is a default instance (see Associated type synonym defaults), the corresponding instance type is not inhabited; i.e., only diverging expressions, such as
undefined
, can assume the type. 
Although it is unusual, there (currently) can be multiple instances for an associated family in a single instance declaration. For example, this is legitimate:
instance GMapKey Flob where data GMap Flob [v] = G1 v data GMap Flob Int = G2 Int ...
Here we give two data instance declarations, one in which the last parameter is
[v]
, and one for which it isInt
. Since you cannot give any subsequent instances for(GMap Flob ...)
, this facility is most useful when the free indexed parameter is of a kind with a finite number of alternatives (unlikeType
). 
When
ExplicitForAll
is enabled, type and kind variables can be explicily bound in associated data or type family instances in the same way (and with the same restrictions) as Data instance declarations or Type instance declarations. For example, adapting the above, the following is accepted:instance Eq (Elem [e]) => Collects [e] where type forall e. Elem [e] = e
10.9.4.2. Associated type synonym defaults
It is possible for the class defining the associated type to specify a default for associated type instances. So for example, this is OK:
class IsBoolMap v where type Key v type instance Key v = Int lookupKey :: Key v > v > Maybe Bool instance IsBoolMap [(Int, Bool)] where lookupKey = lookup
In an instance
declaration for the class, if no explicit type instance
declaration is given for the associated type, the default declaration is used instead, just as with default class methods.
Note the following points:
 The
instance
keyword is optional.  There can be at most one default declaration for an associated type synonym.
 A default declaration is not permitted for an associated data type.

The default declaration must mention only type variables on the left hand side, and type variables may not be repeated on the lefthand side. The right hand side must mention only type variables that are explicitly bound on the left hand side. This restriction is relaxed for kind variables, however, as the right hand side is allowed to mention kind variables that are implicitly bound on the left hand side.
Like with Associated instances, it is possible to explicitly bind type and kind variables in default declarations with a
forall
by using theExplicitForAll
language extension.  Unlike the associated type family declaration itself, the type variables of the default instance are independent of those of the parent class.
Here are some examples:
class C (a :: Type) where type F1 a :: Type type instance F1 a = [a]  OK type instance F1 a = a>a  BAD; only one default instance is allowed type F2 b a  OK; note the family has more type  variables than the class type instance F2 c d = c>d  OK; you don't have to use 'a' in the type instance type F3 a type F3 [b] = b  BAD; only type variables allowed on the LHS, and the argument to F3 is instantiated to [b], which is not a bare type variable type F4 x y type F4 x x = x  BAD; the type variable x is repeated on the LHS type F5 a type F5 b = a  BAD; 'a' is not in scope in the RHS type F6 a :: [k] type F6 a = ('[] :: [x])  OK; the kind variable x is implicitly bound by an invisible kind pattern on the LHS type F7 a type F7 a = Proxy ('[] :: [x])  BAD; the kind variable x is not bound, even by an invisible kind pattern type F8 (x :: a) :: [a] type F8 x = ('[] :: [a])  OK; the kind variable a is implicitly bound by the kind signature of the LHS type pattern type F9 (a :: k) type F9 a = Maybe a  BAD; the kind variable k is instantiated to Type, which is not a bare kind variable type F10 (a :: j) (b :: k) type F10 (a :: z) (b :: z) = Proxy a  BAD; the kind variable z is repeated,  as both j and k are instantiated to z type F11 a b type forall a b. F11 a b = a  OK; LHS type variables can be explicitly bound with 'forall' type F12 (a :: k) type F12 @k a = Proxy a  OK; visible kind application syntax is permitted in default declarations
10.9.4.3. Scoping of class parameters
The visibility of class parameters in the righthand side of associated family instances depends solely on the parameters of the family. As an example, consider the simple class declaration
class C a b where data T a
Only one of the two class parameters is a parameter to the data family. Hence, the following instance declaration is invalid:
instance C [c] d where data T [c] = MkT (c, d)  WRONG!! 'd' is not in scope
Here, the righthand side of the data instance mentions the type variable d
that does not occur in its lefthand side. We cannot admit such data instances as they would compromise type safety.
Bear in mind that it is also possible for the righthand side of an associated family instance to contain kind parameters (by using the PolyKinds
extension). For instance, this class and instance are perfectly admissible:
class C k where type T :: k instance C (Maybe a) where type T = (Nothing :: Maybe a)
Here, although the righthand side (Nothing :: Maybe a)
mentions a kind variable a
which does not occur on the lefthand side, this is acceptable, because a
is implicitly bound by T
’s kind pattern.
A kind variable can also be bound implicitly in a LHS type pattern, as in this example:
class C a where type T (x :: a) :: [a] instance C (Maybe a) where type T x = ('[] :: [Maybe a])
In ('[] :: [Maybe a])
, the kind variable a
is implicitly bound by the kind signature of the LHS type pattern x
.
10.9.4.4. Instance contexts and associated type and data instances
Associated type and data instance declarations do not inherit any context specified on the enclosing instance. For type instance declarations, it is unclear what the context would mean. For data instance declarations, it is unlikely a user would want the context repeated for every data constructor. The only place where the context might likely be useful is in a deriving
clause of an associated data instance. However, even here, the role of the outer instance context is murky. So, for clarity, we just stick to the rule above: the enclosing instance context is ignored. If you need to use a nontrivial context on a derived instance, use a standalone deriving
clause (at the top level).
10.9.5. Import and export
The rules for export lists (Haskell Report Section 5.2) needs adjustment for type families:
 The form
T(..)
, whereT
is a data family, names the familyT
and all the inscope constructors (whether in scope qualified or unqualified) that are data instances ofT
.  The form
T(.., ci, .., fj, ..)
, whereT
is a data family, namesT
and the specified constructorsci
and fieldsfj
as usual. The constructors and field names must belong to some data instance ofT
, but are not required to belong to the same instance.  The form
C(..)
, whereC
is a class, names the classC
and all its methods and associated types.  The form
C(.., mi, .., type Tj, ..)
, whereC
is a class, names the classC
, and the specified methodsmi
and associated typesTj
. The types need a keyword “type
” to distinguish them from data constructors.  Whenever there is no export list and a data instance is defined, the corresponding data family type constructor is exported along with the new data constructors, regardless of whether the data family is defined locally or in another module.
10.9.5.1. Examples
Recall our running GMapKey
class example:
class GMapKey k where data GMap k :: Type > Type insert :: GMap k v > k > v > GMap k v lookup :: GMap k v > k > Maybe v empty :: GMap k v instance (GMapKey a, GMapKey b) => GMapKey (Either a b) where data GMap (Either a b) v = GMapEither (GMap a v) (GMap b v) ...method declarations...
Here are some export lists and their meaning:

module GMap( GMapKey )
Exports just the class name.

module GMap( GMapKey(..) )
Exports the class, the associated type
GMap
and the member functionsempty
,lookup
, andinsert
. The data constructors ofGMap
(in this caseGMapEither
) are not exported. 
module GMap( GMapKey( type GMap, empty, lookup, insert ) )
Same as the previous item. Note the “
type
” keyword. 
module GMap( GMapKey(..), GMap(..) )
Same as previous item, but also exports all the data constructors for
GMap
, namelyGMapEither
. 
module GMap ( GMapKey( empty, lookup, insert), GMap(..) )
Same as previous item.

module GMap ( GMapKey, empty, lookup, insert, GMap(..) )
Same as previous item.
Two things to watch out for:
 You cannot write
GMapKey(type GMap(..))
— i.e., subcomponent specifications cannot be nested. To specifyGMap
’s data constructors, you have to list it separately. 
Consider this example:
module X where data family D module Y where import X data instance D Int = D1  D2
Module
Y
exports all the entities defined inY
, namely the data constructorsD1
andD2
, and implicitly the data familyD
, even though it’s defined inX
. This means you can writeimport Y( D(D1,D2) )
without giving an explicit export list like this:module Y( D(..) ) where ... or module Y( module Y, D ) where ...
10.9.5.2. Instances
Family instances are implicitly exported, just like class instances. However, this applies only to the heads of instances, not to the data constructors an instance defines.
10.9.6. Type families and instance declarations
Type families require us to extend the rules for the form of instance heads, which are given in Relaxed rules for the instance head. Specifically:
 Data type families may appear in an instance head
 Type synonym families may not appear (at all) in an instance head
The reason for the latter restriction is that there is no way to check for instance matching. Consider
type family F a type instance F Bool = Int class C a instance C Int instance C (F a)
Now a constraint (C (F Bool))
would match both instances. The situation is especially bad because the type instance for F Bool
might be in another module, or even in a module that is not yet written.
However, type class instances of instances of data families can be defined much like any other data type. For example, we can say
data instance T Int = T1 Int  T2 Bool instance Eq (T Int) where (T1 i) == (T1 j) = i==j (T2 i) == (T2 j) = i==j _ == _ = False
Note that class instances are always for particular instances of a data family and never for an entire family as a whole. This is for essentially the same reasons that we cannot define a toplevel function that performs pattern matching on the data constructors of different instances of a single type family. It would require a form of extensible case construct.
Data instance declarations can also have deriving
clauses. For example, we can write
data GMap () v = GMapUnit (Maybe v) deriving Show
which implicitly defines an instance of the form
instance Show v => Show (GMap () v) where ...
10.9.7. Injective type families

TypeFamilyDependencies

 Implies
 Since

8.0.1
Allow functional dependency annotations on type families. This allows one to define injective type families.
Starting with GHC 8.0 type families can be annotated with injectivity information. This information is then used by GHC during type checking to resolve type ambiguities in situations where a type variable appears only under type family applications. Consider this contrived example:
type family Id a type instance Id Int = Int type instance Id Bool = Bool id :: Id t > Id t id x = x
Here the definition of id
will be rejected because type variable t
appears only under type family applications and is thus ambiguous. But this code will be accepted if we tell GHC that Id
is injective, which means it will be possible to infer t
at call sites from the type of the argument:
type family Id a = r  r > a
Injective type families are enabled with XTypeFamilyDependencies
language extension. This extension implies XTypeFamilies
.
For full details on injective type families refer to Haskell Symposium 2015 paper Injective type families for Haskell.
10.9.7.1. Syntax of injectivity annotation
The injectivity annotation is added after the type family head and consists of two parts:
 a type variable that names the result of a type family. Syntax:
= tyvar
or= (tyvar :: kind)
. The type variable must be fresh.  an injectivity annotation of the form
 A > B
, whereA
is the result type variable (see previous bullet) andB
is a list of argument type and kind variables in which type family is injective. It is possible to omit some variables if the type family is not injective in them.
Examples:
type family Id a = result  result > a where type family F a b c = d  d > a c b type family G (a :: k) b c = foo  foo > k b where
For open and closed type families it is OK to name the result but skip the injectivity annotation. This is not the case for associated type synonyms, where the named result without injectivity annotation will be interpreted as associated type synonym default.
10.9.7.2. Verifying the injectivity annotation against type family equations
Once the user declares type family to be injective GHC must verify that this declaration is correct, i.e., that type family equations don’t violate the injectivity annotation. A general idea is that if at least one equation (bullets (1), (2) and (3) below) or a pair of equations (bullets (4) and (5) below) violates the injectivity annotation then a type family is not injective in a way the user claims and an error is reported. In the bullets below RHS refers to the righthand side of the type family equation being checked for injectivity. LHS refers to the arguments of that type family equation. Below are the rules followed when checking injectivity of a type family:
 If a RHS of a type family equation is a type family application GHC reports that the type family is not injective.
 If a RHS of a type family equation is a bare type variable we require that all LHS variables (including implicit kind variables) are also bare. In other words, this has to be a sole equation of that type family and it has to cover all possible patterns. If the patterns are not covering GHC reports that the type family is not injective.
 If a LHS type variable that is declared as injective is not mentioned on injective position in the RHS GHC reports that the type family is not injective. Injective position means either argument to a type constructor or injective argument to a type family. Type inference can potentially loop when looking under injective type families in the RHS, so this requires
UndecidableInstances
; GHC suggests enabling the flag when it is necessary. 
Open type families Open type families are typechecked incrementally. This means that when a module is imported type family instances contained in that module are checked against instances present in already imported modules.
A pair of an open type family equations is checked by attempting to unify their RHSs. If the RHSs don’t unify this pair does not violate injectivity annotation. If unification succeeds with a substitution then LHSs of unified equations must be identical under that substitution. If they are not identical then GHC reports that the type family is not injective.

In a closed type family all equations are ordered and in one place. Equations are also checked pairwise but this time an equation has to be paired with all the preceeding equations. Of course a singleequation closed type family is trivially injective (unless (1), (2) or (3) above holds).
When checking a pair of closed type family equations GHC tried to unify their RHSs. If they don’t unify this pair of equations does not violate injectivity annotation. If the RHSs can be unified under some substitution (possibly empty) then either the LHSs unify under the same substitution or the LHS of the latter equation is subsumed by earlier equations. If neither condition is met GHC reports that a type family is not injective.
Note that for the purpose of injectivity check in bullets (4) and (5) GHC uses a special variant of unification algorithm that treats type family applications as possibly unifying with anything.
10.10. Datatype promotion

DataKinds

 Since

7.4.1
Allow promotion of data types to kind level.
This section describes data type promotion, an extension to the kind system that complements kind polymorphism. It is enabled by DataKinds
, and described in more detail in the paper Giving Haskell a Promotion, which appeared at TLDI 2012.
10.10.1. Motivation
Standard Haskell has a rich type language. Types classify terms and serve to avoid many common programming mistakes. The kind language, however, is relatively simple, distinguishing only regular types (kind Type
) and type constructors (e.g. kind Type > Type > Type
). In particular when using advanced type system features, such as type families (Type families) or GADTs (Generalised Algebraic Data Types (GADTs)), this simple kind system is insufficient, and fails to prevent simple errors. Consider the example of typelevel natural numbers, and lengthindexed vectors:
data Ze data Su n data Vec :: Type > Type > Type where Nil :: Vec a Ze Cons :: a > Vec a n > Vec a (Su n)
The kind of Vec
is Type > Type > Type
. This means that, e.g., Vec Int Char
is a wellkinded type, even though this is not what we intend when defining lengthindexed vectors.
With DataKinds
, the example above can then be rewritten to:
data Nat = Ze  Su Nat data Vec :: Type > Nat > Type where Nil :: Vec a 'Ze Cons :: a > Vec a n > Vec a ('Su n)
With the improved kind of Vec
, things like Vec Int Char
are now illkinded, and GHC will report an error.
10.10.2. Overview
With DataKinds
, GHC automatically promotes every datatype to be a kind and its (value) constructors to be type constructors. The following types
data Nat = Zero  Succ Nat data List a = Nil  Cons a (List a) data Pair a b = Pair a b data Sum a b = L a  R b
give rise to the following kinds and type constructors (where promoted constructors are prefixed by a tick '
):
Nat :: Type 'Zero :: Nat 'Succ :: Nat > Nat List :: Type > Type 'Nil :: forall k. List k 'Cons :: forall k. k > List k > List k Pair :: Type > Type > Type 'Pair :: forall k1 k2. k1 > k2 > Pair k1 k2 Sum :: Type > Type > Type 'L :: k1 > Sum k1 k2 'R :: k2 > Sum k1 k2
Virtually all data constructors, even those with rich kinds, can be promoted. There are only a couple of exceptions to this rule:
 Data family instance constructors cannot be promoted at the moment. GHC’s type theory just isn’t up to the task of promoting data families, which requires full dependent types.

Data constructors with contexts that contain nonequality constraints cannot be promoted. For example:
data Foo :: Type > Type where MkFoo1 :: a ~ Int => Foo a  promotable MkFoo2 :: a ~~ Int => Foo a  promotable MkFoo3 :: Show a => Foo a  not promotable
MkFoo1
andMkFoo2
can be promoted, since their contexts only involve equalityoriented constraints. However,MkFoo3
’s context contains a nonequality constraintShow a
, and thus cannot be promoted.
10.10.3. Distinguishing between types and constructors
In the examples above, all promoted constructors are prefixed with a single quote mark '
. This mark tells GHC to look in the data constructor namespace for a name, not the type (constructor) namespace. Consider
data P = MkP  1 data Prom = P  2
We can thus distinguish the type P
(which has a constructor MkP
) from the promoted data constructor 'P
(of kind Prom
).
As a convenience, GHC allows you to omit the quote mark when the name is unambiguous. However, our experience has shown that the quote mark helps to make code more readable and less errorprone. GHC thus supports Wuntickedpromotedconstructors
that will warn you if you use a promoted data constructor without a preceding quote mark.
Just as in the case of Template Haskell (Syntax), GHC gets confused if you put a quote mark before a data constructor whose second character is a quote mark. In this case, just put a space between the promotion quote and the data constructor:
data T = A' type S = 'A'  ERROR: looks like a character type R = ' A'  OK: promoted `A'`
10.10.4. Promoted list and tuple types
With DataKinds
, Haskell’s list and tuple types are natively promoted to kinds, and enjoy the same convenient syntax at the type level, albeit prefixed with a quote:
data HList :: [Type] > Type where HNil :: HList '[] HCons :: a > HList t > HList (a ': t) data Tuple :: (Type,Type) > Type where Tuple :: a > b > Tuple '(a,b) foo0 :: HList '[] foo0 = HNil foo1 :: HList '[Int] foo1 = HCons (3::Int) HNil foo2 :: HList [Int, Bool] foo2 = ...
For typelevel lists of two or more elements, such as the signature of foo2
above, the quote may be omitted because the meaning is unambiguous. But for lists of one or zero elements (as in foo0
and foo1
), the quote is required, because the types []
and [Int]
have existing meanings in Haskell.
Note
The declaration for HCons
also requires TypeOperators
because of infix type operator (':)
10.10.5. Promoting existential data constructors
Note that we do promote existential data constructors that are otherwise suitable. For example, consider the following:
data Ex :: Type where MkEx :: forall a. a > Ex
Both the type Ex
and the data constructor MkEx
get promoted, with the polymorphic kind 'MkEx :: forall k. k > Ex
. Somewhat surprisingly, you can write a type family to extract the member of a typelevel existential:
type family UnEx (ex :: Ex) :: k type instance UnEx (MkEx x) = x
At first blush, UnEx
seems poorlykinded. The return kind k
is not mentioned in the arguments, and thus it would seem that an instance would have to return a member of k
for any k
. However, this is not the case. The type family UnEx
is a kindindexed type family. The return kind k
is an implicit parameter to UnEx
. The elaborated definitions are as follows (where implicit parameters are denoted by braces):
type family UnEx {k :: Type} (ex :: Ex) :: k type instance UnEx {k} (MkEx @k x) = x
Thus, the instance triggers only when the implicit parameter to UnEx
matches the implicit parameter to MkEx
. Because k
is actually a parameter to UnEx
, the kind is not escaping the existential, and the above code is valid.
See also #7347.
10.11. Kind polymorphism

TypeInType

 Implies
 Since

8.0.1
The extension
TypeInType
is now deprecated: its sole effect is to switch onPolyKinds
(and henceKindSignatures
) andDataKinds
.

PolyKinds

 Implies
 Since

7.4.1
Allow kind polymorphic types.
This section describes GHC’s kind system, as it appears in version 8.0 and beyond. The kind system as described here is always in effect, with or without extensions, although it is a conservative extension beyond standard Haskell. The extensions above simply enable syntax and tweak the inference algorithm to allow users to take advantage of the extra expressiveness of GHC’s kind system.
10.11.1. Overview of kind polymorphism
Consider inferring the kind for
data App f a = MkApp (f a)
In Haskell 98, the inferred kind for App
is (Type > Type) > Type >
Type
. But this is overly specific, because another suitable Haskell 98 kind for App
is ((Type > Type) > Type) > (Type > Type) > Type
, where the kind assigned to a
is Type > Type
. Indeed, without kind signatures (KindSignatures
), it is necessary to use a dummy constructor to get a Haskell compiler to infer the second kind. With kind polymorphism (PolyKinds
), GHC infers the kind forall k. (k > Type) > k >
Type
for App
, which is its most general kind.
Thus, the chief benefit of kind polymorphism is that we can now infer these most general kinds and use App
at a variety of kinds:
App Maybe Int  `k` is instantiated to Type data T a = MkT (a Int)  `a` is inferred to have kind (Type > Type) App T Maybe  `k` is instantiated to (Type > Type)
10.11.2. Overview of TypeinType
GHC 8 extends the idea of kind polymorphism by declaring that types and kinds are indeed one and the same. Nothing within GHC distinguishes between types and kinds. Another way of thinking about this is that the type Bool
and the “promoted kind” Bool
are actually identical. (Note that term True
and the type 'True
are still distinct, because the former can be used in expressions and the latter in types.) This lack of distinction between types and kinds is a hallmark of dependently typed languages. Full dependently typed languages also remove the difference between expressions and types, but doing that in GHC is a story for another day.
One simplification allowed by combining types and kinds is that the type of Type
is just Type
. It is true that the Type :: Type
axiom can lead to nontermination, but this is not a problem in GHC, as we already have other means of nonterminating programs in both types and expressions. This decision (among many, many others) does mean that despite the expressiveness of GHC’s type system, a “proof” you write in Haskell is not an irrefutable mathematical proof. GHC promises only partial correctness, that if your programs compile and run to completion, their results indeed have the types assigned. It makes no claim about programs that do not finish in a finite amount of time.
To learn more about this decision and the design of GHC under the hood please see the paper introducing this kind system to GHC/Haskell.
10.11.3. Principles of kind inference
Generally speaking, when PolyKinds
is on, GHC tries to infer the most general kind for a declaration. In many cases (for example, in a datatype declaration) the definition has a righthand side to inform kind inference. But that is not always the case. Consider
type family F a
Type family declarations have no righthand side, but GHC must still infer a kind for F
. Since there are no constraints, it could infer F :: forall k1 k2. k1 > k2
, but that seems too polymorphic. So GHC defaults those entirelyunconstrained kind variables to Type
and we get F :: Type > Type
. You can still declare F
to be kindpolymorphic using kind signatures:
type family F1 a  F1 :: Type > Type type family F2 (a :: k)  F2 :: forall k. k > Type type family F3 a :: k  F3 :: forall k. Type > k type family F4 (a :: k1) :: k2  F4 :: forall k1 k2. k1 > k2
The general principle is this:
 When there is a righthand side, GHC infers the most polymorphic kind consistent with the righthand side. Examples: ordinary data type and GADT declarations, class declarations. In the case of a class declaration the role of “right hand side” is played by the class method signatures.
 When there is no right hand side, GHC defaults argument and result kinds to
Type
, except when directed otherwise by a kind signature. Examples: data and open type family declarations.
This rule has occasionallysurprising consequences (see #10132).
class C a where  Class declarations are generalised  so C :: forall k. k > Constraint data D1 a  No right hand side for these two family type F1 a  declarations, but the class forces (a :: k)  so D1, F1 :: forall k. k > Type data D2 a  No righthand side so D2 :: Type > Type type F2 a  No righthand side so F2 :: Type > Type
The kindpolymorphism from the class declaration makes D1
kindpolymorphic, but not so D2
; and similarly F1
, F1
.
10.11.4. Inferring the order of variables in a type/class declaration
It is possible to get intricate dependencies among the type variables introduced in a type or class declaration. Here is an example:
data T a (b :: k) c = MkT (a c)
After analysing this declaration, GHC will discover that a
and c
can be kindpolymorphic, with a :: k2 > Type
and c :: k2
. We thus infer the following kind:
T :: forall {k2 :: Type} (k :: Type). (k2 > Type) > k > k2 > Type
Note that k2
is placed before k
, and that k
is placed before a
. Also, note that k2
is written here in braces. As explained with TypeApplications
(Inferred vs. specified type variables), type and kind variables that GHC generalises over, but not written in the original program, are not available for visible type application. (These are called inferred variables.) Such variables are written in braces with fprintexplicitforalls
enabled.
The general principle is this:
 Variables not available for type application come first.
 Then come variables the user has written, implicitly brought into scope in a type variable’s kind.
 Lastly come the normal type variables of a declaration.
 Variables not given an explicit ordering by the user are sorted according to ScopedSort (Ordering of specified variables).
With the T
example above, we could bind k
after a
; doing so would not violate dependency concerns. However, it would violate our general principle, and so k
comes first.
Sometimes, this ordering does not respect dependency. For example:
data T2 k (a :: k) (c :: Proxy '[a, b])
It must be that a
and b
have the same kind. Note also that b
is implicitly declared in c
’s kind. Thus, according to our general principle, b
must come before k
. However, b
depends on k
. We thus reject T2
with a suitable error message.
In associated types, we order the type variables as if the type family was a toplevel declaration, ignoring the visibilities of the class’s type variable binders. Here is an example:
class C (a :: k) b where type F (c :: j) (d :: Proxy m) a b
We infer these kinds:
C :: forall {k1 :: Type} (k :: Type). k > k1 > Constraint F :: forall {k1 :: Type} {k2 :: Type} {k3 :: Type} j (m :: k1). j > Proxy m > k2 > k3 > Type
Note that the kind of a
is specified in the kind of C
but inferred in the kind of F
.
The “general principle” described here is meant to make all this more predictable for users. It would not be hard to extend GHC to relax this principle. If you should want a change here, consider writing a proposal to do so.
10.11.5. Complete usersupplied kind signatures and polymorphic recursion

CUSKs

 Since

8.10.1
NB! This is a legacy feature, see StandaloneKindSignatures
for the modern replacement.
Just as in type inference, kind inference for recursive types can only use monomorphic recursion. Consider this (contrived) example:
data T m a = MkT (m a) (T Maybe (m a))  GHC infers kind T :: (Type > Type) > Type > Type
The recursive use of T
forced the second argument to have kind Type
. However, just as in type inference, you can achieve polymorphic recursion by giving a complete usersupplied kind signature (or CUSK) for T
. A CUSK is present when all argument kinds and the result kind are known, without any need for inference. For example:
data T (m :: k > Type) :: k > Type where MkT :: m a > T Maybe (m a) > T m a
The complete usersupplied kind signature specifies the polymorphic kind for T
, and this signature is used for all the calls to T
including the recursive ones. In particular, the recursive use of T
is at kind Type
.
What exactly is considered to be a “complete usersupplied kind signature” for a type constructor? These are the forms:

For a datatype, every type variable must be annotated with a kind. In a GADTstyle declaration, there may also be a kind signature (with a toplevel
::
in the header), but the presence or absence of this annotation does not affect whether or not the declaration has a complete signature.data T1 :: (k > Type) > k > Type where ...  Yes; T1 :: forall k. (k>Type) > k > Type data T2 (a :: k > Type) :: k > Type where ...  Yes; T2 :: forall k. (k>Type) > k > Type data T3 (a :: k > Type) (b :: k) :: Type where ...  Yes; T3 :: forall k. (k>Type) > k > Type data T4 (a :: k > Type) (b :: k) where ...  Yes; T4 :: forall k. (k>Type) > k > Type data T5 a (b :: k) :: Type where ...  No; kind is inferred data T6 a b where ...  No; kind is inferred

For a datatype with a toplevel
::
: all kind variables introduced after the::
must be explicitly quantified.data T1 :: k > Type  No CUSK: `k` is not explicitly quantified data T2 :: forall k. k > Type  CUSK: `k` is bound explicitly data T3 :: forall (k :: Type). k > Type  still a CUSK

For a newtype, the rules are the same as they are for a data type unless
UnliftedNewtypes
is enabled. WithUnliftedNewtypes
, the type constructor only has a CUSK if a kind signature is present. As with a datatype with a toplevel::
, all kind variables must introduced after the::
must be explicitly quantified{# LANGUAGE UnliftedNewtypes #} newtype N1 where  No; missing kind signature newtype N2 :: TYPE 'IntRep where  Yes; kind signature present newtype N3 (a :: Type) where  No; missing kind signature newtype N4 :: k > Type where  No; `k` is not explicitly quantified newtype N5 :: forall (k :: Type). k > Type where  Yes; good signature
 For a class, every type variable must be annotated with a kind.

