Author:
Tobias Lindahl <tobias(dot)lindahl(at)it(dot)uu(dot)se> , Kostis Sagonas <kostis(at)it(dot)uu(dot)se>
Status:
Draft
Type:
Standards Track
Created:
2-Dec-2007
Erlang-Version:
R12B

EEP 8: Types and function specifications #

Abstract #

This EEP describes an extension to the Erlang language for declaring sets of Erlang terms to form a particular type, effectively forming a specific subtype of the set of all Erlang terms. Subsequently, these types can be used to specify types of record fields and argument and return values of functions.

Rationale #

Type information can be used to document function interfaces, provide more information for bug detection tools such as Dialyzer, and can be exploited by documentation tools such as Edoc for generating program documentation of various forms. It is expected that the type language described in this document will supersede and eventually replace the purely comment-based @type and @spec declarations used by Edoc.

Specification #

Types and their syntax #

Types describe sets of Erlang terms. Types consist and are built from a set of predefined types (e.g. integer(), atom(), pid(), …) described below. Predefined types represent a typically infinite set of Erlang terms which belong to this type. For example, the type atom() stands for the set of all Erlang atoms.

For integers and atoms, we allow for singleton types (e.g. the integers -1 and 42 or the atoms 'foo' and 'bar').

All other types are built using unions of either predefined types or singleton types. In a type union between a type and one of its subtypes the subtype is absorbed by the supertype and the union is subsequently treated as if the subtype was not a constituent of the union. For example, the type union:

atom() | 'bar' | integer() | 42

describes the same set of terms as the type union:

atom() | integer()

Because of subtype relations that exist between types, types form a lattice where the topmost element, any(), denotes the set of all Erlang terms and the bottommost element, none(), denotes the empty set of terms.

The set of predefined types and the syntax for types is given below:

Type :: any()            %% The top type, the set of all Erlang terms.
      | none()           %% The bottom type, contains no terms.
      | pid()
      | port()
      | ref()
      | []               %% nil
      | Atom
      | Binary
      | float()
      | Fun
      | Integer
      | List
      | Tuple
      | Union
      | UserDefined      %% described in Section 2

Union :: Type1 | Type2

Atom :: atom()
      | Erlang_Atom      %% 'foo', 'bar', ...

Binary :: binary()                        %% <<_:_ * 8>>
        | <<>>
        | <<_:Erlang_Integer>>            %% Base size
        | <<_:_*Erlang_Integer>>          %% Unit size
        | <<_:Erlang_Integer, _:_*Erlang_Integer>>

Fun :: fun()                              %% any function
     | fun((...) -> Type)                 %% any arity, returning Type
     | fun(() -> Type)
     | fun((TList) -> Type)

Integer :: integer()
         | Erlang_Integer                 %% ..., -1, 0, 1, ... 42 ...
         | Erlang_Integer..Erlang_Integer %% specifies an integer range

List :: list(Type)                        %% Proper list ([]-terminated)
      | improper_list(Type1, Type2)       %% Type1=contents, Type2=termination
      | maybe_improper_list(Type1, Type2) %% Type1 and Type2 as above

Tuple :: tuple()                          %% stands for a tuple of any size
       | {}
       | {TList}

TList :: Type
       | Type, TList

Because lists are commonly used, they have shorthand type notations. The type list(T) has the shorthand [T]. The shorthand [T,...] stands for the set of non-empty proper lists whose elements are of type T. The only difference between the two shorthands is that [T] may be an empty list but [T,...] may not.

Notice that the shorthand for list(), i.e. the list of elements of unknown type, is [_] (or [any()]), not []. The notation [] specifies the singleton type for the empty list.

For convenience, the following types are also built-in. They can be thought as predefined aliases for the type unions also shown in the table. (Some type unions below slightly abuse the syntax of types.)

==========================  =====================================
      Built-in type                   Stands for
==========================  =====================================
``term()``                  ``any()``
``bool()``                  ``'false' | 'true'``
``byte()``                  ``0..255``
``char()``                  ``0..16#10ffff``
``non_neg_integer()``       ``0..``
``pos_integer()``           ``1..``
``neg_integer()``           ``..-1``
``number()``                ``integer() | float()``
``list()``                  ``[any()]``
``maybe_improper_list()``   ``maybe_improper_list(any(), any())``
``maybe_improper_list(T)``  ``maybe_improper_list(T, any())``
``string()``                ``[char()]``
``nonempty_string()``       ``[char(),...]``
``iolist()``                ``maybe_improper_list(``
                                ``char() | binary() |``
                                ``iolist(), binary() | [])``
``module()``                ``atom()``
``mfa()``                   ``{atom(),atom(),byte()}``
``node()``                  ``atom()``
``timeout()``               ``'infinity' | non_neg_integer()``
``no_return()``             ``none()``
==========================  =====================================

Users are not allowed to define types with the same names as the predefined or built-in ones. This is checked by the compiler and its violation results in a compilation error. (For bootstrapping purposes, it can also result to just a warning if this involves a built-in type which has just been introduced.)

