Accumulator factory
You are encouraged to solve this task according to the task description, using any language you may know.
A problem posed by Paul Graham is that of creating a function that takes a single (numeric) argument and which returns another function that is an accumulator. The returned accumulator function in turn also takes a single numeric argument, and returns the sum of all the numeric values passed in so far to that accumulator (including the initial value passed when the accumulator was created).
- Rules
The detailed rules are at http://paulgraham.com/accgensub.html and are reproduced here for simplicity (with additions in small italic text).
- Before you submit an example, make sure the function
- Takes a number n and returns a function (lets call it g), that takes a number i, and returns n incremented by the accumulation of i from every call of function g(i).
Although these exact function and parameter names need not be used - Works for any numeric type-- i.e. can take both ints and floats and returns functions that can take both ints and floats. (It is not enough simply to convert all input to floats. An accumulator that has only seen integers must return integers.) (i.e., if the language doesn't allow for numeric polymorphism, you have to use overloading or something like that)
- Generates functions that return the sum of every number ever passed to them, not just the most recent. (This requires a piece of state to hold the accumulated value, which in turn means that pure functional languages can't be used for this task.)
- Returns a real function, meaning something that you can use wherever you could use a function you had defined in the ordinary way in the text of your program. (Follow your language's conventions here.)
- Doesn't store the accumulated value or the returned functions in a way that could cause them to be inadvertently modified by other code. (No global variables or other such things.)
- Takes a number n and returns a function (lets call it g), that takes a number i, and returns n incremented by the accumulation of i from every call of function g(i).
- E.g. if after the example, you added the following code (in a made-up language) where the factory function is called foo:
x = foo(1); x(5); foo(3); print x(2.3);
- It should print 8.3. (There is no need to print the form of the accumulator function returned by foo(3); it's not part of the task at all.)
- Task
Create a function that implements the described rules.
It need not handle any special error cases not described above. The simplest way to implement the task as described is typically to use a closure, providing the language supports them.
Where it is not possible to hold exactly to the constraints above, describe the deviations.
11l
F accumulator(n)
T Accumulator
Float s
F (Float n)
.s = n
F ()(Float n)
.s += n
R .s
R Accumulator(n)
V x = accumulator(1)
print(x(5))
print(x(2.3))
V x2 = accumulator(3)
print(x2(5))
print(x2(3.3))
print(x2(0))
- Output:
6 8.3 8 11.3 11.3
8th
\ RossetaCode 'accumulator factory'
\ The 'accumulate' word stores the accumulated value in an array, because arrays
\ are mutable:
: accumulate \ n [m] -- n+m \ [m] -> [n+m]
a:pop rot n:+
tuck a:push swap ;
\ To comply with the rules, this takes a number and wraps it in an array, and
\ then curries it. Since 'curry:' is "immediate", we need to postpone its
\ action using 'p:.
: make-accumulator
1 a:close
' accumulate
p: curry: ;
\ We 'curry' the initial value along with 'accumulate' to create
\ a new word, '+10', which will give us the accumulated values
10 make-accumulator +10
\ This loop will add 1, then 2, then 3, to the accumulator, which prints the
\ results 11,13,16:
( +10 . cr ) 1 3 loop
bye
- Output:
11 13 16
ABAP
ABAP, unfortunately, has no first order functions, nor does its OO paradigm implement method overloading. One potential solution to this problem is to use classes to maintain the state, with the import/export parameters being defined as type 'any', so that the resultant type is calculated dynamically.
Another possible solution would be to use the languages in-built JavaScript processing capabilities to dynamically construct a JS source at run-time, which implements the JS Accumulator factory.
Object Oriented Solution
report z_accumulator
class acc definition.
public section.
methods:
call importing iv_i type any default 0 exporting ev_r type any,
constructor importing iv_d type f.
private section.
data a_sum type f.
endclass.
class acc implementation.
method call.
add iv_i to a_sum.
ev_r = a_sum.
endmethod.
start-of-selection.
data: cl_acc type ref to acc,
lv_ret2 type f,
lv_ret1 type i.
create object cl_acc exporting iv_d = 1.
cl_acc->call( exporting iv_i = 5 ).
cl_acc->call( exporting iv_i = '2.3' importing ev_r = lv_ret2 ).
cl_acc->call( exporting iv_i = 2 importing ev_r = lv_ret1 ).
write : / lv_ret2 decimals 2 exponent 0 left-justified, / lv_ret1 left-justified.
- Output:
8.30 10
JavaScript Solution
data: lv_source type string,
cl_processor type ref to cl_java_script,
lv_ret type string.
cl_processor = cl_java_script=>create( ).
concatenate
'function acc(sum) { '
' return function(n) { '
' return sum += n;'
' }; '
' }; '
' var x = acc(1); '
' x(5);'
' var ret = acc(3).toString();'
' ret = ret + x(2.3);'
into lv_source.
lv_ret = cl_processor->evaluate( lv_source ).
if cl_processor->last_condition_code <> cl_java_script=>cc_ok.
write cl_processor->last_error_message.
else.
write lv_ret.
write / 'Done'.
endif.
#function (n) {# return sum += n;#}#8.3
ActionScript
Closures work the same in ActionScript as in JavaScript. ActionScript will transparently convert integers to reals if the function is given a real argument, but the typeof operator must be used to ensure the function isn't sent invalid arguments, such as strings (which would silently convert the accumulated number to a string without throwing an error).
//Throw an error if a non-number argument is used. (typeof evaluates to
// "number" for both integers and reals)
function checkType(obj:Object):void {
if(typeof obj != "number")
throw new ArgumentError("Expected integer or float argument. Recieved " + typeof obj);
}
function accumulator(sum:Object):Function {
checkType(sum);
return function(n:Object):Object {checkType(n); return sum += n};
}
var acc:Function=accumulator(2);
trace(acc(10));
trace(acc(4));
trace(acc("123")); //This causes an ArgumentError to be thrown.
Ada
with Accumulator;
with Ada.Text_IO; use Ada.Text_IO;
procedure Example is
package A is new Accumulator;
package B is new Accumulator;
begin
Put_Line (Integer'Image (A.The_Function (5)));
Put_Line (Integer'Image (B.The_Function (3)));
Put_Line (Float'Image (A.The_Function (2.3)));
end;
generic package Accumulator is
-- This Ada generic package represents an accumulator factory.
-- The required function is provided as The_Function.
-- The first call to The_Function sets the initial value.
-- (Marius Amado-Alves)
function The_Function (X : Integer) return Integer;
function The_Function (X : Integer) return Float;
function The_Function (X : Float) return Float;
end;
package body Accumulator is
-- The accumulator lives through three states. It is in Virgin_State
-- before any use of The_Function. It changes to Integer_State or
-- Float_State, according to the input type used. The accumulation is
-- memorized in variable I or F, according to the state. Float_State,
-- once reached, is never left. A Float output on an Integer_State is
-- simply a conversion, sans effect on state. (Marius Amado-Alves)
type State_T is (Virgin_State, Integer_State, Float_State);
State : State_T := Virgin_State;
I : Integer;
F : Float;
function The_Function (X : Float) return Float is
begin
case State is
when Virgin_State =>
State := Float_State;
F := X;
return F;
when Integer_State =>
State := Float_State;
F := Float (I) + X;
return F;
when Float_State =>
F := F + X;
return F;
end case;
end;
function The_Function (X : Integer) return Float is
begin
case State is
when Virgin_State =>
State := Integer_State;
I := X;
return Float (I);
when Integer_State =>
I := I + X;
return Float (I);
when Float_State =>
F := F + Float (X);
return F;
end case;
end;
function The_Function (X : Integer) return Integer is
begin
case State is
when Virgin_State =>
State := Integer_State;
I := X;
return I;
when Integer_State =>
I := I + X;
return I;
when Float_State =>
F := F + Float (X);
return Integer (F);
end case;
end;
end;
Aikido
function accumulator (sum:real) {
return function(n:real) { return sum += n }
}
var x = accumulator(1)
x(5)
println (accumulator)
println (x(2.3))
- Output:
accumulator 8.3
Aime
af(list l, object o)
{
l[0] = l[0] + o;
}
main(void)
{
object (*f)(object);
f = af.apply(list(1));
f(5);
af.apply(list(3));
o_(f(2.3), "\n");
0;
}
- Output:
8.3
The type is properly preserved over summing:
f = af.apply(list(5));
f(-6);
f(7);
o_form("~: ~\n", f(0).__type, f(0));
f = af.apply(list(8));
f(-6.6);
f(4.2);
o_form("~: /d1/\n", f(0).__type, f(0));
- Output:
integer: 6 real: 5.6
ALGOL 68
Note: Standard ALGOL 68's scoping rules forbids exporting a procedure (or format) out of it's scope (closure). Hence this specimen will run on ELLA ALGOL 68, but is non-standard. For a discussion of first-class functions in ALGOL 68 consult "The Making of Algol 68" - C.H.A. Koster (1993).
MODE NUMBER = UNION(INT,REAL,COMPL);
PROC plus = (NUMBER in a, in b)NUMBER: (
CASE in a IN
(INT a): CASE in b IN (INT b): a+b, (REAL b): a+b, (COMPL b): a+b ESAC,
(REAL a): CASE in b IN (INT b): a+b, (REAL b): a+b, (COMPL b): a+b ESAC,
(COMPL a): CASE in b IN (INT b): a+b, (REAL b): a+b, (COMPL b): a+b ESAC
ESAC
);
main: (
# now override the + and +:= OPerators #
OP + = (NUMBER a, b)NUMBER: plus(a,b);
OP +:= = (REF NUMBER lhs, NUMBER rhs)NUMBER:
lhs := lhs + rhs;
PROC accumulator = (REF NUMBER sum)PROC(NUMBER)NUMBER:
(NUMBER n)NUMBER:
sum +:= n;
PROC (NUMBER)NUMBER x = accumulator(LOC NUMBER := 1);
x(5);
print(("x:",x(2.3), new line));
PROC (NUMBER)NUMBER y = accumulator(LOC NUMBER := 100);
y(500);
print(("y:",y(230), new line));
print(("x:",x(0), new line))
)
- Output:
x: +.830000000000000e +1 y: +830 x: +.830000000000000e +1
AppleScript
This has one deviation. AppleScript needs a script object for the closure on the sum n
. So this factory returns a script object, not a handler by itself. One must call the handler through its script object, as in x's call(1)
.
on accumulator(n)
-- Returns a new script object
-- containing a handler.
script
on call(i)
set n to n + i -- Returns n.
