Chapter 18 Type Classes
- 18.1 Class and Instance declarations
- 18.2 Binding classes
- 18.3 Parameterized Instances
- 18.4 Sections and contexts
- 18.5 Building hierarchies
- 18.6 Summary of the commands
This chapter presents a quick reference of the commands related to type classes. For an actual introduction to type classes, there is a description of the system [127] and the literature on type classes in Haskell which also applies.
18.1 Class and Instance declarations
The syntax for class and instance declarations is the same as record syntax of Coq:
|
|
The αi : τi variables are called the parameters of the class and the fk : typek are called the methods. Each class definition gives rise to a corresponding record declaration and each instance is a regular definition whose name is given by ident and type is an instantiation of the record type.
We’ll use the following example class in the rest of the chapter:
Coq < eqb : A -> A -> bool ;
Coq < eqb_leibniz : forall x y, eqb x y = true -> x = y }.
This class implements a boolean equality test which is compatible with Leibniz equality on some type. An example implementation is:
Coq < { eqb x y := true ;
Coq < eqb_leibniz x y H :=
Coq < match x, y return x = y with tt, tt => refl_equal tt end }.
If one does not give all the members in the Instance declaration, Coq enters the proof-mode and the user is asked to build inhabitants of the remaining fields, e.g.:
Coq < { eqb x y := if x then y else negb y }.
Coq < Proof. intros x y H.
1 subgoal
x : bool
y : bool
H : (if x then y else negb y) = true
============================
x = y
Coq < destruct x ; destruct y ; (discriminate || reflexivity).
Proof completed.
Coq < Defined.
refine {|
eqb := fun x y : bool => if x then y else negb y;
eqb_leibniz := _
:forall x y : bool,
(if x then y else negb y) = true -> x = y |}.
intros x y H.
destruct x; destruct y; discriminate || reflexivity.
eq_bool is defined
One has to take care that the transparency of every field is determined by the transparency of the Instance proof. One can use alternatively the Program Instance variant which has richer facilities for dealing with obligations.
18.2 Binding classes
Once a type class is declared, one can use it in class binders:
neqb is defined
When one calls a class method, a constraint is generated that is satisfied only in contexts where the appropriate instances can be found. In the example above, a constraint EqDec A is generated and satisfied by eqa : EqDec A. In case no satisfying constraint can be found, an error is raised:
Toplevel input, characters 47-50:
> Definition neqb’ (A : Type) (x y : A) := negb (eqb x y).
> ^^^
Error: Cannot infer the implicit parameter EqDec of
eqb.
Could not find an instance for "EqDec A" in environment:
A : Type
x : A
y : A
The algorithm used to solve constraints is a variant of the eauto tactic that does proof search with a set of lemmas (the instances). It will use local hypotheses as well as declared lemmas in the typeclass_instances database. Hence the example can also be written:
neqb’ is defined
However, the generalizing binders should be used instead as they have particular support for type classes:
- They automatically set the maximally implicit status for type class arguments, making derived functions as easy to use as class methods. In the example above, A and eqa should be set maximally implicit.
- They support implicit quantification on partialy applied type classes.
- They also support implicit quantification on superclasses (§18.5.1)
Following the previous example, one can write:
neqb_impl is defined
Here A is implicitly generalized, and the resulting function is equivalent to the one above.
The parsing of generalized type-class binders is different from regular binders:
- Implicit arguments of the class type are ignored.
- Superclasses arguments are automatically generalized.
- Any remaining arguments not given as part of a type class binder will be automatically generalized. In other words, the rightmost parameters are automatically generalized if not mentionned.
One can switch off this special treatment using the ! mark in front of the class name (see example below).
18.3 Parameterized Instances
One can declare parameterized instances as in Haskell simply by giving the constraints as a binding context before the instance, e.g.:
Coq < { eqb x y := match x, y with
Coq < | (la, ra), (lb, rb) => andb (eqb la lb) (eqb ra rb)
Coq < end }.
1 subgoal
A : Type
EA : EqDec A
B : Type
EB : EqDec B
============================
forall x y : A * B,
(let (la, ra) := x in let (lb, rb) := y in (eqb la lb && eqb ra rb)%bool) =
true -> x = y
These instances are used just as well as lemmas in the instance hint database.
18.4 Sections and contexts
To ease the parametrization of developments by type classes, we provide a new way to introduce variables into section contexts, compatible with the implicit argument mechanism. The new command works similarly to the Variables vernacular (see 1.3.1), except it accepts any binding context as argument. For example:
Coq < Context ‘{EA : EqDec A}.
A is assumed
EA is assumed
Coq < { eqb x y := match x, y with
Coq < | Some x, Some y => eqb x y
Coq < | None, None => true
Coq < | _, _ => false
Coq < end }.
Coq < About option_eqb.
option_eqb : forall A : Type, EqDec A -> EqDec (option A)
Arguments A, EA are implicit and maximally inserted
Argument scopes are [type_scope _]
option_eqb is transparent
Expands to: Constant Top.option_eqb
Here the Global modifier redeclares the instance at the end of the section, once it has been generalized by the context variables it uses.
18.5 Building hierarchies
18.5.1 Superclasses
One can also parameterize classes by other classes, generating a hierarchy of classes and superclasses. In the same way, we give the superclasses as a binding context:
Coq < { le : A -> A -> bool }.