For a type synonym, every type variable and the result type must all be annotated with kinds:
type S1 (a :: k) = (a :: k)  Yes S1 :: forall k. k > k type S2 (a :: k) = a  No kind is inferred type S3 (a :: k) = Proxy a  No kind is inferred
Note that in
S2
andS3
, the kind of the righthand side is rather apparent, but it is still not considered to have a complete signature – no inference can be done before detecting the signature. 
An unassociated open type or data family declaration always has a CUSK; unannotated type variables default to kind
Type
:data family D1 a  D1 :: Type > Type data family D2 (a :: k)  D2 :: forall k. k > Type data family D3 (a :: k) :: Type  D3 :: forall k. k > Type type family S1 a :: k > Type  S1 :: forall k. Type > k > Type

An associated type or data family declaration has a CUSK precisely if its enclosing class has a CUSK.
class C a where  no CUSK type AT a b  no CUSK, b is defaulted class D (a :: k) where  yes CUSK type AT2 a b  yes CUSK, b is defaulted
 A closed type family has a complete signature when all of its type variables are annotated and a return kind (with a toplevel
::
) is supplied.
It is possible to write a datatype that syntactically has a CUSK (according to the rules above) but actually requires some inference. As a very contrived example, consider
data Proxy a  Proxy :: forall k. k > Type data X (a :: Proxy k)
According to the rules above X
has a CUSK. Yet, the kind of k
is undetermined. It is thus quantified over, giving X
the kind forall k1 (k :: k1). Proxy k > Type
.
The detection of CUSKs is enabled by the CUSKs
flag, which is switched on by default. This extension is scheduled for deprecation to be replaced with StandaloneKindSignatures
.
10.11.6. Standalone kind signatures and polymorphic recursion

StandaloneKindSignatures

 Implies
 Since

8.10.1
Just as in type inference, kind inference for recursive types can only use monomorphic recursion. Consider this (contrived) example:
data T m a = MkT (m a) (T Maybe (m a))  GHC infers kind T :: (Type > Type) > Type > Type
The recursive use of T
forced the second argument to have kind Type
. However, just as in type inference, you can achieve polymorphic recursion by giving a standalone kind signature for T
:
type T :: (k > Type) > k > Type data T m a = MkT (m a) (T Maybe (m a))
The standalone kind signature specifies the polymorphic kind for T
, and this signature is used for all the calls to T
including the recursive ones. In particular, the recursive use of T
is at kind Type
.
While a standalone kind signature determines the kind of a type constructor, it does not determine its arity. This is of particular importance for type families and type synonyms, as they cannot be partially applied. See Type family declarations for more information about arity.
The arity can be specified using explicit binders and inline kind annotations:
 arity F0 = 0 type F0 :: forall k. k > Type type family F0 :: forall k. k > Type  arity F1 = 1 type F1 :: forall k. k > Type type family F1 :: k > Type  arity F2 = 2 type F2 :: forall k. k > Type type family F2 a :: Type
In absence of an inline kind annotation, the inferred arity includes all explicitly bound parameters and all immediately following invisible parameters:
 arity FD1 = 1 type FD1 :: forall k. k > Type type FD1  arity FD2 = 2 type FD2 :: forall k. k > Type type FD2 a
Note that F0
, F1
, F2
, FD1
, and FD2
all have identical standalone kind signatures. The arity is inferred from the type family header.
10.11.7. Standalone kind signatures and declaration headers
GHC requires that in the presence of a standalone kind signature, data declarations must bind all their inputs. For example:
type Prox1 :: k > Type data Prox1 a = MkProx1  OK. type Prox2 :: k > Type data Prox2 = MkProx2  Error:  • Expected a type, but found something with kind ‘k > Type’  • In the data type declaration for ‘Prox2’
GADTstyle data declarations may either bind their inputs or use an inline signature in addition to the standalone kind signature:
type GProx1 :: k > Type data GProx1 a where MkGProx1 :: GProx1 a  OK. type GProx2 :: k > Type data GProx2 where MkGProx2 :: GProx2 a  Error:  • Expected a type, but found something with kind ‘k > Type’  • In the data type declaration for ‘GProx2’ type GProx3 :: k > Type data GProx3 :: k > Type where MkGProx3 :: GProx3 a  OK. type GProx4 :: k > Type data GProx4 :: w where MkGProx4 :: GProx4 a  OK, w ~ (k > Type)
Classes are subject to the same rules:
type C1 :: Type > Constraint class C1 a  OK. type C2 :: Type > Constraint class C2  Error:  • Couldn't match expected kind ‘Constraint’  with actual kind ‘Type > Constraint’  • In the class declaration for ‘C2’
On the other hand, type families are exempt from this rule:
type F :: Type > Type type family F  OK.
Data families are tricky territory. Their headers are exempt from this rule, but their instances are not:
type T :: k > Type data family T  OK. data instance T Int = MkT1  OK. data instance T = MkT3  Error:  • Expecting one more argument to ‘T’  Expected a type, but ‘T’ has kind ‘k0 > Type’  • In the data instance declaration for ‘T’
This also applies to GADTstyle data instances:
data instance T (a :: Nat) where MkN4 :: T 4 MKN9 :: T 9  OK. data instance T :: Symbol > Type where MkSN :: T "Neptune" MkSJ :: T "Jupiter"  OK. data instance T where MkT4 :: T x  Error:  • Expecting one more argument to ‘T’  Expected a type, but ‘T’ has kind ‘k0 > Type’  • In the data instance declaration for ‘T’
10.11.8. Kind inference in closed type families
Although all open type families are considered to have a complete usersupplied kind signature, we can relax this condition for closed type families, where we have equations on which to perform kind inference. GHC will infer kinds for the arguments and result types of a closed type family.
GHC supports kindindexed type families, where the family matches both on the kind and type. GHC will not infer this behaviour without a complete usersupplied kind signature, as doing so would sometimes infer nonprincipal types. Indeed, we can see kindindexing as a form of polymorphic recursion, where a type is used at a kind other than its most general in its own definition.
For example:
type family F1 a where F1 True = False F1 False = True F1 x = x  F1 fails to compile: kindindexing is not inferred type family F2 (a :: k) where F2 True = False F2 False = True F2 x = x  F2 fails to compile: no complete signature type family F3 (a :: k) :: k where F3 True = False F3 False = True F3 x = x  OK
10.11.9. Kind inference in class instance declarations
Consider the following example of a polykinded class and an instance for it:
class C a where type F a instance C b where type F b = b > b
In the class declaration, nothing constrains the kind of the type a
, so it becomes a polykinded type variable (a :: k)
. Yet, in the instance declaration, the righthand side of the associated type instance b > b
says that b
must be of kind Type
. GHC could theoretically propagate this information back into the instance head, and make that instance declaration apply only to type of kind Type
, as opposed to types of any kind. However, GHC does not do this.
In short: GHC does not propagate kind information from the members of a class instance declaration into the instance declaration head.
This lack of kind inference is simply an engineering problem within GHC, but getting it to work would make a substantial change to the inference infrastructure, and it’s not clear the payoff is worth it. If you want to restrict b
’s kind in the instance above, just use a kind signature in the instance head.
10.11.10. Kind inference in type signatures
When kindchecking a type, GHC considers only what is written in that type when figuring out how to generalise the type’s kind.
For example, consider these definitions (with ScopedTypeVariables
):
data Proxy a  Proxy :: forall k. k > Type p :: forall a. Proxy a p = Proxy :: Proxy (a :: Type)
GHC reports an error, saying that the kind of a
should be a kind variable k
, not Type
. This is because, by looking at the type signature forall a. Proxy a
, GHC assumes a
’s kind should be generalised, not restricted to be Type
. The function definition is then rejected for being more specific than its type signature.
10.11.11. Explicit kind quantification
Enabled by PolyKinds
, GHC supports explicit kind quantification, as in these examples:
data Proxy :: forall k. k > Type f :: (forall k (a :: k). Proxy a > ()) > Int
Note that the second example has a forall
that binds both a kind k
and a type variable a
of kind k
. In general, there is no limit to how deeply nested this sort of dependency can work. However, the dependency must be wellscoped: forall (a :: k) k. ...
is an error.
10.11.12. Implicit quantification in type synonyms and type family instances
Consider the scoping rules for type synonyms and type family instances, such as these:
type TS a (b :: k) = <rhs> type instance TF a (b :: k) = <rhs>
The basic principle is that all variables mentioned on the right hand side <rhs>
must be bound on the left hand side:
type TS a (b :: k) = (k, a, Proxy b)  accepted type TS a (b :: k) = (k, a, Proxy b, z)  rejected: z not in scope
But there is one exception: free variables mentioned in the outermost kind signature on the right hand side are quantified implicitly. Thus, in the following example the variables a
, b
, and k
are all in scope on the right hand side of S
:
type S a b = <rhs> :: k > k
The reason for this exception is that there may be no other way to bind k
. For example, suppose we wanted S
to have the the following kind with an invisible parameter k
:
S :: forall k. Type > Type > k > k
In this case, we could not simply bind k
on the lefthand side, as k
would become a visible parameter:
type S k a b = <rhs> :: k > k S :: forall k > Type > Type > k > k
Note that we only look at the outermost kind signature to decide which variables to quantify implicitly. As a counterexample, consider M1
:
type M1 = 'Just ('Nothing :: Maybe k)  rejected: k not in scope
Here, the kind signature is hidden inside 'Just
, and there is no outermost kind signature. We can fix this example by providing an outermost kind signature:
type M2 = 'Just ('Nothing :: Maybe k) :: Maybe (Maybe k)
Here, k
is brought into scope by :: Maybe (Maybe k)
.
A kind signature is considered to be outermost regardless of redundant parentheses:
type P = 'Nothing :: Maybe a  accepted type P = ((('Nothing :: Maybe a)))  accepted
Closed type family instances are subject to the same rules:
type family F where F = 'Nothing :: Maybe k  accepted type family F where F = 'Just ('Nothing :: Maybe k)  rejected: k not in scope type family F where F = 'Just ('Nothing :: Maybe k) :: Maybe (Maybe k)  accepted type family F :: Maybe (Maybe k) where F = 'Just ('Nothing :: Maybe k)  rejected: k not in scope type family F :: Maybe (Maybe k) where F @k = 'Just ('Nothing :: Maybe k)  accepted
Kind variables can also be quantified in visible positions. Consider the following two examples:
data ProxyKInvis (a :: k) data ProxyKVis k (a :: k)
In the first example, the kind variable k
is an invisible argument to ProxyKInvis
. In other words, a user does not need to instantiate k
explicitly, as kind inference automatically determines what k
should be. For instance, in ProxyKInvis True
, k
is inferred to be Bool
. This is reflected in the kind of ProxyKInvis
:
ProxyKInvis :: forall k. k > Type
In the second example, k
is a visible argument to ProxyKVis
. That is to say, k
is an argument that users must provide explicitly when applying ProxyKVis
. For example, ProxyKVis Bool True
is a well formed type.
What is the kind of ProxyKVis
? One might say forall k. Type > k > Type
, but this isn’t quite right, since this would allow incorrect things like ProxyKVis Bool Int
, which should be rejected due to the fact that Int
is not of kind Bool
. The key observation is that the kind of the second argument depend on the first argument. GHC indicates this dependency in the syntax that it gives for the kind of ProxyKVis
:
ProxyKVis :: forall k > k > Type
This kind is similar to the kind of ProxyKInvis
, but with a key difference: the type variables quantified by the forall
are followed by an arrow (>
), not a dot (.
). This is a visible, dependent quantifier. It is visible in that it the user must pass in a type for k
explicitly, and it is dependent in the sense that k
appears later in the kind of ProxyKVis
. As a counterpart, the k
binder in forall k. k > Type
can be thought of as an invisible, dependent quantifier.
GHC permits writing kinds with this syntax, provided that the ExplicitForAll
and PolyKinds
language extensions are enabled. Just like the invisible forall
, one can put explicit kind signatures on visibly bound kind variables, so the following is syntactically valid:
ProxyKVis :: forall (k :: Type) > k > Type
Currently, the ability to write visible, dependent quantifiers is limited to kinds. Consequently, visible dependent quantifiers are rejected in any context that is unambiguously the type of a term. They are also rejected in the types of data constructors.
10.11.13. Kindindexed GADTs
Consider the type
data G (a :: k) where GInt :: G Int GMaybe :: G Maybe
This datatype G
is GADTlike in both its kind and its type. Suppose you have g :: G a
, where a :: k
. Then pattern matching to discover that g
is in fact GMaybe
tells you both that k ~ (Type > Type)
and a ~ Maybe
. The definition for G
requires that PolyKinds
be in effect, but patternmatching on G
requires no extension beyond GADTs
. That this works is actually a straightforward extension of regular GADTs and a consequence of the fact that kinds and types are the same.
Note that the datatype G
is used at different kinds in its body, and therefore that kindindexed GADTs use a form of polymorphic recursion. It is thus only possible to use this feature if you have provided a complete usersupplied kind signature for the datatype (Complete usersupplied kind signatures and polymorphic recursion).
10.11.14. Higherrank kinds
In concert with RankNTypes
, GHC supports higherrank kinds. Here is an example:
 Heterogeneous propositional equality data (a :: k1) :~~: (b :: k2) where HRefl :: a :~~: a class HTestEquality (t :: forall k. k > Type) where hTestEquality :: forall k1 k2 (a :: k1) (b :: k2). t a > t b > Maybe (a :~~: b)
Note that hTestEquality
takes two arguments where the type variable t
is applied to types of different kinds. That type variable must then be polykinded. Accordingly, the kind of HTestEquality
(the class) is (forall k. k > Type) > Constraint
, a higherrank kind.
A big difference with higherrank kinds as compared with higherrank types is that forall
s in kinds cannot be moved. This is best illustrated by example. Suppose we want to have an instance of HTestEquality
for (:~~:)
.
instance HTestEquality ((:~~:) a) where hTestEquality HRefl HRefl = Just HRefl
With the declaration of (:~~:)
above, it gets kind forall k1 k2. k1 > k2 > Type
. Thus, the type (:~~:) a
has kind k2 > Type
for some k2
. GHC cannot then regeneralize this kind to become forall k2. k2 > Type
as desired. Thus, the instance is rejected as illkinded.
To allow for such an instance, we would have to define (:~~:)
as follows:
data (:~~:) :: forall k1. k1 > forall k2. k2 > Type where HRefl :: a :~~: a
In this redefinition, we give an explicit kind for (:~~:)
, deferring the choice of k2
until after the first argument (a
) has been given. With this declaration for (:~~:)
, the instance for HTestEquality
is accepted.
Another difference between higherrank kinds and types can be found in their treatment of inferred and userspecified type variables. Consider the following program:
newtype Foo (f :: forall k. k > Type) = MkFoo (f Int) data Proxy a = Proxy foo :: Foo Proxy foo = MkFoo Proxy
The kind of Foo
’s parameter is forall k. k > Type
, but the kind of Proxy
is forall {k}. k > Type
, where {k}
denotes that the kind variable k
is to be inferred, not specified by the user. (See Visible type application for more discussion on the inferredspecified distinction). GHC does not consider forall k. k > Type
and forall {k}. k > Type
to be equal at the kind level, and thus rejects Foo Proxy
as illkinded.
10.11.15. Constraints in kinds
As kinds and types are the same, kinds can (with TypeInType
) contain type constraints. However, only equality constraints are supported.
Here is an example of a constrained kind:
type family IsTypeLit a where IsTypeLit Nat = 'True IsTypeLit Symbol = 'True IsTypeLit a = 'False data T :: forall a. (IsTypeLit a ~ 'True) => a > Type where MkNat :: T 42 MkSymbol :: T "Don't panic!"
The declarations above are accepted. However, if we add MkOther :: T Int
, we get an error that the equality constraint is not satisfied; Int
is not a type literal. Note that explicitly quantifying with forall a
is necessary in order for T
to typecheck (see Complete usersupplied kind signatures and polymorphic recursion).
10.11.16. The kind Type

StarIsType

 Since

8.6.1
Treat the unqualified uses of the
*
type operator as nullary and desugar toData.Kind.Type
.
The kind Type
(imported from Data.Kind
) classifies ordinary types. With StarIsType
(currently enabled by default), *
is desugared to Type
, but using this legacy syntax is not recommended due to conflicts with TypeOperators
. This also applies to ★
, the Unicode variant of *
.
10.11.17. Inferring dependency in datatype declarations
If a type variable a
in a datatype, class, or type family declaration depends on another such variable k
in the same declaration, two properties must hold:

a
must appear afterk
in the declaration, and 
k
must appear explicitly in the kind of some type variable in that declaration.
The first bullet simply means that the dependency must be wellscoped. The second bullet concerns GHC’s ability to infer dependency. Inferring this dependency is difficult, and GHC currently requires the dependency to be made explicit, meaning that k
must appear in the kind of a type variable, making it obvious to GHC that dependency is intended. For example:
data Proxy k (a :: k)  OK: dependency is "obvious" data Proxy2 k a = P (Proxy k a)  ERROR: dependency is unclear
In the second declaration, GHC cannot immediately tell that k
should be a dependent variable, and so the declaration is rejected.
It is conceivable that this restriction will be relaxed in the future, but it is (at the time of writing) unclear if the difficulties around this scenario are theoretical (inferring this dependency would mean our type system does not have principal types) or merely practical (inferring this dependency is hard, given GHC’s implementation). So, GHC takes the easy way out and requires a little help from the user.
10.11.18. Inferring dependency in userwritten forall
s
A programmer may use forall
in a type to introduce new quantified type variables. These variables may depend on each other, even in the same forall
. However, GHC requires that the dependency be inferrable from the body of the forall
. Here are some examples:
data Proxy k (a :: k) = MkProxy  just to use below f :: forall k a. Proxy k a  This is just fine. We see that (a :: k). f = undefined g :: Proxy k a > ()  This is to use below. g = undefined data Sing a h :: forall k a. Sing k > Sing a > ()  No obvious relationship between k and a h _ _ = g (MkProxy :: Proxy k a)  This fails. We didn't know that a should have kind k.
Note that in the last example, it’s impossible to learn that a
depends on k
in the body of the forall
(that is, the Sing k > Sing a > ()
). And so GHC rejects the program.
10.11.19. Kind defaulting without PolyKinds
Without PolyKinds
, GHC refuses to generalise over kind variables. It thus defaults kind variables to Type
when possible; when this is not possible, an error is issued.
Here is an example of this in action:
{# LANGUAGE PolyKinds #} import Data.Kind (Type) data Proxy a = P  inferred kind: Proxy :: k > Type data Compose f g x = MkCompose (f (g x))  inferred kind: Compose :: (b > Type) > (a > b) > a > Type  separate module having imported the first {# LANGUAGE NoPolyKinds, DataKinds #} z = Proxy :: Proxy 'MkCompose
In the last line, we use the promoted constructor 'MkCompose
, which has kind
forall (a :: Type) (b :: Type) (f :: b > Type) (g :: a > b) (x :: a). f (g x) > Compose f g x
Now we must infer a type for z
. To do so without generalising over kind variables, we must default the kind variables of 'MkCompose
. We can easily default a
and b
to Type
, but f
and g
would be illkinded if defaulted. The definition for z
is thus an error.
10.11.20. Prettyprinting in the presence of kind polymorphism
With kind polymorphism, there is quite a bit going on behind the scenes that may be invisible to a Haskell programmer. GHC supports several flags that control how types are printed in error messages and at the GHCi prompt. See the discussion of type prettyprinting options for further details. If you are using kind polymorphism and are confused as to why GHC is rejecting (or accepting) your program, we encourage you to turn on these flags, especially fprintexplicitkinds
.
10.12. Levity polymorphism
In order to allow full flexibility in how kinds are used, it is necessary to use the kind system to differentiate between boxed, lifted types (normal, everyday types like Int
and [Bool]
) and unboxed, primitive types (Unboxed types and primitive operations) like Int#
. We thus have socalled levity polymorphism.
Here are the key definitions, all available from GHC.Exts
:
TYPE :: RuntimeRep > Type  highly magical, built into GHC data RuntimeRep = LiftedRep  for things like `Int`  UnliftedRep  for things like `Array#`  IntRep  for `Int#`  TupleRep [RuntimeRep]  unboxed tuples, indexed by the representations of the elements  SumRep [RuntimeRep]  unboxed sums, indexed by the representations of the disjuncts  ... type Type = TYPE LiftedRep  Type is just an ordinary type synonym
The idea is that we have a new fundamental type constant TYPE
, which is parameterised by a RuntimeRep
. We thus get Int# :: TYPE 'IntRep
and Bool :: TYPE 'LiftedRep
. Anything with a type of the form TYPE x
can appear to either side of a function arrow >
. We can thus say that >
has type TYPE r1 > TYPE r2 > TYPE 'LiftedRep
. The result is always lifted because all functions are lifted in GHC.
10.12.1. No levitypolymorphic variables or arguments
If GHC didn’t have to compile programs that run in the real world, that would be the end of the story. But representation polymorphism can cause quite a bit of trouble for GHC’s code generator. Consider
bad :: forall (r1 :: RuntimeRep) (r2 :: RuntimeRep) (a :: TYPE r1) (b :: TYPE r2). (a > b) > a > b bad f x = f x
This seems like a generalisation of the standard $
operator. If we think about compiling this to runnable code, though, problems appear. In particular, when we call bad
, we must somehow pass x
into bad
. How wide (that is, how many bits) is x
? Is it a pointer? What kind of register (floatingpoint or integral) should x
go in? It’s all impossible to say, because x
’s type, a :: TYPE r1
is levity polymorphic. We thus forbid such constructions, via the following straightforward rule:
No variable may have a levitypolymorphic type.
This eliminates bad
because the variable x
would have a representationpolymorphic type.
However, not all is lost. We can still do this:
($) :: forall r (a :: Type) (b :: TYPE r). (a > b) > a > b f $ x = f x
Here, only b
is levity polymorphic. There are no variables with a levitypolymorphic type. And the code generator has no trouble with this. Indeed, this is the true type of GHC’s $
operator, slightly more general than the Haskell 98 version.
Because the code generator must store and move arguments as well as variables, the logic above applies equally well to function arguments, which may not be levitypolymorphic.
10.12.2. Levitypolymorphic bottoms
We can use levity polymorphism to good effect with error
and undefined
, whose types are given here:
undefined :: forall (r :: RuntimeRep) (a :: TYPE r). HasCallStack => a error :: forall (r :: RuntimeRep) (a :: TYPE r). HasCallStack => String > a
These functions do not bind a levitypolymorphic variable, and so are accepted. Their polymorphism allows users to use these to conveniently stub out functions that return unboxed types.
10.12.3. Printing levitypolymorphic types

fprintexplicitruntimereps

Print
RuntimeRep
parameters as they appear; otherwise, they are defaulted to'LiftedRep
.
Most GHC users will not need to worry about levity polymorphism or unboxed types. For these users, seeing the levity polymorphism in the type of $
is unhelpful. And thus, by default, it is suppressed, by supposing all type variables of type RuntimeRep
to be 'LiftedRep
when printing, and printing TYPE 'LiftedRep
as Type
(or *
when StarIsType
is on).
Should you wish to see levity polymorphism in your types, enable the flag fprintexplicitruntimereps
. For example,
ghci> :t ($) ($) :: (a > b) > a > b ghci> :set fprintexplicitruntimereps ghci> :t ($) ($) :: forall (r :: GHC.Types.RuntimeRep) a (b :: TYPE r). (a > b) > a > b
10.13. TypeLevel Literals
GHC supports numeric and string literals at the type level, giving convenient access to a large number of predefined typelevel constants. Numeric literals are of kind Nat
, while string literals are of kind Symbol
. This feature is enabled by the DataKinds
language extension.
The kinds of the literals and all other lowlevel operations for this feature are defined in module GHC.TypeLits
. Note that the module defines some typelevel operators that clash with their valuelevel counterparts (e.g. (+)
). Import and export declarations referring to these operators require an explicit namespace annotation (see Explicit namespaces in import/export).
Here is an example of using typelevel numeric literals to provide a safe interface to a lowlevel function:
import GHC.TypeLits import Data.Word import Foreign newtype ArrPtr (n :: Nat) a = ArrPtr (Ptr a) clearPage :: ArrPtr 4096 Word8 > IO () clearPage (ArrPtr p) = ...
Here is an example of using typelevel string literals to simulate simple record operations:
data Label (l :: Symbol) = Get class Has a l b  a l > b where from :: a > Label l > b data Point = Point Int Int deriving Show instance Has Point "x" Int where from (Point x _) _ = x instance Has Point "y" Int where from (Point _ y) _ = y example = from (Point 1 2) (Get :: Label "x")
10.13.1. Runtime Values for TypeLevel Literals
Sometimes it is useful to access the valuelevel literal associated with a typelevel literal. This is done with the functions natVal
and symbolVal
. For example:
GHC.TypeLits> natVal (Proxy :: Proxy 2) 2
These functions are overloaded because they need to return a different result, depending on the type at which they are instantiated.
natVal :: KnownNat n => proxy n > Integer  instance KnownNat 0  instance KnownNat 1  instance KnownNat 2  ...
GHC discharges the constraint as soon as it knows what concrete typelevel literal is being used in the program. Note that this works only for literals and not arbitrary type expressions. For example, a constraint of the form KnownNat (a + b)
will not be simplified to (KnownNat a, KnownNat b)
; instead, GHC will keep the constraint as is, until it can simplify a + b
to a constant value.
It is also possible to convert a runtime integer or string value to the corresponding typelevel literal. Of course, the resulting type literal will be unknown at compiletime, so it is hidden in an existential type. The conversion may be performed using someNatVal
for integers and someSymbolVal
for strings:
someNatVal :: Integer > Maybe SomeNat SomeNat :: KnownNat n => Proxy n > SomeNat
The operations on strings are similar.
10.13.2. Computing With TypeLevel Naturals
GHC 7.8 can evaluate arithmetic expressions involving typelevel natural numbers. Such expressions may be constructed using the typefamilies (+), (*), (^)
for addition, multiplication, and exponentiation. Numbers may be compared using (<=?)
, which returns a promoted boolean value, or (<=)
, which compares numbers as a constraint. For example:
GHC.TypeLits> natVal (Proxy :: Proxy (2 + 3)) 5
At present, GHC is quite limited in its reasoning about arithmetic: it will only evaluate the arithmetic type functions and compare the results— in the same way that it does for any other type function. In particular, it does not know more general facts about arithmetic, such as the commutativity and associativity of (+)
, for example.
However, it is possible to perform a bit of “backwards” evaluation. For example, here is how we could get GHC to compute arbitrary logarithms at the type level:
lg :: Proxy base > Proxy (base ^ pow) > Proxy pow lg _ _ = Proxy GHC.TypeLits> natVal (lg (Proxy :: Proxy 2) (Proxy :: Proxy 8)) 3
10.14. Equality constraints, Coercible, and the kind Constraint
10.14.1. Equality constraints
A type context can include equality constraints of the form t1 ~ t2
, which denote that the types t1
and t2
need to be the same. In the presence of type families, whether two types are equal cannot generally be decided locally. Hence, the contexts of function signatures may include equality constraints, as in the following example:
sumCollects :: (Collects c1, Collects c2, Elem c1 ~ Elem c2) => c1 > c2 > c2
where we require that the element type of c1
and c2
are the same. In general, the types t1
and t2
of an equality constraint may be arbitrary monotypes; i.e., they may not contain any quantifiers, independent of whether higherrank types are otherwise enabled.
Equality constraints can also appear in class and instance contexts. The former enable a simple translation of programs using functional dependencies into programs using family synonyms instead. The general idea is to rewrite a class declaration of the form
class C a b  a > b
to
class (F a ~ b) => C a b where type F a
That is, we represent every functional dependency (FD) a1 .. an > b
by an FD type family F a1 .. an
and a superclass context equality F a1 .. an ~ b
, essentially giving a name to the functional dependency. In class instances, we define the type instances of FD families in accordance with the class head. Method signatures are not affected by that process.
10.14.2. Heterogeneous equality
GHC also supports kindheterogeneous equality, which relates two types of potentially different kinds. Heterogeneous equality is spelled ~~
. Here are the kinds of ~
and ~~
to better understand their difference:
(~) :: forall k. k > k > Constraint (~~) :: forall k1 k2. k1 > k2 > Constraint
Users will most likely want ~
, but ~~
is available if GHC cannot know, a priori, that the two types of interest have the same kind. Evidence that (a :: k1) ~~ (b :: k2)
tells GHC both that k1
and k2
are the same and that a
and b
are the same.
Because ~
is the more common equality relation, GHC prints out ~~
like ~
unless fprintequalityrelations
is set.
10.14.3. Unlifted heterogeneous equality
Internal to GHC is yet a third equality relation (~#)
. It is heterogeneous (like ~~
) and is used only internally. It may appear in error messages and other output only when fprintequalityrelations
is enabled.
10.14.4. The Coercible
constraint
The constraint Coercible t1 t2
is similar to t1 ~ t2
, but denotes representational equality between t1
and t2
in the sense of Roles (Roles). It is exported by Data.Coerce, which also contains the documentation. More details and discussion can be found in the paper “Safe Coercions”.
10.14.5. The Constraint
kind

ConstraintKinds

 Since

7.4.1
Allow types of kind
Constraint
to be used in contexts.
Normally, constraints (which appear in types to the left of the =>
arrow) have a very restricted syntax. They can only be:
 Class constraints, e.g.
Show a

Implicit parameter
constraints, e.g.?x::Int
(with theImplicitParams
extension) 
Equality constraints, e.g.
a ~ Int
(with theTypeFamilies
orGADTs
extensions)
With the ConstraintKinds
extension, GHC becomes more liberal in what it accepts as constraints in your program. To be precise, with this flag any type of the new kind Constraint
can be used as a constraint. The following things have kind Constraint
:
 Anything which is already valid as a constraint without the flag: saturated applications to type classes, implicit parameter and equality constraints.
 Tuples, all of whose component types have kind
Constraint
. So for example the type(Show a, Ord a)
is of kindConstraint
. 
Anything whose form is not yet known, but the user has declared to have kind
Constraint
(for which they need to import it fromData.Kind
). So for exampletype Foo (f :: Type > Constraint) = forall b. f b => b > b
is allowed, as well as examples involving type families:type family Typ a b :: Constraint type instance Typ Int b = Show b type instance Typ Bool b = Num b func :: Typ a b => a > b > b func = ...
Note that because constraints are just handled as types of a particular kind, this extension allows type constraint synonyms:
type Stringy a = (Read a, Show a) foo :: Stringy a => a > (String, String > a) foo x = (show x, read)
Presently, only standard constraints, tuples and type synonyms for those two sorts of constraint are permitted in instance contexts and superclasses (without extra flags). The reason is that permitting more general constraints can cause type checking to loop, as it would with these two programs:
type family Clsish u a type instance Clsish () a = Cls a class Clsish () a => Cls a where
class OkCls a where type family OkClsish u a type instance OkClsish () a = OkCls a instance OkClsish () a => OkCls a where
You may write programs that use exotic sorts of constraints in instance contexts and superclasses, but to do so you must use UndecidableInstances
to signal that you don’t mind if the type checker fails to terminate.
10.15. Quantified constraints