end call
end script
end accumulator
set x to accumulator(10)
log x's call(1)
set y to accumulator(5)
log y's call(2)
log x's call(3.5)
-- Event Log: (*11*) (*7*) (*14.5*)
Or, to match the task spec and output a little more closely:
on run
set x to foo(1)
x's |λ|(5)
foo(3)
x's |λ|(2.3)
end run
-- foo :: Int -> Script
on foo(sum)
script
on |λ|(n)
set sum to sum + n
end |λ|
end script
end foo
- Output:
8.3
Argile
use std, array
let A = accumulator 42
print(A 0)
print(A 1)
print(A 10)
print(A 100)
let B = accumulator 4.2
print(B 0)
print(B 1)
print(B 10.0)
print(B 100.4)
~A ; ~B
(: use dbg; check mem leak :)
(: accumulator call :)
=: <accumulator a> <num x> := -> (a.t)
call ((a.func) as function(any)(a.t)->(a.t)) with (a.data) ((Cgen x) as a.t)
(: accumulator constructors :)
.: accumulator <int x> :. -> int accumulator
(val (int accumulator) A).init(x)
(A as Accumulator).func = ( .:<int& accu, int x>:. ->int {accu += x; accu} )
A
.: accumulator <real x> :. -> real accumulator
(val (real accumulator) A).init(x)
(A as Accumulator).func = ( .:<real&accu,real x>:. ->real{accu += x; accu} )
A
=: <accumulator& a>.init <num x> :=
a = new (Accumulator)
a.data = (new array of 1 a.t)
*(a.data as (a.t*)) = Cgen x
(: accumulator destructor :)
.: del Accumulator <Accumulator a>:.
free a.data
free a
=: ~ <accumulator a> := {del Accumulator a}
(: accumulator type :)
class Accumulator
function func
any data
=: [<type t=(int)>] accumulator := -> type
Accumulator.prefix
Accumulator.suffix
autocast accumulator<->Accumulator
Astro
fun accumulator(var sum): :: Real -> _
n => sum += n
let f = accumulator!(5)
print f(5) # 10
print f(10) # 20
print f(2.4) # 22.4
BBC BASIC
This code works by copying the function FNdummy() onto the heap and returning a pointer to it.
x = FNaccumulator(1)
dummy = FN(x)(5)
dummy = FNaccumulator(3)
PRINT FN(x)(2.3)
END
DEF FNaccumulator(sum)
LOCAL I%, P%, Q%
DIM P% 53 : Q% = !^FNdummy()
FOR I% = 0 TO 49 : P%?I% = Q%?I% : NEXT
P%!I% = P% : sum = FN(P%+I%)(sum)
= P%+I%
DEF FNdummy(n)
PRIVATE sum
sum += n
= sum
Bracmat
Until 2023 Bracmat had no facility for handling floating point numbers. This solution handles only rational numbers.
( ( accumulator
=
.
' ( add sum object
. (object=add=$arg+!arg)
& !(object.add):?sum
& '($($sum)+!arg):(=?(object.add))
& !sum
)
)
& accumulator$1:(=?x)
& x$5
& accumulator$3
& out$(x$23/10)
)
Output:
83/10
The following solution uses UFP (UnIfancyfied Floating Point) objects to handle the terms in case not both are rational numbers.
( ( accumulator
=
.
' ( add sum object addFunction
. ( addFunction
= A B
. !arg:(?A.?B)
& ( !A:#
& !B:#
& "If both values are recognized as integer or fractional values, just use '+'."
& !A+!B
| "Otherwise, create an object for adding two C doubles and let that run."
& ( new
$ (UFP,'(.$($A)+$($B)))
. go
)
$
)
)
& ( object
= add
= addFunction$($arg.!arg)
)
& !(object.add):?sum
& 'addFunction$($($sum).!arg)
: (=?(object.add))
& !sum
)
)
& accumulator$1:(=?x)
& x$5
& accumulator$1:(=?y)
& y$"5.0"
& out$(x$23/10)
& out$(y$"2.3")
)
Output
83/10 8.3000000000000007E+00
Brat
accumulator = { sum |
{ n | sum = sum + n }
}
x = accumulator 1
x 5
accumulator 3 #Does not affect x
p x 2.3 #Prints 8.3 (1 + 5 + 2.3)
BQN
Ported from Ruby.
Acc ← {
𝕊 sum:
{sum+↩𝕩}
}
x ← Acc 1
X 5
Acc 3
X 2.3
C
Deviation: Not in standard C, but several compilers include the typeof operator as an extension which can be used like a typedef. Functions must be defined outside of the main program body and they retain the same type throughout their life. C11 is supposed to give us some Type-generic macro expressions.
#include <stdio.h>
//~ Take a number n and return a function that takes a number i
#define ACCUMULATOR(name,n) __typeof__(n) name (__typeof__(n) i) { \
static __typeof__(n) _n=n; LOGIC; }
//~ have it return n incremented by the accumulation of i
#define LOGIC return _n+=i
ACCUMULATOR(x,1.0)
ACCUMULATOR(y,3)
ACCUMULATOR(z,'a')
#undef LOGIC
int main (void) {
printf ("%f\n", x(5)); /* 6.000000 */
printf ("%f\n", x(2.3)); /* 8.300000 */
printf ("%i\n", y(5.0)); /* 8 */
printf ("%i\n", y(3.3)); /* 11 */
printf ("%c\n", z(5)); /* f */
return 0;
}
C#
using System;
class Program
{
static Func<dynamic, dynamic> Foo(dynamic n)
{
return i => n += i;
}
static void Main(string[] args)
{
var x = Foo(1);
x(5);
Foo(3);
Console.WriteLine(x(2.3));
}
}
C++
First solution has a deviation: The return type is wrong when the accumulator is called with an integer argument after is has been called with a float argument. Later it is explained how to correct this.
#include <iostream>
class Acc
{
public:
Acc(int init)
: _type(intType)
, _intVal(init)
{}
Acc(float init)
: _type(floatType)
, _floatVal(init)
{}
int operator()(int x)
{
if( _type == intType )
{
_intVal += x;
return _intVal;
}
else
{
_floatVal += x;
return static_cast<int>(_floatVal);
}
}
float operator()(float x)
{
if( _type == intType )
{
_floatVal = _intVal + x;
_type = floatType;
return _floatVal;
}
else
{
_floatVal += x;
return _floatVal;
}
}
private:
enum {floatType, intType} _type;
float _floatVal;
int _intVal;
};
int main()
{
Acc a(1);
a(5);
Acc(3);
std::cout << a(2.3f);
return 0;
}
The following is similar to the above, using lambda functions from C++11. Note that we declared the lambda mutable
, which allows us to modify variables that were captured by value. This feature allows us to maintain mutable state, which is essential for an accumulator.
It suffers from the same deviation as the former, where the return type is wrong when the accumulator is called with a float argument after is has been called with an integer argument.
#include <iostream>
#include <functional>
template <typename T>
std::function<T(T)> makeAccumulator(T sum) {
return [=](T increment) mutable {
return sum += increment;
};
}
int main() {
auto acc = makeAccumulator<float>(1);
acc(5);
makeAccumulator(3);
std::cout << acc(2.3) << std::endl;
return 0;
}
The deviation stems from two sources. First, a C++ object (such as the accumulator) has an immutable type. To correct this, we must separate the accumulator from the cumulant value it holds. For example:
struct CumulantBase_
{
virtual ~CumulantBase_();
virtual std::ostream& Write(std::ostream& dst) const = 0;
};
template<class T_> struct Cumulant_ : CumulantBase_
{
T_ val_;
Cumulant_(const T_& val) : val_(val) {}
std::ostream& Write(std::ostream& dst) const override
{
return dst << val_;
}
};
struct Accumulator_
{
std::unique_ptr<CumulantBase_> val_;
template<class T_> Accumulator_(const T_& val) { Set(val); }
template<class T_> void Set(const T_& val) { val_.reset(new Cumulant_<T_>(val)); }
(This is Coplien's "State" pattern.)
The second issue is that the built-in operator + is a multimethod, implementing a compile-time dispatch and promotion which we must manually reproduce.
// still inside struct Accumulator_
// various operator() implementations provide a de facto multimethod
Accumulator_& operator()(int more)
{
if (auto i = CoerceInt(*val_))
Set(+i + more);
else if (auto d = CoerceDouble(*val_))
Set(+d + more);
else
THROW("Accumulate(int) failed");
return *this;
}
Accumulator_& operator()(double more)
{
if (auto d = CoerceDouble(*val_))
Set(+d + more);
else
THROW("Accumulate(double) failed");
return *this;
}
Accumulator_& operator()(const String_& more)
{
if (auto s = CoerceString(*val_))
Set(+s + more);
else
THROW("Accumulate(string) failed");
return *this;
}
};
These rely on coercion functions which switch on the so-far-accumulated type:
// recognize cumulants by type
boost::optional<int> CoerceInt(const CumulantBase_& c)
{
if (auto p = dynamic_cast<const Cumulant_<int>*>(&c))
return p->val_;
return boost::optional<int>();
}
boost::optional<double> CoerceDouble(const CumulantBase_& c)
{
if (auto p = dynamic_cast<const Cumulant_<double>*>(&c))
return p->val_;
if (auto i = CoerceInt(c))
return boost::optional<double>(i);
return boost::optional<double>();
}
boost::optional<String_> CoerceString(const CumulantBase_& c)
{
if (auto p = dynamic_cast<const Cumulant_<String_>*>(&c))
return p->val_;
return boost::optional<String_>();
}
All that remains is to write to the stream:
std::ostream& operator<<(std::ostream& dst, const Accumulator_& acc)
{
return acc.val_->Write(dst);
}
Ceylon
shared void run() {
Integer|Float accumulator
(variable Integer|Float n)
(Integer|Float i)
=> switch (i)
case (is Integer)
(n = n.plusInteger(i))
case (is Float)
(n = i + (switch(prev = n)
case (is Float) prev
case (is Integer) prev.float));
value x = accumulator(1);
print(x(5));
print(accumulator(3));
print(x(2.3));
}
- Output:
6 <Integer|Float>(Integer|Float) 8.3
Clay
To my knowledge Clay does not admit of an elegant solution to this problem, although it should be stated that I am still exploring the language. But a clean solution mirroring that for other static languages is quite simple (one in which the operative numeric type is constrained by the original call to acc):
acc(n) {
return (m) => {
n = n + m;
return n;
};
}
main() {
var x = acc(1.0);
x(5);
acc(3);
println(x(2.3)); // Prints “8.300000000000001”.
}
Although statically typed, due to Clay’s everywhere-genericity this has the advantage of working out of the box for any type that defines addition:
var y = acc(Vector[Char]("Hello"));
println(y(" World!")); // Prints "Hello World!”.
But you could constrain the function to numeric types were you so inclined:
[N | Numeric?(N)] acc(n: N) {
return (m) => {
n = n + m;
return n;
};
}
One could go crazy with tagged unions and runtime dispatching to rig something up that adhered more closely to the problem’s specification. But I know of no easier way to “change types” in the fashion necessary.
