Special Features of SML/NJ

Extra top-level structures

Standard ML of New Jersey provides the complete SML'97 Basis Library, and in addition provides several other library modules.

The SMLofNJ structure contains substructures that provide:

access to compiler internals,

weak pointers,

lazy suspensions,

first-class continuations,

interval timers, and

information about the underlying system.

Other top-level structures provide:

unsafe operations (such as unchecked array access),

general operating system signal values,

unix-specific signal values,

control of automatic compiler-generated polling,

hooks for cleanup operations at ML exit and restart, and

interface to runtime-system functions.

SML/NJ also provides access to the internal phases of the compiler itself. There are substructures for:

execution profiling,

user-customizable pretty printing,

control of compiler error-message printing, and

control of warning messages.

Other features of the compiler provide access to the SML/NJ "visible compiler", including environments, syntax trees (Ast), abstract syntax (Absyn), parsing, and other basic compilation operations. These "visible compiler" structures are not yet documented, other than by source code for the signatures.

Vector expressions and patterns

Vectors are homogeneous, immutable arrays (see the Vector structure). Vectors are a standard feature of SML'97, but SML/NJ also has special syntax for vector expressions and vector patterns. In SML'97, vectors can be created only by calling functions from the Vector structure, and cannot be pattern-matched.

The vector expression

  #[exp0, exp1, ..., expn-1]

(where n >= 0) creates a vector of length n whose elements are the values of the corresponding subexpressions. As with other aggregate expressions, the element expressions are evaluated from left to right. Vectors may be pattern-matched by vector patterns of the form

  #[pat0, pat1, ..., patn-1]

Such a pattern will only match a vector value of the same length.

Vector expressions and vector patterns are more compact and efficient than lists, and are comparable in cost to records.

Or-patterns

SML/NJ has also extended the syntax of patterns to allow ``or-patterns.'' The basic syntax is:

  (apat1 | ... | apatn)

where the apat_i are atomic patterns. The other restriction is that the variables bound in each apat_i must be the same, and have the same type. A simple example is:

  fun f ("y"|"yes") = true
    | f _ = false;

which has the same meaning as:

  fun f "y" = true
    | f "yes" = true
    | f _ = false;

First-class continuations

A set of primitives has been added to ML to give access to continuations:

  structure SMLofNJ.Cont : sig
    type 'a cont
    val callcc : ('a cont -> 'a) -> 'a
    val throw : 'a cont -> 'a -> 'b
    . . .
  end

The continuation of an expression is an abstraction of what the system will do with the value of the expression.

The use of callcc is described with structure SMLofNJ.Cont.

Quote/Antiquote

An early use of ML was as a MetaLanguage for manipulating terms in an object language. Edinburgh LCF/ML had features to parse one particular object language (called OL). Standard ML of New Jersey has support for arbitrary object languages, with user supplied object-language parsing.

Higher-order Modules

The module system of Standard ML has supported first-order parametric modules in the form of functors. But there are occasions when one like to parameterize over functors as well as structures, which requires a truly higher-order module system (see, for instance, the powerset functor example. SML/NJ now provides a higher-order extension of the module system.

Parameterization over functors can be provided in a straightforward way by allowing functors to be components of structures. Syntactically this can be accomplished merely by allowing functor declarations inside of structure bodies, and by providing syntax for functor specifications in signatures. Functor specifications were already part of the module syntax of the Definition of Standard ML (Figure 8, p. 14), so we have implemented that syntax and added it to the spec class (Figure 7, p. 13). In addition, it is convenient to have a way of declaring functor signatures and some syntactic sugar for curried functor definitions and partial application of curried functors, so these have also been provided. This extension is an ``upward-compatible'' enrichment of the language that should break no existing programs.

