2 Funs

2.1 Example 1 - map

If we want to double every element in a list, we could write a function named double:

double([H|T]) -> [2*H|double(T)];
double([])    -> [].

This function obviously doubles the argument entered as input as follows:

> double([1,2,3,4]).
[2,4,6,8]

We now add the function add_one, which adds one to every element in a list:

add_one([H|T]) -> [H+1|add_one(T)];
add_one([])    -> [].

These functions, double and add_one, have a very similar structure. We can exploit this fact and write a function map which expresses this similarity:


map(F, [H|T]) -> [F(H)|map(F, T)];
map(F, [])    -> [].

We can now express the functions double and add_one in terms of map as follows:

double(L)  -> map(fun(X) -> 2*X end, L).
add_one(L) -> map(fun(X) -> 1 + X end, L).

map(F, List) is a function which takes a function F and a list L as arguments and returns the new list which is obtained by applying F to each of the elements in L.

The process of abstracting out the common features of a number of different programs is called procedural abstraction. Procedural abstraction can be used in order to write several different functions which have a similar structure, but differ only in some minor detail. This is done as follows:

write one function which represents the common features of these functions
parameterize the difference in terms of functions which are passed as arguments to the common function.

2.2 Example 2 - foreach

This example illustrates procedural abstraction. Initially, we show the following two examples written as conventional functions:

all elements of a list are printed onto a stream
a message is broadcast to a list of processes.

print_list(Stream, [H|T]) ->
    io:format(Stream, "~p~n", [H]),
    print_list(Stream, T);
print_list(Stream, []) ->
    true.

broadcast(Msg, [Pid|Pids]) ->
    Pid ! Msg,
    broadcast(Msg, Pids);
broadcast(_, []) ->
    true.

Both these functions have a very similar structure. They both iterate over a list doing something to each element in the list. The "something" has to be carried round as an extra argument to the function which does this.

The function foreach expresses this similarity:


foreach(F, [H|T]) ->
    F(H),
    foreach(F, T);
foreach(F, []) ->
    ok.

Using foreach, print_list becomes:

foreach(fun(H) -> io:format(S, "~p~n",[H]) end, L)

broadcast becomes:

foreach(fun(Pid) -> Pid ! M end, L)

foreach is evaluated for its side-effect and not its value. foreach(Fun ,L) calls Fun(X) for each element X in L and the processing occurs in the order in which the elements were defined in L. map does not define the order in which its elements are processed.

2.3 The Syntax of Funs

Funs are written with the syntax:

F = fun (Arg1, Arg2, ... ArgN) ->
        ...
    end

This creates an anonymous function of N arguments and binds it to the variable F.

If we have already written a function in the same module and wish to pass this function as an argument, we can use the following syntax:

F = fun FunctionName/Arity

With this form of function reference, the function which is referred to does not need to be exported from the module.

We can also refer to a function defined in a different module with the following syntax:

F = {Module, FunctionName}

In this case, the function must be exported from the module in question.

The follow program illustrates the different ways of creating funs:


-module(fun_test).
-export([t1/0, t2/0]).
-import(lists, [map/2]).

t1() -> map(fun(X) -> 2 * X end, [1,2,3,4,5]).

t2() -> map(fun double/1, [1,2,3,4,5]).

double(X) -> X * 2.

We can evaluate the fun F with the syntax:

F(Arg1, Arg2, ..., Argn)

To check whether a term is a fun, use the test is_function/1 in a guard. Example:

f(F, Args) when is_function(F) ->
   apply(F, Args);
f(N, _) when is_integer(N) ->
   N.

Funs are a distinct type. The BIFs erlang:fun_info/1,2 can be used to retrieve information about a fun, and the BIF erlang:fun_to_list/1 returns a textual representation of a fun. The check_process_code/2 BIF returns true if the process contains funs that depend on the old version of a module.

Note

In OTP R5 and earlier releases, funs were represented using tuples.