Clojure
The atom function creates an atomically updatable identity holding a value. The swap! function atomically updates the atom's value, returning the new value. The function returned from an accum call satisfies all the requirements.
(defn accum [n]
(let [acc (atom n)]
(fn [m] (swap! acc + m))))
Similarly, a ref could be used.
(defn accum [n]
(let [acc (ref n)]
#(dosync (alter acc + %))))
CoffeeScript
accumulator = (sum) ->
(n) -> sum += n
f = accumulator(1)
console.log f(5)
console.log f(2.3)
Common Lisp
(defun accumulator (sum)
(lambda (n)
(incf sum n)))
Example usage:
(defvar x (accumulator 1))
(funcall x 5)
(accumulator 3)
(funcall x 2.3)
- Output:
X 6 #<CLOSURE :LAMBDA (N) (SETF SUM (+ SUM N))> 8.3
Crystal
# Make types a bit easier with an alias
alias Num = Int32 | Int64 | Float32 | Float64
def accumulator(sum : Num)
# This proc is very similar to a Ruby lambda
->(n : Num){ sum += n }
end
x = accumulator(5)
puts x.call(5) #=> 10
puts x.call(10) #=> 20
puts x.call(2.4) #=> 22.4
D
import std.stdio;
void main() {
auto x = acc(1);
x(5);
acc(3);
writeln(x(2.3));
}
auto acc(U = real, T)(T initvalue) { // U is type of the accumulator
auto accum = cast(U)initvalue ;
return (U n) { return accum += n ; } ;
}
Dart
The =>
operator is Dart's special syntax for single line closures. When you use it the value of the expression is automatically returned without the return statement.
num
is base type for int
and double
.
Implementation with dynamic typing:
makeAccumulator(s) => (n) => s += n;
Implementation with static typing (preferred in Dart 2):
typedef Accumulator = num Function(num);
Accumulator makeAccumulator(num s) => (num n) => s += n;
Verbose version:
typedef Accumulator = num Function(num);
Accumulator makeAccumulator(num initial) {
num s = initial;
num accumulator(num n) {
s += n;
return s;
}
return accumulator;
}
Usage example for any of above:
void main() {
var x = makeAccumulator(1);
x(5);
makeAccumulator(3);
print(x(2.3));
}
- Output:
8.3
Type checking:
void main() {
var x = makeAccumulator(1);
print(x(5).runtimeType); // int
print(x(2.3).runtimeType); // double
print(x(4).runtimeType); // double
}
Déjà Vu
accum n:
labda i:
set :n + n i
n
local :x accum 1
drop x 5
drop accum 3
!print x 2.3
Delphi
program Accumulator_factory;
{$APPTYPE CONSOLE}
uses
System.SysUtils,
System.Variants;
type
TFn = TFunc<variant, variant>;
function Foo(n: variant): TFn;
begin
Result :=
function(i: variant): variant
begin
n:= n + i;
Result := n;
end;
end;
begin
var x := Foo(1);
x(5);
Foo(3); // do nothing
Writeln(x(2.3));
Readln;
end.
E
def foo(var x) {
return fn y { x += y }
}
EchoLisp
(define-syntax-rule (inc x v) (set! x (+ x v)))
(define (accumulator (sum 0)) (lambda(x) (inc sum x) sum))
(define x (accumulator 1)) → x
(x 5) → 6
;; another closure
(accumulator 3) → (🔒 λ (_x) (📝 #set! sum (#+ sum _x)) sum)
(x 2.3) → 8.3
Elena
ELENA 6.x :
function(acc)
= (n => acc.append(n));
accumulator(n)
= function(new Variable(n));
public program()
{
var x := accumulator(1);
x(5);
var y := accumulator(3);
console.write(x(2.3r))
}
- Output:
8.3
Elixir
Elixir provides Agents to simplify creating a process to maintain state where mutable variables aren't allowed.
defmodule AccumulatorFactory do
def new(initial) do
{:ok, pid} = Agent.start_link(fn() -> initial end)
fn (a) ->
Agent.get_and_update(pid, fn(old) -> {a + old, a + old} end)
end
end
end
The passing test to exercise the Accumulator and show usage:
ExUnit.start
defmodule AccumulatorFactoryTest do
use ExUnit.Case
test "Accumulator basic function" do
foo = AccumulatorFactory.new(1)
foo.(5)
bar = AccumulatorFactory.new(3)
assert bar.(4) == 7
assert foo.(2.3) == 8.3
end
end
- Output:
. Finished in 0.06 seconds (0.06s on load, 0.00s on tests) 1 test, 0 failures Randomized with seed 587000
EMal
in Org:RosettaCode
^|EMal has a mechanism to force the type system to allow nulls on types
|that are usually not nullable, such as int or real.
|In the following code we are telling EMal that int and real implement
|the Number virtual interface, so that it can only
|accept null (because it is an interface), int, and real values.
|^
type Number allows int, real
type AccumulatorUsingNumber
fun foo ← fun by Number n
fun g ← Number by Number i
return n += i
end
return g
end
type AccumulatorUsingVar
^|EMal has an universal supertype Variable (var) that can be used.
|Some manual type checks are required.
|^
fun checkType ← void by var value
if generic!value ≠ real and generic!value ≠ int
error(1, "Only real and int values can be used")
end
end
fun foo ← fun by var n
checkType(n)
fun g ← var by var i
checkType(i)
return n += i
end
return g
end
type Main
^|we have developed two solutions,
|it is time to create a list holding both data types.
|We iterate over the solutions in order to test them.
|^
List solutions ← generic[AccumulatorUsingNumber, AccumulatorUsingVar]
for int i ← 0; i < solutions.length; ++i
generic solution ← solutions[i]
writeLine("=== solution " + (i + 1) + " ===")
fun x ← :solution.foo(1)
x(5)
:solution.foo(3)
watch(x(2.3))
fun y ← :solution.foo(1)
y(5)
:solution.foo(3)
watch(y(2))
end
- Output:
=== solution 1 === Org:RosettaCode:Number, Real: <8.3> Org:RosettaCode:Number, Integer: <8> === solution 2 === Variable, Real: <8.3> Variable, Integer: <8>
Erlang
Erlang doesn't allow for mutable variables, but does have variable capture in closures. By spawning a process which loops endlessly, incrementing the sum and returning it to the caller, this mutable state can be imitated.
-module(acc_factory).
-export([loop/1,new/1]).
loop(N)->
receive
{P,I}->
S =N+I, P!S, loop(S)
end.
new(N)->
P=spawn(acc_factory,loop,[N]),
fun(I)->
P!{self(),I},
receive
V-> V
end
end.
ERRE
PROGRAM ACCUMULATOR
PROCEDURE ACCUMULATOR(SUM,N,A->SUM)
IF NOT A THEN SUM=N ELSE SUM=SUM+N
END PROCEDURE
BEGIN
PRINT(CHR$(12);) ! CLS
ACCUMULATOR(X,1,FALSE->X) ! INIT FIRST ACCUMULATOR
ACCUMULATOR(X,-15,TRUE->X)
ACCUMULATOR(X,2.3,TRUE->X)
ACCUMULATOR(Z,3,FALSE->Z) ! INIT SECOND ACCUMULATOR
ACCUMULATOR(Z,5,TRUE->Z)
ACCUMULATOR(Z,2.3,TRUE->Z)
PRINT(X,Z)
END PROGRAM
- Output:
-11.7 10.3
F#
A statically typed version is not possible, but it is quite easy to write dynamically typed functions in F#:
// dynamically typed add
let add (x: obj) (y: obj) =
match x, y with
| (:? int as m), (:? int as n) -> box(m+n)
| (:? int as n), (:? float as x)
| (:? float as x), (:? int as n) -> box(x + float n)
| (:? float as x), (:? float as y) -> box(x + y)
| _ -> failwith "Run-time type error"
let acc init =
let state = ref (box init)
fun y ->
state := add !state (box y)
!state
do
let x : obj -> obj = acc 1
printfn "%A" (x 5) // prints "6"
acc 3 |> ignore
printfn "%A" (x 2.3) // prints "8.3"
Actually, it is possible to create a statically typed version by using an inline accumulator creation function.
let inline makeAccumulator init =
let acc = ref init
fun i ->
acc := !acc + i
!acc
do
let acc = makeAccumulator 1.0 // create a float accumulator
acc 5.0 |> ignore
let _ = makeAccumulator 3 // create an unused integer accumulator
printfn "%A" (acc 2.3)
- Output:
8.3
Factor
USE: locals
:: accumulator ( n! -- quot ) [ n + dup n! ] ;
1 accumulator
[ 5 swap call drop ]
[ drop 3 accumulator drop ]
[ 2.3 swap call ] tri .
Fantom
The accumulator function is a little unwieldy using multiple ifs to maintain the type of 'sum' until forced to change. Again, a result of the three concrete Num types, Int, Float and Decimal, all being separated in the API.
class AccumulatorFactory
{
static |Num -> Num| accumulator (Num sum)
{
return |Num a -> Num|
{ // switch on type of sum
if (sum is Int)
{ // and then type of a
if (a is Int)
return sum = sum->plus(a)
else if (a is Float)
return sum = sum->plusFloat(a)
else
return sum = sum->plusDecimal(a)
}
else if (sum is Float)
{
if (a is Int)
return sum = sum->plusInt(a)
else if (a is Float)
return sum = sum->plus(a)
else
return sum = sum->plusDecimal(a)
}
else // if (sum is Decimal)
{
if (a is Int)
return sum = sum->plusInt(a)
else if (a is Float)
return sum = sum->plusFloat(a)
else
return sum = sum->plus(a)
}
}
}
public static Void main ()
{
x := accumulator (3.1)
y := accumulator (3f)
echo (x(5)) // the Decimal sum combines with an Int
echo (x(2))
echo (y(5.1)) // the Float sum combines with a Decimal
x = accumulator (1)
x (5)
accumulator (3)
echo (x(2.3)) // the Int sum is now a Decimal
}
}
Forth
Forth is untyped; this works on integers.
: accumulator
create ( n -- ) ,
does> ( n -- acc+n ) tuck +! @ ;
0 accumulator foo
1 foo . \ 1
2 foo . \ 3
3 foo . \ 6
The idiomatic way to deal with floats is to have a float version of this code; for a mixture of integers and floats, you decide at the start to use a float accumulator, and convert integers to floats explicitly:
: faccumulator ( r "name" -- )
create falign f,
does> ( r1 -- r2 )
faligned dup f@ f+ fdup f! ;
1 s>f faccumulator x
5 s>f x fdrop
3 s>f faccumulator y \ unused
2.3e x f.
Fortran
Fortran does not have functions as first class objects, and can not create functions at runtime.
