「SF-QC」2 TypeClasses

Considerring printing different types with this common idiom:

showBool : bool → string
showNat : nat → string
showList : {A : Type} (A → string) → (list A) → string
showPair : {A B : Type} (A → string) → (B → string) → A * B → string

Definition showListOfPairsOfNats := showList (showPair showNat showNat)   (* LOL *)

The designers of Haskell addressed this clunkiness through typeclasses, a mechanism by which the typechecker is instructed to automatically construct “type-driven” functions [Wadler and Blott 1989].

Coq followed Haskell’s lead as well, but

because Coq’s type system is so much richer than that of Haskell, and because typeclasses in Coq are used to automatically construct not only programs but also proofs, Coq’s presentation of typeclasses is quite a bit less “transparent”

Basics

Classes and Instances

Class Show A : Type := {
  show : A → string
}.

Instance showBool : Show bool := {
  show := fun b:bool ⇒ if b then "true" else "false"
}.

Comparing with Haskell:

class Show a where
  show :: a -> string

--  you cannot override a `instance` so in reality you need a `newtype` wrapper to do this
instance Show Bool where
  show b = if b then "True" else "Fasle"

The show function is sometimes said to be overloaded, since it can be applied to arguments of many types, with potentially radically different behavior depending on the type of its argument.

Next, we can define functions that use the overloaded function show like this:

Definition showOne {A : Type} `{Show A} (a : A) : string :=
  "The value is " ++ show a.
  
Compute (showOne true).
Compute (showOne 42).

Definition showTwo {A B : Type}
           `{Show A} `{Show B} (a : A) (b : B) : string :=
  "First is " ++ show a ++ " and second is " ++ show b.

Compute (showTwo true 42).
Compute (showTwo Red Green).

The parameter `{Show A} is a class constraint, which states that the function showOne is expected to be applied only to types A that belong to the Show class.

Concretely, this constraint should be thought of as an extra parameter to showOne supplying evidence that A is an instance of Show — i.e., it is essentially just a show function for A, which is implicitly invoked by the expression show a.

读时猜测（后来发现接下来有更正确的解释）：show 在 name resolution 到 class Show 时就可以根据其参数的 type（比如 T）infer 出「我们需要一个 Show T 的实现（instance，其实就是个 table）」，在 Haskell/Rust 中这个 table 会在 lower 到 IR 时才 made explicit，而 Coq 这里的语法就已经强调了这里需要 implicitly-and-inferred {} 一个 table，这个 table 的名字其实不重要，只要其 type 是被 A parametrized 的 Show 就好了，类似 ML 的 functor 或者 Java 的 generic interface。

This is Ad-hoc polymorphism.

Missing Constraint

What if we forget the class constrints:

Error:
Unable to satisfy the following constraints:
In environment:
A : Type
a : A

?Show : "Show A"

Class `Eq`

Class Eq A :=
  {
    eqb: A → A → bool;
  }.

Notation "x =? y" := (eqb x y) (at level 70).

Instance eqBool : Eq bool :=
  {
    eqb := fun (b c : bool) ⇒ 
       match b, c with
         | true, true ⇒ true
         | true, false ⇒ false
         | false, true ⇒ false
         | false, false ⇒ true
       end
  }.

Instance eqNat : Eq nat :=
  {
    eqb := Nat.eqb
  }.

Why should we need to define a typeclass for boolean equality when Coq’s propositional equality (x = y) is completely generic? while it makes sense to claim that two values x and y are equal no matter what their type is, it is not possible to write a decidable equality checker for arbitrary types. In particular, equality at types like nat → nat is undecidable.

x = y 返回一个需要去证的 Prop (relational) 而非 executable Fixpoint (functional)
因为 function 的 equality 有时候会 undeciable，所以才需要加 Functional Extensionality Axiom（见 LF-06）

Instance eqBoolArrowBool: Eq (bool -> bool) :=
  {
    eqb := fun (f1 f2 : bool -> bool) =>
      (f1 true) =? (f2 true) && (f1 false) =? (f2 false) 
  }.

Compute (id =? id).      (* ==> true *)
Compute (negb =? negb).  (* ==> true *)
Compute (id =? negb).    (* ==> false *)

这里这个 eqb 的定义也是基于 extensionality 的定义，如果考虑到 effects（divergence、IO）是很容易 break 的（类似 parametricity）

Parameterized Instances: New Typeclasses from Old

Structural recursion

Instance showPair {A B : Type} `{Show A} `{Show B} : Show (A * B) :=
  {
    show p :=
      let (a,b) := p in 
        "(" ++ show a ++ "," ++ show b ++ ")"
  }.
Compute (show (true,42)).