2.4 Variable Bindings Within a Fun

The scope rules for variables which occur in funs are as follows:

All variables which occur in the head of a fun are assumed to be "fresh" variables.
Variables which are defined before the fun, and which occur in function calls or guard tests within the fun, have the values they had outside the fun.
No variables may be exported from a fun.

The following examples illustrate these rules:

print_list(File, List) ->
    {ok, Stream} = file:open(File, write),
    foreach(fun(X) -> io:format(Stream,"~p~n",[X]) end, List),
    file:close(Stream).

In the above example, the variable X which is defined in the head of the fun is a new variable. The value of the variable Stream which is used within within the fun gets its value from the file:open line.

Since any variable which occurs in the head of a fun is considered a new variable it would be equally valid to write:

print_list(File, List) ->
    {ok, Stream} = file:open(File, write),
    foreach(fun(File) -> 
                io:format(Stream,"~p~n",[File]) 
            end, List),
    file:close(Stream).

In this example, File is used as the new variable instead of X. This is rather silly since code in the body of the fun cannot refer to the variable File which is defined outside the fun. Compiling this example will yield the diagnostic:

./FileName.erl:Line: Warning: variable 'File' 
      shadowed in 'lambda head'

This reminds us that the variable File which is defined inside the fun collides with the variable File which is defined outside the fun.

The rules for importing variables into a fun has the consequence that certain pattern matching operations have to be moved into guard expressions and cannot be written in the head of the fun. For example, we might write the following code if we intend the first clause of F to be evaluated when the value of its argument is Y:

f(...) ->
    Y = ...
    map(fun(X) when X == Y ->
             ;
           (_) ->
             ...
        end, ...)
    ...

instead of

f(...) ->
    Y = ...
    map(fun(Y) ->
             ;
           (_) ->
             ...
        end, ...)
    ...

2.5 Funs and the Module Lists

The following examples show a dialogue with the Erlang shell. All the higher order functions discussed are exported from the module lists.

map


map(F, [H|T]) -> [F(H)|map(F, T)];
map(F, [])    -> [].

map takes a function of one argument and a list of terms. It returns the list obtained by applying the function to every argument in the list.

> Double = fun(X) -> 2 * X end.
#Fun<erl_eval.6.72228031>
> lists:map(Double, [1,2,3,4,5]).
[2,4,6,8,10]

When a new fun is defined in the shell, the value of the Fun is printed as Fun#<erl_eval>.

any


any(Pred, [H|T]) ->
    case Pred(H) of
        true  ->  true;
        false ->  any(Pred, T)
    end;
any(Pred, []) ->
    false.

any takes a predicate P of one argument and a list of terms. A predicate is a function which returns true or false. any is true if there is a term X in the list such that P(X) is true.

We define a predicate Big(X) which is true if its argument is greater that 10.

> Big =  fun(X) -> if X > 10 -> true; true -> false end end.
#Fun<erl_eval.6.72228031>
> lists:any(Big, [1,2,3,4]).
false
> lists:any(Big, [1,2,3,12,5]).
true

all


all(Pred, [H|T]) ->
    case Pred(H) of
        true  ->  all(Pred, T);
        false ->  false
    end;
all(Pred, []) ->
    true.

all has the same arguments as any. It is true if the predicate applied to all elements in the list is true.

> lists:all(Big, [1,2,3,4,12,6]).   
false
> lists:all(Big, [12,13,14,15]).       
true

foreach


foreach(F, [H|T]) ->
    F(H),
    foreach(F, T);
foreach(F, []) ->
    ok.

foreach takes a function of one argument and a list of terms. The function is applied to each argument in the list. foreach returns ok. It is used for its side-effect only.

> lists:foreach(fun(X) -> io:format("~w~n",[X]) end, [1,2,3,4]). 
1
2
3
4
ok

foldl


foldl(F, Accu, [Hd|Tail]) ->
    foldl(F, F(Hd, Accu), Tail);
foldl(F, Accu, []) -> Accu.

foldl takes a function of two arguments, an accumulator and a list. The function is called with two arguments. The first argument is the successive elements in the list, the second argument is the accumulator. The function must return a new accumulator which is used the next time the function is called.