Fortran77
Fortran77 does not support objects and overloading and thus the user must declare the type of the function to generate. The following are noted:
The code uses CPP which is at least available on the GNU compiler with the -cpp directive.
The code uses the semicolon as command separators. This was not standard in Fortran77 but was accepted by many compilers (some used colon instead).
The "data" command implies that the variables are static. This was not standard in Fortran77 but was accepted by virtually all compilers.
#define foo(type,g,nn) \
typex function g(i);\
typex i,s,n;\
data s,n/0,nn/;\
s=s+i;\
g=s+n;\
end
foo(real,x,1)
foo(integer,y,3)
program acc
real x, temp
integer y, itemp
temp = x(5.0)
print *, x(2.3)
itemp = y(5)
print *, y(2)
stop
end
- Output:
8.30000019 10
Fortran2003
Fortran2003 and later supports objects and overloading. The overloaded functions are encapsulated in an object.
module modAcc
implicit none
private
integer, public, parameter :: KRL = selected_real_kind(14)
type, public :: AccType
real(KRL), private :: dn, dsum
complex(KRL), private :: fn, fsum
integer, private :: jn, jsum, icod
contains
procedure, private :: initd, initf, initi
generic, public :: init => initd, initf, initi
procedure, private :: dfun, ffun, jfun
generic, public :: fun => dfun, jfun, ffun
end type AccType
contains
subroutine initd(self, dd)
class(AccType), intent(inout) :: self
real(KRL), intent(in) :: dd
self%dn = dd
self%icod = 1
end subroutine initd
subroutine initf(self, ff)
class(AccType), intent(inout) :: self
complex(KRL), intent(in) :: ff
self%fn = ff
self%icod = 2
end subroutine initf
subroutine initi(self, jj)
class(AccType), intent(inout) :: self
integer, intent(in) :: jj
self%jn = jj
self%icod = 3
end subroutine initi
real(KRL) function dfun(self, di)
class(AccType), intent(inout) :: self
real(KRL), intent(in) :: di
self%dsum = self%dsum + di
dfun = self%dn + self%dsum
end function dfun
complex(KRL) function ffun(self, fi)
class(AccType), intent(inout) :: self
complex(KRL), intent(in) :: fi
self%fsum = self%fsum + fi
ffun = self%fn + self%fsum
end function ffun
integer function jfun(self, ji)
class(AccType), intent(inout) :: self
integer, intent(in) :: ji
self%jsum = self%jsum + ji
jfun = self%jn + self%jsum
end function jfun
end module modAcc
program test
use modAcc
implicit none
type(AccType) :: x, y
integer :: itemp
real(KRL) :: temp
call x%init(1.0_KRL)
temp = x%fun(5.0_KRL)
call y%init(3)
print *, x%fun(2.3_KRL)
itemp = y%fun(5)
print *, y%fun(2)
end program test
- Output:
8.3000000000000007 10
FreeBASIC
It doesn't appear to be possible to program this task in FreeBASIC in the precise way it is posed.
The problem is that FB doesn't support closures and, whilst we can manufacture an equivalent object, we'd then have the further problem that you can't pass pointers to object methods, only to static procedures.
To get around this restriction we'd normally wrap the object method in a static procedure and pass an object pointer to that followed by any other arguments required by the method. However, this won't work here because the task specifies that the method can take only a single number argument and the object pointer would be internal to 'foo' in any case.
Probably the best we can do is for 'foo' to return the object and then to call the method 'g' directly on that:
' FB 1.05.0 Win64
' uses overloaded methods to deal with the integer/float aspect (long and single are both 4 bytes)
Type Bar
Public:
Declare Constructor(As Long)
Declare Constructor(As Single)
Declare Function g(As Long) As Long
Declare Function g(As Single) As Single
Private:
As Single sum_ '' can't be altered by external code
End Type
Constructor Bar(i As Long)
sum_ = i
End Constructor
Constructor Bar(s As Single)
sum_ = s
End Constructor
Function Bar.g(i As Long) As Long
sum_ += i
Return sum_ '' would round down to a Long if non-integral Singles had been added previously
End Function
Function Bar.g(s As Single) As Single
sum_ += s
Return sum_
End Function
Function foo Overload(i As Long) As Bar '' returns a Bar object rather than a pointer to Bar.g
Dim b As Bar = Bar(i)
Return b
End Function
Function foo Overload(s As Single) As Bar '' overload of foo to deal with Single argument
Dim b As Bar = Bar(s)
Return b
End Function
Dim x As Bar = foo(1) '' assigns Bar object to x
x.g(5) '' calls the Long overload of g on the Bar object
foo(3) '' creates a separate Bar object which is unused
print x.g(2.3) '' calls the Single overload of g on the Bar object and should print 1 + 5 + 2.3 = 8.3
Print
Print "Press any key to quit"
Sleep
- Output:
8.3
Go
Small deviation on condition 2. The task specifies to handle all numeric types, and only int and float64 are shown here. The technique would extend to all types just as easily, but Go has lots of numeric types and the program would be big.
package main
import "fmt"
func accumulator(sum interface{}) func(interface{}) interface{} {
return func(nv interface{}) interface{} {
switch s := sum.(type) {
case int:
switch n := nv.(type) {
case int:
sum = s + n
case float64:
sum = float64(s) + n
}
case float64:
switch n := nv.(type) {
case int:
sum = s + float64(n)
case float64:
sum = s + n
}
default:
sum = nv
}
return sum
}
}
func main() {
x := accumulator(1)
x(5)
accumulator(3)
fmt.Println(x(2.3))
}
- Output:
8.3
Golo
#!/usr/bin/env golosh
----
An accumulator factory example for Rosetta Code.
This one uses the box function to create an AtomicReference.
----
module rosetta.AccumulatorFactory
function accumulator = |n| {
let number = box(n)
return |i| -> number: accumulateAndGet(i, |a, b| -> a + b)
}
function main = |args| {
let acc = accumulator(3)
println(acc(1))
println(acc(1.1))
println(acc(10))
println(acc(100.101))
}
Groovy
Solution:
def accumulator = { Number n ->
def value = n;
{ it = 0 -> value += it}
}
Test:
def x = accumulator(1)
println x()
assert x() instanceof Integer
println x(5)
assert x() instanceof Integer
def y = accumulator(3)
println y()
assert y() instanceof Integer
println x(2.3)
assert x() instanceof BigDecimal
println y(10)
assert y() instanceof Integer
println y(200L)
assert y() instanceof Long
println y(2.25D)
assert y() instanceof Double
- Output:
1 6 3 8.3 13 213 215.25
Haskell
import Control.Monad.ST
import Data.STRef
accumulator :: (Num a) => a -> ST s (a -> ST s a)
accumulator sum0 = do
sum <- newSTRef sum0
return $ \n -> do
modifySTRef sum (+ n)
readSTRef sum
main :: IO ()
main = print foo
where foo = runST $ do
x <- accumulator 1
x 5
accumulator 3
x 2.3
- Output:
8.3
Note The accumulator
function could be written in applicative style:
accumulator = newSTRef >=> return . factory
where factory s n = modifySTRef s (+ n) >> readSTRef s
Icon and Unicon
At first glance you might expect the example below to run under Icon; however, as the co-expression calling sequence is Unicon specific.
Strictly speaking, genAcc(n) returns a co-expression, not a function. However, the invocation syntax here is indistinguishable from calling a function.
procedure main()
a := genAcc(3)
b := genAcc(5)
write(" " ,center("a",5), " ", center("b", 5))
write("genAcc: ", right(a(4),5), " ", right(b(4), 5))
write("genAcc: ", right(a(2),5), " ", right(b(3),5))
write("genAcc: ", right(a(4.5),5)," ", right(b(1.3),5))
end
procedure genAcc(n) # The generator factory
return makeProc { while i := (n@&source)[1] do n +:= i }
end
procedure makeProc(A) # A Programmer-Defined Control Operation
return (@A[1],A[1])
end
This example produces the output:
a b genAcc: 7 9 genAcc: 9 12 genAcc: 13.5 13.3
To adapt the above for use in Icon, the function-syntax for activating co-expressions (e.g. a(4)) available in Unicon would have to be replaced with the activation operator (e.g. [4]@a). The use of a list as the value passed through activation is to retain compatibility with the Unicon approach.
Io
accumulator := method(sum,
block(x, sum = sum + x) setIsActivatable(true)
)
x := accumulator(1)
x(5)
accumulator(3)
x(2.3) println // --> 8.3000000000000007
J
See j:Guides/Lexical_Closure, including the dissent section.
oleg=:1 :0
a=. cocreate''
n__a=: m
a&(4 : 'n__x=: n__x + y')
)
Example use:
F=: 10 oleg
F 11
21
F 12
33
F 11
44
Java
Java has no first-class functions, so an accumulator can't use the x(5)
syntax. The standard syntactic workaround is to use a standard method name, like x.call(5)
or x.apply(5)
. This is a deviation from task.
Our accumulator sums with long integers as far as possible before switching to floats. This requires the use of the Number
class. The code needs Java 5 to autobox primitive values 1
or 2.3
into instances of Number. The apply
method is ready to implement interface UnaryOperator in Java 8.
public class Accumulator
//implements java.util.function.UnaryOperator<Number> // Java 8
{
private Number sum;
public Accumulator(Number sum0) {
sum = sum0;
}
public Number apply(Number n) {
// Acts like sum += n, but chooses long or double.
// Converts weird types (like BigInteger) to double.
return (longable(sum) && longable(n)) ?
(sum = sum.longValue() + n.longValue()) :
(sum = sum.doubleValue() + n.doubleValue());
}
private static boolean longable(Number n) {
return n instanceof Byte || n instanceof Short ||
n instanceof Integer || n instanceof Long;
}
public static void main(String[] args) {
Accumulator x = new Accumulator(1);
x.apply(5);
new Accumulator(3);
System.out.println(x.apply(2.3));
}
}
- Output:
8.3
A printed Accumulator would look like Accumulator@42e816
Java 8 added the lambda syntax. A lambda is an anonymous inner class that implements a one-method interface. We can make the accumulator as a lambda, but it must store the sum in another object. We use a one-element array.
import java.util.function.UnaryOperator;
public class AccumulatorFactory {
public static UnaryOperator<Number> accumulator(Number sum0) {
// Allows sum[0] = ... inside lambda.
Number[] sum = { sum0 };
// Acts like n -> sum[0] += n, but chooses long or double.
// Converts weird types (like BigInteger) to double.
return n -> (longable(sum[0]) && longable(n)) ?