QuantifiedConstraints

 Since

8.6.1
Allow constraints to quantify over types.
The extension QuantifiedConstraints
introduces quantified constraints, which give a new level of expressiveness in constraints. For example, consider
data Rose f a = Branch a (f (Rose f a)) instance (Eq a, ???) => Eq (Rose f a) where (Branch x1 c1) == (Branch x2 c2) = x1==x1 && c1==c2
From the x1==x2
we need Eq a
, which is fine. From c1==c2
we need Eq (f (Rose f a))
which is not fine in Haskell today; we have no way to solve such a constraint.
QuantifiedConstraints
lets us write this
instance (Eq a, forall b. (Eq b) => Eq (f b)) => Eq (Rose f a) where (Branch x1 c1) == (Branch x2 c2) = x1==x1 && c1==c2
Here, the quantified constraint forall b. (Eq b) => Eq (f b)
behaves a bit like a local instance declaration, and makes the instance typeable.
The paper Quantified class constraints (by Bottu, Karachalias, Schrijvers, Oliveira, Wadler, Haskell Symposium 2017) describes this feature in technical detail, with examples, and so is a primary reference source for this feature.
10.15.1. Motivation
Introducing quantified constraints offers two main benefits:

Firstly, they enable terminating resolution where this was not possible before. Consider for instance the following instance declaration for the general rose datatype
data Rose f x = Rose x (f (Rose f x)) instance (Eq a, forall b. Eq b => Eq (f b)) => Eq (Rose f a) where (Rose x1 rs1) == (Rose x2 rs2) = x1 == x2 && rs1 == rs2
This extension allows us to write constraints of the form
forall b. Eq b => Eq (f b)
, which is needed to solve theEq (f (Rose f x))
constraint arising from the second usage of the(==)
method. 
Secondly, quantified constraints allow for more concise and precise specifications. As an example, consider the MTL type class for monad transformers:
class Trans t where lift :: Monad m => m a > (t m) a
The developer knows that a monad transformer takes a monad
m
into a new monadt m
. But this property is not formally specified in the above declaration. This omission becomes an issue when defining monad transformer composition:newtype (t1 * t2) m a = C { runC :: t1 (t2 m) a } instance (Trans t1, Trans t2) => Trans (t1 * t2) where lift = C . lift . lift
The goal here is to
lift
from monadm
tot2 m
and thenlift
this again intot1 (t2 m)
. However, this secondlift
can only be accepted when(t2 m)
is a monad and there is no way of establishing that this fact universally holds.Quantified constraints enable this property to be made explicit in the
Trans
class declaration:class (forall m. Monad m => Monad (t m)) => Trans t where lift :: Monad m => m a > (t m) a
This idea is very old; see Seciton 7 of Derivable type classes.
10.15.2. Syntax changes
Haskell 2010 defines a context
(the bit to the left of =>
in a type) like this
context ::= class  ( class1, ..., classn ) class ::= qtycls tyvar  qtycls (tyvar atype1 ... atypen)
We to extend class
(warning: this is a rather confusingly named nonterminal symbol) with two extra forms, namely precisely what can appear in an instance declaration
class ::= ...  [context =>] qtycls inst  [context =>] tyvar inst
The definition of inst
is unchanged from the Haskell Report (roughly, just a type). The context =>
part is optional. That is the only syntactic change to the language.
Notes:

Where GHC allows extensions instance declarations we allow exactly the same extensions to this new form of
class
. Specifically, withExplicitForAll
andMultiParamTypeClasses
the syntax becomesclass ::= ...  [forall tyavrs .] [context =>] qtycls inst1 ... instn  [forall tyavrs .] [context =>] tyvar inst1 ... instn
Note that an explicit
forall
is often absolutely essential. Consider the rosetree exampleinstance (Eq a, forall b. Eq b => Eq (f b)) => Eq (Rose f a) where ...
Without the
forall b
, the type variableb
would be quantified over the whole instance declaration, which is not what is intended. 
One of these new quantified constraints can appear anywhere that any other constraint can, not just in instance declarations. Notably, it can appear in a type signature for a value binding, data constructor, or expression. For example
f :: (Eq a, forall b. Eq b => Eq (f b)) => Rose f a > Rose f a > Bool f t1 t2 = not (t1 == t2)

The form with a type variable at the head allows this:
instance (forall xx. c (Free c xx)) => Monad (Free c) where Free f >>= g = f g
See Iceland Jack’s summary. The key point is that the bit to the right of the
=>
may be headed by a type variable (c
in this case), rather than a class. It should not be one of the forall’d variables, though.(NB: this goes beyond what is described in the paper, but does not seem to introduce any new technical difficulties.)
10.15.3. Typing changes
See the paper.
10.15.4. Superclasses
Suppose we have:
f :: forall m. (forall a. Ord a => Ord (m a)) => m Int > Bool f x = x == x
From the x==x
we need an Eq (m Int)
constraint, but the context only gives us a way to figure out Ord (m a)
constraints. But from the given constraint forall a. Ord a => Ord (m a)
we derive a second given constraint forall a. Ord a => Eq (m a)
, and from that we can readily solve Eq (m Int)
. This process is very similar to the way that superclasses already work: given an Ord a
constraint we derive a second given Eq a
constraint.
NB: This treatment of superclasses goes beyond the paper, but is specifically desired by users.
10.15.5. Overlap
Quantified constraints can potentially lead to overlapping local axioms. Consider for instance the following example:
class A a where {} class B a where {} class C a where {} class (A a => C a) => D a where {} class (B a => C a) => E a where {} class C a => F a where {} instance (B a, D a, E a) => F a where {}
When type checking the instance declaration for F a
, we need to check that the superclass C
of F
holds. We thus try to entail the constraint C a
under the theory containing:
 The instance axioms :
(B a, D a, E a) => F a
 The local axioms from the instance context :
B a
,D a
andE a
 The closure of the superclass relation over these local axioms :
A a => C a
andB a => C a
However, the A a => C a
and B a => C a
axioms both match the wanted constraint C a
. There are several possible approaches for handling these overlapping local axioms:

Pick first. We can simply select the first matching axiom we encounter. In the above example, this would be
A a => C a
. We’d then need to entailA a
, for which we have no matching axioms available, causing the above program to be rejected.But suppose we made a slight adjustment to the order of the instance context, putting
E a
beforeD a
:instance (B a, E a, D a) => F a where {}
The first matching axiom we encounter while entailing
C a
, isB a => C a
. We have a local axiomB a
available, so now the program is suddenly accepted. This behaviour, where the ordering of an instance context determines whether or not the program is accepted, seems rather confusing for the developer.  Reject if in doubt. An alternative approach would be to check for overlapping axioms, when solving a constraint. When multiple matching axioms are discovered, we reject the program. This approach is a bit conservative, in that it may reject working programs. But it seem much more transparent towards the developer, who can be presented with a clear message, explaining why the program is rejected.

Backtracking. Lastly, a simple form of backtracking could be introduced. We simply select the first matching axiom we encounter and when the entailment fails, we backtrack and look for other axioms that might match the wanted constraint.
This seems the most intuitive and transparent approach towards the developer, who no longer needs to concern himself with the fact that his code might contain overlapping axioms or with the ordering of his instance contexts. But backtracking would apply equally to ordinary instance selection (in the presence of overlapping instances), so it is a much more pervasive change, with substantial consequences for the type inference engine.
GHC adopts Reject if in doubt for now. We can see how painful it is in practice, and try something more ambitious if necessary.
10.15.6. Instance lookup
In the light of the overlap decision, instance lookup works like this when trying to solve a class constraint C t
 First see if there is a given unquantified constraint
C t
. If so, use it to solve the constraint.  If not, look at all the available given quantified constraints; if exactly one one matches
C t
, choose it; if more than one matches, report an error.  If no quantified constraints match, look up in the global instances, as described in Instance resolution and Overlapping instances.
10.15.7. Termination
GHC uses the Paterson Conditions to ensure that instance resolution terminates. How are those rules modified for quantified constraints? In two ways.
 Each quantified constraint, taken by itself, must satisfy the termination rules for an instance declaration.
 After “for each class constraint
(C t1 ... tn)
”, add “or each quantified constraint(forall as. context => C t1 .. tn)
”
Note that the second item only at the head of the quantified constraint, not its context. Reason: the head is the new goal that has to be solved if we use the instance declaration.
Of course, UndecidableInstances
lifts the Paterson Conditions, as now.
10.15.8. Coherence
Although quantified constraints are a little like local instance declarations, they differ in one big way: the local instances are written by the compiler, not the user, and hence cannot introduce incoherence. Consider
f :: (forall a. Eq a => Eq (f a)) => f b > f Bool f x = ...rhs...
In ...rhs...
there is, in effect a local instance for Eq (f a)
for any a
. But at a call site for f
the compiler itself produces evidence to pass to f
. For example, if we called f Nothing
, then f
is Maybe
and the compiler must prove (at the call site) that forall a. Eq a => Eq (Maybe a)
holds. It can do this easily, by appealing to the existing instance declaration for Eq (Maybe a)
.
In short, quantifed constraints do not introduce incoherence.
10.16. Extensions to type signatures
10.16.1. Explicit universal quantification (forall)

ExplicitForAll

 Since

6.12.1
Allow use of the
forall
keyword in places where universal quantification is implicit.
Haskell type signatures are implicitly quantified. When the language option ExplicitForAll
is used, the keyword forall
allows us to say exactly what this means. For example:
g :: b > b
means this:
g :: forall b. (b > b)
The two are treated identically, except that the latter may bring type variables into scope (see Lexically scoped type variables).
This extension also enables explicit quantification of type and kind variables in Data instance declarations, Type instance declarations, Closed type families, Associated instances, and Rewrite rules.
Notes:

As well in type signatures, you can also use an explicit
forall
in an instance declaration:instance forall a. Eq a => Eq [a] where ...

If the
Wunusedforalls
flag is enabled, a warning will be emitted when you write a type variable in an explicitforall
statement that is otherwise unused. For instance:g :: forall a b. (b > b)
would warn about the unused type variable
a
.
10.16.2. The context of a type signature
The FlexibleContexts
extension lifts the Haskell 98 restriction that the typeclass constraints in a type signature must have the form (class typevariable) or (class (typevariable type1 type2 … typen)). With FlexibleContexts
these type signatures are perfectly okay
g :: Eq [a] => ... g :: Ord (T a ()) => ...
The flag FlexibleContexts
also lifts the corresponding restriction on class declarations (The superclasses of a class declaration) and instance declarations (Relaxed rules for instance contexts).
10.16.3. Ambiguous types and the ambiguity check

AllowAmbiguousTypes

 Since

7.8.1
Allow type signatures which appear that they would result in an unusable binding.
Each userwritten type signature is subjected to an ambiguity check. The ambiguity check rejects functions that can never be called; for example:
f :: C a => Int
The idea is there can be no legal calls to f
because every call will give rise to an ambiguous constraint. Indeed, the only purpose of the ambiguity check is to report functions that cannot possibly be called. We could soundly omit the ambiguity check on type signatures entirely, at the expense of delaying ambiguity errors to call sites. Indeed, the language extension AllowAmbiguousTypes
switches off the ambiguity check.
Ambiguity can be subtle. Consider this example which uses functional dependencies:
class D a b  a > b where .. h :: D Int b => Int
The Int
may well fix b
at the call site, so that signature should not be rejected. Moreover, the dependencies might be hidden. Consider
class X a b where ... class D a b  a > b where ... instance D a b => X [a] b where... h :: X a b => a > a
Here h
’s type looks ambiguous in b
, but here’s a legal call:
...(h [True])...
That gives rise to a (X [Bool] beta)
constraint, and using the instance means we need (D Bool beta)
and that fixes beta
via D
’s fundep!
Behind all these special cases there is a simple guiding principle. Consider
f :: type f = ...blah... g :: type g = f
You would think that the definition of g
would surely typecheck! After all f
has exactly the same type, and g=f
. But in fact f
’s type is instantiated and the instantiated constraints are solved against the constraints bound by g
’s signature. So, in the case an ambiguous type, solving will fail. For example, consider the earlier definition f :: C a => Int
:
f :: C a => Int f = ...blah... g :: C a => Int g = f
In g
’s definition, we’ll instantiate to (C alpha)
and try to deduce (C alpha)
from (C a)
, and fail.
So in fact we use this as our definition of ambiguity: a type ty
is ambiguous if and only if ((undefined :: ty) :: ty)
would fail to typecheck. We use a very similar test for inferred types, to ensure that they too are unambiguous.
Switching off the ambiguity check. Even if a function has an ambiguous type according the “guiding principle”, it is possible that the function is callable. For example:
class D a b where ... instance D Bool b where ... strange :: D a b => a > a strange = ...blah... foo = strange True
Here strange
’s type is ambiguous, but the call in foo
is OK because it gives rise to a constraint (D Bool beta)
, which is soluble by the (D Bool b)
instance.
Another way of getting rid of the ambiguity at the call site is to use the TypeApplications
extension to specify the types. For example:
class D a b where h :: b instance D Int Int where ... main = print (h @Int @Int)
Here a
is ambiguous in the definition of D
but later specified to be Int
using type applications.
AllowAmbiguousTypes
allows you to switch off the ambiguity check. However, even with ambiguity checking switched off, GHC will complain about a function that can never be called, such as this one:
f :: (Int ~ Bool) => a > a
Sometimes AllowAmbiguousTypes
does not mix well with RankNTypes
. For example:
foo :: forall r. (forall i. (KnownNat i) => r) > r foo f = f @1 boo :: forall j. (KnownNat j) => Int boo = .... h :: Int h = foo boo
This program will be rejected as ambiguous because GHC will not unify the type variables j
and i
.
Unlike the previous examples, it is not currently possible to resolve the ambiguity manually by using TypeApplications
.
Note
A historical note. GHC used to impose some more restrictive and less principled conditions on type signatures. For type forall tv1..tvn (c1, ...,cn) => type
GHC used to require
 that each universally quantified type variable
tvi
must be “reachable” fromtype
, and  that every constraint
ci
mentions at least one of the universally quantified type variablestvi
. These adhoc restrictions are completely subsumed by the new ambiguity check.
10.16.4. Explicitlykinded quantification

KindSignatures

 Since

6.8.1
Allow explicit kind signatures on type variables.
Haskell infers the kind of each type variable. Sometimes it is nice to be able to give the kind explicitly as (machinechecked) documentation, just as it is nice to give a type signature for a function. On some occasions, it is essential to do so. For example, in his paper “Restricted Data Types in Haskell” (Haskell Workshop 1999) John Hughes had to define the data type:
data Set cxt a = Set [a]  Unused (cxt a > ())
The only use for the Unused
constructor was to force the correct kind for the type variable cxt
.
GHC now instead allows you to specify the kind of a type variable directly, wherever a type variable is explicitly bound, with the extension KindSignatures
.
This extension enables kind signatures in the following places:

data
declarations:data Set (cxt :: Type > Type) a = Set [a]

type
declarations:type T (f :: Type > Type) = f Int

class
declarations:class (Eq a) => C (f :: Type > Type) a where ...

forall
’s in type signatures:f :: forall (cxt :: Type > Type). Set cxt Int
The parentheses are required.
As part of the same extension, you can put kind annotations in types as well. Thus:
f :: (Int :: Type) > Int g :: forall a. a > (a :: Type)
The syntax is
atype ::= '(' ctype '::' kind ')
The parentheses are required.
10.17. Lexically scoped type variables

ScopedTypeVariables

 Implies
 Since

6.8.1
Enable lexical scoping of type variables explicitly introduced with
forall
.
Tip
ScopedTypeVariables
breaks GHC’s usual rule that explicit forall
is optional and doesn’t affect semantics. For the Declaration type signatures (or Expression type signatures) examples in this section, the explicit forall
is required. (If omitted, usually the program will not compile; in a few cases it will compile but the functions get a different signature.) To trigger those forms of ScopedTypeVariables
, the forall
must appear against the toplevel signature (or outer expression) but not against nested signatures referring to the same type variables.
Explicit forall
is not always required – see pattern signature equivalent for the example in this section, or Pattern type signatures.
GHC supports lexically scoped type variables, without which some type signatures are simply impossible to write. For example:
f :: forall a. [a] > [a] f xs = ys ++ ys where ys :: [a] ys = reverse xs
The type signature for f
brings the type variable a
into scope, because of the explicit forall
(Declaration type signatures). The type variables bound by a forall
scope over the entire definition of the accompanying value declaration. In this example, the type variable a
scopes over the whole definition of f
, including over the type signature for ys
. In Haskell 98 it is not possible to declare a type for ys
; a major benefit of scoped type variables is that it becomes possible to do so.
An equivalent form for that example, avoiding explicit forall
uses Pattern type signatures:
f :: [a] > [a] f (xs :: [aa]) = xs ++ ys where ys :: [aa] ys = reverse xs
Unlike the forall
form, type variable a
from f
’s signature is not scoped over f
’s equation(s). Type variable aa
bound by the pattern signature is scoped over the righthand side of f
’s equation. (Therefore there is no need to use a distinct type variable; using a
would be equivalent.)
10.17.1. Overview
The design follows the following principles
 A scoped type variable stands for a type variable, and not for a type. (This is a change from GHC’s earlier design.)
 Furthermore, distinct lexical type variables stand for distinct type variables. This means that every programmerwritten type signature (including one that contains free scoped type variables) denotes a rigid type; that is, the type is fully known to the type checker, and no inference is involved.
 Lexical type variables may be alpharenamed freely, without changing the program.
A lexically scoped type variable can be bound by:
 A declaration type signature (Declaration type signatures)
 An expression type signature (Expression type signatures)
 A pattern type signature (Pattern type signatures)
 Class and instance declarations (Class and instance declarations)
In Haskell, a programmerwritten type signature is implicitly quantified over its free type variables (Section 4.1.2 of the Haskell Report). Lexically scoped type variables affect this implicit quantification rules as follows: any type variable that is in scope is not universally quantified. For example, if type variable a
is in scope, then
(e :: a > a) means (e :: a > a) (e :: b > b) means (e :: forall b. b>b) (e :: a > b) means (e :: forall b. a>b)
10.17.2. Declaration type signatures
A declaration type signature that has explicit quantification (using forall
) brings into scope the explicitlyquantified type variables, in the definition of the named function. For example:
f :: forall a. [a] > [a] f (x:xs) = xs ++ [ x :: a ]
The “forall a
” brings “a
” into scope in the definition of “f
”.
This only happens if:

The quantification in
f
’s type signature is explicit. For example:g :: [a] > [a] g (x:xs) = xs ++ [ x :: a ]
This program will be rejected, because “
a
” does not scope over the definition of “g
”, so “x::a
” means “x::forall a. a
” by Haskell’s usual implicit quantification rules. 
The type variable is quantified by the single, syntactically visible, outermost
forall
of the type signature. For example, GHC will reject all of the following examples:f1 :: forall a. forall b. a > [b] > [b] f1 _ (x:xs) = xs ++ [ x :: b ] f2 :: forall a. a > forall b. [b] > [b] f2 _ (x:xs) = xs ++ [ x :: b ] type Foo = forall b. [b] > [b] f3 :: Foo f3 (x:xs) = xs ++ [ x :: b ]
In
f1
andf2
, the type variableb
is not quantified by the outermostforall
, so it is not in scope over the bodies of the functions. Neither isb
in scope over the body off3
, as theforall
is tucked underneath theFoo
type synonym. 
The signature gives a type for a function binding or a bare variable binding, not a pattern binding. For example:
f1 :: forall a. [a] > [a] f1 (x:xs) = xs ++ [ x :: a ]  OK f2 :: forall a. [a] > [a] f2 = \(x:xs) > xs ++ [ x :: a ]  OK f3 :: forall a. [a] > [a] Just f3 = Just (\(x:xs) > xs ++ [ x :: a ])  Not OK!
f1
is a function binding, andf2
binds a bare variable; in both cases the type signature bringsa
into scope. However the binding forf3
is a pattern binding, and sof3
is a fresh variable brought into scope by the pattern, not connected with top levelf3
. Then type variablea
is not in scope of the righthand side ofJust f3 = ...
.
10.17.3. Expression type signatures
An expression type signature that has explicit quantification (using forall
) brings into scope the explicitlyquantified type variables, in the annotated expression. For example:
f = runST ( (op >>= \(x :: STRef s Int) > g x) :: forall s. ST s Bool )
Here, the type signature forall s. ST s Bool
brings the type variable s
into scope, in the annotated expression (op >>= \(x :: STRef s Int) > g x)
.
10.17.4. Pattern type signatures
A type signature may occur in any pattern; this is a pattern type signature. For example:
 f and g assume that 'a' is already in scope f = \(x::Int, y::a) > x g (x::a) = x h ((x,y) :: (Int,Bool)) = (y,x)
In the case where all the type variables in the pattern type signature are already in scope (i.e. bound by the enclosing context), matters are simple: the signature simply constrains the type of the pattern in the obvious way.
Unlike expression and declaration type signatures, pattern type signatures are not implicitly generalised. The pattern in a pattern binding may only mention type variables that are already in scope. For example:
f :: forall a. [a] > (Int, [a]) f xs = (n, zs) where (ys::[a], n) = (reverse xs, length xs)  OK (zs::[a]) = xs ++ ys  OK Just (v::b) = ...  Not OK; b is not in scope
Here, the pattern signatures for ys
and zs
are fine, but the one for v
is not because b
is not in scope.
However, in all patterns other than pattern bindings, a pattern type signature may mention a type variable that is not in scope; in this case, the signature brings that type variable into scope. For example:
 same f and g as above, now assuming that 'a' is not already in scope f = \(x::Int, y::a) > x  'a' is in scope on RHS of > g (x::a) = x :: a hh (Just (v :: b)) = v :: b
The pattern type signature makes the type variable available on the righthand side of the equation.
Bringing type variables into scope is particularly important for existential data constructors. For example:
data T = forall a. MkT [a] k :: T > T k (MkT [t::a]) = MkT t3 where (t3::[a]) = [t,t,t]
Here, the pattern type signature [t::a]
mentions a lexical type variable that is not already in scope. Indeed, it must not already be in scope, because it is bound by the pattern match. The effect is to bring it into scope, standing for the existentiallybound type variable.
It does seem odd that the existentiallybound type variable must not be already in scope. Contrast that usually namebindings merely shadow (make a ‘hole’) in a samenamed outer variable’s scope. But we must have some way to bring such type variables into scope, else we could not name existentiallybound type variables in subsequent type signatures.
Compare the two (identical) definitions for examples f
, g
; they are both legal whether or not a
is already in scope. They differ in that if a
is already in scope, the signature constrains the pattern, rather than the pattern binding the variable.
10.17.5. Class and instance declarations
The type variables in the head of a class
or instance
declaration scope over the methods defined in the where
part. You do not even need an explicit forall
(although you are allowed an explicit forall
in an instance
declaration; see Explicit universal quantification (forall)). For example:
class C a where op :: [a] > a op xs = let ys::[a] ys = reverse xs in head ys instance C b => C [b] where op xs = reverse (head (xs :: [[b]]))
10.18. Bindings and generalisation
10.18.1. Switching off the dreaded Monomorphism Restriction

NoMonomorphismRestriction

 Default

on
 Since

6.8.1
Prevents the compiler from applying the monomorphism restriction to bindings lacking explicit type signatures.
Haskell’s monomorphism restriction (see Section 4.5.5 of the Haskell Report) can be completely switched off by NoMonomorphismRestriction
. Since GHC 7.8.1, the monomorphism restriction is switched off by default in GHCi’s interactive options (see Setting options for interactive evaluation only).
10.18.2. Letgeneralisation

MonoLocalBinds

 Since

6.12.1
Infer less polymorphic types for local bindings by default.
An MLstyle language usually generalises the type of any let
bound or where
bound variable, so that it is as polymorphic as possible. With the extension MonoLocalBinds
GHC implements a slightly more conservative policy, using the following rules:

A variable is closed if and only if
 the variable is letbound

one of the following holds:
 the variable has an explicit type signature that has no free type variables, or
 its binding group is fully generalised (see next bullet)

A binding group is fully generalised if and only if
 each of its free variables is either imported or closed, and
 the binding is not affected by the monomorphism restriction (Haskell Report, Section 4.5.5)
For example, consider
f x = x + 1 g x = let h y = f y * 2 k z = z+x in h x + k x
Here f
is generalised because it has no free variables; and its binding group is unaffected by the monomorphism restriction; and hence f
is closed. The same reasoning applies to g
, except that it has one closed free variable, namely f
. Similarly h
is closed, even though it is not bound at top level, because its only free variable f
is closed. But k
is not closed, because it mentions x
which is not closed (because it is not letbound).
Notice that a toplevel binding that is affected by the monomorphism restriction is not closed, and hence may in turn prevent generalisation of bindings that mention it.
The rationale for this more conservative strategy is given in the papers “Let should not be generalised” and “Modular type inference with local assumptions”, and a related blog post.
The extension MonoLocalBinds
is implied by TypeFamilies
and GADTs
. You can switch it off again with NoMonoLocalBinds
but type inference becomes less predictable if you do so. (Read the papers!)
10.19. Visible type application

TypeApplications

 Since

8.0.1
Allow the use of type application syntax.
The TypeApplications
extension allows you to use visible type application in expressions. Here is an example: show (read @Int "5")
. The @Int
is the visible type application; it specifies the value of the type variable in read
’s type.
A visible type application is preceded with an @
sign. (To disambiguate the syntax, the @
must be preceded with a nonidentifier letter, usually a space. For example, read@Int 5
would not parse.) It can be used whenever the full polymorphic type of the function is known. If the function is an identifier (the common case), its type is considered known only when the identifier has been given a type signature. If the identifier does not have a type signature, visible type application cannot be used.
GHC also permits visible kind application, where users can declare the kind arguments to be instantiated in kindpolymorphic cases. Its usage parallels visible type application in the term level, as specified above.
10.19.1. Inferred vs. specified type variables
GHC tracks a distinction between what we call inferred and specified type variables. Only specified type variables are available for instantiation with visible type application. An example illustrates this well:
f :: (Eq b, Eq a) => a > b > Bool f x y = (x == x) && (y == y) g x y = (x == x) && (y == y)
The functions f
and g
have the same body, but only f
is given a type signature. When GHC is figuring out how to process a visible type application, it must know what variable to instantiate. It thus must be able to provide an ordering to the type variables in a function’s type.
If the user has supplied a type signature, as in f
, then this is easy: we just take the ordering from the type signature, going left to right and using the first occurrence of a variable to choose its position within the ordering. Thus, the variables in f
will be b
, then a
.
In contrast, there is no reliable way to do this for g
; we will not know whether Eq a
or Eq b
will be listed first in the constraint in g
’s type. In order to have visible type application be robust between releases of GHC, we thus forbid its use with g
.
We say that the type variables in f
are specified, while those in g
are inferred. The general rule is this: if the user has written a type variable in the source program, it is specified; if not, it is inferred.
This rule applies in datatype declarations, too. For example, if we have data Proxy a = Proxy
(and PolyKinds
is enabled), then a
will be assigned kind k
, where k
is a fresh kind variable. Because k
was not written by the user, it will be unavailable for type application in the type of the constructor Proxy
; only the a
will be available.
When fprintexplicitforalls
is enabled, inferred variables are printed in braces. Thus, the type of the data constructor Proxy
from the previous example would be forall {k} (a :: k). Proxy a
. We can observe this behavior in a GHCi session:
> :set XTypeApplications fprintexplicitforalls > let myLength1 :: Foldable f => f a > Int; myLength1 = length > :type +v myLength1 myLength1 :: forall (f :: * > *) a. Foldable f => f a > Int > let myLength2 = length > :type +v myLength2 myLength2 :: forall {a} {t :: * > *}. Foldable t => t a > Int > :type +v myLength2 @[] <interactive>:1:1: error: • Cannot apply expression of type ‘t0 a0 > Int’ to a visible type argument ‘[]’ • In the expression: myLength2 @[]
Notice that since myLength1
was defined with an explicit type signature, :type +v
reports that all of its type variables are available for type application. On the other hand, myLength2
was not given a type signature. As a result, all of its type variables are surrounded with braces, and trying to use visible type application with myLength2
fails.
Also note the use of :type +v
in the GHCi session above instead of :type
. This is because :type
gives you the type that would be inferred for a variable assigned to the expression provided (that is, the type of x
in let x = <expr>
). As we saw above with myLength2
, this type will have no variables available to visible type application. On the other hand, :type +v
gives you the actual type of the expression provided. To illustrate this:
> :type myLength1 myLength1 :: forall {a} {f :: * > *}. Foldable f => f a > Int > :type myLength2 myLength2 :: forall {a} {t :: * > *}. Foldable t => t a > Int
Using :type
might lead one to conclude that none of the type variables in myLength1
’s type signature are available for type application. This isn’t true, however! Be sure to use :type +v
if you want the most accurate information with respect to visible type application properties.
10.19.2. Ordering of specified variables
In the simple case of the previous section, we can say that specified variables appear in lefttoright order. However, not all cases are so simple. Here are the rules in the subtler cases:
 If an identifier’s type has a
forall
, then the order of type variables as written in theforall
is retained. 
If any of the variables depend on other variables (that is, if some of the variables are kind variables), the variables are reordered so that kind variables come before type variables, preserving the lefttoright order as much as possible. That is, GHC performs a stable topological sort on the variables. Example:
h :: Proxy (a :: (j, k)) > Proxy (b :: Proxy a) > ()  as if h :: forall j k a b. ...
In this example,
a
depends onj
andk
, andb
depends ona
. Even thougha
appears lexically beforej
andk
,j
andk
are quantified first, becausea
depends onj
andk
. Note further thatj
andk
are not reordered with respect to each other, even though doing so would not violate dependency conditions.A “stable topological sort” here, we mean that we perform this algorithm (which we call ScopedSort):
 Work lefttoright through the input list of type variables, with a cursor.
 If variable
v
at the cursor is depended on by any earlier variablew
, movev
immediately before the leftmost suchw
.
 Class methods’ type arguments include the class type variables, followed by any variables an individual method is polymorphic in. So,
class Monad m where return :: a > m a
means thatreturn
’s type arguments arem, a
.  With the
RankNTypes
extension (Lexically scoped type variables), it is possible to declare type arguments somewhere other than the beginning of a type. For example, we can havepair :: forall a. a > forall b. b > (a, b)
and then saypair @Bool True @Char
which would have typeChar > (Bool, Char)
.  Partial type signatures (Partial Type Signatures) work nicely with visible type application. If you want to specify only the second type argument to
wurble
, then you can saywurble @_ @Int
. The first argument is a wildcard, just like in a partial type signature. However, if used in a visible type application/visible kind application, it is not necessary to specifyPartialTypeSignatures
and your code will not generate a warning informing you of the omitted type.
The section in this manual on kind polymorphism describes how variables in type and class declarations are ordered (Inferring the order of variables in a type/class declaration).
10.20. Implicit parameters

ImplicitParams

 Since

6.8.1
Allow definition of functions expecting implicit parameters.
Implicit parameters are implemented as described in [Lewis2000] and enabled with the option ImplicitParams
. (Most of the following, still rather incomplete, documentation is due to Jeff Lewis.)