If we have a list of lists L = ["I","like","Erlang"], then we can sum the lengths of all the strings in L as follows:

> L = ["I","like","Erlang"].
["I","like","Erlang"]
10> lists:foldl(fun(X, Sum) -> length(X) + Sum end, 0, L).                    
11

foldl works like a while loop in an imperative language:

L =  ["I","like","Erlang"],
Sum = 0,
while( L != []){
    Sum += length(head(L)),
    L = tail(L)
end

mapfoldl


mapfoldl(F, Accu0, [Hd|Tail]) ->
    {R,Accu1} = F(Hd, Accu0),
    {Rs,Accu2} = mapfoldl(F, Accu1, Tail),
    {[R|Rs], Accu2};
mapfoldl(F, Accu, []) -> {[], Accu}.

mapfoldl simultaneously maps and folds over a list. The following example shows how to change all letters in L to upper case and count them.

First upcase:

> Upcase =  fun(X) when $a =< X,  X =< $z -> X + $A - $a;
(X) -> X 
end.
#Fun<erl_eval.6.72228031>
> Upcase_word = 
fun(X) -> 
lists:map(Upcase, X) 
end.
#Fun<erl_eval.6.72228031>
> Upcase_word("Erlang").
"ERLANG"
> lists:map(Upcase_word, L).
["I","LIKE","ERLANG"]

Now we can do the fold and the map at the same time:

> lists:mapfoldl(fun(Word, Sum) ->
{Upcase_word(Word), Sum + length(Word)}
end, 0, L).
{["I","LIKE","ERLANG"],11}

filter


filter(F, [H|T]) ->
    case F(H) of
        true  -> [H|filter(F, T)];
        false -> filter(F, T)
    end;
filter(F, []) -> [].

filter takes a predicate of one argument and a list and returns all element in the list which satisfy the predicate.

> lists:filter(Big, [500,12,2,45,6,7]).
[500,12,45]

When we combine maps and filters we can write very succinct code. For example, suppose we want to define a set difference function. We want to define diff(L1, L2) to be the difference between the lists L1 and L2. This is the list of all elements in L1 which are not contained in L2. This code can be written as follows:

diff(L1, L2) -> 
    filter(fun(X) -> not member(X, L2) end, L1).

The AND intersection of the list L1 and L2 is also easily defined:

intersection(L1,L2) -> filter(fun(X) -> member(X,L1) end, L2).

takewhile


takewhile(Pred, [H|T]) ->
    case Pred(H) of
        true  -> [H|takewhile(Pred, T)];
        false -> []
    end;
takewhile(Pred, []) ->
    [].

takewhile(P, L) takes elements X from a list L as long as the predicate P(X) is true.

> lists:takewhile(Big, [200,500,45,5,3,45,6]).  
[200,500,45]

dropwhile


dropwhile(Pred, [H|T]) ->
    case Pred(H) of
        true  -> dropwhile(Pred, T);
        false -> [H|T]
    end;
dropwhile(Pred, []) ->
    [].

dropwhile is the complement of takewhile.

> lists:dropwhile(Big, [200,500,45,5,3,45,6]).
[5,3,45,6]

splitwith


splitwith(Pred, L) ->
    splitwith(Pred, L, []).

splitwith(Pred, [H|T], L) ->
    case Pred(H) of 
        true  -> splitwith(Pred, T, [H|L]);
        false -> {reverse(L), [H|T]}
    end;
splitwith(Pred, [], L) ->
    {reverse(L), []}.

splitwith(P, L) splits the list L into the two sub-lists {L1, L2}, where L = takewhile(P, L) and L2 = dropwhile(P, L).