(sum[0] = sum[0].longValue() + n.longValue()) :
(sum[0] = sum[0].doubleValue() + n.doubleValue());
}
private static boolean longable(Number n) {
return n instanceof Byte || n instanceof Short ||
n instanceof Integer || n instanceof Long;
}
public static void main(String[] args) {
UnaryOperator<Number> x = accumulator(1);
x.apply(5);
accumulator(3);
System.out.println(x.apply(2.3));
}
}
JavaScript
ES5
function accumulator(sum) {
return function(n) {
return sum += n;
}
}
var x = accumulator(1);
x(5);
console.log(accumulator(3).toString() + '<br>');
console.log(x(2.3));
- Output:
function (n) { return sum += n; } 8.3
ES6
let accumulator = sum => (n => sum += n);
let x = accumulator(1);
console.log(x(5));
accumulator(3);
console.log(x(2.3));
- Output:
6 8.3
JavaScript 1.8 (SpiderMonkey Only)
function accumulator(sum) function(n) sum += n;
var x = accumulator(1);
x(5);
console.log(accumulator(3).toSource());
console.log(x(2.3));
- Output:
(function (n) sum += n) 8.3
Jsish
From Javascript ES5 entry.
/* Accumulator factory, in Jsish */
function accumulator(sum) {
return function(n) {
return sum += n;
};
}
provide('accumulatorFactory', '0.6');
if (Interp.conf('unitTest')) {
var x,y;
;x = accumulator(1);
;accumulator;
;x;
;x(5);
;accumulator(3);
;x(2.3);
;y = accumulator(0);
;y;
;x(1);
;y(2);
;x(3);
;y(4);
;x(5);
}
/*
=!EXPECTSTART!=
x = accumulator(1) ==> "function(n) {\n return sum += n;\n }"
accumulator ==> "function accumulator(sum) {\n return function(n) {\n return sum += n;\n };\n}"
x ==> "function(n) {\n return sum += n;\n }"
x(5) ==> 6
accumulator(3) ==> "function(n) {\n return sum += n;\n }"
x(2.3) ==> 8.3
y = accumulator(0) ==> "function(n) {\n return sum += n;\n }"
y ==> "function(n) {\n return sum += n;\n }"
x(1) ==> 9.3
y(2) ==> 2
x(3) ==> 12.3
y(4) ==> 6
x(5) ==> 17.3
=!EXPECTEND!=
*/
- Output:
prompt$ jsish -u accumulatorFactory.jsi [PASS] accumulatorFactory.jsi
Julia
function accumulator(i)
f(n) = i += n
return f
end
x = accumulator(1)
@show x(5)
accumulator(3)
@show x(2.3)
- Output:
x(5) = 6 x(2.3) = 8.3
Kotlin
Overloads would be needed for all six primitive numeric types but, in the interests of brevity, only two overloads of 'foo' have been coded:
// version 1.1
fun foo(n: Double): (d: Double) -> Double {
var nn = n
return { nn += it; nn }
}
fun foo(n: Int): (i: Int) -> Int {
var nn = n
return { nn += it; nn }
}
fun main(args: Array<String>) {
val x = foo(1.0) // calls 'Double' overload
x(5.0)
foo(3.0)
println(x(2.3))
val y = foo(1) // calls 'Int' overload
y(5)
foo(5)
println(y(2))
}
- Output:
8.3 8
Lambdatalk
Lambdatlk is a functional programming language without closures but with mutable arrays.
{def acc
{lambda {:a :n}
{+ {A.toS {A.addlast! :n :a}}}}}
-> acc
1) using a global:
{def A {A.new 1}}
-> A
{acc {A} 5}
-> 6
{acc {A} 2.3}
-> 8.3
2) inside a local context:
{let { {:a {A.new 1}}
} {br}{acc :a 5}
{br}{acc :a 2.3}
} ->
6
8.3
Lang
Lang does not support closures. The use of combinator functions and pointers allows a function to store state.
fp.accumulator = ($sum) -> {
$sumPtr = $[sum]
fp.f = ($sumPtr, $n) -> {
$*sumPtr += $n
return $*sumPtr
}
return fn.argCnt1(fn.combA2(fp.f, $sumPtr))
}
$x = fp.accumulator(1)
fn.println($x(5))
fp.accumulator(3)
fn.println($x(2.3))
fn.println()
$y = fp.accumulator(1.)
fn.println($y(5))
fn.println($y(2.3))
- Output:
6 8.3 6.0 8.3
LFE
LFE doesn't support mutable data (nor global variables); as such, this task requires a work-around. There are two ways to accomplish it: via closure on anonymous function, or closure on a process.
Traditional closure
(defun accum (m)
(lambda (n)
(let ((sum (+ m n)))
`(#(func ,(accum sum))
#(sum ,sum)))))
Since we want to use both the returned function as well as the data for the call, we return a tuple containing both. Using standard LFE pattern matching, we can extract these.
Usage (in the REPL):
> (set x (accum 1)) #Fun<lfe_eval.12.122728658> > (set `(#(func ,x) ,_) (funcall x 5)) (#(func #Fun<lfe_eval.12.122728658>) #(sum 6)) > (funcall x 3) (#(func #Fun<lfe_eval.12.122728658>) #(sum 9)) > (set `(#(func ,x) ,_) (funcall x 2.3)) (#(func #Fun<lfe_eval.12.122728658>) #(sum 8.3))
Note that we want to re-set the variable x
with each call in order to use its updated state (since LFE is a functional programming language which doesn't support mutable global variables.
Process closure
We can creating a looping process which provides the same functionality as the self-calling function in the "traditional closure" approach:
(defun loop (m)
(receive
(`#(,caller ,n)
(let ((sum (+ m n)))
(! caller sum)
(loop sum)))))
(defun accum (m)
(let ((loop-pid (spawn (lambda () (loop m)))))
(lambda (n)
(! loop-pid `#(,(self) ,n))
(receive
(sum sum)))))
Usage (in the REPL):
> (accum 1) #Fun<lfe_eval.12.122728658> > (set x (accum 1)) #Fun<lfe_eval.12.122728658> > (funcall x 5) 6 > (accum 3) #Fun<lfe_eval.12.122728658> > (funcall x 2.3) 8.3
Since we're using a looping process to track state, there's no need to re-set the x
variable with each call.
Lua
A simple implementation:
function acc(init)
init = init or 0
return function(delta)
init = init + (delta or 0)
return init
end
end
An expanded example of similar but more complex functionality:
do
local accSum = 0; -- accumulator factory 'upvalue'
function acc(v) -- the accumulator factory
accSum = accSum + (v or 0) -- increment factory sum
local closuredSum = accSum; -- new 'upvalue' at each factory call
return function (w) -- the produced accumulator function
closuredSum = closuredSum + (w or 0) -- increment product 'upvalue'
return closuredSum -- return 'upvalue'
end, accSum -- end of product closure
end--acc
end--end of factory closure
Usage example:
x = acc(1) -- x stores the product with initial value = 1
x(5) -- add 5 to x's sum
acc(3) -- add 3 to factory's sum
print (x(2.3)) --> 8.3 -- add 2.3 to x's sum then print the result
y = acc() -- create new function with factory's sum as initial value
print (y()) --> 4 -- print the accumulated value inside the product y
M2000 Interpreter
\\ M2000 Interpreter
\\ accumulator factory
foo=lambda acc=0 (n as double=0) -> {
\\ interpreter place this: read n as double=0 as first line of lambda function
if n=0 then =acc : exit
acc+=n
\\ acc passed as a closuer to lambda (a copy of acc in the result lambda function)
=lambda acc -> {
' if stack of values is empty then return a copy of acc
if empty then =acc : exit
read x
\\ x has no type here, can be any numeric type (also can be an object too)
\\ accumulator is double, and is a closure (a copy of acc in foo)
acc+=x
\\ any variable in M2000 hold first type
\\ if x is an object then we get error, except if object use this operator
x=acc
\\ so we return x type
=x
}
}
x=foo(1&) ' 1& is long type (32bit)
call void x(5) ' 5 is double type (the default type for M2000)
call void foo(3#) ' void tell to interpreter to throw result, 3# is Currency type
print x(2.3@) ' print 8.3, 2.3@ is Decimal type
print foo()=4 ' print true
def ExpType$(z)=type$(z)
print ExpType$(foo())="Double"
print ExpType$(x(0&))="Long"
print ExpType$(x(0@))="Decimal"
print ExpType$(x())="Double"
print ExpType$(foo(20))="lambda"
Maple
This creates a procedure closed over the local variable total in the factory procedure. The initial value, if not passed to the factory procedure, is taken to be 0 and, if the generated accumulator is given no value, it increments the total by 1.
AccumulatorFactory := proc( initial := 0 )
local total := initial;
proc( val := 1 ) total := total + val end
end proc:
Running this, we get:
> acc := AccumulatorFactory( 1 ):
> acc( 5 );
6
> AccumulatorFactory( 3 ):
> acc( 2.3 );
8.3
> acc(); # use the default increment of 1
9.3
> acc( 3 - 4 * I ); # also handles complex numbers
12.3 - 4. I
> acc( I ); # add the imaginary unit
12.3 - 3. I
Mathematica / Wolfram Language
accFactory[initial_] :=
Module[{total = initial},
Function[x, total += x]
]
x=accFactory[1];
x[5.0];
accFactory[3];
x[2.3]
- Output:
8.3
Mercury
Strict-adherence-to-the-task solution
Deviations:
1. this doesn't work with "any numerical type" out of the box, but requires that users add numerical types to a typeclass.
2. this likely violates some hidden taste requirements of the task, as used by Paul Graham to dismiss Forth solutions. Certainly, this is not really an example of Mercury that anyone would want to use in a Mercury project.
:- module accum.
:- interface.
:- typeclass addable(T) where [
func T + T = T
].
:- impure func gen(T) = (impure (func(T)) = T) <= addable(T).
:- implementation.
:- import_module bt_array, univ, int.
:- mutable(states, bt_array(univ), make_empty_array(0), ground, [untrailed]).
gen(N) = F :-
some [!S] (
semipure get_states(!:S),
size(!.S, Size),
resize(!.S, 0, Size + 1, univ(N), !:S),
impure set_states(!.S)
),
F = (impure (func(Add)) = M :-
some [!SF] (
semipure get_states(!:SF),
!.SF ^ elem(Size) = U,
det_univ_to_type(U, M0),
M = M0 + Add,
!SF ^ elem(Size) := univ(M),
impure set_states(!.SF)
)).
As used:
:- module accumuser.
:- interface.
:- import_module io.
:- pred main(io::di, io::uo) is det.
:- implementation.
:- import_module accum, list, string, int, float.
:- instance addable(int) where [
A + B = int.(A + B)
].
:- instance addable(float) where [
A + B = float.(A + B)
].