Contrary to Haskell, we have no special syntax for superclasses, but this declaration is morally equivalent to:
Class `(E : EqDec A) => Ord A := { le : A -> A -> bool }.
This declaration means that any instance of the Ord class must have an instance of EqDec. The parameters of the subclass contain at least all the parameters of its superclasses in their order of appearance (here A is the only one). As we have seen, Ord is encoded as a record type with two parameters: a type A and an E of type EqDec A. However, one can still use it as if it had a single parameter inside generalizing binders: the generalization of superclasses will be done automatically.
In some cases, to be able to specify sharing of structures, one may want to give explicitly the superclasses. It is possible to do it directly in regular generalized binders, and using the ! modifier in class binders. For example:
Coq < andb (le x y) (neqb x y).
The ! modifier switches the way a binder is parsed back to the regular interpretation of Coq. In particular, it uses the implicit arguments mechanism if available, as shown in the example.
18.5.2 Substructures
Substructures are components of a class which are instances of a class themselves. They often arise when using classes for logical properties, e.g.:
Coq < reflexivity : forall x, R x x.
Coq < Class Transitive (A : Type) (R : relation A) :=
Coq < transitivity : forall x y z, R x y -> R y z -> R x z.
This declares singleton classes for reflexive and transitive relations, (see 1 for an explanation). These may be used as part of other classes:
Coq < { PreOrder_Reflexive :> Reflexive A R ;
Coq < PreOrder_Transitive :> Transitive A R }.
The syntax :> indicates that each PreOrder can be seen as a Reflexive relation. So each time a reflexive relation is needed, a preorder can be used instead. This is very similar to the coercion mechanism of Structure declarations. The implementation simply declares each projection as an instance.
One can also declare existing objects or structure projections using the Existing Instance command to achieve the same effect.
18.6 Summary of the commands
18.6.1 Class ident binder1 … bindern : sort:= { field1 ; …; fieldk }.
The Class command is used to declare a type class with parameters binder1 to bindern and fields field1 to fieldk.
Variants:
- Class ident binder1 …bindern : sort:= ident1 : type1. This variant declares a singleton class whose only method is ident1. This singleton class is a so-called definitional class, represented simply as a definition ident binder1 …bindern := type1 and whose instances are themselves objects of this type. Definitional classes are not wrapped inside records, and the trivial projection of an instance of such a class is convertible to the instance itself. This can be useful to make instances of existing objects easily and to reduce proof size by not inserting useless projections. The class constant itself is declared rigid during resolution so that the class abstraction is maintained.
- Existing Class ident. This variant declares a class a posteriori from a constant or inductive definition. No methods or instances are defined.
18.6.2 Instance ident binder1 …bindern : Class t1 …tn [| priority] := { field1 := b1 ; …; fieldi := bi }
The Instance command is used to declare a type class instance named ident of the class Class with parameters t1 to tn and fields b1 to bi, where each field must be a declared field of the class. Missing fields must be filled in interactive proof mode.
An arbitrary context of the form binder1 …bindern can be put after the name of the instance and before the colon to declare a parameterized instance. An optional priority can be declared, 0 being the highest priority as for auto hints.
Variants:
- Instance ident binder1 …bindern : forall bindern+1 …binderm, Class t1 …tn [| priority] := term This syntax is used for declaration of singleton class instances or for directly giving an explicit term of type forall bindern+1 …binderm, Class t1 …tn. One need not even mention the unique field name for singleton classes.
- Global Instance One can use the Global modifier on instances declared in a section so that their generalization is automatically redeclared after the section is closed.
- Program Instance Switches the type-checking to Program (chapter 22) and uses the obligation mechanism to manage missing fields.
- Declare Instance In a Module Type, this command states that a corresponding concrete instance should exist in any implementation of this Module Type. This is similar to the distinction between Parameter vs. Definition, or between Declare Module and Module.
Besides the Class and Instance vernacular commands, there are a few other commands related to type classes.
18.6.3 Existing Instance ident
This commands adds an arbitrary constant whose type ends with an applied type class to the instance database. It can be used for redeclaring instances at the end of sections, or declaring structure projections as instances. This is almost equivalent to Hint Resolve ident : typeclass_instances.
18.6.4 Context binder1 …bindern
Declares variables according to the given binding context, which might use implicit generalization (see 18.4).
18.6.5 Typeclasses Transparent, Opaque ident1 …identn
This commands defines the transparency of ident1 …identn during type class resolution. It is useful when some constants prevent some unifications and make resolution fail. It is also useful to declare constants which should never be unfolded during proof-search, like fixpoints or anything which does not look like an abbreviation. This can additionally speed up proof search as the typeclass map can be indexed by such rigid constants (see 8.13.1). By default, all constants and local variables are considered transparent. One should take care not to make opaque any constant that is used to abbreviate a type, like relation A := A -> A -> Prop.
This is equivalent to Hint Transparent,Opaque ident : typeclass_instances.
18.6.6 Typeclasses eauto := [debug] [dfs | bfs] [depth]
This commands allows to customize the type class resolution tactic, based on a variant of eauto. The flags semantics are:
- debug In debug mode, the trace of successfully applied tactics is printed.
- dfs, bfs This sets the search strategy to depth-first search (the default) or breadth-first search.
- depth This sets the depth of the search (the default is 100).