> lists:splitwith(Big, [200,500,45,5,3,45,6]).
{[200,500,45],[5,3,45,6]}

2.6 Funs Which Return Funs

So far, this section has only described functions which take funs as arguments. It is also possible to write more powerful functions which themselves return funs. The following examples illustrate these type of functions.

Simple Higher Order Functions

Adder(X) is a function which, given X, returns a new function G such that G(K) returns K + X.

> Adder = fun(X) -> fun(Y) -> X + Y end end.
#Fun<erl_eval.6.72228031>
> Add6 = Adder(6).
#Fun<erl_eval.6.72228031>
> Add6(10).
16

Infinite Lists

The idea is to write something like:

-module(lazy).
-export([ints_from/1]).
ints_from(N) ->
    fun() ->
            [N|ints_from(N+1)]
    end.

Then we can proceed as follows:

> XX = lazy:ints_from(1).
#Fun<lazy.0.29874839>
> XX().
[1|#Fun<lazy.0.29874839>]
> hd(XX()).
1
> Y = tl(XX()).
#Fun<lazy.0.29874839>
> hd(Y()).
2

etc. - this is an example of "lazy embedding".

Parsing

The following examples show parsers of the following type:

Parser(Toks) -> {ok, Tree, Toks1} | fail

Toks is the list of tokens to be parsed. A successful parse returns {ok, Tree, Toks1}, where Tree is a parse tree and Toks1 is a tail of Tree which contains symbols encountered after the structure which was correctly parsed. Otherwise fail is returned.

The example which follows illustrates a simple, functional parser which parses the grammar:

(a | b) & (c | d)

The following code defines a function pconst(X) in the module funparse, which returns a fun which parses a list of tokens.


pconst(X) ->
    fun (T) ->
       case T of
           [X|T1] -> {ok, {const, X}, T1};
           _      -> fail
       end
    end.

This function can be used as follows:

> P1 = funparse:pconst(a).
#Fun<funparse.0.22674075>
> P1([a,b,c]).
{ok,{const,a},[b,c]}
> P1([x,y,z]).     
fail

Next, we define the two higher order functions pand and por which combine primitive parsers to produce more complex parsers. Firstly pand:


pand(P1, P2) ->
    fun (T) ->
        case P1(T) of
            {ok, R1, T1} ->
                case P2(T1) of
                    {ok, R2, T2} ->
                        {ok, {'and', R1, R2}};
                    fail ->
                        fail
                end;
            fail ->
                fail
        end
    end.

Given a parser P1 for grammar G1, and a parser P2 for grammar G2, pand(P1, P2) returns a parser for the grammar which consists of sequences of tokens which satisfy G1 followed by sequences of tokens which satisfy G2.

por(P1, P2) returns a parser for the language described by the grammar G1 or G2.


por(P1, P2) ->
    fun (T) ->
        case P1(T) of
            {ok, R, T1} -> 
                {ok, {'or',1,R}, T1};
            fail -> 
                case P2(T) of
                    {ok, R1, T1} ->
                        {ok, {'or',2,R1}, T1};
                    fail ->
                        fail
                end
        end
    end.

The original problem was to parse the grammar (a | b) & (c | d). The following code addresses this problem:


grammar() ->
    pand(
         por(pconst(a), pconst(b)),
         por(pconst(c), pconst(d))).

The following code adds a parser interface to the grammar:


parse(List) ->
    (grammar())(List).

We can test this parser as follows:

> funparse:parse([a,c]).
{ok,{'and',{'or',1,{const,a}},{'or',1,{const,c}}}}
> funparse:parse([a,d]). 
{ok,{'and',{'or',1,{const,a}},{'or',2,{const,d}}}}
> funparse:parse([b,c]).   
{ok,{'and',{'or',2,{const,b}},{'or',1,{const,c}}}}
> funparse:parse([b,d]). 
{ok,{'and',{'or',2,{const,b}},{'or',2,{const,d}}}}
> funparse:parse([a,b]).   
fail