:- pragma promise_pure main/2.
main(!IO) :-
impure F = accum.gen(1),
impure N1 = impure_apply(F, 1),
impure N2 = impure_apply(F, 1),
impure G = accum.gen(500.0),
impure R1 = impure_apply(G, -10.0),
impure R2 = impure_apply(G, -50.0),
io.format("%d, %d\n", [i(N1), i(N2)], !IO),
io.format("%.0f, %.0f\n", [f(R1), f(R2)], !IO).
- Output:
2, 3 490, 440
Realistic solution
Deviations:
1. This still requires addition of numeric types to a typeclass, for a generic +
2. This doesn't return a closure with mutable state, but the state itself, which the caller can thread through rules that apply to them.
:- module accum2.
:- interface.
:- typeclass addable(T) where [
func T + T = T
].
:- type accum(T).
% init(N) = Acc
% Return an accumulator with initial value of N
%
:- func init(T) = accum(T)
<= addable(T).
% bump(By, N, !Acc)
% Add By to accumulator !Acc, yielding the next number as N
%
:- pred bump(T::in, T::out, accum(T)::in, accum(T)::out) is det
<= addable(T).
:- implementation.
:- type accum(T) == T.
init(N) = N.
bump(X, N, N0, N) :-
N = X + N0.
As used, with the same output:
:- module accumuser2.
:- interface.
:- import_module io.
:- pred main(io::di, io::uo) is det.
:- implementation.
:- import_module accum2, list, string, int, float.
:- instance addable(int) where [
A + B = int.(A + B)
].
:- instance addable(float) where [
A + B = float.(A + B)
].
main(!IO) :-
some [!A1] (
!:A1 = accum2.init(1),
accum2.bump(1, N1, !A1),
accum2.bump(1, N2, !.A1, _)
),
some [!A2] (
!:A2 = accum2.init(500.0),
accum2.bump(-10.0, R1, !A2),
accum2.bump(-50.0, R2, !.A2, _)
),
io.format("%d, %d\n", [i(N1), i(N2)], !IO),
io.format("%.0f, %.0f\n", [f(R1), f(R2)], !IO).
MiniScript
Accumulator = function(n)
adder = {"sum": n}
adder.plus = function(n)
self.sum += n
return self.sum
end function
adder.getSum = function(n)
obj = self
_sum = function(n)
return obj.plus(n)
end function
return @_sum
end function
return adder.getSum
end function
acc1 = Accumulator(0)
print acc1(10) // prints 10
print acc1(2) // prints 12
acc2 = Accumulator(1)
print acc2(100) // prints 101
print acc1(0) // prints 12
- Output:
miniscript.exe accumulator.ms 10 12 101 12
Nemerle
Nemerle doesn't have a dynamic type, but we can use matching to bind types to objects.
def Foo(n) {
mutable value : object = n;
fun (i : object) {
match(i) {
|x is int => match(value) {
|y is int => value = x + y;
|y is double => value = x + y;
}
|x is double => match(value) {
|y is int => value = x + (y :> double);
|y is double => value = x + y;
}
}
value
}
}
def x = Foo(1);
def y = Foo(2.2);
x(5);
System.Console.WriteLine(x(2.3));
System.Console.WriteLine(y(3));
Output:
8.3 5.2
NewLisp
(define (sum (x 0)) (inc 0 x))
- Output:
> (define (sum (x 0)) (inc 0 x)) (lambda ((x 0)) (inc 0 x)) > (sum 1) 1 > (sum 1) 2 > (sum 1) 3 > (sum 1.4) 4.4 > (sum 1.4) 5.8 > (sum 1.8) 7.6 >
NGS
{
F Acc(start:Int) {
sum = start
F acc(i:Int) {
sum = sum + i
sum
}
}
acc = Acc(10)
echo(acc(5))
echo(acc(2))
}
- Output:
15 17
Nim
Nim being a static typed language, if the accumulator function was created with an integer, it will always return an integer. So, it isn’t possible to fulfill the second requirement, at least by using standard types.
Three solutions are possible:
- – convert all values to float;
- – give the accumulator type the type of the initial value provided at the creation;
- – use a customized type.
We provide the code for the three solutions.
Using float accumulator
Argument to the factory function may be any signed integer, unsigned integer or float. Argument to the created accumulator function must be float. Result is always float.
proc accumulator[T: SomeNumber](x: T): auto =
var sum = float(x)
result = proc (n: float): float =
sum += n
result = sum
let acc = accumulator(1)
echo acc(5) # 6
discard accumulator(3) # Create another accumulator.
echo acc(2.3) # 8.3
- Output:
6.0 8.300000000000001
Fixed accumulator type
Argument to the factory function nay be any signed integer, unsigned integer or float. Argument to the accumulator function must be of the same type. Result of the accumulator function is also of the same type.
proc accumulator[T: SomeNumber](x: T): auto =
var sum = x
result = proc (n: T): T =
sum += n
result = sum
let x = accumulator(1)
echo x(5) # 6
echo x(2) # 8
let y = accumulator(3.5)
echo y(2) # 5.5
echo y(3) # 8.5
- Output:
6 8 5.5 8.5
Customized number type
Argument to the factory function must be "int" or "float" (extension to other types is possible). Argument of the accumulator function is of the customized type "Number" but may be "int" or "float" thanks to the converters. Result of the accumulator function is of type "Number" but will be displayed either as "int" or "float" according to the actual contents.
This solution fulfills the requirements.
type
# Kind of numbers. We limit this example to "int" and "float".
NumberKind = enum kInt, kFloat
# Customized number type (using variants).
Number = object
case kind: NumberKind
of kInt:
ival: int
of kFloat:
fval: float
# The converters allow transparent conversion from int or float to Number.
converter toNumber(n: int): Number = Number(kind: kInt, ival: n)
converter toNumber(n: float): Number = Number(kind: kFloat, fval: n)
#---------------------------------------------------------------------------------------------------
proc accumulator[T: int|float](x: T): auto =
## Factory procedure.
# Allocate the accumulator storage.
when T is int:
var sum = Number(kind: kInt, ival: x)
elif T is float:
var sum = Number(kind: kFloat, fval: x)
# Create the accumulator procedure.
result = proc (n: Number): Number =
# Create the accumulator procedure.
result = proc (n: Number): Number =
case sum.kind
of kInt:
case n.kind
of kInt:
# Add an int to an int.
sum.ival += n.ival
of kFloat:
# Add a float to an int => change the kind of accumulator to float.
sum = Number(kind: kFloat, fval: sum.ival.toFloat + n.fval)
of kFloat:
case n.kind
of kInt:
# Add an int to a float.
sum.fval += n.ival.toFloat
of kFloat:
# Add a float to a float.
sum.fval += n.fval
result = sum
#---------------------------------------------------------------------------------------------------
proc `$`(n: Number): string =
## Display the accumulator contents as an int or a float depending of its kind.
case n.kind
of kInt: $n.ival
of kFloat: $n.fval
#---------------------------------------------------------------------------------------------------
let acc = accumulator(1)
echo acc(5) # 6
discard accumulator(3) # Create another accumulator.
echo acc(2.3) # 8.3
- Output:
6 8.300000000000001
Nit
Source: the official Nit repository
# The `accumulator factory` task.
#
# Nit has no first-class function.
# A class is used to store the state.
module accumulator_factory
class Accumulator
# The accumulated sum
# Numeric is used, so Int and Float are accepted
private var sum: Numeric
fun call(n: Numeric): Numeric
do
# `add` is the safe `+` method on Numeric
sum = sum.add(n)
return sum
end
end
var x = new Accumulator(1)
x.call(5)
var y = new Accumulator(3)
print x.call(2.3)
Output:
8.3
Objeck
Uses objects instead of first class functions.
bundle Default {
class Accumulator {
@sum : Float;
New(sum : Float) {
@sum := sum;
}
method : public : Call(n : Float) ~ Float {
@sum += n;
return @sum;
}
function : Main(args : String[]) ~ Nil {
x := Accumulator->New(1.0);
x->Call(5.0 );
x->Call(2.3)->PrintLine();
}
}
}
Objective-C
#import <Foundation/Foundation.h>
typedef double (^Accumulator)(double);
Accumulator accumulator_factory(double initial) {
__block double sum = initial;
Accumulator acc = ^(double n){
return sum += n;
};
return acc;
}
int main (int argc, const char * argv[]) {
@autoreleasepool {
Accumulator x = accumulator_factory(1);
x(5);
accumulator_factory(3);
NSLog(@"%f", x(2.3));
}
return 0;
}
- Output:
8.300000
OCaml
Deviations: An accumulator instance can take either integers or floats, but not both mixed (due to lack of runtime polymorphism).
let accumulator sum0 =
let sum = ref sum0 in
fun n ->
sum := !sum +. n;
!sum;;
let _ =
let x = accumulator 1.0 in
ignore (x 5.0);
let _ = accumulator 3.0 in
Printf.printf "%g\n" (x 2.3)
;;
- Output:
8.3
Octave
# not a function file:
1;
function fun = foo(init)
currentSum = init;
fun = @(add) currentSum = currentSum + add; currentSum;
endfunction
x = foo(1);
x(5);
foo(3);
disp(x(2.3));
Oforth
Oforth can only returns blocks, not functions, but a block can be used wherever a function is used.
The block returned by foo (a closure), when performed, retrieves the current value from the closure parameter, adds the top of stack, and stores the result back to the closure's parameter. The result is dup, so it is also returned.
: foo( n -- bl )
#[ n swap + dup ->n ] ;
Usage :
: testfoo
| x y z |
1 foo ->x
5 x perform .