Lewis2000

“Implicit parameters: dynamic scoping with static types”, J Lewis, MB Shields, E Meijer, J Launchbury, 27th ACM Symposium on Principles of Programming Languages (POPL’00), Boston, Jan 2000.
A variable is called dynamically bound when it is bound by the calling context of a function and statically bound when bound by the callee’s context. In Haskell, all variables are statically bound. Dynamic binding of variables is a notion that goes back to Lisp, but was later discarded in more modern incarnations, such as Scheme. Dynamic binding can be very confusing in an untyped language, and unfortunately, typed languages, in particular HindleyMilner typed languages like Haskell, only support static scoping of variables.
However, by a simple extension to the type class system of Haskell, we can support dynamic binding. Basically, we express the use of a dynamically bound variable as a constraint on the type. These constraints lead to types of the form (?x::t') => t
, which says “this function uses a dynamicallybound variable ?x
of type t'
”. For example, the following expresses the type of a sort function, implicitly parameterised by a comparison function named cmp
.
sort :: (?cmp :: a > a > Bool) => [a] > [a]
The dynamic binding constraints are just a new form of predicate in the type class system.
An implicit parameter occurs in an expression using the special form ?x
, where x
is any valid identifier (e.g. ord ?x
is a valid expression). Use of this construct also introduces a new dynamicbinding constraint in the type of the expression. For example, the following definition shows how we can define an implicitly parameterised sort function in terms of an explicitly parameterised sortBy
function:
sortBy :: (a > a > Bool) > [a] > [a] sort :: (?cmp :: a > a > Bool) => [a] > [a] sort = sortBy ?cmp
10.20.1. Implicitparameter type constraints
Dynamic binding constraints behave just like other type class constraints in that they are automatically propagated. Thus, when a function is used, its implicit parameters are inherited by the function that called it. For example, our sort
function might be used to pick out the least value in a list:
least :: (?cmp :: a > a > Bool) => [a] > a least xs = head (sort xs)
Without lifting a finger, the ?cmp
parameter is propagated to become a parameter of least
as well. With explicit parameters, the default is that parameters must always be explicit propagated. With implicit parameters, the default is to always propagate them.
An implicitparameter type constraint differs from other type class constraints in the following way: All uses of a particular implicit parameter must have the same type. This means that the type of (?x, ?x)
is (?x::a) => (a,a)
, and not (?x::a, ?x::b) => (a, b)
, as would be the case for type class constraints.
You can’t have an implicit parameter in the context of a class or instance declaration. For example, both these declarations are illegal:
class (?x::Int) => C a where ... instance (?x::a) => Foo [a] where ...
Reason: exactly which implicit parameter you pick up depends on exactly where you invoke a function. But the “invocation” of instance declarations is done behind the scenes by the compiler, so it’s hard to figure out exactly where it is done. Easiest thing is to outlaw the offending types.
Implicitparameter constraints do not cause ambiguity. For example, consider:
f :: (?x :: [a]) => Int > Int f n = n + length ?x g :: (Read a, Show a) => String > String g s = show (read s)
Here, g
has an ambiguous type, and is rejected, but f
is fine. The binding for ?x
at f
’s call site is quite unambiguous, and fixes the type a
.
10.20.2. Implicitparameter bindings
An implicit parameter is bound using the standard let
or where
binding forms. For example, we define the min
function by binding cmp
.
min :: Ord a => [a] > a min = let ?cmp = (<=) in least
A group of implicitparameter bindings may occur anywhere a normal group of Haskell bindings can occur, except at top level. That is, they can occur in a let
(including in a list comprehension, or donotation, or pattern guards), or a where
clause. Note the following points:
 An implicitparameter binding group must be a collection of simple bindings to implicitstyle variables (no functionstyle bindings, and no type signatures); these bindings are neither polymorphic or recursive.
 You may not mix implicitparameter bindings with ordinary bindings in a single
let
expression; use two nestedlet
s instead. (In the case ofwhere
you are stuck, since you can’t nestwhere
clauses.) 
You may put multiple implicitparameter bindings in a single binding group; but they are not treated as a mutually recursive group (as ordinary
let
bindings are). Instead they are treated as a nonrecursive group, simultaneously binding all the implicit parameter. The bindings are not nested, and may be reordered without changing the meaning of the program. For example, consider:f t = let { ?x = t; ?y = ?x+(1::Int) } in ?x + ?y
The use of
?x
in the binding for?y
does not “see” the binding for?x
, so the type off
isf :: (?x::Int) => Int > Int
10.20.3. Implicit parameters and polymorphic recursion
Consider these two definitions:
len1 :: [a] > Int len1 xs = let ?acc = 0 in len_acc1 xs len_acc1 [] = ?acc len_acc1 (x:xs) = let ?acc = ?acc + (1::Int) in len_acc1 xs  len2 :: [a] > Int len2 xs = let ?acc = 0 in len_acc2 xs len_acc2 :: (?acc :: Int) => [a] > Int len_acc2 [] = ?acc len_acc2 (x:xs) = let ?acc = ?acc + (1::Int) in len_acc2 xs
The only difference between the two groups is that in the second group len_acc
is given a type signature. In the former case, len_acc1
is monomorphic in its own righthand side, so the implicit parameter ?acc
is not passed to the recursive call. In the latter case, because len_acc2
has a type signature, the recursive call is made to the polymorphic version, which takes ?acc
as an implicit parameter. So we get the following results in GHCi:
Prog> len1 "hello" 0 Prog> len2 "hello" 5
Adding a type signature dramatically changes the result! This is a rather counterintuitive phenomenon, worth watching out for.
10.20.4. Implicit parameters and monomorphism
GHC applies the dreaded Monomorphism Restriction (section 4.5.5 of the Haskell Report) to implicit parameters. For example, consider:
f :: Int > Int f v = let ?x = 0 in let y = ?x + v in let ?x = 5 in y
Since the binding for y
falls under the Monomorphism Restriction it is not generalised, so the type of y
is simply Int
, not (?x::Int) => Int
. Hence, (f 9)
returns result 9
. If you add a type signature for y
, then y
will get type (?x::Int) => Int
, so the occurrence of y
in the body of the let
will see the inner binding of ?x
, so (f 9)
will return 14
.
10.21. Arbitraryrank polymorphism

RankNTypes

 Implies
 Since

6.8.1
Allow types of arbitrary rank.

Rank2Types

 Since

6.8.1
A deprecated alias of
RankNTypes
.
GHC’s type system supports arbitraryrank explicit universal quantification in types. For example, all the following types are legal:
f1 :: forall a b. a > b > a g1 :: forall a b. (Ord a, Eq b) => a > b > a f2 :: (forall a. a>a) > Int > Int g2 :: (forall a. Eq a => [a] > a > Bool) > Int > Int f3 :: ((forall a. a>a) > Int) > Bool > Bool f4 :: Int > (forall a. a > a)
Here, f1
and g1
are rank1 types, and can be written in standard Haskell (e.g. f1 :: a>b>a
). The forall
makes explicit the universal quantification that is implicitly added by Haskell.
The functions f2
and g2
have rank2 types; the forall
is on the left of a function arrow. As g2
shows, the polymorphic type on the left of the function arrow can be overloaded.
The function f3
has a rank3 type; it has rank2 types on the left of a function arrow.
The language option RankNTypes
(which implies ExplicitForAll
) enables higherrank types. That is, you can nest forall
s arbitrarily deep in function arrows. For example, a foralltype (also called a “type scheme”), including a typeclass context, is legal:
 On the left or right (see
f4
, for example) of a function arrow  As the argument of a constructor, or type of a field, in a data type declaration. For example, any of the
f1, f2, f3, g1, g2
above would be valid field type signatures.  As the type of an implicit parameter
 In a pattern type signature (see Lexically scoped type variables)
The RankNTypes
option is also required for any type with a forall
or context to the right of an arrow (e.g. f :: Int > forall a. a>a
, or g :: Int > Ord a => a > a
). Such types are technically rank 1, but are clearly not Haskell98, and an extra extension did not seem worth the bother.
In particular, in data
and newtype
declarations the constructor arguments may be polymorphic types of any rank; see examples in Examples. Note that the declared types are nevertheless always monomorphic. This is important because by default GHC will not instantiate type variables to a polymorphic type (Impredicative polymorphism).
The obsolete language option Rank2Types
is a synonym for RankNTypes
. They used to specify finer distinctions that GHC no longer makes.
10.21.1. Examples
These are examples of data
and newtype
declarations whose data constructors have polymorphic argument types:
data T a = T1 (forall b. b > b > b) a data MonadT m = MkMonad { return :: forall a. a > m a, bind :: forall a b. m a > (a > m b) > m b } newtype Swizzle = MkSwizzle (forall a. Ord a => [a] > [a])
The constructors have rank2 types:
T1 :: forall a. (forall b. b > b > b) > a > T a MkMonad :: forall m. (forall a. a > m a) > (forall a b. m a > (a > m b) > m b) > MonadT m MkSwizzle :: (forall a. Ord a => [a] > [a]) > Swizzle
In earlier versions of GHC, it was possible to omit the forall
in the type of the constructor if there was an explicit context. For example:
newtype Swizzle' = MkSwizzle' (Ord a => [a] > [a])
Since GHC 8.0 declarations such as MkSwizzle'
will cause an outofscope error.
As for type signatures, implicit quantification happens for nonoverloaded types too. So if you write this:
f :: (a > a) > a
it’s just as if you had written this:
f :: forall a. (a > a) > a
That is, since the type variable a
isn’t in scope, it’s implicitly universally quantified.
You construct values of types T1, MonadT, Swizzle
by applying the constructor to suitable values, just as usual. For example,
a1 :: T Int a1 = T1 (\xy>x) 3 a2, a3 :: Swizzle a2 = MkSwizzle sort a3 = MkSwizzle reverse a4 :: MonadT Maybe a4 = let r x = Just x b m k = case m of Just y > k y Nothing > Nothing in MkMonad r b mkTs :: (forall b. b > b > b) > a > [T a] mkTs f x y = [T1 f x, T1 f y]
The type of the argument can, as usual, be more general than the type required, as (MkSwizzle reverse)
shows. (reverse
does not need the Ord
constraint.)
When you use pattern matching, the bound variables may now have polymorphic types. For example:
f :: T a > a > (a, Char) f (T1 w k) x = (w k x, w 'c' 'd') g :: (Ord a, Ord b) => Swizzle > [a] > (a > b) > [b] g (MkSwizzle s) xs f = s (map f (s xs)) h :: MonadT m > [m a] > m [a] h m [] = return m [] h m (x:xs) = bind m x $ \y > bind m (h m xs) $ \ys > return m (y:ys)
In the function h
we use the record selectors return
and bind
to extract the polymorphic bind and return functions from the MonadT
data structure, rather than using pattern matching.
10.21.2. Type inference
In general, type inference for arbitraryrank types is undecidable. GHC uses an algorithm proposed by Odersky and Laufer (“Putting type annotations to work”, POPL’96) to get a decidable algorithm by requiring some help from the programmer. We do not yet have a formal specification of “some help” but the rule is this:
For a lambdabound or casebound variable, x, either the programmer provides an explicit polymorphic type for x, or GHC’s type inference will assume that x’s type has no foralls in it.
What does it mean to “provide” an explicit type for x? You can do that by giving a type signature for x directly, using a pattern type signature (Lexically scoped type variables), thus:
\ f :: (forall a. a>a) > (f True, f 'c')
Alternatively, you can give a type signature to the enclosing context, which GHC can “push down” to find the type for the variable:
(\ f > (f True, f 'c')) :: (forall a. a>a) > (Bool,Char)
Here the type signature on the expression can be pushed inwards to give a type signature for f. Similarly, and more commonly, one can give a type signature for the function itself:
h :: (forall a. a>a) > (Bool,Char) h f = (f True, f 'c')
You don’t need to give a type signature if the lambda bound variable is a constructor argument. Here is an example we saw earlier:
f :: T a > a > (a, Char) f (T1 w k) x = (w k x, w 'c' 'd')
Here we do not need to give a type signature to w
, because it is an argument of constructor T1
and that tells GHC all it needs to know.
10.21.3. Implicit quantification
GHC performs implicit quantification as follows. At the outermost level (only) of userwritten types, if and only if there is no explicit forall
, GHC finds all the type variables mentioned in the type that are not already in scope, and universally quantifies them. For example, the following pairs are equivalent:
f :: a > a f :: forall a. a > a g (x::a) = let h :: a > b > b h x y = y in ... g (x::a) = let h :: forall b. a > b > b h x y = y in ...
Notice that GHC always adds implicit quantifiers at the outermost level of a userwritten type; it does not find the innermost possible quantification point. For example:
f :: (a > a) > Int  MEANS f :: forall a. (a > a) > Int  NOT f :: (forall a. a > a) > Int g :: (Ord a => a > a) > Int  MEANS g :: forall a. (Ord a => a > a) > Int  NOT g :: (forall a. Ord a => a > a) > Int
If you want the latter type, you can write your forall
s explicitly. Indeed, doing so is strongly advised for rank2 types.
Sometimes there is no “outermost level”, in which case no implicit quantification happens:
data PackMap a b s t = PackMap (Monad f => (a > f b) > s > f t)
This is rejected because there is no “outermost level” for the types on the RHS (it would obviously be terrible to add extra parameters to PackMap
), so no implicit quantification happens, and the declaration is rejected (with “f
is out of scope”). Solution: use an explicit forall
:
data PackMap a b s t = PackMap (forall f. Monad f => (a > f b) > s > f t)
10.22. Impredicative polymorphism

ImpredicativeTypes

 Implies
 Since

6.10.1
Allow impredicative polymorphic types.
In general, GHC will only instantiate a polymorphic function at a monomorphic type (one with no foralls). For example,
runST :: (forall s. ST s a) > a id :: forall b. b > b foo = id runST  Rejected
The definition of foo
is rejected because one would have to instantiate id
’s type with b := (forall s. ST s a) > a
, and that is not allowed. Instantiating polymorphic type variables with polymorphic types is called impredicative polymorphism.
GHC has extremely flaky support for impredicative polymorphism, enabled with ImpredicativeTypes
. If it worked, this would mean that you could call a polymorphic function at a polymorphic type, and parameterise data structures over polymorphic types. For example:
f :: Maybe (forall a. [a] > [a]) > Maybe ([Int], [Char]) f (Just g) = Just (g [3], g "hello") f Nothing = Nothing
Notice here that the Maybe
type is parameterised by the polymorphic type (forall a. [a] > [a])
. However the extension should be considered highly experimental, and certainly unsupported. You are welcome to try it, but please don’t rely on it working consistently, or working the same in subsequent releases. See this wiki page for more details.
If you want impredicative polymorphism, the main workaround is to use a newtype wrapper. The id runST
example can be written using this workaround like this:
runST :: (forall s. ST s a) > a id :: forall b. b > b newtype Wrap a = Wrap { unWrap :: (forall s. ST s a) > a } foo :: (forall s. ST s a) > a foo = unWrap (id (Wrap runST))  Here id is called at monomorphic type (Wrap a)
10.23. Typed Holes
Typed holes are a feature of GHC that allows special placeholders written with a leading underscore (e.g., “_
”, “_foo
”, “_bar
”), to be used as expressions. During compilation these holes will generate an error message that describes which type is expected at the hole’s location, information about the origin of any free type variables, and a list of local bindings that might help fill the hole and bindings in scope that fit the type of the hole that might help fill the hole with actual code. Typed holes are always enabled in GHC.
The goal of typed holes is to help with writing Haskell code rather than to change the type system. Typed holes can be used to obtain extra information from the type checker, which might otherwise be hard to get. Normally, using GHCi, users can inspect the (inferred) type signatures of all toplevel bindings. However, this method is less convenient with terms that are not defined on toplevel or inside complex expressions. Holes allow the user to check the type of the term they are about to write.
For example, compiling the following module with GHC:
f :: a > a f x = _
will fail with the following error:
hole.hs:2:7: Found hole `_' with type: a Where: `a' is a rigid type variable bound by the type signature for f :: a > a at hole.hs:1:6 In the expression: _ In an equation for `f': f x = _ Relevant bindings include x :: a (bound at hole.hs:2:3) f :: a > a (bound at hole.hs:2:1) Valid hole fits include x :: a (bound at hole.hs:2:3)
Here are some more details:

A “
Found hole
” error usually terminates compilation, like any other type error. After all, you have omitted some code from your program. Nevertheless, you can run and test a piece of code containing holes, by using thefdefertypedholes
flag. This flag defers errors produced by typed holes until runtime, and converts them into compiletime warnings. These warnings can in turn be suppressed entirely byWnotypedholes
.The same behaviour for “
Variable out of scope
” errors, it terminates compilation by default. You can defer such errors by using thefdeferoutofscopevariables
flag. This flag defers errors produced by out of scope variables until runtime, and converts them into compiletime warnings. These warnings can in turn be suppressed entirely byWnodeferredoutofscopevariables
.The result is that a hole or a variable will behave like
undefined
, but with the added benefits that it shows a warning at compile time, and will show the same message if it gets evaluated at runtime. This behaviour follows that of thefdefertypeerrors
option, which impliesfdefertypedholes
andfdeferoutofscopevariables
. See Deferring type errors to runtime. 
All unbound identifiers are treated as typed holes, whether or not they start with an underscore. The only difference is in the error message:
cons z = z : True : _x : y
yields the errors
Foo.hs:3:21: error: Found hole: _x :: Bool Or perhaps ‘_x’ is misspelled, or not in scope In the first argument of ‘(:)’, namely ‘_x’ In the second argument of ‘(:)’, namely ‘_x : y’ In the second argument of ‘(:)’, namely ‘True : _x : y’ Relevant bindings include z :: Bool (bound at Foo.hs:3:6) cons :: Bool > [Bool] (bound at Foo.hs:3:1) Valid hole fits include z :: Bool (bound at mpt.hs:2:6) otherwise :: Bool (imported from ‘Prelude’ at mpt.hs:1:810 (and originally defined in ‘GHC.Base’)) False :: Bool (imported from ‘Prelude’ at mpt.hs:1:810 (and originally defined in ‘GHC.Types’)) True :: Bool (imported from ‘Prelude’ at mpt.hs:1:810 (and originally defined in ‘GHC.Types’)) maxBound :: forall a. Bounded a => a with maxBound @Bool (imported from ‘Prelude’ at mpt.hs:1:810 (and originally defined in ‘GHC.Enum’)) minBound :: forall a. Bounded a => a with minBound @Bool (imported from ‘Prelude’ at mpt.hs:1:810 (and originally defined in ‘GHC.Enum’)) Foo.hs:3:26: error: Variable not in scope: y :: [Bool]
More information is given for explicit holes (i.e. ones that start with an underscore), than for outofscope variables, because the latter are often unintended typos, so the extra information is distracting. If you want the detailed information, use a leading underscore to make explicit your intent to use a hole.

Unbound identifiers with the same name are never unified, even within the same function, but shown individually. For example:
cons = _x : _x
results in the following errors:
unbound.hs:1:8: Found hole '_x' with type: a Where: `a' is a rigid type variable bound by the inferred type of cons :: [a] at unbound.hs:1:1 In the first argument of `(:)', namely `_x' In the expression: _x : _x In an equation for `cons': cons = _x : _x Relevant bindings include cons :: [a] (bound at unbound.hs:1:1) unbound.hs:1:13: Found hole: _x :: [a] Where: ‘a’ is a rigid type variable bound by the inferred type of cons :: [a] at unbound.hs:3:112 Or perhaps ‘_x’ is misspelled, or not in scope In the second argument of ‘(:)’, namely ‘_x’ In the expression: _x : _x In an equation for ‘cons’: cons = _x : _x Relevant bindings include cons :: [a] (bound at unbound.hs:3:1) Valid hole fits include cons :: forall a. [a] with cons @a (defined at mpt.hs:3:1) mempty :: forall a. Monoid a => a with mempty @[a] (imported from ‘Prelude’ at mpt.hs:1:810 (and originally defined in ‘GHC.Base’))
Notice the two different types reported for the two different occurrences of
_x
. 
No language extension is required to use typed holes. The lexeme “
_
” was previously illegal in Haskell, but now has a more informative error message. The lexeme “_x
” is a perfectly legal variable, and its behaviour is unchanged when it is in scope. For examplef _x = _x + 1
does not elict any errors. Only a variable that is not in scope (whether or not it starts with an underscore) is treated as an error (which it always was), albeit now with a more informative error message.
 Unbound data constructors used in expressions behave exactly as above. However, unbound data constructors used in patterns cannot be deferred, and instead bring compilation to a halt. (In implementation terms, they are reported by the renamer rather than the type checker.)

The list of valid hole fits is found by checking which bindings in scope would fit into the hole. As an example, compiling the following module with GHC:
import Data.List (inits) g :: [String] g = _ "hello, world"
yields the errors:
• Found hole: _ :: [Char] > [String] • In the expression: _ In the expression: _ "hello, world" In an equation for ‘g’: g = _ "hello, world" • Relevant bindings include g :: [String] (bound at mpt.hs:6:1) Valid hole fits include lines :: String > [String] (imported from ‘Prelude’ at mpt.hs:3:89 (and originally defined in ‘base4.11.0.0:Data.OldList’)) words :: String > [String] (imported from ‘Prelude’ at mpt.hs:3:89 (and originally defined in ‘base4.11.0.0:Data.OldList’)) inits :: forall a. [a] > [[a]] with inits @Char (imported from ‘Data.List’ at mpt.hs:4:1923 (and originally defined in ‘base4.11.0.0:Data.OldList’)) repeat :: forall a. a > [a] with repeat @String (imported from ‘Prelude’ at mpt.hs:3:89 (and originally defined in ‘GHC.List’)) fail :: forall (m :: * > *). Monad m => forall a. String > m a with fail @[] @String (imported from ‘Prelude’ at mpt.hs:3:89 (and originally defined in ‘GHC.Base’)) return :: forall (m :: * > *). Monad m => forall a. a > m a with return @[] @String (imported from ‘Prelude’ at mpt.hs:3:89 (and originally defined in ‘GHC.Base’)) pure :: forall (f :: * > *). Applicative f => forall a. a > f a with pure @[] @String (imported from ‘Prelude’ at mpt.hs:3:89 (and originally defined in ‘GHC.Base’)) read :: forall a. Read a => String > a with read @[String] (imported from ‘Prelude’ at mpt.hs:3:89 (and originally defined in ‘Text.Read’)) mempty :: forall a. Monoid a => a with mempty @([Char] > [String]) (imported from ‘Prelude’ at mpt.hs:3:89 (and originally defined in ‘GHC.Base’))
There are a few flags for controlling the amount of context information shown for typed holes:

fshowholeconstraints

When reporting typed holes, also print constraints that are in scope. Example:
f :: Eq a => a > Bool f x = _
results in the following message:
show_constraints.hs:4:7: error: • Found hole: _ :: Bool • In the expression: _ In an equation for ‘f’: f x = _ • Relevant bindings include x :: a (bound at show_constraints.hs:4:3) f :: a > Bool (bound at show_constraints.hs:4:1) Constraints include Eq a (from show_constraints.hs:3:122) Valid hole fits include otherwise :: Bool False :: Bool True :: Bool maxBound :: forall a. Bounded a => a with maxBound @Bool minBound :: forall a. Bounded a => a with minBound @Bool
10.23.1. Valid Hole Fits
GHC sometimes suggests valid hole fits for typed holes, which is configurable by a few flags.

fnoshowvalidholefits

 Default

off
This flag can be toggled to turn off the display of valid hole fits entirely.

fmaxvalidholefits=⟨n⟩

 Default

6
The list of valid hole fits is limited by displaying up to 6 hole fits per hole. The number of hole fits shown can be set by this flag. Turning the limit off with
fnomaxvalidholefits
displays all found hole fits.

fshowtypeofholefits

 Default

on
By default, the hole fits show the type of the hole fit. This can be turned off by the reverse of this flag.

fshowtypeappofholefits

 Default

on
By default, the hole fits show the type application needed to make this hole fit fit the type of the hole, e.g. for the hole
(_ :: Int > [Int])
,mempty
is a hole fit withmempty @(Int > [Int])
. This can be toggled off with the reverse of this flag.

fshowdocsofholefits

 Default

off
It can sometime be the case that the name and type of a valid hole fit is not enough to realize what the fit stands for. This flag adds the documentation of the fit to the message, if the documentation is available (and the module from which the function comes was compiled with the
haddock
flag).

fshowtypeappvarsofholefits

 Default

on
By default, the hole fits show the type application needed to make this hole fit fit the type of the hole, e.g. for the hole
(_ :: Int > [Int])
,mempty :: Monoid a => a
is a hole fit withmempty @(Int > [Int])
. This flag toggles whether to showa ~ (Int > [Int])
instead ofmempty @(Int > [Int])
in the where clause of the valid hole fit message.

fshowprovenanceofholefits

 Default

on
By default, each hole fit shows the provenance information of its hole fit, i.e. where it was bound or defined, and what module it was originally defined in if it was imported. This can be toggled off using the reverse of this flag.

funcluttervalidholefits

 Default

off
This flag can be toggled to decrease the verbosity of the valid hole fit suggestions by not showing the provenance nor type application of the suggestions.
10.23.1.1. Refinement Hole Fits
When the flag frefinementlevelholefits=⟨n⟩
is set to an n
larger than 0
, GHC will offer up a list of valid refinement hole fits, which are valid hole fits that need up to n
levels of additional refinement to be complete, where each level represents an additional hole in the hole fit that requires filling in. As an example, consider the hole in
f :: [Integer] > Integer f = _
When the refinement level is not set, it will only offer valid hole fits suggestions:
Valid hole fits include f :: [Integer] > Integer head :: forall a. [a] > a with head @Integer last :: forall a. [a] > a with last @Integer maximum :: forall (t :: * > *). Foldable t => forall a. Ord a => t a > a with maximum @[] @Integer minimum :: forall (t :: * > *). Foldable t => forall a. Ord a => t a > a with minimum @[] @Integer product :: forall (t :: * > *). Foldable t => forall a. Num a => t a > a with product @[] @Integer sum :: forall (t :: * > *). Foldable t => forall a. Num a => t a > a with sum @[] @Integer
However, with frefinementlevelholefits=⟨n⟩
set to e.g. 1
, it will additionally offer up a list of refinement hole fits, in this case:
Valid refinement hole fits include foldl1 (_ :: Integer > Integer > Integer) with foldl1 @[] @Integer where foldl1 :: forall (t :: * > *). Foldable t => forall a. (a > a > a) > t a > a foldr1 (_ :: Integer > Integer > Integer) with foldr1 @[] @Integer where foldr1 :: forall (t :: * > *). Foldable t => forall a. (a > a > a) > t a > a const (_ :: Integer) with const @Integer @[Integer] where const :: forall a b. a > b > a ($) (_ :: [Integer] > Integer) with ($) @'GHC.Types.LiftedRep @[Integer] @Integer where ($) :: forall a b. (a > b) > a > b fail (_ :: String) with fail @((>) [Integer]) @Integer where fail :: forall (m :: * > *). Monad m => forall a. String > m a return (_ :: Integer) with return @((>) [Integer]) @Integer where return :: forall (m :: * > *). Monad m => forall a. a > m a (Some refinement hole fits suppressed; use fmaxrefinementholefits=N or fnomaxrefinementholefits)
Which shows that the hole could be replaced with e.g. foldl1 _
. While not fixing the hole, this can help users understand what options they have.

frefinementlevelholefits=⟨n⟩

 Default

off
The list of valid refinement hole fits is generated by considering hole fits with a varying amount of additional holes. The amount of holes in a refinement can be set by this flag. If the flag is set to 0 or not set at all, no valid refinement hole fits will be suggested.

fabstractrefinementholefits

 Default

off
Valid list of valid refinement hole fits can often grow large when the refinement level is
>= 2
, with holes likehead _ _
orfst _ _
, which are valid refinements, but which are unlikely to be relevant since one or more of the holes are still completely open, in that neither the type nor kind of those holes are constrained by the proposed identifier at all. By default, such holes are not reported. By turning this flag on, such holes are included in the list of valid refinement hole fits.

fmaxrefinementholefits=⟨n⟩

 Default

6
The list of valid refinement hole fits is limited by displaying up to 6 hole fits per hole. The number of hole fits shown can be set by this flag. Turning the limit off with
fnomaxrefinementholefits
displays all found hole fits.

fshowholematchesofholefits

 Default

on
The types of the additional holes in refinement hole fits are displayed in the output, e.g.
foldl1 (_ :: a > a > a)
is a refinement for the hole_ :: [a] > a
. If this flag is toggled off, the output will display onlyfoldl1 _
, which can be used as a direct replacement for the hole, without requiringXScopedTypeVariables
.
10.23.1.2. Sorting Valid Hole Fits
There are currently two ways to sort valid hole fits. Sorting can be toggled with fsortvalidholefits

fnosortvalidholefits

 Default

off
By default the valid hole fits are sorted to show the most relevant hole fits at the top of the list of valid hole fits. This can be toggled off with this flag.

