In programming languages with Hindley-Milner type inference and imperative features, in particular the ML programming language family, the value restriction means that declarations are only polymorphically generalized if they are syntactic values (also called non-expansive). The value restriction prevents reference cells from holding values of different types and preserves type safety.
In the Hindley–Milner type system, expressions can be given multiple types through parametric polymorphism. But naively giving multiple types to references breaks type safety. The following are typing rules for references and related operators in ML-like languages.
\begin{align} ref&:\forall\alpha.\alpha\to(\alpha ref)\\ !&:\forall\alpha.(\alpha ref)\to\alpha\\ :=&:\forall\alpha.(\alpha ref)\to\alpha\tounit \end{align}
The operators have the following semantics: takes a value and creates a reference containing that value, (dereference) takes a reference and reads the value in that reference, and (assignment) updates a reference to contain a new value and returns a value of the unit type. Given these, the following program[1] unsoundly applies a function meant for integers to a Boolean value.c
is given the type , which is then instantiated to be of the type when typing the assignment c := (fn x => x + 1)
, and ref when typing the dereference !c true
.
Under the value restriction, the types of let bound expressions are only generalized if the expressions are syntactic values. In his paper, Wright considers the following to be syntactic values: constants, variables, -expressions and constructors applied to values. The function and operator applications are not considered values. In particular, applications of the
ref
The above example is rejected by the type checker under the value restriction as follows.
c
is given the type . This type is not generalized and is a free variable in the typing context for the body of the let binding.c
is modified in the typing context to be of type via unification.!c
is typed asint\toint