3 foo ->y
2.3 x perform dup . ", x accumulator value is a" . class .cr
10 y perform dup . ", y accumulator value is a" . class .cr
"aaa" foo ->z
"bbb" z perform dup . ", z accumulator value is a" . class .cr
;
- Output:
>testfoo 6 8.3 , x accumulator value is a #Float 13 , y accumulator value is a #Integer aaabbb , z accumulator value is a #String ok
ooRexx
ooRexx does not have functions that can maintain state between calls. The standard work around is to use an object instance and a defined method name.
x = .accumulator~new(1) -- new accumulator with initial value of "1"
x~call(5)
x~call(2.3)
say "Accumulator value is now" x -- displays current value
-- an accumulator class instance can be instantiated and
-- used to sum up a series of numbers
::class accumulator
::method init -- instance initializer...sets the accumulator initial value
expose sum
use strict arg sum = 0 -- sets default sum value if not specified
-- perform the accumulator function
::method call
expose sum
use strict arg n
sum += n -- bump the accumulator
return sum -- return the new value
-- extra credit...display the current accumulator value
::method string
expose sum
return sum
OxygenBasic
Class AccumFactory '================= double v method constructor() end method method destructor() end method method Accum(double n) as AccumFactory new AccumFactory af af.v=v+n return af end method method FloatValue() as double return v end method method IntValue() as sys return v end method method StringValue(sys dp=16) as string return str v,dp end method end class '======================= 'TESTS (all results: PI) '======================= new AccumFactory af 'GENERATE ACCUMULATORS let a=af.Accum(1) 'integer let b=a.Accum(pi) 'float let c=b.Accum("-1") 'string 'STRING OUTPUT print c.StringValue(4) ' show 4 decimal places 'FLOAT OUTPUT print c.FloatValue 'USE FUNCTIONS IN EXPRESSION print 10 * c.FloatValue() / ( 10 * a.IntValue() ) 'FINISH del af : del a : del b : del c
Oz
A bit unwieldy because the '+' operator does not allow mixed type operands. The implementation is thread-safe (atomic Exchange operation).
declare
fun {Acc Init}
State = {NewCell Init}
in
fun {$ X}
OldState
in
{Exchange State OldState} = {Sum OldState X}
end
end
fun {Sum A B}
if {All [A B] Int.is} then A+B
else {ToFloat A}+{ToFloat B}
end
end
fun {ToFloat X}
if {Float.is X} then X
elseif {Int.is X} then {Int.toFloat X}
end
end
X = {Acc 1}
in
{X 5 _}
{Acc 3 _}
{Show {X 2.3}}
PARI/GP
stack = List([1]);
factory(b,c=0) = my(a=stack[1]++);listput(stack,c);(b)->stack[a]+=b;
foo(f) = factory(0, f); \\ initialize the factory
Run the factory:
gp > x = foo(1); gp > x(5); gp > y = foo(3); gp > print(x(2.3)); 8.300000000000 gp > print(y(1)); 4 gp > print(x(1)); 9.300000000000 gp > print(y(1/3)); 13/3
PascalABC.NET
function foo(n: real): real -> real :=
i -> begin
n += i;
Result := n;
end;
begin
var x := foo(1);
x(5);
foo(3);
print(x(2.3));
end.
- Output:
8.3
Perl
There's a little deviation: the syntax $x->(5)
differs from the usual x(5)
.
sub accumulator {
my $sum = shift;
sub { $sum += shift }
}
my $x = accumulator(1);
$x->(5);
accumulator(3);
print $x->(2.3), "\n";
- Output:
8.3
Phix
Emulated. There is nothing clever about this - both the answer and the task requirements!
Numeric polymorphism is inherently supported in phix. While technically this does not return
a function, the following demonstrates how the "standard_function" can be invoked in exactly
the same manner as a result from the factory, without the caller knowing which is which, and
I would guess that is one of the more important motivations for the original task. But it is
worth stating there are much easier ways to do this, hence generally speaking this approach
is not particularly recommended or advocated.
A variation on Closures/Value_capture#Phix, only in this case the inner function is kept in the returned variable and for simplicity there are no partial args - but it would be easy enough to add that sort of flexibility here if needed.
Rule#5 is deliberately ignored: if rogue code can corrupt the accumulators variable, it can just as easily corrupt the "closure" it would otherwise be held in, however well-hidden some other programming language would like to pretend that is, and of course the latter sort of corruption would be significantly harder to debug. Obviously, for safety, you would normally make the accumulators variable private(/non-global) in a separate source file, along with accumulate/accumulate_factory/call_function, and if you really don't like accumulators being visible (??) I suppose you could always just allocate a bit of memory in accumulator_factory() and return a pointer to that instead of an id/length.
sequence accumulators = {} function accumulate(integer id, atom v) accumulators[id] += v return accumulators[id] end function constant r_accumulate = routine_id("accumulate") function accumulator_factory(atom initv=0) accumulators = append(accumulators,initv) return {r_accumulate,length(accumulators)} end function function call_function(object rid, object args) if sequence(rid) then {rid, integer id} = rid args = id&args end if return call_func(rid,args) end function function standard_function() return "standard function" end function constant r_standard_function = routine_id("standard_function") constant x = accumulator_factory(1), y = accumulator_factory(3) {} = call_function(x,5) {} = call_function(y,3) ?call_function(x,2.3) ?call_function(y,4) ?call_function(r_standard_function,{})
- Output:
8.3 10 "standard function"
PHP
<?php
function accumulator($start){
return create_function('$x','static $v='.$start.';return $v+=$x;');
}
$acc = accumulator(5);
echo $acc(5), "\n"; //prints 10
echo $acc(10), "\n"; //prints 20
?>
<?php
function accumulator($sum){
return function ($x) use (&$sum) { return $sum += $x; };
}
$acc = accumulator(5);
echo $acc(5), "\n"; //prints 10
echo $acc(10), "\n"; //prints 20
?>
PicoLisp
(de accumulator (Sum)
(curry (Sum) (N)
(inc 'Sum N) ) )
(def 'a (accumulator 7))
(a 1) # Output: -> 8
(a 2) # Output: -> 10
(a -5) # Output: -> 5
Pony
use "assert"
class Accumulator
var value:(I64|F64)
new create(v:(I64|F64))=>
value=v
fun ref apply(v:(I64|F64)=I64(0)):(I64|F64)=>
value=match value
| let x:I64=>match v
| let y:I64=>x+y
| let y:F64=>x.f64()+y
end
| let x:F64=>match v
| let y:I64=>x+y.f64()
| let y:F64=>x+y
end
end
value
actor Main
new create(env:Env)=>
var r:Accumulator=Accumulator(I64(0))
r(I64(5))
r(I64(2))
try
Fact(match r()
|let x:I64=>x==7
|let y:F64=>y==7.0
end)?
env.out.print("The value I have so far is " + r().string())
else
env.out.print("An error of some sort happened!")
end
r(F64(5.5))
env.out.print("This is okay..." + r().string())
PostScript
/mk-acc { % accumulator generator
{0 add 0 0 2 index put}
7 array copy
dup 0 4 -1 roll put
dup dup 2 exch put
cvx
} def
% Examples (= is a printing command in PostScript):
/a 1 mk-acc def % create accumulator #1, name it a
5 a = % add 5 to 1, print it
10 mk-acc % create accumulator #2, leave it anonymous on the stack
2.71 a = % add 2.71 to 6, print it
dup 3.14 exch exec = % add 3.14 to 10, print it
dup 100 exch exec = % add 100 to 13.14, print it
12 a = % add 12 to 8.71, print it
% accumulator #2 is still available on the stack
PowerShell
Wikipedia says, “In programming languages, a closure is a function or reference to a function together with a referencing environment. A closure—unlike a plain function pointer—allows a function to access those non-local variables even when invoked outside its immediate lexical scope.”
The GetNewClosure method returns a ScriptBlock with captured variables.
function Get-Accumulator ([double]$Start)
{
{param([double]$Plus) return $script:Start += $Plus}.GetNewClosure()
}
$total = Get-Accumulator -Start 1
& $total -Plus 5.0 | Out-Null
& $total -Plus 2.3
- Output:
8.3
Prolog
Uses the module lambda written by Ulrich Neumerkel.
:- use_module(library(lambda)).
define_g(N, G) :-
put_attr(V, user, N),
G = V +\X^Y^(get_attr(V, user, N1),
Y is X + N1,
put_attr(V, user, Y)).
accumulator :-
define_g(1, G),
format('Code of g : ~w~n', [G]),
call(G, 5, S),
writeln(S),
call(G, 2.3, R1),
writeln(R1).
- Output:
8 ?- accumulator. Code of g : _G275+\_G285^_G288^ (get_attr(_G275,user,_G296),_G288 is _G285+_G296,put_attr(_G275,user,_G288)) 6 8.3 true.
Python
>>> def accumulator(sum):
def f(n):
f.sum += n
return f.sum
f.sum = sum
return f
>>> x = accumulator(1)
>>> x(5)
6
>>> x(2.3)
8.3000000000000007
>>> x = accumulator(1)
>>> x(5)
6
>>> x(2.3)
8.3000000000000007
>>> x2 = accumulator(3)
>>> x2(5)
8
>>> x2(3.3)
11.300000000000001
>>> x(0)
8.3000000000000007
>>> x2(0)
11.300000000000001
def accumulator(sum):
def f(n):
nonlocal sum
sum += n
return sum
return f
x = accumulator(1)
x(5)
print(accumulator(3))
print(x(2.3))
- Output:
<function f at 0xb7c2d0ac> 8.3
def accumulator(sum):
while True:
sum += yield sum
x = accumulator(1)
x.send(None)
x.send(5)
print(accumulator(3))
print(x.send(2.3))
- Output:
<generator object accumulator at 0x106555e60> 8.3
Quackery
Dynamic, Lambda
Quackery is untyped. This solution works with bignums. factory
returns a lambda function. (In Quackery terminology, it leaves a nest on the stack.) Nests on the stack are performed (i.e. executed or evaluated) with do
.
[ tuck tally share ]this[ swap ] is accumulate ( n s --> [ n )
[ [ stack ] copy tuck put nested
' accumulate nested join ] is factory ( n --> [ )
- Output:
Let's see this in action in the Quackery shell.
/O> 23 factory ... Stack: [ [ stack 23 ] accumulate ]
23 factory
has returned an accumulator function initialised to 23
.
Now let's put 100
underneath it using swap
, perform the accumulator using do
and then print the top of stack using echo
.
/O> 100 swap do echo ... 123 Stack: [ [ stack 123 ] accumulate ]
The running total has been printed, and the updated accumulator function has remained on the stack. (Everything in Quackery is immutable except for ancillary stacks (created with [ stack ]
), which Quackery has instead of variables. It is rare to embed an ancillary stack in a nest, but this is a good use case.)
Now let's create a second accumulator function with factory
and confirm that the two accumulator functions behave independently of one another by do
-ing first one of them, then the other.
/O> 10 6 ** factory ... Stack: [ [ stack 123 ] accumulate ] [ [ stack 1000000 ] accumulate ] /O> 234567 swap do echo ... 1234567 Stack: [ [ stack 123 ] accumulate ] [ [ stack 1234567 ] accumulate ] /O> swap ... Stack: [ [ stack 1234567 ] accumulate ] [ [ stack 123 ] accumulate ] /O> 123 swap do echo ... 246 Stack: [ [ stack 1234567 ] accumulate ] [ [ stack 246 ] accumulate ]
And since we've finished testing, we should tidy up after ourselves.
/O> empty ... Stack empty.
Static, Named
We can create a named version by extending the Quackery compiler, build
.
This version does not need to leave a lambda function on the stack, as it can be referred to by name.
In accordance with The Building Regulations, it starts with some sanity checks to enable the compiler to fail gracefully. For details see The Book of Quackery.
[ dip
[ -1 split dup [] = if
[ $ "accumulator needs a starting value."
message put bail ]
do dup number? not if
[ $ "accumulator needs a number."
message put bail ]
[ stack ] copy
tuck put nested
' [ tuck tally share ]
join nested join ] ] builds accumulator ( [ $ --> [ $ )
- Output:
First we will check that it complies with The Building Regulations, then we will create two accumulators, foo
and bar
and use them alternately to confirm they do not affect each other.