fsortbysizeholefits

 Default

on
Sorts by how big the types the quantified type variables in the type of the function would have to be in order to match the type of the hole.

fsortbysubsumptionholefits

 Default

off
An alternative sort. Sorts by checking which hole fits subsume other hole fits, such that if hole fit a could be used as hole fits for hole fit b, then b appears before a in the output. It is more precise than the default sort, but also a lot slower, since a subsumption check has to be run for each pair of valid hole fits.
10.24. Partial Type Signatures

PartialTypeSignatures

 Since

7.10.1
Type checker will allow inferred types for holes.
A partial type signature is a type signature containing special placeholders called wildcards. A wildcard is written as an underscore (e.g. “_
”) or, if NamedWildCards
is enabled, any identifier with a leading underscore (e.g. “_foo
”, “_bar
”). Partial type signatures are to type signatures what Typed Holes are to expressions. During compilation these wildcards or holes will generate an error message that describes which type was inferred at the hole’s location, and information about the origin of any free type variables. GHC reports such error messages by default.
Unlike Typed Holes, which make the program incomplete and will generate errors when they are evaluated, this needn’t be the case for holes in type signatures. The type checker is capable (in most cases) of typechecking a binding with or without a type signature. A partial type signature bridges the gap between the two extremes, the programmer can choose which parts of a type to annotate and which to leave over to the typechecker to infer.
By default, the typechecker will report an error message for each hole in a partial type signature, informing the programmer of the inferred type. When the PartialTypeSignatures
extension is enabled, the typechecker will accept the inferred type for each hole, generating warnings instead of errors. Additionally, these warnings can be silenced with the Wnopartialtypesignatures
flag.
However, because GHC must infer the type when part of a type is left out, it is unable to use polymorphic recursion. The same restriction takes place when the type signature is omitted completely.
10.24.1. Syntax
A (partial) type signature has the following form: forall a b .. . (C1, C2, ..) => tau
. It consists of three parts:
 The type variables:
a b ..
 The constraints:
(C1, C2, ..)
 The (mono)type:
tau
We distinguish three kinds of wildcards.
10.24.1.1. Type Wildcards
Wildcards occurring within the monotype (tau) part of the type signature are type wildcards (“type” is often omitted as this is the default kind of wildcard). Type wildcards can be instantiated to any monotype like Bool
or Maybe [Bool]
, including functions and higherkinded types like (Int > Bool)
or Maybe
.
not' :: Bool > _ not' x = not x  Inferred: Bool > Bool maybools :: _ maybools = Just [True]  Inferred: Maybe [Bool] just1 :: _ Int just1 = Just 1  Inferred: Maybe Int filterInt :: _ > _ > [Int] filterInt = filter  has type forall a. (a > Bool) > [a] > [a]  Inferred: (Int > Bool) > [Int] > [Int]
For instance, the first wildcard in the type signature not'
would produce the following error message:
Test.hs:4:17: error: • Found type wildcard ‘_’ standing for ‘Bool’ To use the inferred type, enable PartialTypeSignatures • In the type signature: not' :: Bool > _ • Relevant bindings include not' :: Bool > Bool (bound at Test.hs:5:1)
When a wildcard is not instantiated to a monotype, it will be generalised over, i.e. replaced by a fresh type variable, e.g.
foo :: _ > _ foo x = x  Inferred: forall t. t > t filter' :: _ filter' = filter  has type forall a. (a > Bool) > [a] > [a]  Inferred: (a > Bool) > [a] > [a]
10.24.1.2. Named Wildcards

NamedWildCards

 Since

7.10.1
Allow naming of wildcards (e.g.
_x
) in type signatures.
Type wildcards can also be named by giving the underscore an identifier as suffix, i.e. _a
. These are called named wildcards. All occurrences of the same named wildcard within one type signature will unify to the same type. For example:
f :: _x > _x f ('c', y) = ('d', error "Urk")  Inferred: forall t. (Char, t) > (Char, t)
The named wildcard forces the argument and result types to be the same. Lacking a signature, GHC would have inferred forall a b. (Char, a) > (Char, b)
. A named wildcard can be mentioned in constraints, provided it also occurs in the monotype part of the type signature to make sure that it unifies with something:
somethingShowable :: Show _x => _x > _ somethingShowable x = show x  Inferred type: Show a => a > String somethingShowable' :: Show _x => _x > _ somethingShowable' x = show (not x)  Inferred type: Bool > String
Besides an extraconstraints wildcard (see ExtraConstraints Wildcard), only named wildcards can occur in the constraints, e.g. the _x
in Show _x
.
Named wildcards should not be confused with type variables. Even though syntactically similar, named wildcards can unify with monotypes as well as be generalised over (and behave as type variables).
In the first example above, _x
is generalised over (and is effectively replaced by a fresh type variable a
). In the second example, _x
is unified with the Bool
type, and as Bool
implements the Show
type class, the constraint Show Bool
can be simplified away.
By default, GHC (as the Haskell 2010 standard prescribes) parses identifiers starting with an underscore in a type as type variables. To treat them as named wildcards, the NamedWildCards
extension should be enabled. The example below demonstrated the effect.
foo :: _a > _a foo _ = False
Compiling this program without enabling NamedWildCards
produces the following error message complaining about the type variable _a
no matching the actual type Bool
.
Test.hs:5:9: error: • Couldn't match expected type ‘_a’ with actual type ‘Bool’ ‘_a’ is a rigid type variable bound by the type signature for: foo :: forall _a. _a > _a at Test.hs:4:8 • In the expression: False In an equation for ‘foo’: foo _ = False • Relevant bindings include foo :: _a > _a (bound at Test.hs:5:1)
Compiling this program with NamedWildCards
(as well as PartialTypeSignatures
) enabled produces the following error message reporting the inferred type of the named wildcard _a
.
Test.hs:4:8: warning: [Wpartialtypesignatures] • Found type wildcard ‘_a’ standing for ‘Bool’ • In the type signature: foo :: _a > _a • Relevant bindings include foo :: Bool > Bool (bound at Test.hs:5:1)
10.24.1.3. ExtraConstraints Wildcard
The third kind of wildcard is the extraconstraints wildcard. The presence of an extraconstraints wildcard indicates that an arbitrary number of extra constraints may be inferred during type checking and will be added to the type signature. In the example below, the extraconstraints wildcard is used to infer three extra constraints.
arbitCs :: _ => a > String arbitCs x = show (succ x) ++ show (x == x)  Inferred:  forall a. (Enum a, Eq a, Show a) => a > String  Error: Test.hs:5:12: error: Found constraint wildcard ‘_’ standing for ‘(Show a, Eq a, Enum a)’ To use the inferred type, enable PartialTypeSignatures In the type signature: arbitCs :: _ => a > String
An extraconstraints wildcard shouldn’t prevent the programmer from already listing the constraints he knows or wants to annotate, e.g.
 Also a correct partial type signature: arbitCs' :: (Enum a, _) => a > String arbitCs' x = arbitCs x  Inferred:  forall a. (Enum a, Show a, Eq a) => a > String  Error: Test.hs:9:22: error: Found constraint wildcard ‘_’ standing for ‘()’ To use the inferred type, enable PartialTypeSignatures In the type signature: arbitCs' :: (Enum a, _) => a > String
An extraconstraints wildcard can also lead to zero extra constraints to be inferred, e.g.
noCs :: _ => String noCs = "noCs"  Inferred: String  Error: Test.hs:13:9: error: Found constraint wildcard ‘_’ standing for ‘()’ To use the inferred type, enable PartialTypeSignatures In the type signature: noCs :: _ => String
As a single extraconstraints wildcard is enough to infer any number of constraints, only one is allowed in a type signature and it should come last in the list of constraints.
Extraconstraints wildcards cannot be named.
10.24.2. Where can they occur?
Partial type signatures are allowed for bindings, pattern and expression signatures, except that extraconstraints wildcards are not supported in pattern or expression signatures. In the following example a wildcard is used in each of the three possible contexts.
{# LANGUAGE ScopedTypeVariables #} foo :: _ foo (x :: _) = (x :: _)  Inferred: forall w_. w_ > w_
Anonymous and named wildcards can occur on the left hand side of a type or data instance declaration; see Wildcards on the LHS of data and type family instances.
Anonymous wildcards are also allowed in visible type applications/ visible kind applications (Visible type application). If you want to specify only the second type argument to wurble
, then you can say wurble @_ @Int
where the first argument is a wildcard.
Standalone deriving
declarations permit the use of a single, extraconstraints wildcard, like so:
deriving instance _ => Eq (Foo a)
This denotes a derived Eq (Foo a)
instance where the context is inferred, in much the same way that ordinary deriving
clauses do. Any other use of wildcards in a standalone deriving
declaration is prohibited.
In all other contexts, type wildcards are disallowed, and a named wildcard is treated as an ordinary type variable. For example:
class C _ where ...  Illegal instance Eq (T _)  Illegal (currently; would actually make sense) instance Eq _a => Eq (T _a)  Perfectly fine, same as Eq a => Eq (T a)
Partial type signatures can also be used in Template Haskell splices.

Declaration splices: partial type signature are fully supported.
{# LANGUAGE TemplateHaskell, NamedWildCards #} $( [d foo :: _ => _a > _a > _ foo x y = x == y] )

Expression splices: anonymous and named wildcards can be used in expression signatures. Extraconstraints wildcards are not supported, just like in regular expression signatures.
{# LANGUAGE TemplateHaskell, NamedWildCards #} $( [e foo = (Just True :: _m _) ] )
 Typed expression splices: the same wildcards as in (untyped) expression splices are supported.

Pattern splices: anonymous and named wildcards can be used in pattern signatures. Note that
ScopedTypeVariables
has to be enabled to allow pattern signatures. Extraconstraints wildcards are not supported, just like in regular pattern signatures.{# LANGUAGE TemplateHaskell, ScopedTypeVariables #} foo $( [p (x :: _) ] ) = x

Type splices: only anonymous wildcards are supported in type splices. Named and extraconstraints wildcards are not.
{# LANGUAGE TemplateHaskell #} foo :: $( [t _ ] ) > a foo x = x
10.25. Custom compiletime errors
When designing embedded domain specific languages in Haskell, it is useful to have something like error
at the type level. In this way, the EDSL designer may show a type error that is specific to the DSL, rather than the standard GHC type error.
For example, consider a type class that is not intended to be used with functions, but the user accidentally used it at a function type, perhaps because they missed an argument to some function. Then, instead of getting the standard GHC message about a missing instance, it would be nicer to emit a more friendly message specific to the EDSL. Similarly, the reduction of a typelevel function may get stuck due to an error, at which point it would be nice to report an EDSL specific error, rather than a generic error about an ambiguous type.
To solve this, GHC provides a single typelevel function,
type family TypeError (msg :: ErrorMessage) :: k
along with a small typelevel language (via DataKinds
) for constructing prettyprinted error messages,
 ErrorMessage is intended to be used as a kind data ErrorMessage = Text Symbol  Show this text as is  forall t. ShowType t  Pretty print a type  ErrorMessage :<>: ErrorMessage  Put two chunks of error message next to each other  ErrorMessage :$$: ErrorMessage  Put two chunks of error message above each other
in the GHC.TypeLits module.
For instance, we might use this interface to provide a more useful error message for applications of show
on unsaturated functions like this,
{# LANGUAGE DataKinds #} {# LANGUAGE TypeOperators #} {# LANGUAGE UndecidableInstances #} import GHC.TypeLits instance TypeError (Text "Cannot 'Show' functions." :$$: Text "Perhaps there is a missing argument?") => Show (a > b) where showsPrec = error "unreachable" main = print negate
Which will produce the following compiletime error,
Test.hs:12:8: error: • Cannot 'Show' functions. Perhaps there is a missing argument? • In the expression: print negate In an equation for ‘main’: main = print negate
10.26. Deferring type errors to runtime
While developing, sometimes it is desirable to allow compilation to succeed even if there are type errors in the code. Consider the following case:
module Main where a :: Int a = 'a' main = print "b"
Even though a
is illtyped, it is not used in the end, so if all that we’re interested in is main
it can be useful to be able to ignore the problems in a
.
For more motivation and details please refer to the Wiki page or the original paper.
10.26.1. Enabling deferring of type errors
The flag fdefertypeerrors
controls whether type errors are deferred to runtime. Type errors will still be emitted as warnings, but will not prevent compilation. You can use Wnodeferredtypeerrors
to suppress these warnings.
This flag implies the fdefertypedholes
and fdeferoutofscopevariables
flags, which enables this behaviour for typed holes and variables. Should you so wish, it is possible to enable fdefertypeerrors
without enabling fdefertypedholes
or fdeferoutofscopevariables
, by explicitly specifying fnodefertypedholes
or fnodeferoutofscopevariables
on the commandline after the fdefertypeerrors
flag.
At runtime, whenever a term containing a type error would need to be evaluated, the error is converted into a runtime exception of type TypeError
. Note that type errors are deferred as much as possible during runtime, but invalid coercions are never performed, even when they would ultimately result in a value of the correct type. For example, given the following code:
x :: Int x = 0 y :: Char y = x z :: Int z = y
evaluating z
will result in a runtime TypeError
.
10.26.2. Deferred type errors in GHCi
The flag fdefertypeerrors
works in GHCi as well, with one exception: for “naked” expressions typed at the prompt, type errors don’t get delayed, so for example:
Prelude> fst (True, 1 == 'a') <interactive>:2:12: No instance for (Num Char) arising from the literal `1' Possible fix: add an instance declaration for (Num Char) In the first argument of `(==)', namely `1' In the expression: 1 == 'a' In the first argument of `fst', namely `(True, 1 == 'a')'
Otherwise, in the common case of a simple type error such as typing reverse True
at the prompt, you would get a warning and then an immediatelyfollowing type error when the expression is evaluated.
This exception doesn’t apply to statements, as the following example demonstrates:
Prelude> let x = (True, 1 == 'a') <interactive>:3:16: Warning: No instance for (Num Char) arising from the literal `1' Possible fix: add an instance declaration for (Num Char) In the first argument of `(==)', namely `1' In the expression: 1 == 'a' In the expression: (True, 1 == 'a') Prelude> fst x True
10.26.3. Limitations of deferred type errors
The errors that can be deferred are:
 Out of scope term variables
 Equality constraints; e.g.
ord True
gives rise to an insoluble equality constraintChar ~ Bool
, which can be deferred.  Typeclass and implicitparameter constraints
All other type errors are reported immediately, and cannot be deferred; for example, an illkinded type signature, an instance declaration that is nonterminating or illformed, a typefamily instance that does not obey the declared injectivity constraints, etc etc.
In a few cases, even equality constraints cannot be deferred. Specifically:

Kindequalities cannot be deferred, e.g.
f :: Int Bool > Char
This type signature contains a kind error which cannot be deferred.
 Type equalities under a forall cannot be deferred (c.f. #14605).
10.27. Template Haskell
Template Haskell allows you to do compiletime metaprogramming in Haskell. The background to the main technical innovations is discussed in “Template Metaprogramming for Haskell” (Proc Haskell Workshop 2002).
The Template Haskell page on the GHC Wiki has a wealth of information. You may also consult the Haddock reference documentation <Language.Haskell.TH.>. Many changes to the original design are described in Notes on Template Haskell version 2. Not all of these changes are in GHC, however.
The first example from that paper is set out below (A Template Haskell Worked Example) as a worked example to help get you started.
The documentation here describes the realisation of Template Haskell in GHC. It is not detailed enough to understand Template Haskell; see the Wiki page.
10.27.1. Syntax

TemplateHaskell

 Implies
 Since

6.0. Typed splices introduced in GHC 7.8.1.
Enable Template Haskell’s splice and quotation syntax.

TemplateHaskellQuotes

 Since

8.0.1
Enable only Template Haskell’s quotation syntax.
Template Haskell has the following new syntactic constructions. You need to use the extension TemplateHaskell
to switch these syntactic extensions on. Alternatively, the TemplateHaskellQuotes
extension can be used to enable the quotation subset of Template Haskell (i.e. without splice syntax). The TemplateHaskellQuotes
extension is considered safe under Safe Haskell while TemplateHaskell
is not.

A splice is written
$x
, wherex
is an identifier, or$(...)
, where the “…” is an arbitrary expression. There must be no space between the “$” and the identifier or parenthesis. This use of “$” overrides its meaning as an infix operator, just as “M.x” overrides the meaning of “.” as an infix operator. If you want the infix operator, put spaces around it.A splice can occur in place of
 an expression; the spliced expression must have type
Q Exp
 a pattern; the spliced pattern must have type
Q Pat
 a type; the spliced expression must have type
Q Type
 a list of declarations at top level; the spliced expression must have type
Q [Dec]
Inside a splice you can only call functions defined in imported modules, not functions defined elsewhere in the same module. Note that declaration splices are not allowed anywhere except at top level (outside any other declarations).
 an expression; the spliced expression must have type

A expression quotation is written in Oxford brackets, thus:

[ ... ]
, or[e ... ]
, where the “…” is an expression; the quotation has typeQ Exp
. 
[d ... ]
, where the “…” is a list of toplevel declarations; the quotation has typeQ [Dec]
. 
[t ... ]
, where the “…” is a type; the quotation has typeQ Type
. 
[p ... ]
, where the “…” is a pattern; the quotation has typeQ Pat
.
See Where can they occur? for using partial type signatures in quotations.


A typed expression splice is written
$$x
, wherex
is an identifier, or$$(...)
, where the “…” is an arbitrary expression.A typed expression splice can occur in place of an expression; the spliced expression must have type
Q (TExp a)

A typed expression quotation is written as
[ ... ]
, or[e ... ]
, where the “…” is an expression; if the “…” expression has typea
, then the quotation has typeQ (TExp a)
.Values of type
TExp a
may be converted to values of typeExp
using the functionunType :: TExp a > Exp
. 
A quasiquotation can appear in a pattern, type, expression, or declaration context and is also written in Oxford brackets:

[varid ... ]
, where the “…” is an arbitrary string; a full description of the quasiquotation facility is given in Template Haskell Quasiquotation.


A name can be quoted with either one or two prefix single quotes:

'f
has typeName
, and names the functionf
. Similarly'C
has typeName
and names the data constructorC
. In general'
⟨thing⟩ interprets ⟨thing⟩ in an expression context.A name whose second character is a single quote (sadly) cannot be quoted in this way, because it will be parsed instead as a quoted character. For example, if the function is called
f'7
(which is a legal Haskell identifier), an attempt to quote it as'f'7
would be parsed as the character literal'f'
followed by the numeric literal7
. There is no current escape mechanism in this (unusual) situation. 
''T
has typeName
, and names the type constructorT
. That is,''
⟨thing⟩ interprets ⟨thing⟩ in a type context.
These
Names
can be used to construct Template Haskell expressions, patterns, declarations etc. They may also be given as an argument to thereify
function. 

It is possible for a splice to expand to an expression that contain names which are not in scope at the site of the splice. As an example, consider the following code:
module Bar where import Language.Haskell.TH add1 :: Int > Q Exp add1 x = [ x + 1 ]
Now consider a splice using
add1
in a separate module:module Foo where import Bar two :: Int two = $(add1 1)
Template Haskell cannot know what the argument to
add1
will be at the function’s definition site, so a lifting mechanism is used to promotex
into a value of typeQ Exp
. This functionality is exposed to the user as theLift
typeclass in theLanguage.Haskell.TH.Syntax
module. If a type has aLift
instance, then any of its values can be lifted to a Template Haskell expression:class Lift t where lift :: t > Q Exp
In general, if GHC sees an expression within Oxford brackets (e.g.,
[ foo bar ]
, then GHC looks up each name within the brackets. If a name is global (e.g., supposefoo
comes from an import or a toplevel declaration), then the fully qualified name is used directly in the quotation. If the name is local (e.g., supposebar
is bound locally in the function definitionmkFoo bar = [ foo bar ]
), then GHC useslift
on it (so GHC pretends[ foo bar ]
actually contains[ foo $(lift bar) ]
). Local names, which are not in scope at splice locations, are actually evaluated when the quotation is processed.The
templatehaskell
library providesLift
instances for many common data types. Furthermore, it is possible to deriveLift
instances automatically by using theDeriveLift
language extension. See Deriving Lift instances for more information. 
You may omit the
$(...)
in a toplevel declaration splice. Simply writing an expression (rather than a declaration) implies a splice. For example, you can writemodule Foo where import Bar f x = x $(deriveStuff 'f)  Uses the $(...) notation g y = y+1 deriveStuff 'g  Omits the $(...) h z = z1
This abbreviation makes toplevel declaration slices quieter and less intimidating.

Pattern splices introduce variable binders but scoping of variables in expressions inside the pattern’s scope is only checked when a splice is run. Note that pattern splices that occur outside of any quotation brackets are run at compile time. Pattern splices occurring inside a quotation bracket are not run at compile time; they are run when the bracket is spliced in, sometime later. For example,
mkPat :: Q Pat mkPat = [p (x, y) ]  in another module: foo :: (Char, String) > String foo $(mkPat) = x : z bar :: Q Exp bar = [ \ $(mkPat) > x : w ]
will fail with
z
being out of scope in the definition offoo
but it will not fail withw
being out of scope in the definition ofbar
. That will only happen whenbar
is spliced. 
A pattern quasiquoter may generate binders that scope over the righthand side of a definition because these binders are in scope lexically. For example, given a quasiquoter
haskell
that parses Haskell, in the following code, they
in the righthand side off
refers to they
bound by thehaskell
pattern quasiquoter, not the toplevely = 7
.y :: Int y = 7 f :: Int > Int > Int f n = \ [haskelly] > y+n

Toplevel declaration splices break up a source file into declaration groups. A declaration group is the group of declarations created by a toplevel declaration splice, plus those following it, down to but not including the next toplevel declaration splice. N.B. only toplevel splices delimit declaration groups, not expression splices. The first declaration group in a module includes all toplevel definitions down to but not including the first toplevel declaration splice.
Each declaration group is mutually recursive only within the group. Declaration groups can refer to definitions within previous groups, but not later ones.
Accordingly, the type environment seen by
reify
includes all the toplevel declarations up to the end of the immediately preceding declaration group, but no more.Unlike normal declaration splices, declaration quasiquoters do not cause a break. These quasiquoters are expanded before the rest of the declaration group is processed, and the declarations they generate are merged into the surrounding declaration group. Consequently, the type environment seen by
reify
from a declaration quasiquoter will not include anything from the quasiquoter’s declaration group.Concretely, consider the following code
module M where import ... f x = x $(th1 4) h y = k y y $(blah1) [qqblah] k x y z = x + y + z $(th2 10) w z = $(blah2)
In this example, a
reify
inside… The splice
$(th1 ...)
would see the definition off
 the splice is toplevel and thus all definitions in the previous declaration group are visible (that is, all definitions in the module upto, but not including, the splice itself).  The splice
$(blah1)
cannot refer to the functionw
w
is part of a later declaration group, and thus invisible, similarly,$(blah1)
cannot see the definition ofh
(since it is part of the same declaration group as$(blah1)
. However, the splice$(blah1)
can see the definition off
(since it is in the immediately preceding declaration group).  The splice
$(th2 ...)
would see the definition off
, all the bindings created by$(th1 ...)
, the definition ofh
and all bindings created by[qqblah]
(they are all in previous declaration groups).  The body of
h
can refer to the functionk
appearing on the other side of the declaration quasiquoter, as quasiquoters do not cause a declaration group to be broken up.  The
qq
quasiquoter would be able to see the definition off
from the preceding declaration group, but not the definitions ofh
ork
, or any definitions from subsequent declaration groups.  The splice
$(blah2)
would see the same definitions as the splice$(th2 ...)
(but not any bindings it creates).
Note that since an expression splice is unable to refer to declarations in the same declaration group, we can introduce a toplevel (empty) splice to break up the declaration group
module M where data D = C1  C2 f1 = $(th1 ...) $(return []) f2 = $(th2 ...)
Here
 The splice
$(th1 ...)
cannot refer toD
 it is in the same declaration group.  The declaration group containing
D
is terminated by the empty toplevel declaration splice$(return [])
(recall,Q
is a Monad, so we may simplyreturn
the empty list of declarations).  Since the declaration group containing
D
is in the previous declaration group, the splice$(th2 ...)
can refer toD
.
 The splice

Expression quotations accept most Haskell language constructs. However, there are some GHCspecific extensions which expression quotations currently do not support, including
(Compared to the original paper, there are many differences of detail. The syntax for a declaration splice uses “$
” not “splice
”. The type of the enclosed expression must be Q [Dec]
, not [Q Dec]
. Typed expression splices and quotations are supported.)

fenablethsplicewarnings

Template Haskell splices won’t be checked for warnings, because the code causing the warning might originate from a thirdparty library and possibly was not written by the user. If you want to have warnings for splices anyway, pass
fenablethsplicewarnings
.
10.27.2. Using Template Haskell
 The data types and monadic constructor functions for Template Haskell are in the library
Language.Haskell.TH.Syntax
.  You can only run a function at compile time if it is imported from another module. That is, you can’t define a function in a module, and call it from within a splice in the same module. (It would make sense to do so, but it’s hard to implement.)

You can only run a function at compile time if it is imported from another module that is not part of a mutuallyrecursive group of modules that includes the module currently being compiled. Furthermore, all of the modules of the mutuallyrecursive group must be reachable by nonSOURCE imports from the module where the splice is to be run.
For example, when compiling module A, you can only run Template Haskell functions imported from B if B does not import A (directly or indirectly). The reason should be clear: to run B we must compile and run A, but we are currently typechecking A.
 If you are building GHC from source, you need at least a stage2 bootstrap compiler to run Template Haskell splices and quasiquotes. A stage1 compiler will only accept regular quotes of Haskell. Reason: TH splices and quasiquotes compile and run a program, and then looks at the result. So it’s important that the program it compiles produces results whose representations are identical to those of the compiler itself.
Template Haskell works in any mode (make
, interactive
, or fileatatime). There used to be a restriction to the former two, but that restriction has been lifted.
10.27.3. Viewing Template Haskell generated code
The flag ddumpsplices
shows the expansion of all toplevel declaration splices, both typed and untyped, as they happen. As with all dump flags, the default is for this output to be sent to stdout. For a nontrivial program, you may be interested in combining this with the ddumptofile
flag (see Dumping out compiler intermediate structures. For each file using Template Haskell, this will show the output in a .dumpsplices
file.
The flag dthdecfile
dumps the expansions of all toplevel TH declaration splices, both typed and untyped, in the file M.th.hs
for each module M
being compiled. Note that other types of splices (expressions, types, and patterns) are not shown. Application developers can check this into their repository so that they can grep for identifiers that were defined in Template Haskell. This is similar to using ddumptofile
with ddumpsplices
but it always generates a file instead of being coupled to ddumptofile
. The format is also different: it does not show code from the original file, instead it only shows generated code and has a comment for the splice location of the original file.
Below is a sample output of ddumpsplices
TH_pragma.hs:(6,4)(8,26): Splicing declarations [d foo :: Int > Int foo x = x + 1 ] ======> foo :: Int > Int foo x = (x + 1)
Below is the output of the same sample using dthdecfile
 TH_pragma.hs:(6,4)(8,26): Splicing declarations foo :: Int > Int foo x = (x + 1)
10.27.4. A Template Haskell Worked Example
To help you get over the confidence barrier, try out this skeletal worked example. First cut and paste the two modules below into Main.hs
and Printf.hs
:
{ Main.hs } module Main where  Import our template "pr" import Printf ( pr )  The splice operator $ takes the Haskell source code  generated at compile time by "pr" and splices it into  the argument of "putStrLn". main = putStrLn ( $(pr "Hello") ) { Printf.hs } module Printf where  Skeletal printf from the paper.  It needs to be in a separate module to the one where  you intend to use it.  Import some Template Haskell syntax import Language.Haskell.TH  Describe a format string data Format = D  S  L String  Parse a format string. This is left largely to you  as we are here interested in building our first ever  Template Haskell program and not in building printf. parse :: String > [Format] parse s = [ L s ]  Generate Haskell source code from a parsed representation  of the format string. This code will be spliced into  the module which calls "pr", at compile time. gen :: [Format] > Q Exp gen [D] = [ \n > show n ] gen [S] = [ \s > s ] gen [L s] = stringE s  Here we generate the Haskell code for the splice  from an input format string. pr :: String > Q Exp pr s = gen (parse s)
Now run the compiler,
$ ghc make XTemplateHaskell main.hs o main
Run main
and here is your output:
$ ./main Hello
10.27.5. Using Template Haskell with Profiling
Template Haskell relies on GHC’s builtin bytecode compiler and interpreter to run the splice expressions. The bytecode interpreter runs the compiled expression on top of the same runtime on which GHC itself is running; this means that the compiled code referred to by the interpreted expression must be compatible with this runtime, and in particular this means that object code that is compiled for profiling cannot be loaded and used by a splice expression, because profiled object code is only compatible with the profiling version of the runtime.
This causes difficulties if you have a multimodule program containing Template Haskell code and you need to compile it for profiling, because GHC cannot load the profiled object code and use it when executing the splices.
Fortunately GHC provides two workarounds.
The first option is to compile the program twice:
 Compile the program or library first the normal way, without
prof
.  Then compile it again with
prof
, and additionally useosuf p_o
to name the object files differently (you can choose any suffix that isn’t the normal object suffix here). GHC will automatically load the object files built in the first step when executing splice expressions. If you omit theosuf ⟨suffix⟩
flag when building withprof
and Template Haskell is used, GHC will emit an error message.
The second option is to add the flag fexternalinterpreter
(see Running the interpreter in a separate process), which runs the interpreter in a separate process, wherein it can load and run the profiled code directly. There’s no need to compile the code twice, just add fexternalinterpreter
and it should just work. (this option is experimental in GHC 8.0.x, but it may become the default in future releases).
10.27.6. Template Haskell Quasiquotation

QuasiQuotes

 Since

6.10.1
Enable Template Haskell Quasiquotation syntax.
Quasiquotation allows patterns and expressions to be written using programmerdefined concrete syntax; the motivation behind the extension and several examples are documented in “Why It’s Nice to be Quoted: Quasiquoting for Haskell” (Proc Haskell Workshop 2007). The example below shows how to write a quasiquoter for a simple expression language.
Here are the salient features

A quasiquote has the form
[quoter string ]
. The ⟨quoter⟩ must be the name of an imported quoter, either qualified or unqualified; it cannot be an arbitrary expression.
 The ⟨quoter⟩ cannot be “
e
”, “t
”, “d
”, or “p
”, since those overlap with Template Haskell quotations.  There must be no spaces in the token
[quoter
.  The quoted ⟨string⟩ can be arbitrary, and may contain newlines.
 The quoted ⟨string⟩ finishes at the first occurrence of the twocharacter sequence
"]"
. Absolutely no escaping is performed. If you want to embed that character sequence in the string, you must invent your own escape convention (such as, say, using the string"~]"
instead), and make your quoter function interpret"~]"
as"]"
. One way to implement this is to compose your quoter with a preprocessing pass to perform your escape conversion. See the discussion in #5348 for details.