Structural equality

Instance eqPair {A B : Type} `{Eq A} `{Eq B} : Eq (A * B) :=
  {
    eqb p1 p2 :=
      let (p1a,p1b) := p1 in
      let (p2a,p2b) := p2 in
      andb (p1a =? p2a) (p1b =? p2b)
  }.

Slightly more complicated example: typical list:

(* the book didn't use any from ListNotation *)
Fixpoint showListAux {A : Type} (s : A → string) (l : list A) : string :=
  match l with
    | nil ⇒ ""
    | cons h nil ⇒ s h
    | cons h t ⇒ append (append (s h) ", ") (showListAux s t)
  end.
Instance showList {A : Type} `{Show A} : Show (list A) :=
  {
    show l := append "[" (append (showListAux show l) "]")
  }.
  
(* I used them though *)
Fixpoint eqListAux {A : Type} `{Eq A} (l1 l2 : list A) : bool :=
  match l1, l2 with
    | nil, nil => true
    | (h1::t1), (h2::t2) => (h1 =? h2) && (eqListAux t1 t2)
    | _, _ => false
  end.

Instance eqList {A : Type} `{Eq A} : Eq (list A) :=
  {
    eqb l1 l2 := eqListAux l1 l2
  }.

Class Hierarchies

we might want a typeclass Ord for “ordered types” that support both equality and a less-or-equal comparison operator.

A bad way would be declare a new class with two func eq and le.

It’s better to establish dependencies between typeclasses, similar with OOP class inheritence and subtyping (but better!), this gave good code reuses.

We often want to organize typeclasses into hierarchies.

Class Ord A `{Eq A} : Type :=
  {
    le : A → A → bool
  }.
Check Ord. (* ==>
Ord
     : forall A : Type, Eq A -> Type
*)

class Eq is a “super(type)class” of Ord (not to be confused with OOP superclass)

This is Sub-typeclassing.

Fixpoint listOrdAux {A : Type} `{Ord A} (l1 l2 : list A) : bool := 
  match l1, l2 with
  | [], _ => true
  | _, [] => false
  | h1::t1, h2::t2 => if (h1 =? h2)
                     then (listOrdAux t1 t2)
                     else (le h1 h2)
  end.

Instance listOrd {A : Type} `{Ord A} : Ord (list A) :=
  {
    le l1 l2 := listOrdAux l1 l2
  }.

(* truthy *)
Compute (le [1] [2]).
Compute (le [1;2] [2;2]).
Compute (le [1;2;3] [2]).

(* falsy *)
Compute (le [1;2;3] [1]).
Compute (le [2] [1;2;3]).

How It works

Implicit Generalization

所以 `{...} 这个 “backtick” notation is called implicit generalization，比 implicit {} 多做了一件自动 generalize 泛化 free varabile 的事情。

that was added to Coq to support typeclasses but that can also be used to good effect elsewhere.

Definition showOne1 `{Show A} (a : A) : string :=
  "The value is " ++ show a.

Print showOne1.
(* ==>
    showOne1 = 
      fun (A : Type) (H : Show A) (a : A) => "The value is " ++ show a
           : forall A : Type, Show A -> A -> string

    Arguments A, H are implicit and maximally inserted
*)

notice that the occurrence of A inside the `{...} is unbound and automatically insert the binding that we wrote explicitly before.

The “implicit and maximally generalized” annotation on the last line means that the automatically inserted bindings are treated (注：printed) as if they had been written with {...}, rather than (...).

The “implicit” part means that the type argument A and the Show witness H are usually expected to be left implicit
whenever we write showOne1, Coq will automatically insert two unification variables as the first two arguments.

This automatic insertion can be disabled by writing @, so a bare occurrence of showOne1 means the same as @showOne1 _ _

这里的 witness H 即 A implements Show 的 evidence，本质就是个 table or record，可以 written more explicitly:

Definition showOne2 `{_ : Show A} (a : A) : string :=
  "The value is " ++ show a.

Definition showOne3 `{H : Show A} (a : A) : string :=
  "The value is " ++ show a.