NOTE: The following built-in list types also exist, but they are expected to be rarely used. Hence, they have long names:

nonempty_maybe_improper_list(Type) :: nonempty_maybe_improper_list(Type, any())
nonempty_maybe_improper_list() :: nonempty_maybe_improper_list(any())

where the following two types

nonempty_improper_list(Type1, Type2)
nonempty_maybe_improper_list(Type1, Type2)

define the set of Erlang terms one would expect.

Also for convenience, we allow for record notation to be used. Records are just shorthands for the corresponding tuples:

Record :: #Erlang_Atom{}
        | #Erlang_Atom{Fields}

Records have been extended to possibly contain type information. This is described in Section 3 below.

Type declarations of user-defined types #

As seen, the basic syntax of a type is an atom followed by closed parentheses. New types are declared using 'type' compiler attributes as in the following:

-type my_type() :: Type.

where the type name is an atom ('my_type' in the above) followed by parenthesis. Type is a type as defined in the previous section. A current restriction is that Type can contain only predefined types or user-defined types which have been previously defined. This restriction is enforced by the compiler and results in a compilation error. (A similar restriction currently exists for records).

This means that general recursive types cannot be defined. Lifting this restriction is future work.

Type declarations can also be parametrized by including type variables between the parentheses. The syntax of type variables is the same as Erlang variables (starts with an upper case letter). Naturally, these variables can - and should - appear on the RHS of the definition. A concrete example appears below:

-type orddict(Key, Val) :: [{Key, Val}].

Type information in record declarations #

The types of record fields can be specified in the declaration of the record. The syntax for this is:

-record(rec, {field1 :: Type1, field2, field3 :: Type3}).

For fields without type annotations, their type defaults to any(). I.e., the above is a shorthand for:

-record(rec, {field1 :: Type1, field2 :: any(), field3 :: Type3}).

In the presence of initial values for fields, the type must be declared after the initialisation as in the following:

-record(rec, {field1 = [] :: Type1, field2, field3 = 42 :: Type3}).

Naturally, the initial values for fields should be compatible with (i.e. a member of) the corresponding types. This is checked by the compiler and results in a compilation error if a violation is detected. For fields without initial values, the singleton type 'undefined' is added to all declared types. In other words, the following two record declarations have identical effects:

-record(rec, {f1 = 42 :: integer(),
              f2      :: float(),
              f3      :: 'a' | 'b').

-record(rec, {f1 = 42 :: integer(),
              f2      :: 'undefined' | float(),
              f3      :: 'undefined' | 'a' | 'b').

For this reason, it is recommended that records contain initializers, whenever possible.

Any record, containing type information or not, once defined, can be used as a type using the syntax:

#rec{}

In addition, the record fields can be further specified when using a record type by adding type information about the field in the following manner:

#rec{some_field :: Type}

Any unspecified fields are assumed to have the type in the original record declaration.

Specifications (contracts) for functions #

A contract (or specification) for a function is given using the new compiler attribute 'spec'. The basic format is as follows:

-spec Module:Function(ArgType1, ..., ArgTypeN) -> ReturnType.

The arity of the function has to match the number of arguments, or else a compilation error occurs.

This form can also be used in header files (.hrl) to declare type information for exported functions. Then these header files can be included in files that (implicitly or explicitly) import these functions.

For most uses within a given module, the following shorthand is allowed:

-spec Function(ArgType1, ..., ArgTypeN) -> ReturnType.

Also, for documentation purposes, argument names can be given:

-spec Function(ArgName1 :: Type1, ..., ArgNameN :: TypeN) -> RT.

A function specification can be overloaded. That is, it can have several types, separated by a semicolon (;):

-spec foo(T1, T2) -> T3
       ; (T4, T5) -> T6.

A current restriction, which currently results in a warning (OBS: not an error) by the compiler, is that the domains of the argument types cannot be overlapping. For example, the following specification results in a warning:

-spec foo(pos_integer()) -> pos_integer()
       ; (integer()) -> integer().

Type variables can be used in specifications to specify relations for the input and output arguments of a function. For example, the following specification defines the type of a polymorphic identity function:

-spec id(X) -> X.

However, note that the above specification does not restrict the input and output type in any way. We can constrain these types by guard-like subtype constraints:

-spec id(X) -> X when is_subtype(X, tuple()).

and provide bounded quantification. Currently, the is_subtype/2 guard is the only guard which can be used in a 'spec' attribute.

The scope of an is_subtype/2 constraint is the (...) -> RetType specification after which it appears. To avoid confusion, we suggest that different variables are used in different constituents of an overloaded contract as in the example below:

-spec foo({X, integer()}) -> X when is_subtype(X, atom())
       ; ([Y]) -> Y when is_subtype(Y, number()).

Some functions in Erlang are not meant to return; either because they define servers or because they are used to throw exceptions as the function below:

my_error(Err) -> erlang:throw({error, Err}).

For such functions we recommend the use of the special no_return() type for their “return”, via a contract of the form:

-spec my_error(term()) -> no_return().

Current limitations #

The main limitation is the inability to define recursive types.

Copyright #

This document has been placed in the public domain.