A quasiquote may appear in place of
 An expression
 A pattern
 A type
 A toplevel declaration
(Only the first two are described in the paper.)

A quoter is a value of type
Language.Haskell.TH.Quote.QuasiQuoter
, which is defined thus:data QuasiQuoter = QuasiQuoter { quoteExp :: String > Q Exp, quotePat :: String > Q Pat, quoteType :: String > Q Type, quoteDec :: String > Q [Dec] }
That is, a quoter is a tuple of four parsers, one for each of the contexts in which a quasiquote can occur.
 A quasiquote is expanded by applying the appropriate parser to the string enclosed by the Oxford brackets. The context of the quasiquote (expression, pattern, type, declaration) determines which of the parsers is called.
 Unlike normal declaration splices of the form
$(...)
, declaration quasiquotes do not cause a declaration group break. See Syntax for more information.
Warning
QuasiQuotes
introduces an unfortunate ambiguity with list comprehension syntax. Consider the following,
let x = [v v < [0..10]]
Without QuasiQuotes
this is parsed as a list comprehension. With QuasiQuotes
this is parsed as a quasiquote; however, this parse will fail due to the lack of a closing ]
. See #11679.
The example below shows quasiquotation in action. The quoter expr
is bound to a value of type QuasiQuoter
defined in module Expr
. The example makes use of an antiquoted variable n
, indicated by the syntax 'int:n
(this syntax for antiquotation was defined by the parser’s author, not by GHC). This binds n
to the integer value argument of the constructor IntExpr
when pattern matching. Please see the referenced paper for further details regarding antiquotation as well as the description of a technique that uses SYB to leverage a single parser of type String > a
to generate both an expression parser that returns a value of type Q Exp
and a pattern parser that returns a value of type Q Pat
.
Quasiquoters must obey the same stage restrictions as Template Haskell, e.g., in the example, expr
cannot be defined in Main.hs
where it is used, but must be imported.
{  file Main.hs  } module Main where import Expr main :: IO () main = do { print $ eval [expr1 + 2] ; case IntExpr 1 of { [expr'int:n] > print n ; _ > return () } } {  file Expr.hs  } module Expr where import qualified Language.Haskell.TH as TH import Language.Haskell.TH.Quote data Expr = IntExpr Integer  AntiIntExpr String  BinopExpr BinOp Expr Expr  AntiExpr String deriving(Show, Typeable, Data) data BinOp = AddOp  SubOp  MulOp  DivOp deriving(Show, Typeable, Data) eval :: Expr > Integer eval (IntExpr n) = n eval (BinopExpr op x y) = (opToFun op) (eval x) (eval y) where opToFun AddOp = (+) opToFun SubOp = () opToFun MulOp = (*) opToFun DivOp = div expr = QuasiQuoter { quoteExp = parseExprExp, quotePat = parseExprPat }  Parse an Expr, returning its representation as  either a Q Exp or a Q Pat. See the referenced paper  for how to use SYB to do this by writing a single  parser of type String > Expr instead of two  separate parsers. parseExprExp :: String > Q Exp parseExprExp ... parseExprPat :: String > Q Pat parseExprPat ...
Now run the compiler:
$ ghc make XQuasiQuotes Main.hs o main
Run “main” and here is your output:
$ ./main 3 1
10.28. Arrow notation

Arrows

 Since

6.8.1
Enable arrow notation.
Arrows are a generalisation of monads introduced by John Hughes. For more details, see
 “Generalising Monads to Arrows”, John Hughes, in Science of Computer Programming 37, pp. 67–111, May 2000. The paper that introduced arrows: a friendly introduction, motivated with programming examples.
 “A New Notation for Arrows”, Ross Paterson, in ICFP, Sep 2001. Introduced the notation described here.
 “Arrows and Computation”, Ross Paterson, in The Fun of Programming, Palgrave, 2003.
 “Programming with Arrows”, John Hughes, in 5th International Summer School on Advanced Functional Programming, Lecture Notes in Computer Science vol. 3622, Springer, 2004. This paper includes another introduction to the notation, with practical examples.
 “Type and Translation Rules for Arrow Notation in GHC”, Ross Paterson and Simon Peyton Jones, September 16, 2004. A terse enumeration of the formal rules used (extracted from comments in the source code).
 The arrows web page at
http://www.haskell.org/arrows/
<http://www.haskell.org/arrows/>`__.
With the Arrows
extension, GHC supports the arrow notation described in the second of these papers, translating it using combinators from the Control.Arrow module. What follows is a brief introduction to the notation; it won’t make much sense unless you’ve read Hughes’s paper.
The extension adds a new kind of expression for defining arrows:
exp10 ::= ...  proc apat > cmd
where proc
is a new keyword. The variables of the pattern are bound in the body of the proc
expression, which is a new sort of thing called a command. The syntax of commands is as follows:
cmd ::= exp10 < exp  exp10 << exp  cmd0
with ⟨cmd⟩^{0} up to ⟨cmd⟩^{9} defined using infix operators as for expressions, and
cmd10 ::= \ apat ... apat > cmd  let decls in cmd  if exp then cmd else cmd  case exp of { calts }  do { cstmt ; ... cstmt ; cmd }  fcmd fcmd ::= fcmd aexp  ( cmd )  ( aexp cmd ... cmd ) cstmt ::= let decls  pat < cmd  rec { cstmt ; ... cstmt [;] }  cmd
where ⟨calts⟩ are like ⟨alts⟩ except that the bodies are commands instead of expressions.
Commands produce values, but (like monadic computations) may yield more than one value, or none, and may do other things as well. For the most part, familiarity with monadic notation is a good guide to using commands. However the values of expressions, even monadic ones, are determined by the values of the variables they contain; this is not necessarily the case for commands.
A simple example of the new notation is the expression
proc x > f < x+1
We call this a procedure or arrow abstraction. As with a lambda expression, the variable x
is a new variable bound within the proc
expression. It refers to the input to the arrow. In the above example, <
is not an identifier but a new reserved symbol used for building commands from an expression of arrow type and an expression to be fed as input to that arrow. (The weird look will make more sense later.) It may be read as analogue of application for arrows. The above example is equivalent to the Haskell expression
arr (\ x > x+1) >>> f
That would make no sense if the expression to the left of <
involves the bound variable x
. More generally, the expression to the left of <
may not involve any local variable, i.e. a variable bound in the current arrow abstraction. For such a situation there is a variant <<
, as in
proc x > f x << x+1
which is equivalent to
arr (\ x > (f x, x+1)) >>> app
so in this case the arrow must belong to the ArrowApply
class. Such an arrow is equivalent to a monad, so if you’re using this form you may find a monadic formulation more convenient.
10.28.1. donotation for commands
Another form of command is a form of do
notation. For example, you can write
proc x > do y < f < x+1 g < 2*y let z = x+y t < h < x*z returnA < t+z
You can read this much like ordinary do
notation, but with commands in place of monadic expressions. The first line sends the value of x+1
as an input to the arrow f
, and matches its output against y
. In the next line, the output is discarded. The arrow returnA
is defined in the Control.Arrow module as arr id
. The above example is treated as an abbreviation for
arr (\ x > (x, x)) >>> first (arr (\ x > x+1) >>> f) >>> arr (\ (y, x) > (y, (x, y))) >>> first (arr (\ y > 2*y) >>> g) >>> arr snd >>> arr (\ (x, y) > let z = x+y in ((x, z), z)) >>> first (arr (\ (x, z) > x*z) >>> h) >>> arr (\ (t, z) > t+z) >>> returnA
Note that variables not used later in the composition are projected out. After simplification using rewrite rules (see Rewrite rules) defined in the Control.Arrow module, this reduces to
arr (\ x > (x+1, x)) >>> first f >>> arr (\ (y, x) > (2*y, (x, y))) >>> first g >>> arr (\ (_, (x, y)) > let z = x+y in (x*z, z)) >>> first h >>> arr (\ (t, z) > t+z)
which is what you might have written by hand. With arrow notation, GHC keeps track of all those tuples of variables for you.
Note that although the above translation suggests that let
bound variables like z
must be monomorphic, the actual translation produces Core, so polymorphic variables are allowed.
It’s also possible to have mutually recursive bindings, using the new rec
keyword, as in the following example:
counter :: ArrowCircuit a => a Bool Int counter = proc reset > do rec output < returnA < if reset then 0 else next next < delay 0 < output+1 returnA < output
The translation of such forms uses the loop
combinator, so the arrow concerned must belong to the ArrowLoop
class.
10.28.2. Conditional commands
In the previous example, we used a conditional expression to construct the input for an arrow. Sometimes we want to conditionally execute different commands, as in
proc (x,y) > if f x y then g < x+1 else h < y+2
which is translated to
arr (\ (x,y) > if f x y then Left x else Right y) >>> (arr (\x > x+1) >>> g)  (arr (\y > y+2) >>> h)
Since the translation uses 
, the arrow concerned must belong to the ArrowChoice
class.
There are also case
commands, like
case input of [] > f < () [x] > g < x+1 x1:x2:xs > do y < h < (x1, x2) ys < k < xs returnA < y:ys
The syntax is the same as for case
expressions, except that the bodies of the alternatives are commands rather than expressions. The translation is similar to that of if
commands.
10.28.3. Defining your own control structures
As we’re seen, arrow notation provides constructs, modelled on those for expressions, for sequencing, value recursion and conditionals. But suitable combinators, which you can define in ordinary Haskell, may also be used to build new commands out of existing ones. The basic idea is that a command defines an arrow from environments to values. These environments assign values to the free local variables of the command. Thus combinators that produce arrows from arrows may also be used to build commands from commands. For example, the ArrowPlus
class includes a combinator
ArrowPlus a => (<+>) :: a b c > a b c > a b c
so we can use it to build commands:
expr' = proc x > do returnA < x <+> do symbol Plus < () y < term < () expr' < x + y <+> do symbol Minus < () y < term < () expr' < x  y
(The do
on the first line is needed to prevent the first <+> ...
from being interpreted as part of the expression on the previous line.) This is equivalent to
expr' = (proc x > returnA < x) <+> (proc x > do symbol Plus < () y < term < () expr' < x + y) <+> (proc x > do symbol Minus < () y < term < () expr' < x  y)
We are actually using <+>
here with the more specific type
ArrowPlus a => (<+>) :: a (e,()) c > a (e,()) c > a (e,()) c
It is essential that this operator be polymorphic in e
(representing the environment input to the command and thence to its subcommands) and satisfy the corresponding naturality property
arr (first k) >>> (f <+> g) = (arr (first k) >>> f) <+> (arr (first k) >>> g)
at least for strict k
. (This should be automatic if you’re not using seq
.) This ensures that environments seen by the subcommands are environments of the whole command, and also allows the translation to safely trim these environments. (The second component of the input pairs can contain unnamed input values, as described in the next section.) The operator must also not use any variable defined within the current arrow abstraction.
We could define our own operator
untilA :: ArrowChoice a => a (e,s) () > a (e,s) Bool > a (e,s) () untilA body cond = proc x > do b < cond < x if b then returnA < () else do body < x untilA body cond < x
and use it in the same way. Of course this infix syntax only makes sense for binary operators; there is also a more general syntax involving special brackets:
proc x > do y < f < x+1 (untilA (increment < x+y) (within 0.5 < x))
10.28.4. Primitive constructs
Some operators will need to pass additional inputs to their subcommands. For example, in an arrow type supporting exceptions, the operator that attaches an exception handler will wish to pass the exception that occurred to the handler. Such an operator might have a type
handleA :: ... => a (e,s) c > a (e,(Ex,s)) c > a (e,s) c
where Ex
is the type of exceptions handled. You could then use this with arrow notation by writing a command
body `handleA` \ ex > handler
so that if an exception is raised in the command body
, the variable ex
is bound to the value of the exception and the command handler
, which typically refers to ex
, is entered. Though the syntax here looks like a functional lambda, we are talking about commands, and something different is going on. The input to the arrow represented by a command consists of values for the free local variables in the command, plus a stack of anonymous values. In all the prior examples, we made no assumptions about this stack. In the second argument to handleA
, the value of the exception has been added to the stack input to the handler. The command form of lambda merely gives this value a name.
More concretely, the input to a command consists of a pair of an environment and a stack. Each value on the stack is paired with the remainder of the stack, with an empty stack being ()
. So operators like handleA
that pass extra inputs to their subcommands can be designed for use with the notation by placing the values on the stack paired with the environment in this way. More precisely, the type of each argument of the operator (and its result) should have the form
a (e, (t1, ... (tn, ())...)) t
where ⟨e⟩ is a polymorphic variable (representing the environment) and ⟨ti⟩ are the types of the values on the stack, with ⟨t1⟩ being the “top”. The polymorphic variable ⟨e⟩ must not occur in ⟨a⟩, ⟨ti⟩ or ⟨t⟩. However the arrows involved need not be the same. Here are some more examples of suitable operators:
bracketA :: ... => a (e,s) b > a (e,(b,s)) c > a (e,(c,s)) d > a (e,s) d runReader :: ... => a (e,s) c > a' (e,(State,s)) c runState :: ... => a (e,s) c > a' (e,(State,s)) (c,State)
We can supply the extra input required by commands built with the last two by applying them to ordinary expressions, as in
proc x > do s < ... (runReader (do { ... })) s
which adds s
to the stack of inputs to the command built using runReader
.
The command versions of lambda abstraction and application are analogous to the expression versions. In particular, the beta and eta rules describe equivalences of commands. These three features (operators, lambda abstraction and application) are the core of the notation; everything else can be built using them, though the results would be somewhat clumsy. For example, we could simulate do
notation by defining
bind :: Arrow a => a (e,s) b > a (e,(b,s)) c > a (e,s) c u `bind` f = returnA &&& u >>> f bind_ :: Arrow a => a (e,s) b > a (e,s) c > a (e,s) c u `bind_` f = u `bind` (arr fst >>> f)
We could simulate if
by defining
cond :: ArrowChoice a => a (e,s) b > a (e,s) b > a (e,(Bool,s)) b cond f g = arr (\ (e,(b,s)) > if b then Left (e,s) else Right (e,s)) >>> f  g
10.28.5. Differences with the paper
 Instead of a single form of arrow application (arrow tail) with two translations, the implementation provides two forms
<
(firstorder) and<<
(higherorder).  Userdefined operators are flagged with banana brackets instead of a new
form
keyword.  In the paper and the previous implementation, values on the stack were paired to the right of the environment in a single argument, but now the environment and stack are separate arguments.
10.28.6. Portability
Although only GHC implements arrow notation directly, there is also a preprocessor (available from the arrows web page) that translates arrow notation into Haskell 98 for use with other Haskell systems. You would still want to check arrow programs with GHC; tracing type errors in the preprocessor output is not easy. Modules intended for both GHC and the preprocessor must observe some additional restrictions:
 The module must import Control.Arrow.
 The preprocessor cannot cope with other Haskell extensions. These would have to go in separate modules.
 Because the preprocessor targets Haskell (rather than Core),
let
bound variables are monomorphic.
10.29. Bang patterns and Strict Haskell
In highperformance Haskell code (e.g. numeric code) eliminating thunks from an inner loop can be a huge win. GHC supports three extensions to allow the programmer to specify use of strict (callbyvalue) evaluation rather than lazy (callbyneed) evaluation.
 Bang patterns (
BangPatterns
) makes pattern matching and let bindings stricter.  Strict data types (
StrictData
) makes constructor fields strict by default, on a permodule basis.  Strict pattern (
Strict
) makes all patterns and let bindings strict by default, on a permodule basis.
The latter two extensions are simply a way to avoid littering highperformance code with bang patterns, making it harder to read.
Bang patterns and strict matching do not affect the type system in any way.
10.29.1. Bang patterns

BangPatterns

 Since

6.8.1
Allow use of bang pattern syntax.
GHC supports an extension of pattern matching called bang patterns, written !pat
. Bang patterns are under consideration for Haskell Prime. The Haskell prime feature description contains more discussion and examples than the material below.
The main idea is to add a single new production to the syntax of patterns:
pat ::= !pat
Matching an expression e
against a pattern !p
is done by first evaluating e
(to WHNF) and then matching the result against p
. Example:
f1 !x = True
This definition makes f1
is strict in x
, whereas without the bang it would be lazy. Bang patterns can be nested of course:
f2 (!x, y) = [x,y]
Here, f2
is strict in x
but not in y
.
Note the following points:

A bang only really has an effect if it precedes a variable or wildcard pattern:
f3 !(x,y) = [x,y] f4 (x,y) = [x,y]
Here,
f3
andf4
are identical; putting a bang before a pattern that forces evaluation anyway does nothing. 
A bang pattern is allowed in a let or where clause, and makes the binding strict. For example:
let !x = e in body let !(p,q) = e in body
In both cases
e
is evaluated before starting to evaluatebody
.However, nested bangs in a let/where pattern binding behave uniformly with all other forms of pattern matching. For example
let (!x,[y]) = e in b
is equivalent to this:
let { t = case e of (x,[y]) > x `seq` (x,y) x = fst t y = snd t } in b
The binding is lazy, but when either
x
ory
is evaluated byb
the entire pattern is matched, including forcing the evaluation ofx
.See Semantics of let bindings with bang patterns for the detailed semantics.
 A pattern with a bang at the outermost level is not allowed at the top level of a module.

Bang patterns work in
case
expressions too, of course:g5 x = let y = f x in body g6 x = case f x of { y > body } g7 x = case f x of { !y > body }
The functions
g5
andg6
mean exactly the same thing. Butg7
evaluates(f x)
, bindsy
to the result, and then evaluatesbody
. 
There is one problem with syntactic ambiguity. Consider:
f !x = 3
Is this a definition of the infix function “
(!)
”, or of the “f
” with a bang pattern? GHC resolves this ambiguity in favour of the latter. If you want to define(!)
with bangpatterns enabled, you have to do so using prefix notation:(!) f x = 3
10.29.2. Strictbydefault data types

StrictData

 Since

8.0.1
Make fields of data types defined in the current module strict by default.
Informally the StrictData
language extension switches data type declarations to be strict by default allowing fields to be lazy by adding a ~
in front of the field.
When the user writes
data T = C a data T' = C' ~a
we interpret it as if they had written
data T = C !a data T' = C' a
The extension only affects definitions in this module.
10.29.3. Strictbydefault pattern bindings

Strict

 Implies
 Since

8.0.1
Make bindings in the current module strict by default.
Informally the Strict
language extension switches functions, data types, and bindings to be strict by default, allowing optional laziness by adding ~
in front of a variable. This essentially reverses the present situation where laziness is default and strictness can be optionally had by adding !
in front of a variable.
Strict
implies StrictData.

Function definitions
When the user writes
f x = ...
we interpret it as if they had written
f !x = ...
Adding
~
in front ofx
gives the regular lazy behavior.Turning patterns into irrefutable ones requires
~(~p)
or(~ ~p)
whenStrict
is enabled. 
Let/where bindings
When the user writes
let x = ... let pat = ...
we interpret it as if they had written
let !x = ... let !pat = ...
Adding
~
in front ofx
gives the regular lazy behavior. The general rule is that we add an implicit bang on the outermost pattern, unless disabled with~
. 
Pattern matching in case expressions, lambdas, donotation, etc
The outermost pattern of all pattern matches gets an implicit bang, unless disabled with
~
. This applies to case expressions, patterns in lambda, donotation, list comprehension, and so on. For examplecase x of (a,b) > rhs
is interpreted as
case x of !(a,b) > rhs
Since the semantics of pattern matching in case expressions is strict, this usually has no effect whatsoever. But it does make a difference in the degenerate case of variables and newtypes. So
case x of y > rhs
is lazy in Haskell, but with
Strict
is interpreted ascase x of !y > rhs
which evaluates
x
. Similarly, ifnewtype Age = MkAge Int
, thencase x of MkAge i > rhs
is lazy in Haskell; but with
Strict
the added bang makes it strict.Similarly
\ x > body do { x < rhs; blah } [ e  x < rhs; blah }
all get implicit bangs on the
x
pattern. 
Nested patterns
Notice that we do not put bangs on nested patterns. For example
let (p,q) = if flob then (undefined, undefined) else (True, False) in ...
will behave like
let !(p,q) = if flob then (undefined, undefined) else (True,False) in ...
which will strictly evaluate the right hand side, and bind
p
andq
to the components of the pair. But the pair itself is lazy (unless we also compile thePrelude
withStrict
; see Modularity below). Sop
andq
may end up bound to undefined. See also Dynamic semantics of bang patterns below. 
Top level bindings
are unaffected by
Strict
. For example:x = factorial 20 (y,z) = if x > 10 then True else False
Here
x
and the pattern binding(y,z)
remain lazy. Reason: there is no good moment to force them, until first use. 
Newtypes
There is no effect on newtypes, which simply rename existing types. For example:
newtype T = C a f (C x) = rhs1 g !(C x) = rhs2
In ordinary Haskell,
f
is lazy in its argument and hence inx
; andg
is strict in its argument and hence also strict inx
. WithStrict
, both become strict becausef
’s argument gets an implicit bang.
10.29.4. Modularity
Strict
and StrictData
only affects definitions in the module they are used in. Functions and data types imported from other modules are unaffected. For example, we won’t evaluate the argument to Just
before applying the constructor. Similarly we won’t evaluate the first argument to Data.Map.findWithDefault
before applying the function.
This is crucial to preserve correctness. Entities defined in other modules might rely on laziness for correctness (whether functional or performance).
Tuples, lists, Maybe
, and all the other types from Prelude
continue to have their existing, lazy, semantics.
10.29.5. Dynamic semantics of bang patterns
The semantics of Haskell pattern matching is described in Section 3.17.2 of the Haskell Report. To this description add one extra item 10, saying:

Matching the pattern
!pat
against a valuev
behaves as follows: if
v
is bottom, the match diverges  otherwise,
pat
is matched againstv
 if
Similarly, in Figure 4 of Section 3.17.3, add a new case (t):
case v of { !pat > e; _ > e' } = v `seq` case v of { pat > e; _ > e' }
That leaves let expressions, whose translation is given in Section 3.12 of the Haskell Report. Replace the “Translation” there with the following one. Given let { bind1 ... bindn } in body
:
FORCE
Replace any binding !p = e
with v = case e of p > (x1, ..., xn); (x1, ..., xn) = v
and replace body
with v seq body
, where v
is fresh. This translation works fine if p
is already a variable x
, but can obviously be optimised by not introducing a fresh variable v
.
SPLIT
Replace any binding p = e
, where p
is not a variable, with v = e; x1 = case v of p > x1; ...; xn = case v of p > xn
, where v
is fresh and x1
.. xn
are the bound variables of p
. Again if e
is a variable, this can be optimised by not introducing a fresh variable.
The result will be a (possibly) recursive set of bindings, binding only simple variables on the left hand side. (One could go one step further, as in the Haskell Report and make the recursive bindings nonrecursive using fix
, but we do not do so in Core, and it only obfuscates matters, so we do not do so here.)
The translation is carefully crafted to make bang patterns meaningful for recursive and polymorphic bindings as well as straightforward nonrecursive bindings.
Here are some examples of how this translation works. The first expression of each sequence is Haskell source; the subsequent ones are Core.
Here is a simple nonrecursive case:
let x :: Int  Nonrecursive !x = factorial y in body ===> (FORCE) let x = factorial y in x `seq` body ===> (inline seq) let x = factorial y in case x of x > body ===> (inline x) case factorial y of x > body
Same again, only with a pattern binding:
let !(Just x, Left y) = e in body ===> (FORCE) let v = case e of (Just x, Left y) > (x,y) (x,y) = v in v `seq` body ===> (SPLIT) let v = case e of (Just x, Left y) > (x,y) x = case v of (x,y) > x y = case v of (x,y) > y in v `seq` body ===> (inline seq, float x,y bindings inwards) let v = case e of (Just x, Left y) > (x,y) in case v of v > let x = case v of (x,y) > x y = case v of (x,y) > y in body ===> (fluff up v's pattern; this is a standard Core optimisation) let v = case e of (Just x, Left y) > (x,y) in case v of v@(p,q) > let x = case v of (x,y) > x y = case v of (x,y) > y in body ===> (case of known constructor) let v = case e of (Just x, Left y) > (x,y) in case v of v@(p,q) > let x = p y = q in body ===> (inline x,y, v) case (case e of (Just x, Left y) > (x,y) of (p,q) > body[p/x, q/y] ===> (case of case) case e of (Just x, Left y) > body[p/x, q/y]
The final form is just what we want: a simple case expression.
Here is a recursive case
letrec xs :: [Int]  Recursive !xs = factorial y : xs in body ===> (FORCE) letrec xs = factorial y : xs in xs `seq` body ===> (inline seq) letrec xs = factorial y : xs in case xs of xs > body ===> (eliminate case of value) letrec xs = factorial y : xs in body
and a polymorphic one:
let f :: forall a. [a] > [a]  Polymorphic !f = fst (reverse, True) in body ===> (FORCE) let f = /\a. fst (reverse a, True) in f `seq` body ===> (inline seq, inline f) case (/\a. fst (reverse a, True)) of f > body
Notice that the seq
is added only in the translation to Core If we did it in Haskell source, thus
let f = ... in f `seq` body
then f
’s polymorphic type would get instantiated, so the Core translation would be
let f = ... in f Any `seq` body
When overloading is involved, the results might be slightly counter intuitive:
let f :: forall a. Eq a => a > [a] > Bool  Overloaded !f = fst (member, True) in body ===> (FORCE) let f = /\a \(d::Eq a). fst (member, True) in f `seq` body ===> (inline seq, case of value) let f = /\a \(d::Eq a). fst (member, True) in body
Note that the bang has no effect at all in this case
10.30. Assertions
If you want to make use of assertions in your standard Haskell code, you could define a function like the following:
assert :: Bool > a > a assert False x = error "assertion failed!" assert _ x = x
which works, but gives you back a less than useful error message – an assertion failed, but which and where?
One way out is to define an extended assert
function which also takes a descriptive string to include in the error message and perhaps combine this with the use of a preprocessor which inserts the source location where assert
was used.
GHC offers a helping hand here, doing all of this for you. For every use of assert
in the user’s source:
kelvinToC :: Double > Double kelvinToC k = assert (k >= 0.0) (k273.15)
GHC will rewrite this to also include the source location where the assertion was made,
assert pred val ==> assertError "Main.hs15" pred val
The rewrite is only performed by the compiler when it spots applications of Control.Exception.assert
, so you can still define and use your own versions of assert
, should you so wish. If not, import Control.Exception
to make use assert
in your code.
GHC ignores assertions when optimisation is turned on with the O
flag. That is, expressions of the form assert pred e
will be rewritten to e
. You can also disable assertions using the fignoreasserts
option. The option fnoignoreasserts
allows enabling assertions even when optimisation is turned on.
Assertion failures can be caught, see the documentation for the Control.Exception library for the details.
10.31. Static pointers