甚至

Definition showOne4 `{Show} a : string :=
  "The value is " ++ show a.

showOne = 
fun (A : Type) (H : Show A) (a : A) => "The value is " ++ show a
     : forall A : Type, Show A -> A -> string

Set Printing Implicit.

showOne = 
fun (A : Type) (H : Show A) (a : A) => "The value is " ++ @show A H a     (* <-- 注意这里 *)
     : forall A : Type, Show A -> A -> string

vs. Haskell

顺便，Haskell 的话，Show 是可以直接 inferred from the use of show 得

Prelude> showOne a = show a
Prelude> :t showOne
showOne :: Show a => a -> String

但是 Coq 不行，会退化上「上一个定义的 instance Show」，还挺奇怪的（

Definition showOne5 a : string :=  (* not generalized *)
  "The value is " ++ show a.

Free Superclass Instance

``{Ord A} led Coq to fill in both A and H : Eq A because it's the superclass of Ord` (appears as the second argument).

Definition max1 `{Ord A} (x y : A) :=
  if le x y then y else x.

Set Printing Implicit.
Print max1.
(* ==>
     max1 = 
       fun (A : Type) (H : Eq A) (H0 : @Ord A H) (x y : A) =>
         if @le A H H0 x y then y else x

   : forall (A : Type) (H : Eq A), 
       @Ord A H -> A -> A -> A    
*)
Check Ord.
(* ==> Ord : forall A : Type, Eq A -> Type *)

Ord type 写详细的话可以是：

Ord : forall (A : Type), (H: Eq A) -> Type

Other usages of `{}

Implicit generalized Prop mentioning free vars.

Generalizable Variables x y.

Lemma commutativity_property : `{x + y = y + x}.
Proof. intros. omega. Qed.

Check commutativity_property.

Implicit generalized fun/λ, however…

Definition implicit_fun := `{x + y}.
Compute (implicit_fun 2 3)  (* ==> Error *)
Compute (@implicit_fun 2 3)

Implicitly-generalized but inserted as explicit via `(...)

Definition implicit_fun := `(x + y).
Compute (implicit_fun 2 3)

这里可以看到 Coq 的所有语法都是正交的（非常牛逼……）

()/{} 控制是否是 implicit argument
`-prefix 控制是否做 implicit generalization
- N.B. 可能你忘记了但是 → is degenerated ∀ (Π)，所以 generalization 自然会生成 fun

Records are Products

Record types must be declared before they are used. For example:

Record Point :=
  Build_Point
    {
      px : nat;
      py : nat
    }.

(* built with constructor *)
Check (Build_Point 2 4).

(* built with record syntax *)
Check {| px := 2; py := 4 |}.
Check {| py := 2; px := 4 |}.

(* field access, with a clunky "dot notation" *)
Definition r : Point := {| px := 2; py := 4 |}.
Compute (r.(px) + r.(py)).

和 OCaml 一样是 nominal typing 而非 structural typing。类似于 OCaml 中的 record 其实到 backend 了就会和 tuple 等价：都会 lower 到 Heap Block）， Coq 中的 Record 其实和 Pair/Product 也是等价：都是 arity 为 2 的 Inductive type：

Inductive Point : Set := 
  | Build_Point : nat → nat → Point.

我仿造 Print px. 输出的定义模拟了一下：

Inductive Point2 : Set := 
  | Build_Point2 (px2:nat) (py2:nat).
Definition px2 := fun p : Point2 => let (px, _) := p in px.
Definition py2 := fun p : Point2 => let (_, py) := p in py.

Definition r2 : Point2 := Build_Point2 2 4.
Compute (r2.(px2) + r2.(py2)).                        (* => 6 *)

Definition r2 : Point2 := {| px2 := 2; py2 := 4 |}.   (* Error: px2 is not a projection *)

可以发现 dot notation 是可以工作的，. 应该只是一个 pipe 但是 {|...|} 不知道为什么这里会认为 px2 不是一个 record projection.

Note that the field names have to be different. Any given field name can belong to only one record type. This greatly simplifies type inference!

Typeclasses are Records

Typeclasses and instances, in turn, are basically just syntactic sugar for record types and values (together with a bit of magic for using proof search to fill in appropriate instances during typechecking…

Internally, a typeclass declaration is elaborated into a parameterized Record declaration:

Class Show A : Type := { show : A → string }.