Functors as structure components. In the extended language, a signature can contain a functor specification:

  signature SIG =
  sig
    type t
    val a : t
    functor F(X: sig type s
		     val b: s
		 end) : sig val x : t * X.s end
  end

To match such a signature, a structure is allowed to contain a functor declaration:

  structure S : SIG =
  struct
    type t = int
    val a = 3
    functor F(X: sig type s val b: s end) = 
      struct val x = (a,X.b) end
  end

This makes it possible higher-order functors by including a functor as a component of a parameter structure or of a result structure. The case of a functor parameter is illustrated by the following example.

  signature MONOID = 
  sig
    type t
    val plus: t*t -> t
    val e: t
  end;

  (* functor signature declaration *)
  funsig PROD (structure M: MONOID
	       structure N: MONOID) = MONOID

  functor Square(structure X: MONOID
		 functor Prod: PROD): MONOID =
      Prod(structure M = X
	   structure N = X);

Note that this example involves the definition of a functor signature PROD. Currently functor signature declarations take one of the following forms:

  funsig funid (strid: sigexp) = sigexp}
  funsig funid (specs) = sigexp

This syntax is viewed as provisional and subject to change.

A common use of functors returning functors in their result is to approximate a curried functor with multiple parameters. Here is how one might define a curried monoid product functor:

  functor CurriedProd (M: MONOID) =
  struct
    functor Prod1 (N: MONOID) : MONOID =
      struct
	type t = M.t * N.t
	val e = (M.e, N.e)
	fun plus((m1,n1),(m2,n2))=(M.plus(m1,m2),N.plus(n1,n2))
      end;
  end

This works, but the partial application of this functor is rather awkward because it requires the explicit creation of an intermediate structure:

  structure IntMonoid =
  struct
    type t = int
    val e = 0
    val plus = (op +): int*int -> int
  end;

  structure Temp = CurriedProd(IntMonoid);

  functor ProdInt = Temp.Prod1;

To simplify the use of this sort of functor, some derived forms provide syntactic sugar for curried functor definition and partial application. Thus the above example can be written:

  functor CurriedProd (M: MONOID) (N: MONOID) : MONOID =
  struct
    type t = M.t * N.t
    val e = (M.e, N.e)
    fun plus((m1,n1),(m2,n2))=(M.plus(m1,m2),N.plus(n1,n2))
  end;

  functor ProdInt = CurriedProd(IntMonoid);

The syntax for curried forms of functor signature and functor declarations and for the corresponding partial applications can be summarized as follows:

  funsig funsigid (par1) ... (parn) = sigexp

  functor funid (par1) ... (parn) = strexp

  functor funid1 = funid2 (arg1) ... (argn)

  structure strid = funid (arg1) ... (argn)

where

  par ::= id : sigexp | specs
  arg ::= strexp | dec

In the case of a partial application defining a functor, it is assumed that the funid2 on the right hand side takes more than n arguments, while in the case of the structure declaration funid should take exactly n arguments. As a degenerate case where n=0 we have identity functor declarations:

  functor funid1 = funid2

There is also a "let" form of functor expression:

  fctexp ::= let dec in fctexp end

which can only be used in functor definitions of the form:

  functor funid = let dec in fctexp end.

The curried functor declaration

  functor F (par1) ... (parn) = strexp

is a derived form that is translated into the following declaration

  functor F (par1) =
  struct
    functor %fct% (par2) ... (parn) = strexp
  end

and the declarations

  structure S = F (arg1) ... (argn)
  functor G = F (arg1) ... (argn)

are derived forms expanding into (respectively):

  local
    structure %hidden% = F (arg1)
  in
    structure S = %hidden%.%fct% (arg2) ... (argn)
  end

and

  local
    structure %hidden% = F (arg1) ... (argn)
  in
    functor G = %hidden%.%fct%
  end

Currently there is no checking that a complete set of arguments is supplied when a curried functor is applied to define a structure, as illustrated by the following example:

  functor Foo (X: sig type s end) (Y: sig type t end) =
  struct type u = X.s * Y.t end

  structure A = struct type s = int end

  structure S = Foo (A)  (* Foo A yields a (useless) structure *)

  functor G = Foo (A)    (* Foo A yields a functor *)

Of course, the structure S defined in this way is useless, since we cannot use the pseudo-identifier %fct% to select its functor component. Arity checking to prevent this sort of error will be added in a future release.