StaticPointers

 Since

7.10.1
Allow use of static pointer syntax.
The language extension StaticPointers
adds a new syntactic form static e
, which stands for a reference to the closed expression ⟨e⟩. This reference is stable and portable, in the sense that it remains valid across different processes on possibly different machines. Thus, a process can create a reference and send it to another process that can resolve it to ⟨e⟩.
With this extension turned on, static
is no longer a valid identifier.
Static pointers were first proposed in the paper Towards Haskell in the cloud, Jeff Epstein, Andrew P. Black and Simon PeytonJones, Proceedings of the 4th ACM Symposium on Haskell, pp. 118129, ACM, 2011.
10.31.1. Using static pointers
Each reference is given a key which can be used to locate it at runtime with GHC.StaticPtr.unsafeLookupStaticPtr which uses a global and immutable table called the Static Pointer Table. The compiler includes entries in this table for all static forms found in the linked modules. The value can be obtained from the reference via GHC.StaticPtr.deRefStaticPtr.
The body e
of a static e
expression must be a closed expression. Where we say an expression is closed when all of its free (type) variables are closed. And a variable is closed if it is letbound to a closed expression and its type is closed as well. And a type is closed if it has no free variables.
All of the following are permissible:
inc :: Int > Int inc x = x + 1 ref1 = static 1 ref2 = static inc ref3 = static (inc 1) ref4 = static ((\x > x + 1) (1 :: Int)) ref5 y = static (let x = 1 in x) ref6 y = let x = 1 in static x
While the following definitions are rejected:
ref7 y = let x = y in static x  x is not closed ref8 y = static (let x = 1 in y)  y is not letbound ref8 (y :: a) = let x = undefined :: a in static x  x has a nonclosed type
Note
While modules loaded in GHCi with the :load
command may use StaticPointers
and static
expressions, statements entered on the REPL may not. This is a limitation of GHCi; see #12356 for details.
Note
The set of keys used for locating static pointers in the Static Pointer Table is not guaranteed to remain stable for different program binaries. Or in other words, only processes launched from the same program binary are guaranteed to use the same set of keys.
10.31.2. Static semantics of static pointers
Informally, if we have a closed expression
e :: forall a_1 ... a_n . t
the static form is of type
static e :: (IsStatic p, Typeable a_1, ... , Typeable a_n) => p t
A static form determines a value of type StaticPtr t
, but just like OverloadedLists
and OverloadedStrings
, this literal expression is overloaded to allow lifting a StaticPtr
into another type implicitly, via the IsStatic
class:
class IsStatic p where fromStaticPtr :: StaticPtr a > p a
The only predefined instance is the obvious one that does nothing:
instance IsStatic StaticPtr where fromStaticPtr sptr = sptr
Furthermore, type t
is constrained to have a Typeable
instance. The following are therefore illegal:
static show  No Typeable instance for (Show a => a > String) static Control.Monad.ST.runST  No Typeable instance for ((forall s. ST s a) > a)
That being said, with the appropriate use of wrapper datatypes, the above limitations induce no loss of generality:
{# LANGUAGE ConstraintKinds #} {# LANGUAGE ExistentialQuantification #} {# LANGUAGE Rank2Types #} {# LANGUAGE StandaloneDeriving #} {# LANGUAGE StaticPointers #} import Control.Monad.ST import Data.Typeable import GHC.StaticPtr data Dict c = c => Dict g1 :: Typeable a => StaticPtr (Dict (Show a) > a > String) g1 = static (\Dict > show) data Rank2Wrapper f = R2W (forall s. f s) deriving Typeable newtype Flip f a s = Flip { unFlip :: f s a } deriving Typeable g2 :: Typeable a => StaticPtr (Rank2Wrapper (Flip ST a) > a) g2 = static (\(R2W f) > runST (unFlip f))
10.32. Pragmas
GHC supports several pragmas, or instructions to the compiler placed in the source code. Pragmas don’t normally affect the meaning of the program, but they might affect the efficiency of the generated code.
Pragmas all take the form {# word ... #}
where ⟨word⟩ indicates the type of pragma, and is followed optionally by information specific to that type of pragma. Case is ignored in ⟨word⟩. The various values for ⟨word⟩ that GHC understands are described in the following sections; any pragma encountered with an unrecognised ⟨word⟩ is ignored. The layout rule applies in pragmas, so the closing #}
should start in a column to the right of the opening {#
.
Certain pragmas are fileheader pragmas:
 A fileheader pragma must precede the
module
keyword in the file.  There can be as many fileheader pragmas as you please, and they can be preceded or followed by comments.
 Fileheader pragmas are read once only, before preprocessing the file (e.g. with cpp).
 The fileheader pragmas are:
{# LANGUAGE #}
,{# OPTIONS_GHC #}
, and{# INCLUDE #}
.
10.32.1. LANGUAGE
pragma

{# LANGUAGE ⟨ext⟩, ⟨ext⟩, ... #}

 Where

file header
Enable or disable a set of language extensions.
The LANGUAGE
pragma allows language extensions to be enabled in a portable way. It is the intention that all Haskell compilers support the LANGUAGE
pragma with the same syntax, although not all extensions are supported by all compilers, of course. The LANGUAGE
pragma should be used instead of OPTIONS_GHC
, if possible.
For example, to enable the FFI and preprocessing with CPP:
{# LANGUAGE ForeignFunctionInterface, CPP #}
LANGUAGE
is a fileheader pragma (see Pragmas).
Every language extension can also be turned into a commandline flag by prefixing it with “X
”; for example XForeignFunctionInterface
. (Similarly, all “X
” flags can be written as LANGUAGE
pragmas.)
A list of all supported language extensions can be obtained by invoking ghc supportedextensions
(see supportedextensions
).
Any extension from the Extension
type defined in Language.Haskell.Extension may be used. GHC will report an error if any of the requested extensions are not supported.
10.32.2. OPTIONS_GHC
pragma

{# OPTIONS_GHC ⟨flags⟩ #}

 Where

file header
The OPTIONS_GHC
pragma is used to specify additional options that are given to the compiler when compiling this source file. See Command line options in source files for details.
Previous versions of GHC accepted OPTIONS
rather than OPTIONS_GHC
, but that is now deprecated.
OPTIONS_GHC
is a fileheader pragma (see Pragmas).
10.32.3. INCLUDE
pragma
The INCLUDE
used to be necessary for specifying header files to be included when using the FFI and compiling via C. It is no longer required for GHC, but is accepted (and ignored) for compatibility with other compilers.
10.32.4. WARNING
and DEPRECATED
pragmas

{# WARNING #}

 Where

declaration
The
WARNING
pragma allows you to attach an arbitrary warning to a particular function, class, or type.

{# DEPRECATED #}

 Where

declaration
A
DEPRECATED
pragma lets you specify that a particular function, class, or type is deprecated.
There are two ways of using these pragmas.

You can work on an entire module thus:
module Wibble {# DEPRECATED "Use Wobble instead" #} where ...
Or:
module Wibble {# WARNING "This is an unstable interface." #} where ...
When you compile any module that import
Wibble
, GHC will print the specified message. 
You can attach a warning to a function, class, type, or data constructor, with the following toplevel declarations:
{# DEPRECATED f, C, T "Don't use these" #} {# WARNING unsafePerformIO "This is unsafe; I hope you know what you're doing" #}
When you compile any module that imports and uses any of the specified entities, GHC will print the specified message.
You can only attach to entities declared at top level in the module being compiled, and you can only use unqualified names in the list of entities. A capitalised name, such as
T
refers to either the type constructorT
or the data constructorT
, or both if both are in scope. If both are in scope, there is currently no way to specify one without the other (c.f. fixities Infix type constructors, classes, and type variables).
Also note that the argument to DEPRECATED
and WARNING
can also be a list of strings, in which case the strings will be presented on separate lines in the resulting warning message,
{# DEPRECATED foo, bar ["Don't use these", "Use gar instead"] #}
Warnings and deprecations are not reported for (a) uses within the defining module, (b) defining a method in a class instance, and (c) uses in an export list. The latter reduces spurious complaints within a library in which one module gathers together and reexports the exports of several others.
You can suppress the warnings with the flag Wnowarningsdeprecations
.
10.32.5. MINIMAL
pragma

{# MINIMAL ⟨name⟩  ⟨name⟩ , ... #}

 Where

in class body
Define the methods needed for a minimal complete instance of a class.
The MINIMAL
pragma is used to specify the minimal complete definition of a class, i.e. specify which methods must be implemented by all instances. If an instance does not satisfy the minimal complete definition, then a warning is generated. This can be useful when a class has methods with circular defaults. For example
class Eq a where (==) :: a > a > Bool (/=) :: a > a > Bool x == y = not (x /= y) x /= y = not (x == y) {# MINIMAL (==)  (/=) #}
Without the MINIMAL
pragma no warning would be generated for an instance that implements neither method.
The syntax for minimal complete definition is:
mindef ::= name  '(' mindef ')'  mindef '' mindef  mindef ',' mindef
A vertical bar denotes disjunction, i.e. one of the two sides is required. A comma denotes conjunction, i.e. both sides are required. Conjunction binds stronger than disjunction.
If no MINIMAL
pragma is given in the class declaration, it is just as if a pragma {# MINIMAL op1, op2, ..., opn #}
was given, where the opi
are the methods that lack a default method in the class declaration (c.f. Wmissingmethods
, Warnings and sanitychecking).
This warning can be turned off with the flag Wnomissingmethods
.
10.32.6. INLINE
and NOINLINE
pragmas
These pragmas control the inlining of function definitions.
10.32.6.1. INLINE
pragma

{# INLINE ⟨name⟩ #}

 Where

toplevel
Force GHC to inline a value.
GHC (with O
, as always) tries to inline (or “unfold”) functions/values that are “small enough,” thus avoiding the call overhead and possibly exposing other morewonderful optimisations. GHC has a set of heuristics, tuned over a long period of time using many benchmarks, that decide when it is beneficial to inline a function at its call site. The heuristics are designed to inline functions when it appears to be beneficial to do so, but without incurring excessive code bloat. If a function looks too big, it won’t be inlined, and functions larger than a certain size will not even have their definition exported in the interface file. Some of the thresholds that govern these heuristic decisions can be changed using flags, see f*: platformindependent flags.
Normally GHC will do a reasonable job of deciding by itself when it is a good idea to inline a function. However, sometimes you might want to override the default behaviour. For example, if you have a key function that is important to inline because it leads to further optimisations, but GHC judges it to be too big to inline.
The sledgehammer you can bring to bear is the INLINE
pragma, used thusly:
key_function :: Int > String > (Bool, Double) {# INLINE key_function #}
The major effect of an INLINE
pragma is to declare a function’s “cost” to be very low. The normal unfolding machinery will then be very keen to inline it. However, an INLINE
pragma for a function “f
” has a number of other effects:

While GHC is keen to inline the function, it does not do so blindly. For example, if you write
map key_function xs
there really isn’t any point in inlining
key_function
to getmap (\x > body) xs
In general, GHC only inlines the function if there is some reason (no matter how slight) to suppose that it is useful to do so.

Moreover, GHC will only inline the function if it is fully applied, where “fully applied” means applied to as many arguments as appear (syntactically) on the LHS of the function definition. For example:
comp1 :: (b > c) > (a > b) > a > c {# INLINE comp1 #} comp1 f g = \x > f (g x) comp2 :: (b > c) > (a > b) > a > c {# INLINE comp2 #} comp2 f g x = f (g x)
The two functions
comp1
andcomp2
have the same semantics, butcomp1
will be inlined when applied to two arguments, whilecomp2
requires three. This might make a big difference if you saymap (not `comp1` not) xs
which will optimise better than the corresponding use of
comp2
. 
It is useful for GHC to optimise the definition of an INLINE function
f
just like any other nonINLINE
function, in case the noninlined version off
is ultimately called. But we don’t want to inline the optimised version off
; a major reason forINLINE
pragmas is to expose functions inf
’s RHS that have rewrite rules, and it’s no good if those functions have been optimised away.So GHC guarantees to inline precisely the code that you wrote, no more and no less. It does this by capturing a copy of the definition of the function to use for inlining (we call this the “inlineRHS”), which it leaves untouched, while optimising the ordinarily RHS as usual. For externallyvisible functions the inlineRHS (not the optimised RHS) is recorded in the interface file.
 An
INLINE
function is not worker/wrappered by strictness analysis. It’s going to be inlined wholesale instead.
GHC ensures that inlining cannot go on forever: every mutuallyrecursive group is cut by one or more loop breakers that is never inlined (see Secrets of the GHC inliner, JFP 12(4) July 2002). GHC tries not to select a function with an INLINE
pragma as a loop breaker, but when there is no choice even an INLINE function can be selected, in which case the INLINE
pragma is ignored. For example, for a selfrecursive function, the loop breaker can only be the function itself, so an INLINE
pragma is always ignored.
Syntactically, an INLINE
pragma for a function can be put anywhere its type signature could be put.
INLINE
pragmas are a particularly good idea for the then
/return
(or bind
/unit
) functions in a monad. For example, in GHC’s own UniqueSupply
monad code, we have:
{# INLINE thenUs #} {# INLINE returnUs #}
See also the NOINLINE
(NOINLINE pragma) and INLINABLE
(INLINABLE pragma) pragmas.
10.32.6.2. INLINABLE
pragma

{# INLINABLE ⟨name⟩ #}

 Where

toplevel
Suggest that the compiler always consider inlining
name
.
An {# INLINABLE f #}
pragma on a function f
has the following behaviour:
 While
INLINE
says “please inline me”, theINLINABLE
says “feel free to inline me; use your discretion”. In other words the choice is left to GHC, which uses the same rules as for pragmafree functions. UnlikeINLINE
, that decision is made at the call site, and will therefore be affected by the inlining threshold, optimisation level etc.  Like
INLINE
, theINLINABLE
pragma retains a copy of the original RHS for inlining purposes, and persists it in the interface file, regardless of the size of the RHS.  One way to use
INLINABLE
is in conjunction with the special functioninline
(Special builtin functions). The callinline f
tries very hard to inlinef
. To make sure thatf
can be inlined, it is a good idea to mark the definition off
asINLINABLE
, so that GHC guarantees to expose an unfolding regardless of how big it is. Moreover, by annotatingf
asINLINABLE
, you ensure thatf
’s original RHS is inlined, rather than whatever random optimised version off
GHC’s optimiser has produced.  The
INLINABLE
pragma also works withSPECIALISE
: if you mark functionf
asINLINABLE
, then you can subsequentlySPECIALISE
in another module (see SPECIALIZE pragma).  Unlike
INLINE
, it is OK to use anINLINABLE
pragma on a recursive function. The principal reason do to so to allow later use ofSPECIALISE
The alternative spelling INLINEABLE
is also accepted by GHC.
10.32.6.3. NOINLINE
pragma

{# NOINLINE ⟨name⟩ #}

 Where

toplevel
Instructs the compiler not to inline a value.
The NOINLINE
pragma does exactly what you’d expect: it stops the named function from being inlined by the compiler. You shouldn’t ever need to do this, unless you’re very cautious about code size.
NOTINLINE
is a synonym for NOINLINE
(NOINLINE
is specified by Haskell 98 as the standard way to disable inlining, so it should be used if you want your code to be portable).
10.32.6.4. CONLIKE
modifier

{# CONLIKE #}

Instructs GHC to consider a value to be especially cheap to inline.
An INLINE
or NOINLINE
pragma may have a CONLIKE
modifier, which affects matching in RULE
s (only). See How rules interact with CONLIKE pragmas.
10.32.6.5. Phase control
Sometimes you want to control exactly when in GHC’s pipeline the INLINE
pragma is switched on. Inlining happens only during runs of the simplifier. Each run of the simplifier has a different phase number; the phase number decreases towards zero. If you use dverbosecore2core
you will see the sequence of phase numbers for successive runs of the simplifier. In an INLINE
pragma you can optionally specify a phase number, thus:
 “
INLINE[k] f
” means: do not inlinef
until phasek
, but from phasek
onwards be very keen to inline it.  “
INLINE[~k] f
” means: be very keen to inlinef
until phasek
, but from phasek
onwards do not inline it.  “
NOINLINE[k] f
” means: do not inlinef
until phasek
, but from phasek
onwards be willing to inline it (as if there was no pragma).  “
NOINLINE[~k] f
” means: be willing to inlinef
until phasek
, but from phasek
onwards do not inline it.
The same information is summarised here:
 Before phase 2 Phase 2 and later {# INLINE [2] f #}  No Yes {# INLINE [~2] f #}  Yes No {# NOINLINE [2] f #}  No Maybe {# NOINLINE [~2] f #}  Maybe No {# INLINE f #}  Yes Yes {# NOINLINE f #}  No No
By “Maybe” we mean that the usual heuristic inlining rules apply (if the function body is small, or it is applied to interestinglooking arguments etc). Another way to understand the semantics is this:
 For both
INLINE
andNOINLINE
, the phase number says when inlining is allowed at all.  The
INLINE
pragma has the additional effect of making the function body look small, so that when inlining is allowed it is very likely to happen.
The same phasenumbering control is available for RULE
s (Rewrite rules).
10.32.7. LINE
pragma

{# LINE ⟨lineno⟩ "⟨file⟩" #}

 Where

anywhere
Generated by preprocessors to convey source line numbers of the original source.
This pragma is similar to C’s #line
pragma, and is mainly for use in automatically generated Haskell code. It lets you specify the line number and filename of the original code; for example
{# LINE 42 "Foo.vhs" #}
if you’d generated the current file from something called Foo.vhs
and this line corresponds to line 42 in the original. GHC will adjust its error messages to refer to the line/file named in the LINE
pragma.
LINE
pragmas generated from Template Haskell set the file and line position for the duration of the splice and are limited to the splice. Note that because Template Haskell splices abstract syntax, the file positions are not automatically advanced.
10.32.8. COLUMN
pragma
This is the analogue of the LINE
pragma and is likewise intended for use in automatically generated Haskell code. It lets you specify the column number of the original code; for example
foo = do {# COLUMN 42 #}pure () pure ()
This adjusts all column numbers immediately after the pragma to start at 42. The presence of this pragma only affects the quality of the diagnostics and does not change the syntax of the code itself.
10.32.9. RULES
pragma
The RULES
pragma lets you specify rewrite rules. It is described in Rewrite rules.
10.32.10. SPECIALIZE
pragma

{# SPECIALIZE ⟨name⟩ :: ⟨type⟩ #}

Ask that GHC specialize a polymorphic value to a particular type.
(UK spelling also accepted.) For key overloaded functions, you can create extra versions (NB: at the cost of larger code) specialised to particular types. Thus, if you have an overloaded function:
hammeredLookup :: Ord key => [(key, value)] > key > value
If it is heavily used on lists with Widget
keys, you could specialise it as follows:
{# SPECIALIZE hammeredLookup :: [(Widget, value)] > Widget > value #}
 A
SPECIALIZE
pragma for a function can be put anywhere its type signature could be put. Moreover, you can alsoSPECIALIZE
an imported function provided it was given anINLINABLE
pragma at its definition site (INLINABLE pragma).  A
SPECIALIZE
has the effect of generating (a) a specialised version of the function and (b) a rewrite rule (see Rewrite rules) that rewrites a call to the unspecialised function into a call to the specialised one. Moreover, given aSPECIALIZE
pragma for a functionf
, GHC will automatically create specialisations for any typeclassoverloaded functions called byf
, if they are in the same module as theSPECIALIZE
pragma, or if they areINLINABLE
; and so on, transitively. 
You can add phase control (Phase control) to the RULE generated by a
SPECIALIZE
pragma, just as you can if you write aRULE
directly. For example:{# SPECIALIZE [0] hammeredLookup :: [(Widget, value)] > Widget > value #}
generates a specialisation rule that only fires in Phase 0 (the final phase). If you do not specify any phase control in the
SPECIALIZE
pragma, the phase control is inherited from the inline pragma (if any) of the function. For example:foo :: Num a => a > a foo = ...blah... {# NOINLINE [0] foo #} {# SPECIALIZE foo :: Int > Int #}
The
NOINLINE
pragma tells GHC not to inlinefoo
until Phase 0; and this property is inherited by the specialisation RULE, which will therefore only fire in Phase 0.The main reason for using phase control on specialisations is so that you can write optimisation RULES that fire early in the compilation pipeline, and only then specialise the calls to the function. If specialisation is done too early, the optimisation rules might fail to fire.

The type in a
SPECIALIZE
pragma can be any type that is less polymorphic than the type of the original function. In concrete terms, if the original function isf
then the pragma{# SPECIALIZE f :: <type> #}
is valid if and only if the definition
f_spec :: <type> f_spec = f
is valid. Here are some examples (where we only give the type signature for the original function, not its code):
f :: Eq a => a > b > b {# SPECIALISE f :: Int > b > b #} g :: (Eq a, Ix b) => a > b > b {# SPECIALISE g :: (Eq a) => a > Int > Int #} h :: Eq a => a > a > a {# SPECIALISE h :: (Eq a) => [a] > [a] > [a] #}
The last of these examples will generate a RULE with a somewhatcomplex lefthand side (try it yourself), so it might not fire very well. If you use this kind of specialisation, let us know how well it works.
10.32.10.1. SPECIALIZE INLINE

{# SPECIALIZE INLINE ⟨name⟩ :: ⟨type⟩ #}

 Where

toplevel
A SPECIALIZE
pragma can optionally be followed with a INLINE
or NOINLINE
pragma, optionally followed by a phase, as described in INLINE and NOINLINE pragmas. The INLINE
pragma affects the specialised version of the function (only), and applies even if the function is recursive. The motivating example is this:
 A GADT for arrays with typeindexed representation data Arr e where ArrInt :: !Int > ByteArray# > Arr Int ArrPair :: !Int > Arr e1 > Arr e2 > Arr (e1, e2) (!:) :: Arr e > Int > e {# SPECIALISE INLINE (!:) :: Arr Int > Int > Int #} {# SPECIALISE INLINE (!:) :: Arr (a, b) > Int > (a, b) #} (ArrInt _ ba) !: (I# i) = I# (indexIntArray# ba i) (ArrPair _ a1 a2) !: i = (a1 !: i, a2 !: i)
Here, (!:)
is a recursive function that indexes arrays of type Arr e
. Consider a call to (!:)
at type (Int,Int)
. The second specialisation will fire, and the specialised function will be inlined. It has two calls to (!:)
, both at type Int
. Both these calls fire the first specialisation, whose body is also inlined. The result is a typebased unrolling of the indexing function.
You can add explicit phase control (Phase control) to SPECIALISE INLINE
pragma, just like on an INLINE
pragma; if you do so, the same phase is used for the rewrite rule and the INLINE control of the specialised function.
Warning
You can make GHC diverge by using SPECIALISE INLINE
on an ordinarilyrecursive function.
10.32.10.2. SPECIALIZE
for imported functions
Generally, you can only give a SPECIALIZE
pragma for a function defined in the same module. However if a function f
is given an INLINABLE
pragma at its definition site, then it can subsequently be specialised by importing modules (see INLINABLE pragma). For example
module Map( lookup, blah blah ) where lookup :: Ord key => [(key,a)] > key > Maybe a lookup = ... {# INLINABLE lookup #} module Client where import Map( lookup ) data T = T1  T2 deriving( Eq, Ord ) {# SPECIALISE lookup :: [(T,a)] > T > Maybe a
Here, lookup
is declared INLINABLE
, but it cannot be specialised for type T
at its definition site, because that type does not exist yet. Instead a client module can define T
and then specialise lookup
at that type.
Moreover, every module that imports Client
(or imports a module that imports Client
, transitively) will “see”, and make use of, the specialised version of lookup
. You don’t need to put a SPECIALIZE
pragma in every module.
Moreover you often don’t even need the SPECIALIZE
pragma in the first place. When compiling a module M
, GHC’s optimiser (when given the O
flag) automatically considers each toplevel overloaded function declared in M
, and specialises it for the different types at which it is called in M
. The optimiser also considers each imported INLINABLE
overloaded function, and specialises it for the different types at which it is called in M
. So in our example, it would be enough for lookup
to be called at type T
:
module Client where import Map( lookup ) data T = T1  T2 deriving( Eq, Ord ) findT1 :: [(T,a)] > Maybe a findT1 m = lookup m T1  A call of lookup at type T
However, sometimes there are no such calls, in which case the pragma can be useful.
10.32.11. SPECIALIZE
instance pragma

{# SPECIALIZE instance ⟨instance head⟩ #}

 Where

instance body
Same idea, except for instance declarations. For example:
instance (Eq a) => Eq (Foo a) where { {# SPECIALIZE instance Eq (Foo [(Int, Bar)]) #} ... usual stuff ... }
The pragma must occur inside the where
part of the instance declaration.
10.32.12. UNPACK
pragma

{# UNPACK #}

 Where

data constructor field
Instructs the compiler to unpack the contents of a constructor field into the constructor itself.
The UNPACK
indicates to the compiler that it should unpack the contents of a constructor field into the constructor itself, removing a level of indirection. For example:
data T = T {# UNPACK #} !Float {# UNPACK #} !Float
will create a constructor T
containing two unboxed floats. This may not always be an optimisation: if the T
constructor is scrutinised and the floats passed to a nonstrict function for example, they will have to be reboxed (this is done automatically by the compiler).
Unpacking constructor fields should only be used in conjunction with O
1, in order to expose unfoldings to the compiler so the reboxing can be removed as often as possible. For example:
f :: T > Float f (T f1 f2) = f1 + f2
The compiler will avoid reboxing f1
and f2
by inlining +
on floats, but only when O
is on.
Any singleconstructor data is eligible for unpacking; for example
data T = T {# UNPACK #} !(Int,Int)
will store the two Int
s directly in the T
constructor, by flattening the pair. Multilevel unpacking is also supported:
data T = T {# UNPACK #} !S data S = S {# UNPACK #} !Int {# UNPACK #} !Int
will store two unboxed Int#
s directly in the T
constructor. The unpacker can see through newtypes, too.
See also the funboxstrictfields
flag, which essentially has the effect of adding {# UNPACK #}
to every strict constructor field.
10.32.13. NOUNPACK
pragma

{# NOUNPACK #}

 Where

toplevel
Instructs the compiler not to unpack a constructor field.
The NOUNPACK
pragma indicates to the compiler that it should not unpack the contents of a constructor field. Example:
data T = T {# NOUNPACK #} !(Int,Int)
Even with the flags funboxstrictfields
and O
, the field of the constructor T
is not unpacked.
10.32.14. SOURCE
pragma

{# SOURCE #}

 Where

after
import
statement
Import a module by
hsboot
file to break a module loop.
The {# SOURCE #}
pragma is used only in import
declarations, to break a module loop. It is described in detail in How to compile mutually recursive modules.
10.32.15. COMPLETE
pragmas

{# COMPLETE #}

 Where

at top level
Specify the set of constructors or pattern synonyms which constitute a total match.
The COMPLETE
pragma is used to inform the pattern match checker that a certain set of patterns is complete and that any function which matches on all the specified patterns is total.
The most common usage of COMPLETE
pragmas is with Pattern synonyms. On its own, the checker is very naive and assumes that any match involving a pattern synonym will fail. As a result, any pattern match on a pattern synonym is regarded as incomplete unless the user adds a catchall case.
For example, the data types 2 * A
and A + A
are isomorphic but some computations are more naturally expressed in terms of one or the other. To get the best of both worlds, we can choose one as our implementation and then provide a set of pattern synonyms so that users can use the other representation if they desire. We can then specify a COMPLETE
pragma in order to inform the pattern match checker that a function which matches on both LeftChoice
and RightChoice
is total.
data Choice a = Choice Bool a pattern LeftChoice :: a > Choice a pattern LeftChoice a = Choice False a pattern RightChoice :: a > Choice a pattern RightChoice a = Choice True a {# COMPLETE LeftChoice, RightChoice #} foo :: Choice Int > Int foo (LeftChoice n) = n * 2 foo (RightChoice n) = n  2
COMPLETE
pragmas are only used by the pattern match checker. If a function definition matches on all the constructors specified in the pragma then the compiler will produce no warning.
COMPLETE
pragmas can contain any data constructors or pattern synonyms which are in scope, but must mention at least one data constructor or pattern synonym defined in the same module. COMPLETE
pragmas may only appear at the top level of a module. Once defined, they are automatically imported and exported from modules. COMPLETE
pragmas should be thought of as asserting a universal truth about a set of patterns and as a result, should not be used to silence context specific incomplete match warnings.
When specifying a COMPLETE
pragma, the result types of all patterns must be consistent with each other. This is a sanity check as it would be impossible to match on all the patterns if the types were inconsistent.
The result type must also be unambiguous. Usually this can be inferred but when all the pattern synonyms in a group are polymorphic in the constructor the user must provide a type signature.
class LL f where go :: f a > () instance LL [] where go _ = () pattern T :: LL f => f a pattern T < (go > ()) {# COMPLETE T :: [] #}  No warning foo :: [a] > Int foo T = 5
10.32.16. OVERLAPPING
, OVERLAPPABLE
, OVERLAPS
, and INCOHERENT
pragmas

{# OVERLAPPING #}

{# OVERLAPPABLE #}

{# OVERLAPS #}

{# INCOHERENT #}

 Where

on instance head
The pragmas OVERLAPPING
, OVERLAPPABLE
, OVERLAPS
, INCOHERENT
are used to specify the overlap behavior for individual instances, as described in Section Overlapping instances. The pragmas are written immediately after the instance
keyword, like this:
instance {# OVERLAPPING #} C t where ...
10.33. Rewrite rules

{# RULES "⟨name⟩" forall ⟨binder⟩ ... . ⟨expr⟩ = ⟨expr⟩ ... #}

 Where

toplevel
Define a rewrite rule to be used to optimize a source program.
The programmer can specify rewrite rules as part of the source program (in a pragma). Here is an example:
{# RULES "map/map" forall f g xs. map f (map g xs) = map (f.g) xs #}
Use the debug flag ddumpsimplstats
to see what rules fired. If you need more information, then ddumprulefirings
shows you each individual rule firing and ddumprulerewrites
also shows what the code looks like before and after the rewrite.

fenablerewriterules

Allow the compiler to apply rewrite rules to the source program.
10.33.1. Syntax
From a syntactic point of view:
 There may be zero or more rules in a
RULES
pragma, separated by semicolons (which may be generated by the layout rule). 
The layout rule applies in a pragma. Currently no new indentation level is set, so if you put several rules in single
RULES
pragma and wish to use layout to separate them, you must lay out the starting in the same column as the enclosing definitions.{# RULES "map/map" forall f g xs. map f (map g xs) = map (f.g) xs "map/append" forall f xs ys. map f (xs ++ ys) = map f xs ++ map f ys #}
Furthermore, the closing
#}
should start in a column to the right of the opening{#
.  Each rule has a name, enclosed in double quotes. The name itself has no significance at all. It is only used when reporting how many times the rule fired.