Print Show.
Record Show (A : Type) : Type := 
    Build_Show { show : A -> string }

Set Printing All.
Print Show.
Variant Show (A : Type) : Type :=
    Build_Show : forall _ : forall _ : A, string, Show A

(* to make it more clear... *)
Inductive Show (A : Type) : Type :=
  | Build_Show : ∀(show : ∀(a : A), string), Show A
  
(* or more GADT looking, i.e., implicit generalized *)
Inductive Show (A : Type) : Type :=
  | Build_Show : (A -> string) -> Show A

Coq actually call a single-field record Variant. Well actually, I found it’s for any single-constructor Inductively constructed type. You can even use Variant nonchangbly with Inductive as a keyword…

Set Printing All.
Print Point.
Variant Point : Set :=
    Build_Point : forall (_ : nat) (_ : nat), Point

Analogously, Instance declarations become record values:

Print showNat.
showNat = {| show := string_of_nat |}
    : Show nat

Similarly, overloaded functions like show are really just record projections, which in turn are just functions that select a particular argument of a one-constructor Inductive type.

Print show.
show = 
  fun (A : Type) (Show0 : Show A) => 
    let (show) := Show0 in show
      : forall A : Type, Show A -> A -> string

Set Printing All.
Print show.
show = 
  fun (A : Type) (Show0 : Show A) =>
    match Show0 return (forall _ : A, string) with
    | Build_Show _ show => show
    end
      : forall (A : Type) (_ : Show A) (_ : A), string

Inferring Instances

appropriate instances are automatically inferred (and/or constructed!) during typechecking.

Definition eg42 := show 42.

Set Printing Implicit.
Print eg42.
eg42 = @show nat showNat 42 : string

different with Compute, Print 居然还可以这么把所有 implicit argument (after inferred) 都给 print 出来……

type inferrence:

show is expanded to @show _ _ 42
obviously it’s @show nat __42
obviously it’s @show nat (?H : Show Nat) 42

Okay now where to find this witness/evidence/instance/record/table/you-name-it ?H

It attempts to find or construct such a value using a variant of the eauto proof search procedure that refers to a “hint database” called typeclass_instances.

Print HintDb typeclass_instances.  (* too much to be useful *)

“hint database” to me is better understood as a reverse of environment or typing context Γ. Though specialized with only Instance there. （这么一看实现一个 Scala 的 Implicit 也不难啊）

Coq can even print what’s happening during this proof search!

Set Typeclasses Debug.
Check (show 42).
(* ==>
     Debug: 1: looking for (Show nat) without backtracking
     Debug: 1.1: exact showNat on (Show nat), 0 subgoal(s)
*)

Check (show (true,42)).
(* ==>
     Debug: 1: looking for (Show (bool * nat)) without backtracking
     Debug: 1.1: simple apply @showPair on (Show (bool * nat)), 2 subgoal(s)
     Debug: 1.1.3 : (Show bool)
     Debug: 1.1.3: looking for (Show bool) without backtracking
     Debug: 1.1.3.1: exact showBool on (Show bool), 0 subgoal(s)
     Debug: 1.1.3 : (Show nat)
     Debug: 1.1.3: looking for (Show nat) without backtracking
     Debug: 1.1.3.1: exact showNat on (Show nat), 0 subgoal(s)      *)
Unset Typeclasses Debug.

In summary, here are the steps again:

show 42
    ===>   { Implicit arguments }
@show _ _ 42
    ===>   { Typing }
@show (?A : Type) (?Show0 : Show ?A) 42
    ===>   { Unification }
@show nat (?Show0 : Show nat) 42
    ===>   { Proof search for Show Nat returns showNat }
@show nat showNat 42

Typeclasses and Proofs

Propositional Typeclass Members

Class EqDec (A : Type) {H : Eq A} := 
  { 
    eqb_eq : ∀ x y, x =? y = true ↔ x = y 
  }.

Instance eqdecNat : EqDec nat := 
  {
    eqb_eq := Nat.eqb_eq
  }.

这里可以用于抽象 LF-07 的 reflection

Substructures

Naturally, it is also possible to have typeclass instances as members of other typeclasses: these are called substructures.

这里的 relation 来自 Prelude 不过和 LF-11 用法一样：

Require Import Coq.Relations.Relation_Definitions.
Class Reflexive (A : Type) (R : relation A) :=
  { 
    reflexivity : ∀ x, R x x
  }.
Class Transitive (A : Type) (R : relation A) :=
  {
    transitivity : ∀ x y z, R x y → R y z → R x z
  }.

