Type inference

From Wikipedia, the free encyclopedia

Type inference is a feature present in some strongly statically typed programming languages. It is often characteristic of — but not limited to — functional programming languages in general. Some languages that include type inference are: Haskell, Cayenne, Clean, ML, OCaml, Epigram, Scala, Nemerle, D, Chrome and Boo. This feature is planned for Fortress, C# 3.0, C++0x and Perl 6. The ability to infer types automatically makes many programming tasks easier, leaving the programmer free to omit type annotations while maintaining type safety.

1 Nontechnical explanation
2 Technical description
3 Example
4 Hindley–Milner type inference algorithm
- 4.1 The Algorithm
  - 4.1.1 Generating the equations
  - 4.1.2 Solving the equations
5 References
6 External links

[edit] Nontechnical explanation

In most programming languages, all values have a type which describes the kind of data a particular value describes. In some languages, the type is known only at runtime; these languages are dynamically typed. In other languages, the type is known at compile time; these languages are statically typed. In statically typed languages, the input and output types of functions and local variables ordinarily must be explicitly provided by type annotations. In C:

int addone(int x) {
    int result;

    result = x+1;
    return result;
}

Note: this is for example purposes only and not how programs are typically written.

The beginning of this function definition, int addone(int x) declares that addone is a function which takes one argument, an integer, and returns an integer. int result; declares that the local variable result is an integer. In a language where type inference is available, the code might be written like this instead:

addone(x) {
    val result;

    result = x+1;
    return result;
}

This looks very similar to how code is written in a dynamically typed language, yet all types are known at compile time. In the imaginary language in which the last example is written, + always takes two integers and returns one integer. (This is how it works in, for example, OCaml.) From this, the type inferencer can infer that the value of x+1 is an integer, therefore result is an integer, therefore the return value of addone is an integer. Similarly, since + requires that both of its arguments be integers, x must be an integer, and therefore addone accepts one integer as an argument.

[edit] Technical description

Type inference refers to the ability to deduce automatically, either partially or fully, the type of the value derived from the eventual evaluation of an expression. As this process is systematically performed at compile time, the compiler is often able to infer the type of a variable or the type signature of a function, without explicit type annotations having been given. In many cases, it is possible to omit type annotations from a program completely if the type inference system is robust enough, or the program or language simple enough.

To obtain the information required to infer correctly the type of an expression lacking an explicit type annotation, the compiler either gathers this information as an aggregate and subsequent reduction of the type annotations given for its subexpressions (which may themselves be variables or functions), or through an implicit understanding of the type of various atomic values (e.g., () : Unit; true : Bool; 42 : Integer; 3.14159 : Real; etc.). It is through recognition of the eventual reduction of expressions to implicitly typed atomic values that the compiler for a type inferring language is able to compile a program completely without type annotations. In the case of highly complex forms of higher order programming and polymorphism, it is not always possible for the compiler to infer as much, however, and type annotations are occasionally necessary for disambiguation.

[edit] Example

For example, let us consider the Haskell function map, which applies a procedure to each element of a list, and may be defined as:

map f [] = []
map f (first:rest) = f first : map f rest

From this, it is evident that the function map takes a list as its second argument, that its first argument f is a function that can be applied to the type of elements of that list, and that the result of map is constructed as a list with elements that are results of f. So we can reliably construct a type signature

map :: (a -> b) -> [a] -> [b]

where the syntax "a->b" denotes a procedure that takes an a as its parameter and produces a b. "a->b->c" is equivalent to "a->(b->c)".

Note that the inferred type of map is parametrically polymorphic: The type of the arguments and results of f are not inferred, but left as type variables, and so map can be applied to functions and lists of various types, as long as the actual types match in each invocation.

[edit] Hindley–Milner type inference algorithm

The common algorithm used to perform the type inference is the one now commonly referred to as Hindley–Milner or Damas–Milner algorithm.

The origin of this algorithm is the type inference algorithm for the simply typed lambda calculus, which was devised by Haskell B. Curry and Robert Feys in 1958. In 1969 Roger Hindley extended this work and proved that their algorithm always inferred the most general type. In 1978 Robin Milner, independently of Hindley's work, provided an equivalent algorithm. In 1985 Luis Damas finally proved that Milner's algorithm is complete and extended it to support systems with polymorphic references.

[edit] The Algorithm

The algorithm proceeds in two steps. First, we need to generate a number of equations to solve, then we need to solve them.

[edit] Generating the equations

The algorithm used for generating the equations is similar to a regular type checker, so let's consider first a regular type checker for the typed lambda calculus given by

$e \, ::= E \mid v \mid (\lambda v:\tau. e) \mid (e\, e)$

and

$\tau \, ::= T \mid \tau \to \tau$

where $E$ is a primitive expression (such as "3") and $T$ is a primitive type (such as "Integer").

We want to construct a function $f$ of type $\epsilon \to t \to \tau$ , where $ε$ is a type environment and $t$ is a term. We assume that this function is already defined on primitives. The other cases are:

$f\, \Gamma\, v = \tau$ whenever the binding $v\, :\,\tau$ is in $Γ$

$f\, \Gamma\, (g\, e) = \tau$ whenever $\tau_1 = \tau_2 \to \tau$ where $\tau_1 = f\, \Gamma\, g$ and $\tau_2 = f\, \Gamma\, e$ .

$f\, \Gamma\, (\lambda v:\tau. e) = \tau \to \tau_e$ where $\tau_e = f\, \Gamma'\, e$ and $Γ'$ is $Γ$ extended by the binding $v \,:\,\tau$ .

Note that so far we do not specify what to do when we fail to meet the various conditions. This is because, in the simple type *checking* algorithm, the check simply fails whenever anything goes wrong.

Now, we develop a more sophisticated algorithm that can deal with type variables and constraints on them. Therefore, we extend the set T of primitive types to include an infinite supply of variables, denoted by lowercase Greek letters $α,β,...$

This is a limited overview. For now, refer to Types and Programming Languages by Benjamin Pierce, Sections 22.1-4 .

[edit] Solving the equations

Solving the equations proceeds by unification. This is - maybe surprisingly - a rather simple algorithm. The function $u$ operates on type equations and returns a structure called a "substitution". A substitution is simply a mapping from type variables to types. Substitutions can be composed and extended in the obvious ways.

Unifying the empty set of equations is easy enough: $u\, \emptyset = \mathbf{i}$ , where $\mathbf{i}$ is the identity substitution.

Unifying a variable with a type goes this way: $u\, ([\alpha = T] \cup C) = u\, (C') \cdot (\alpha \mapsto T)$ , where $\cdot$ is the substitution composition operator, and $C'$ is the set of remaining constraints $C$ with the new substitution $\alpha \mapsto T$ applied to it.