A rule may optionally have a phasecontrol number (see Phase control), immediately after the name of the rule. Thus:
{# RULES "map/map" [2] forall f g xs. map f (map g xs) = map (f.g) xs #}
The
[2]
means that the rule is active in Phase 2 and subsequent phases. The inverse notation[~2]
is also accepted, meaning that the rule is active up to, but not including, Phase 2.Rules support the special phasecontrol notation
[~]
, which means the rule is never active. This feature supports plugins (see Compiler Plugins), by making it possible to define a RULE that is never run by GHC, but is nevertheless parsed, typechecked etc, so that it is available to the plugin.  Each (term) variable mentioned in a rule must either be in scope (e.g.
map
), or bound by theforall
(e.g.f
,g
,xs
). The variables bound by theforall
are called the pattern variables. They are separated by spaces, just like in a typeforall
. 
A pattern variable may optionally have a type signature. If the type of the pattern variable is polymorphic, it must have a type signature. For example, here is the
foldr/build
rule:"fold/build" forall k z (g::forall b. (a>b>b) > b > b) . foldr k z (build g) = g k z
Since
g
has a polymorphic type, it must have a type signature. 
If
ExplicitForAll
is enabled, type/kind variables can also be explicitly bound. For example:{# RULES "id" forall a. forall (x :: a). id @a x = x #}
When a typelevel explicit
forall
is present, each type/kind variable mentioned must now also be either in scope or bound by theforall
. In particular, unlike some other places in Haskell, this means free kind variables will not be implicitly bound. For example:"this_is_bad" forall (c :: k). forall (x :: Proxy c) ... "this_is_ok" forall k (c :: k). forall (x :: Proxy c) ...
When bound type/kind variables are needed, both foralls must always be included, though if no pattern variables are needed, the second can be left empty. For example:
{# RULES "map/id" forall a. forall. map (id @a) = id @[a] #}

The left hand side of a rule must consist of a toplevel variable applied to arbitrary expressions. For example, this is not OK:
"wrong1" forall e1 e2. case True of { True > e1; False > e2 } = e1 "wrong2" forall f. f True = True "wrong3" forall x. Just x = Nothing
In
"wrong1"
, the LHS is not an application; in"wrong2"
, the LHS has a pattern variable in the head. In"wrong3"
, the LHS consists of a constructor, rather than a variable, applied to an argument.  A rule does not need to be in the same module as (any of) the variables it mentions, though of course they need to be in scope.
 All rules are implicitly exported from the module, and are therefore in force in any module that imports the module that defined the rule, directly or indirectly. (That is, if A imports B, which imports C, then C’s rules are in force when compiling A.) The situation is very similar to that for instance declarations.
 Inside a
RULES
“forall
” is treated as a keyword, regardless of any other flag settings. Furthermore, inside aRULES
, the language extensionScopedTypeVariables
is automatically enabled; see Lexically scoped type variables.  Like other pragmas,
RULES
pragmas are always checked for scope errors, and are typechecked. Typechecking means that the LHS and RHS of a rule are typechecked, and must have the same type. However, rules are only enabled if thefenablerewriterules
flag is on (see Semantics).
10.33.2. Semantics
From a semantic point of view:
 Rules are enabled (that is, used during optimisation) by the
fenablerewriterules
flag. This flag is implied byO
, and may be switched off (as usual) byfnoenablerewriterules
. (NB: enablingfenablerewriterules
withoutO
may not do what you expect, though, because withoutO
GHC ignores all optimisation information in interface files; seefignoreinterfacepragmas
). Note thatfenablerewriterules
is an optimisation flag, and has no effect on parsing or typechecking.  Rules are regarded as lefttoright rewrite rules. When GHC finds an expression that is a substitution instance of the LHS of a rule, it replaces the expression by the (appropriatelysubstituted) RHS. By “a substitution instance” we mean that the LHS can be made equal to the expression by substituting for the pattern variables.
 GHC makes absolutely no attempt to verify that the LHS and RHS of a rule have the same meaning. That is undecidable in general, and infeasible in most interesting cases. The responsibility is entirely the programmer’s!

GHC makes no attempt to make sure that the rules are confluent or terminating. For example:
"loop" forall x y. f x y = f y x
This rule will cause the compiler to go into an infinite loop.
 If more than one rule matches a call, GHC will choose one arbitrarily to apply.

GHC currently uses a very simple, syntactic, matching algorithm for matching a rule LHS with an expression. It seeks a substitution which makes the LHS and expression syntactically equal modulo alpha conversion. The pattern (rule), but not the expression, is etaexpanded if necessary. (Etaexpanding the expression can lead to laziness bugs.) But not beta conversion (that’s called higherorder matching).
Matching is carried out on GHC’s intermediate language, which includes type abstractions and applications. So a rule only matches if the types match too. See Specialisation below.

GHC keeps trying to apply the rules as it optimises the program. For example, consider:
let s = map f t = map g in s (t xs)
The expression
s (t xs)
does not match the rule"map/map"
, but GHC will substitute fors
andt
, giving an expression which does match. Ifs
ort
was (a) used more than once, and (b) large or a redex, then it would not be substituted, and the rule would not fire.
10.33.3. How rules interact with INLINE
/NOINLINE
pragmas
Ordinary inlining happens at the same time as rule rewriting, which may lead to unexpected results. Consider this (artificial) example
f x = x g y = f y h z = g True {# RULES "f" f True = False #}
Since f
’s righthand side is small, it is inlined into g
, to give
g y = y
Now g
is inlined into h
, but f
’s RULE has no chance to fire. If instead GHC had first inlined g
into h
then there would have been a better chance that f
’s RULES
might fire.
The way to get predictable behaviour is to use a NOINLINE
pragma, or an INLINE[⟨phase⟩]
pragma, on f
, to ensure that it is not inlined until its RULES
have had a chance to fire. The warning flag Winlineruleshadowing
(see Warnings and sanitychecking) warns about this situation.
10.33.4. How rules interact with CONLIKE
pragmas
GHC is very cautious about duplicating work. For example, consider
f k z xs = let xs = build g in ...(foldr k z xs)...sum xs... {# RULES "foldr/build" forall k z g. foldr k z (build g) = g k z #}
Since xs
is used twice, GHC does not fire the foldr/build rule. Rightly so, because it might take a lot of work to compute xs
, which would be duplicated if the rule fired.
Sometimes, however, this approach is overcautious, and we do want the rule to fire, even though doing so would duplicate redex. There is no way that GHC can work out when this is a good idea, so we provide the CONLIKE
pragma to declare it, thus:
{# INLINE CONLIKE [1] f #} f x = blah
CONLIKE
is a modifier to an INLINE
or NOINLINE
pragma. It specifies that an application of f
to one argument (in general, the number of arguments to the left of the =
sign) should be considered cheap enough to duplicate, if such a duplication would make rule fire. (The name “CONLIKE” is short for “constructorlike”, because constructors certainly have such a property.) The CONLIKE
pragma is a modifier to INLINE
/NOINLINE
because it really only makes sense to match f
on the LHS of a rule if you are sure that f
is not going to be inlined before the rule has a chance to fire.
10.33.5. How rules interact with class methods
Giving a RULE for a class method is a bad idea:
class C a where op :: a > a > a instance C Bool where op x y = ...rhs for op at Bool... {# RULES "f" op True y = False #}
In this example, op
is not an ordinary toplevel function; it is a class method. GHC rapidly rewrites any occurrences of op
usedattypeBool to a specialised function, say opBool
, where
opBool :: Bool > Bool > Bool opBool x y = ..rhs for op at Bool...
So the RULE never has a chance to fire, for just the same reasons as in How rules interact with INLINE/NOINLINE pragmas.
The solution is to define the instancespecific function yourself, with a pragma to prevent it being inlined too early, and give a RULE for it:
instance C Bool where op = opBool opBool :: Bool > Bool > Bool {# NOINLINE [1] opBool #} opBool x y = ..rhs for op at Bool... {# RULES "f" opBool True y = False #}
If you want a RULE that truly applies to the overloaded class method, the only way to do it is like this:
class C a where op_c :: a > a > a op :: C a => a > a > a {# NOINLINE [1] op #} op = op_c {# RULES "reassociate" op (op x y) z = op x (op y z) #}
Now the inlining of op
is delayed until the rule has a chance to fire. The downside is that instance declarations must define op_c
, but all other uses should go via op
.
10.33.6. List fusion
The RULES mechanism is used to implement fusion (deforestation) of common list functions. If a “good consumer” consumes an intermediate list constructed by a “good producer”, the intermediate list should be eliminated entirely.
The following are good producers:
 List comprehensions
 Enumerations of
Int
,Integer
andChar
(e.g.['a'..'z']
).  Explicit lists (e.g.
[True, False]
)  The cons constructor (e.g
3:4:[]
) ++
map

take
,filter

iterate
,repeat

zip
,zipWith
The following are good consumers:
 List comprehensions

array
(on its second argument) 
++
(on its first argument) foldr
map

take
,filter
concat

unzip
,unzip2
,unzip3
,unzip4

zip
,zipWith
(but on one argument only; if both are good producers,zip
will fuse with one but not the other) partition
head

and
,or
,any
,all
sequence_
msum
So, for example, the following should generate no intermediate lists:
array (1,10) [(i,i*i)  i < map (+ 1) [0..9]]
This list could readily be extended; if there are Prelude functions that you use a lot which are not included, please tell us.
If you want to write your own good consumers or producers, look at the Prelude definitions of the above functions to see how to do so.
10.33.7. Specialisation
Rewrite rules can be used to get the same effect as a feature present in earlier versions of GHC. For example, suppose that:
genericLookup :: Ord a => Table a b > a > b intLookup :: Table Int b > Int > b
where intLookup
is an implementation of genericLookup
that works very fast for keys of type Int
. You might wish to tell GHC to use intLookup
instead of genericLookup
whenever the latter was called with type Table Int b > Int > b
. It used to be possible to write
{# SPECIALIZE genericLookup :: Table Int b > Int > b = intLookup #}
This feature is no longer in GHC, but rewrite rules let you do the same thing:
{# RULES "genericLookup/Int" genericLookup = intLookup #}
This slightly oddlooking rule instructs GHC to replace genericLookup
by intLookup
whenever the types match. What is more, this rule does not need to be in the same file as genericLookup
, unlike the SPECIALIZE
pragmas which currently do (so that they have an original definition available to specialise).
It is Your Responsibility to make sure that intLookup
really behaves as a specialised version of genericLookup
!!!
An example in which using RULES
for specialisation will Win Big:
toDouble :: Real a => a > Double toDouble = fromRational . toRational {# RULES "toDouble/Int" toDouble = i2d #} i2d (I# i) = D# (int2Double# i)  uses Glasgow primop directly
The i2d
function is virtually one machine instruction; the default conversion—via an intermediate Rational
is obscenely expensive by comparison.
10.33.8. Controlling what’s going on in rewrite rules
 Use
ddumprules
to see the rules that are defined in this module. This includes rules generated by the specialisation pass, but excludes rules imported from other modules.  Use
ddumpsimplstats
to see what rules are being fired. If you adddpprdebug
you get a more detailed listing.  Use
ddumprulefirings
orddumprulerewrites
to see in great detail what rules are being fired. If you adddpprdebug
you get a still more detailed listing. 
The definition of (say)
build
inGHC/Base.hs
looks like this:build :: forall a. (forall b. (a > b > b) > b > b) > [a] {# INLINE build #} build g = g (:) []
Notice the
INLINE
! That prevents(:)
from being inlined when compilingPrelBase
, so that an importing module will “see” the(:)
, and can match it on the LHS of a rule.INLINE
prevents any inlining happening in the RHS of theINLINE
thing. I regret the delicacy of this.  In
libraries/base/GHC/Base.hs
look at the rules formap
to see how to write rules that will do fusion and yet give an efficient program even if fusion doesn’t happen. More rules inGHC/List.hs
.
10.34. Special builtin functions
GHC has a few builtin functions with special behaviour. In particular:
 GHC.Exts.inline allows control over inlining on a percallsite basis.
 GHC.Exts.lazy restrains the strictness analyser.
 GHC.Exts.oneShot gives a hint to the compiler about how often a function is being called.
10.35. Generic classes
GHC used to have an implementation of generic classes as defined in the paper “Derivable type classes”, Ralf Hinze and Simon Peyton Jones, Haskell Workshop, Montreal Sept 2000, pp. 94105. These have been removed and replaced by the more general support for generic programming.
10.36. Generic programming
Using a combination of DeriveGeneric
, DefaultSignatures
, and DeriveAnyClass
, you can easily do datatypegeneric programming using the GHC.Generics framework. This section gives a very brief overview of how to do it.
Generic programming support in GHC allows defining classes with methods that do not need a user specification when instantiating: the method body is automatically derived by GHC. This is similar to what happens for standard classes such as Read
and Show
, for instance, but now for userdefined classes.
10.36.1. Deriving representations
The first thing we need is generic representations. The GHC.Generics
module defines a couple of primitive types that are used to represent Haskell datatypes:
  Unit: used for constructors without arguments data U1 p = U1   Constants, additional parameters and recursion of kind Type newtype K1 i c p = K1 { unK1 :: c }   Metainformation (constructor names, etc.) newtype M1 i c f p = M1 { unM1 :: f p }   Sums: encode choice between constructors infixr 5 :+: data (:+:) f g p = L1 (f p)  R1 (g p)   Products: encode multiple arguments to constructors infixr 6 :*: data (:*:) f g p = f p :*: g p
The Generic
and Generic1
classes mediate between userdefined datatypes and their internal representation as a sumofproducts:
class Generic a where  Encode the representation of a user datatype type Rep a :: Type > Type  Convert from the datatype to its representation from :: a > (Rep a) x  Convert from the representation to the datatype to :: (Rep a) x > a class Generic1 (f :: k > Type) where type Rep1 f :: k > Type from1 :: f a > Rep1 f a to1 :: Rep1 f a > f a
Generic1
is used for functions that can only be defined over type containers, such as map
. Note that Generic1
ranges over types of kind Type > Type
by default, but if the PolyKinds
extension is enabled, then it can range of types of kind k > Type
, for any kind k
.

DeriveGeneric

 Since

7.2.1
Allow automatic deriving of instances for the
Generic
typeclass.
Instances of these classes can be derived by GHC with the DeriveGeneric
extension, and are necessary to be able to define generic instances automatically.
For example, a userdefined datatype of trees
data UserTree a = Node a (UserTree a) (UserTree a)  Leaf
in a Main
module in a package named foo
will get the following representation:
instance Generic (UserTree a) where  Representation type type Rep (UserTree a) = M1 D ('MetaData "UserTree" "Main" "packagename" 'False) ( M1 C ('MetaCons "Node" 'PrefixI 'False) ( M1 S ('MetaSel 'Nothing 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (K1 R a) :*: M1 S ('MetaSel 'Nothing 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (K1 R (UserTree a)) :*: M1 S ('MetaSel 'Nothing 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (K1 R (UserTree a))) :+: M1 C ('MetaCons "Leaf" 'PrefixI 'False) U1)  Conversion functions from (Node x l r) = M1 (L1 (M1 (M1 (K1 x) :*: M1 (K1 l) :*: M1 (K1 r)))) from Leaf = M1 (R1 (M1 U1)) to (M1 (L1 (M1 (M1 (K1 x) :*: M1 (K1 l) :*: M1 (K1 r))))) = Node x l r to (M1 (R1 (M1 U1))) = Leaf
This representation is generated automatically if a deriving Generic
clause is attached to the datatype. Standalone deriving can also be used.
10.36.2. Writing generic functions
A generic function is defined by creating a class and giving instances for each of the representation types of GHC.Generics
. As an example we show generic serialization:
data Bin = O  I class GSerialize f where gput :: f a > [Bin] instance GSerialize U1 where gput U1 = [] instance (GSerialize a, GSerialize b) => GSerialize (a :*: b) where gput (x :*: y) = gput x ++ gput y instance (GSerialize a, GSerialize b) => GSerialize (a :+: b) where gput (L1 x) = O : gput x gput (R1 x) = I : gput x instance (GSerialize a) => GSerialize (M1 i c a) where gput (M1 x) = gput x instance (Serialize a) => GSerialize (K1 i a) where gput (K1 x) = put x
A caveat: this encoding strategy may not be reliable across different versions of GHC. When deriving a Generic
instance is free to choose any nesting of :+:
and :*:
it chooses, so if GHC chooses (a :+: b) :+: c
, then the encoding for a
would be [O, O]
, b
would be [O, I]
, and c
would be [I]
. However, if GHC chooses a :+: (b :+: c)
, then the encoding for a
would be [O]
, b
would be [I, O]
, and c
would be [I, I]
. (In practice, the current implementation tries to produce a moreorless balanced nesting of :+:
and :*:
so that the traversal of the structure of the datatype from the root to a particular component can be performed in logarithmic rather than linear time.)
Typically this GSerialize
class will not be exported, as it only makes sense to have instances for the representation types.
10.36.3. Unlifted representation types
The data family URec
is provided to enable generic programming over datatypes with certain unlifted arguments. There are six instances corresponding to common unlifted types:
data family URec a p data instance URec (Ptr ()) p = UAddr { uAddr# :: Addr# } data instance URec Char p = UChar { uChar# :: Char# } data instance URec Double p = UDouble { uDouble# :: Double# } data instance URec Int p = UInt { uInt# :: Int# } data instance URec Float p = UFloat { uFloat# :: Float# } data instance URec Word p = UWord { uWord# :: Word# }
Six type synonyms are provided for convenience:
type UAddr = URec (Ptr ()) type UChar = URec Char type UDouble = URec Double type UFloat = URec Float type UInt = URec Int type UWord = URec Word
As an example, this data declaration:
data IntHash = IntHash Int# deriving Generic
results in the following Generic
instance:
instance 'Generic' IntHash where type 'Rep' IntHash = 'D1' ('MetaData "IntHash" "Main" "packagename" 'False) ('C1' ('MetaCons "IntHash" 'PrefixI 'False) ('S1' ('MetaSel 'Nothing 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) 'UInt'))
A user could provide, for example, a GSerialize UInt
instance so that a Serialize IntHash
instance could be easily defined in terms of GSerialize
.
10.36.4. Generic defaults
The only thing left to do now is to define a “frontend” class, which is exposed to the user:
class Serialize a where put :: a > [Bin] default put :: (Generic a, GSerialize (Rep a)) => a > [Bin] put = gput . from
Here we use a default signature to specify that the user does not have to provide an implementation for put
, as long as there is a Generic
instance for the type to instantiate. For the UserTree
type, for instance, the user can just write:
instance (Serialize a) => Serialize (UserTree a)
The default method for put
is then used, corresponding to the generic implementation of serialization. If you are using DeriveAnyClass
, the same instance is generated by simply attaching a deriving Serialize
clause to the UserTree
datatype declaration. For more examples of generic functions please refer to the genericderiving package on Hackage.
10.36.5. More information
For more details please refer to the Haskell Wiki page or the original paper [Generics2010].

Generics2010

Jose Pedro Magalhaes, Atze Dijkstra, Johan Jeuring, and Andres Loeh. A generic deriving mechanism for Haskell. Proceedings of the third ACM Haskell symposium on Haskell (Haskell’2010), pp. 3748, ACM, 2010.
10.37. Roles
Using GeneralizedNewtypeDeriving
(Generalised derived instances for newtypes), a programmer can take existing instances of classes and “lift” these into instances of that class for a newtype. However, this is not always safe. For example, consider the following:
newtype Age = MkAge { unAge :: Int } type family Inspect x type instance Inspect Age = Int type instance Inspect Int = Bool class BadIdea a where bad :: a > Inspect a instance BadIdea Int where bad = (> 0) deriving instance BadIdea Age  not allowed!
If the derived instance were allowed, what would the type of its method bad
be? It would seem to be Age > Inspect Age
, which is equivalent to Age > Int
, according to the type family Inspect
. Yet, if we simply adapt the implementation from the instance for Int
, the implementation for bad
produces a Bool
, and we have trouble.
The way to identify such situations is to have roles assigned to type variables of datatypes, classes, and type synonyms.
Roles as implemented in GHC are a from a simplified version of the work described in Generative type abstraction and typelevel computation, published at POPL 2011.
10.37.1. Nominal, Representational, and Phantom
The goal of the roles system is to track when two types have the same underlying representation. In the example above, Age
and Int
have the same representation. But, the corresponding instances of BadIdea
would not have the same representation, because the types of the implementations of bad
would be different.
Suppose we have two uses of a type constructor, each applied to the same parameters except for one difference. (For example, T Age Bool c
and T Int Bool c
for some type T
.) The role of a type parameter says what we need to know about the two differing type arguments in order to know that the two outer types have the same representation (in the example, what must be true about Age
and Int
in order to show that T Age Bool c
has the same representation as T Int Bool c
).
GHC supports three different roles for type parameters: nominal, representational, and phantom. If a type parameter has a nominal role, then the two types that differ must not actually differ at all: they must be identical (after type family reduction). If a type parameter has a representational role, then the two types must have the same representation. (If T
’s first parameter’s role is representational, then T Age Bool c
and T Int Bool c
would have the same representation, because Age
and Int
have the same representation.) If a type parameter has a phantom role, then we need no further information.
Here are some examples:
data Simple a = MkSimple a  a has role representational type family F type instance F Int = Bool type instance F Age = Char data Complex a = MkComplex (F a)  a has role nominal data Phant a = MkPhant Bool  a has role phantom
The type Simple
has its parameter at role representational, which is generally the most common case. Simple Age
would have the same representation as Simple Int
. The type Complex
, on the other hand, has its parameter at role nominal, because Complex Age
and Complex Int
are not the same. Lastly, Phant Age
and Phant Bool
have the same representation, even though Age
and Bool
are unrelated.
10.37.2. Role inference
What role should a given type parameter should have? GHC performs role inference to determine the correct role for every parameter. It starts with a few base facts: (>)
has two representational parameters; (~)
has two nominal parameters; all type families’ parameters are nominal; and all GADTlike parameters are nominal. Then, these facts are propagated to all places where these types are used. The default role for datatypes and synonyms is phantom; the default role for classes is nominal. Thus, for datatypes and synonyms, any parameters unused in the righthand side (or used only in other types in phantom positions) will be phantom. Whenever a parameter is used in a representational position (that is, used as a type argument to a constructor whose corresponding variable is at role representational), we raise its role from phantom to representational. Similarly, when a parameter is used in a nominal position, its role is upgraded to nominal. We never downgrade a role from nominal to phantom or representational, or from representational to phantom. In this way, we infer the mostgeneral role for each parameter.
Classes have their roles default to nominal to promote coherence of class instances. If a C Int
were stored in a datatype, it would be quite bad if that were somehow changed into a C Age
somewhere, especially if another C Age
had been declared!
There is one particularly tricky case that should be explained:
data Tricky a b = MkTricky (a b)
What should Tricky
’s roles be? At first blush, it would seem that both a
and b
should be at role representational, since both are used in the righthand side and neither is involved in a type family. However, this would be wrong, as the following example shows:
data Nom a = MkNom (F a)  type family F from example above
Is Tricky Nom Age
representationally equal to Tricky Nom Int
? No! The former stores a Char
and the latter stores a Bool
. The solution to this is to require all parameters to type variables to have role nominal. Thus, GHC would infer role representational for a
but role nominal for b
.
10.37.3. Role annotations

RoleAnnotations

 Since

7.8.1
Allow role annotation syntax.
Sometimes the programmer wants to constrain the inference process. For example, the base library contains the following definition:
data Ptr a = Ptr Addr#
The idea is that a
should really be a representational parameter, but role inference assigns it to phantom. This makes some level of sense: a pointer to an Int
really is representationally the same as a pointer to a Bool
. But, that’s not at all how we want to use Ptr
s! So, we want to be able to say
type role Ptr representational data Ptr a = Ptr Addr#
The type role
(enabled with RoleAnnotations
) declaration forces the parameter a
to be at role representational, not role phantom. GHC then checks the usersupplied roles to make sure they don’t break any promises. It would be bad, for example, if the user could make BadIdea
’s role be representational.
As another example, we can consider a type Set a
that represents a set of data, ordered according to a
’s Ord
instance. While it would generally be typesafe to consider a
to be at role representational, it is possible that a newtype
and its base type have different orderings encoded in their respective Ord
instances. This would lead to misbehavior at runtime. So, the author of the Set
datatype would like its parameter to be at role nominal. This would be done with a declaration
type role Set nominal
Role annotations can also be used should a programmer wish to write a class with a representational (or phantom) role. However, as a class with nonnominal roles can quickly lead to class instance incoherence, it is necessary to also specify IncoherentInstances
to allow nonnominal roles for classes.
The other place where role annotations may be necessary are in hsboot
files (How to compile mutually recursive modules), where the righthand sides of definitions can be omitted. As usual, the types/classes declared in an hsboot
file must match up with the definitions in the hs
file, including down to the roles. The default role for datatypes is representational in hsboot
files, corresponding to the common use case.
Role annotations are allowed on data, newtype, and class declarations. A role annotation declaration starts with type role
and is followed by one role listing for each parameter of the type. (This parameter count includes parameters implicitly specified by a kind signature in a GADTstyle data or newtype declaration.) Each role listing is a role (nominal
, representational
, or phantom
) or a _
. Using a _
says that GHC should infer that role. The role annotation may go anywhere in the same module as the datatype or class definition (much like a valuelevel type signature). Here are some examples:
type role T1 _ phantom data T1 a b = MkT1 a  b is not used; annotation is fine but unnecessary type role T2 _ phantom data T2 a b = MkT2 b  ERROR: b is used and cannot be phantom type role T3 _ nominal data T3 a b = MkT3 a  OK: nominal is higher than necessary, but safe type role T4 nominal data T4 a = MkT4 (a Int)  OK, but nominal is higher than necessary type role C representational _  OK, with XIncoherentInstances class C a b where ...  OK, b will get a nominal role type role X nominal type X a = ...  ERROR: role annotations not allowed for type synonyms
10.38. HasCallStack
GHC.Stack.HasCallStack
is a lightweight method of obtaining a partial callstack at any point in the program.
A function can request its callsite with the HasCallStack
constraint and access it as a Haskell value by using callStack
.
One can then use functions from GHC.Stack
to inspect or pretty print (as is done in f
below) the call stack.
f :: HasCallStack => IO () f = putStrLn (prettyCallStack callStack)
g :: HasCallStack => IO () g = f
Evaluating f
directly shows a call stack with a single entry, while evaluating g
, which also requests its callsite, shows two entries, one for each computation “annotated” with HasCallStack
.
ghci> f CallStack (from HasCallStack): f, called at <interactive>:19:1 in interactive:Ghci1 ghci> g CallStack (from HasCallStack): f, called at <interactive>:17:5 in main:Main g, called at <interactive>:20:1 in interactive:Ghci2
The error
function from the Prelude supports printing the call stack that led to the error in addition to the usual error message:
ghci> error "bad" *** Exception: bad CallStack (from HasCallStack): error, called at <interactive>:25:1 in interactive:Ghci5
The call stack here consists of a single entry, pinpointing the source of the call to error
. However, by annotating several computations with HasCallStack
, figuring out the exact circumstances and sequences of calls that lead to a call to error
becomes a lot easier, as demonstrated with the simple example below.
f :: HasCallStack => IO () f = error "bad bad bad" g :: HasCallStack => IO () g = f h :: HasCallStack => IO () h = g
ghci> h *** Exception: bad bad bad CallStack (from HasCallStack): error, called at callstack.hs:4:5 in main:Main f, called at callstack.hs:7:5 in main:Main g, called at callstack.hs:10:5 in main:Main h, called at <interactive>:28:1 in interactive:Ghci1
The CallStack
will only extend as far as the types allow it, for example
myHead :: HasCallStack => [a] > a myHead [] = errorWithCallStack "empty" myHead (x:xs) = x bad :: Int bad = myHead []
ghci> bad *** Exception: empty CallStack (from HasCallStack): errorWithCallStack, called at Bad.hs:8:15 in main:Bad myHead, called at Bad.hs:12:7 in main:Bad
includes the callsite of errorWithCallStack
in myHead
, and of myHead
in bad
, but not the callsite of bad
at the GHCi prompt.
GHC solves HasCallStack
constraints in two steps:
 If there is a
CallStack
in scope – i.e. the enclosing definition has aHasCallStack
constraint – GHC will push the new callsite onto the existingCallStack
.  Otherwise GHC will solve the
HasCallStack
constraint for the singletonCallStack
containing just the current callsite.
Importantly, GHC will never infer a HasCallStack
constraint, you must request it explicitly.
CallStack
is kept abstract, but GHC provides a function
getCallStack :: CallStack > [(String, SrcLoc)]
to access the individual callsites in the stack. The String
is the name of the function that was called, and the SrcLoc
provides the package, module, and file name, as well as the line and column numbers.
GHC.Stack
additionally exports a function withFrozenCallStack
that allows users to freeze the current CallStack
, preventing any future push operations from having an effect. This can be used by library authors to prevent CallStack
s from exposing unnecessary implementation details. Consider the myHead
example above, the errorWithCallStack
line in the printed stack is not particularly enlightening, so we might choose to suppress it by freezing the CallStack
that we pass to errorWithCallStack
.
myHead :: HasCallStack => [a] > a myHead [] = withFrozenCallStack (errorWithCallStack "empty") myHead (x:xs) = x
ghci> myHead [] *** Exception: empty CallStack (from HasCallStack): myHead, called at Bad.hs:12:7 in main:Bad
NOTE: The intrepid user may notice that HasCallStack
is just an alias for an implicit parameter ?callStack :: CallStack
. This is an implementation detail and should not be considered part of the CallStack
API, we may decide to change the implementation in the future.
10.38.1. Compared with other sources of stack traces
HasCallStack
does not interact with the RTS and does not require compilation with prof
. On the other hand, as the CallStack
is built up explicitly via the HasCallStack
constraints, it will generally not contain as much information as the simulated callstacks maintained by the RTS.
© 2002–2007 The University Court of the University of Glasgow. All rights reserved.
Licensed under the Glasgow Haskell Compiler License.
https://downloads.haskell.org/~ghc/8.10.2/docs/html/users_guide/glasgow_exts.html