Class PreOrder (A : Type) (R : relation A) :=
  { PreOrder_Reflexive :> Reflexive A R ;
    PreOrder_Transitive :> Transitive A R }.

The syntax :> indicates that each PreOrder can be seen as a Reflexive and Transitive relation, so that, any time a reflexive relation is needed, a preorder can be used instead.

这里的 :> 方向和 subtyping 的 subsumption 是反着的……跟 SML 的 ascription :> 一样……

subtyping T :> S : value of S can safely be used as value of T
ascription P :> R : value of P can safely be used as value of R

Why?

Some Useful Typeclasses

`Dec`

The ssreflect library defines what it means for a proposition P to be decidable like this…

Require Import ssreflect ssrbool.
Print decidable.
(* ==>
     decidable = fun P : Prop => {P} + {~ P}
*)

.. where {P} + {¬ P} is an “informative disjunction” of P and ¬P.

即两个 evidence（参考 LF-07)

Class Dec (P : Prop) : Type :=
  {
    dec : decidable P
  }.

Monad

In Haskell, one place typeclasses are used very heavily is with the Monad typeclass, especially in conjunction with Haskell’s “do notation” for monadic actions.

Monads are an extremely powerful tool for organizing and streamlining code in a wide range of situations where computations can be thought of as yielding a result along with some kind of “effect.”

说话很严谨「in a wide range of situations where … “effect”」

most older projects simply define their own monads and monadic notations — sometimes typeclass-based, often not — while newer projects use one of several generic libraries for monads. Our current favorite (as of Summer 2017) is the monad typeclasses in Gregory Malecha’s ext-lib package:

https://github.com/coq-ext-lib/coq-ext-lib/blob/v8.5/theories/Structures/Monad.v

Require Export ExtLib.Structures.Monads.
Export MonadNotation.
Open Scope monad_scope.

Class Monad (M : Type → Type) : Type := { 
  ret : ∀ {T : Type}, T → M T ;
  bind : ∀ {T U : Type}, M T → (T → M U) → M U
}.

Instance optionMonad : Monad option := {
  ret T x := Some x ;
  bind T U m f :=
    match m with
      None ⇒ None
    | Some x ⇒ f x
    end
}.

Compare with Haskell:

class Applicative m => Monad (m :: * -> *) where
  return :: a -> m a
  (>>=) :: m a -> (a -> m b) -> m b
  
instance Monad Maybe where
  return = Just
  (>>=)  = (>>=)
  where
    (>>=) :: Maybe a -> (a -> Maybe b) -> Maybe b
    Nothing  >>= _ = Nothing
    (Just x) >>= f = f x

After mimic do notation: (as PLF-11)

Definition sum3 (l : list nat) : option nat :=
  x0 <- nth_opt 0 l ;;
  x1 <- nth_opt 1 l ;;
  x2 <- nth_opt 2 l ;;
  ret (x0 + x1 + x2).

Controlling Instantiation

“Defaulting”

Would better explicitly typed. searching can be stupid

Manipulating the Hint Database

One of the ways in which Coq’s typeclasses differ most from Haskell’s is the lack, in Coq, of an automatic check for “overlapping instances.”

在 Haskell 中一大 use case 是可以做类似 C++ 的 partial specification（偏特化）

Check out this on the pros and cons of overlapping instances in Haskell
Check out [this] (https://www.ibm.com/developerworks/community/blogs/12bb75c9-dfec-42f5-8b55-b669cc56ad76/entry/c__e6_a8_a1_e6_9d_bf__e7_a9_b6_e7_ab_9f_e4_bb_80_e4_b9_88_e6_98_af_e7_89_b9_e5_8c_96?lang=en) on template partial specification in C++

That is, it is completely legal to define a given type to be an instance of a given class in two different ways. When this happens, it is unpredictable which instance will be found first by the instance search process;

Workarounds in Coq when this happen:

removing instances from hint database
priorities

Debugging

TBD.

Instantiation Failures
Nontermination

Alternative Structuring Mechanisms

large-scale structuring mechanisms

Typeclasses are just one of several mechanisms that can be used in Coq for structuring large developments. Others include:

canonical structures

bare dependent records

modules and functors

Module and functors is very familiar!

Quickcheck - A Tutorial on Typeclasses in Coq