/O> accumulator is foobar ... accumulator needs a starting value. Stack empty. /O> $ "this is a string" accumulator is foobar ... accumulator needs a number. Stack empty. /O> 23 accumulator is foo ... [ 10 6 ** ] accumulator is bar ... Stack empty. /O> 100 foo echo ... 123 Stack empty. /O> 234567 bar echo ... 1234567 Stack empty. /O> 123 foo echo ... 246 Stack empty.
R
accumulatorFactory <- function(init) {
currentSum <- init
function(add) {
currentSum <<- currentSum + add
currentSum
}
}
- Output:
> f <- accumulatorFactory(1) > f(5) [1] 6 > f(2.3) [1] 8.3
Racket
#lang racket
(define ((accumulator n) i)
(set! n (+ n i))
n)
Raku
(formerly Perl 6)
sub accum ($n is copy) { sub { $n += $^x } }
#Example use:
my $a = accum 5;
$a(4.5);
say $a(.5); # Prints "10".
# You can also use the "&" sigil to create a function that behaves syntactically
# like any other function (i.e. no sigil nor parentheses needed to call it):
my &b = accum 5;
say b 3; # Prints "8".
REBOL
make-acc-gen: func [start-val] [
use [state] [
state: start-val
func [value] [
state: state + value
]
]
]
- Output:
>> x: make-acc-gen 1 >> x 5 == 6 >> make-acc-gen 3 >> print x 2.3 8.3
Retro
Retro only supports integers.
:acc (ns-)
d:create , [ [ fetch ] [ v:inc ] bi ] does ;
- Output:
#10 'foo acc foo foo foo dump-stack 10 11 12 Ok
REXX
This REXX program is partially modeled after the ooRexx example.
This example will handle any kind of number: integer, floating point.
/*REXX program shows one method an accumulator factory could be implemented. */
x=.accumulator(1) /*initialize accumulator with a 1 value*/
x=call(5)
x=call(2.3)
say ' X value is now' x /*displays the current value of X. */
say 'Accumulator value is now' sum /*displays the current value of accum.*/
exit /*stick a fork in it, we're all done. */
/*──────────────────────────────────────────────────────────────────────────────────────*/
.accumulator: procedure expose sum; if symbol('SUM')=="LIT" then sum=0 /*1st time?*/
sum=sum + arg(1) /*add──►sum*/
return sum
/*──────────────────────────────────────────────────────────────────────────────────────*/
call: procedure expose sum; sum=sum+arg(1); return sum /*add arg1 ──► sum.*/
output
X value is now 8.3 Accumulator value is now 8.3
Ring
oGenerator = new Generator
Func main
oGenerator {
accumulator = generator(1)
see call accumulator(5)
see nl
generator(3)
see call accumulator(2.3)
}
Class Generator
aN = []
func generator i
aN + i
return eval(substr("return func d {
oGenerator {
aN[#id#] += d
return aN[#id#]
}
}","#id#",string(len(aN))))
- Output:
6 8.30
RPL
This implementation complies with all the rules except maybe the last one ("Doesn't store the accumulated value or the returned functions in a way that could cause them to be inadvertently modified by other code"). The accumulated value is actually stored in a global variable, but as its name is generated with the system time, the likelihood of another code guessing it is very low - unless that code deliberately intends to do so.
RPL code | Comment |
---|---|
≪
"M" TIME →STR + SWAP
OVER OBJ→ STO
"≪ '" SWAP + "' STO+ SWAP DROP RCL ≫" + OBJ→
≫ ‘FOO’ STO
|
FOO ( n → ≪ accumulator ≫ )
create a global variable with a timestamp name
initialize variable with n
create lambda function
.
|
Let's check it works:
Command line | Test example |
---|---|
1 FOO 'X' STO 5 X DROP 3 FOO DROP 2.3 X |
x = foo(1); // X contains ≪ 'M17.3741285888' STO+ LASTARG SWAP DROP RCL ≫
x(5);
foo(3);
print x(2.3);
|
- Output:
1: 8.3
Ruby
Ruby deviates from the task because methods and Proc objects have different syntax. So, x = accumulator(1) is valid, but x(5) is an error: the syntax must be x.call(5) or x[5] (with square brackets). Ruby 1.9 also allows x.(5) (with an extra dot).
def accumulator(sum)
lambda {|n| sum += n}
end
# mixing Integer and Float
x = accumulator(1)
x.call(5)
accumulator(3)
puts x.call(2.3) # prints 8.3
The output of p accumulator(3) looks like
#<Proc:0x0000000207ba7f30@/tmp/accumulator.rb:2> # Ruby 1.8.6 #<Proc:0x000002060d1788@/tmp/accumulator.rb:2 (lambda)> # Ruby 1.9.2
This accumulator also works with other types that have a + method.
require 'rational'
require 'complex'
y = accumulator(Rational(2, 3))
puts y[Rational(1, 2)] # 7/6
puts y[4] # 31/6
puts y[Complex(0, 1)] # 31/6+1i
t = accumulator(Time.utc(1999, 8, 7, 6, 5))
# (Ruby 1.8.6) (Ruby 1.9.2)
puts t[4] # Sat Aug 07 06:05:04 UTC 1999 1999-08-07 06:05:04 UTC
puts t[-12 * 60 * 60] # Fri Aug 06 18:05:04 UTC 1999 1999-08-06 18:05:04 UTC
require 'matrix'
m = accumulator(Matrix[[1, 2], [3, 4]])
puts m[Matrix[[5, 6], [7, 8]]] # Matrix[[6, 8], [10, 12]]
If we define x as a method of self, then the syntax x(5)
works, but we deviate more from the task, because x might get "inadvertently modified" by other methods of self.
def accumulator(sum)
lambda {|n| sum += n}
end
class << self
define_method :x, &accumulator(1)
end
x(5)
accumulator(3)
puts x(2.3) # prints 8.3
Rust
This solution is explicitly rejected by the task description. It must be possible to create the accumulator with one type (e.g. int), then accumulate another type (e.g. float) correctly.
Changing "x = foo(1.)" to "x = foo(1)" in the code below should not change the output (it does).
// rustc 1.26.0 or later
use std::ops::Add;
fn foo<Num>(n: Num) -> impl FnMut(Num) -> Num
where Num: Add<Output=Num> + Copy + 'static {
let mut acc = n;
move |i: Num| {
acc = acc + i;
acc
}
}
fn main() {
let mut x = foo(1.);
x(5.);
foo(3.);
println!("{}", x(2.3));
}
- Output:
8.3
Over-engineered Solution
This solution uses a custom number type that can be either an i64 or f64. It also creates a generic struct that is callable using the unstable fn traits, which can be called to add anything that can be added to it's accumulator value.
// Accumulator
#![feature(unboxed_closures, fn_traits)]
pub struct Accumulator<T> {
value: T,
}
impl<T> Accumulator<T> {
pub fn new(value: T) -> Self {
Self { value }
}
}
impl<T, N> FnOnce<(N,)> for Accumulator<T>
where
T: std::ops::AddAssign<N> + Clone,
{
type Output = T;
extern "rust-call" fn call_once(mut self, (n,): (N,)) -> T {
self.value += n;
self.value
}
}
impl<T, N> FnMut<(N,)> for Accumulator<T>
where
T: std::ops::AddAssign<N> + Clone,
{
extern "rust-call" fn call_mut(&mut self, (n,): (N,)) -> T {
self.value += n;
self.value.clone()
}
}
// Number
#[derive(Copy, Clone, Debug)]
pub enum Number {
Int(i64),
Float(f64),
}
impl From<i64> for Number {
fn from(int: i64) -> Number {
Number::Int(int)
}
}
impl From<f64> for Number {
fn from(float: f64) -> Number {
Number::Float(float)
}
}
impl std::ops::AddAssign<i64> for Number {
fn add_assign(&mut self, n: i64) {
match self {
Number::Int(s) => *s += n,
Number::Float(s) => *s += n as f64,
}
}
}
impl std::ops::AddAssign<f64> for Number {
fn add_assign(&mut self, n: f64) {
*self = match *self {
Number::Int(s) => Number::Float(s as f64 + n),
Number::Float(s) => Number::Float(s + n),
}
}
}
impl std::fmt::Display for Number {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
match self {
Number::Int(x) => write!(f, "{}", x),
Number::Float(x) => write!(f, "{}", x),
}
}
}
// Demonstration
fn foo(n: impl Into<Number>) -> Accumulator<Number> {
Accumulator::new(n.into())
}
fn main() {
let mut x = foo(1);
x(5);
foo(3);
println!("{}", x(2.3));
let mut s = Accumulator::new(String::from("rosetta"));
s(" ");
println!("{}", s("code"));
}
- Output:
8.3 rosetta code
Scala
The type of a function can't change in Scala, and there is no "numeric" type that is a supertype of all such types. So, if the accumulator is declared as integer, it can only receive and return integers, and so on.
def AccumulatorFactory[N](n: N)(implicit num: Numeric[N]) = {
import num._
var acc = n
(inc: N) => {
acc = acc + inc
acc
}
}
- Sample:
scala> val x = AccumulatorFactory(1.0) x: (Double) => Double = <function1> scala> x(5.0) res7: Double = 6.0 scala> AccumulatorFactory(3.0) res8: (Double) => Double = <function1> scala> println(x(2.3)) 8.3
Scheme
(define (accumulator sum)
(lambda (n)
(set! sum (+ sum n))
sum))
;; or:
(define ((accumulator sum) n)
(set! sum (+ sum n))
sum)
(define x (accumulator 1))
(x 5)
(display (accumulator 3)) (newline)
(display (x 2.3)) (newline)
- Output:
#<procedure> 8.3
Sidef
class Accumulator(sum) {
method add(num) {
sum += num;
}
}
var x = Accumulator(1);
x.add(5);
Accumulator(3);
say x.add(2.3); # prints: 8.3
The same thing can be achieved by returning a closure from the Accumulator function.
func Accumulator(sum) {
func(num) { sum += num };
}
var x = Accumulator(1);
x(5);
Accumulator(3);
say x(2.3); # prints: 8.3
Simula
BEGIN
! ABSTRACTION FOR SIMULA'S TWO NUMERIC TYPES ;
CLASS NUMBER;
VIRTUAL:
PROCEDURE OUT IS PROCEDURE OUT;;
BEGIN
END NUMBER;
NUMBER CLASS INTEGERNUMBER(INTVAL); INTEGER INTVAL;
BEGIN
PROCEDURE OUT; OUTINT(INTVAL, 10);
END INTEGERNUMBER;
NUMBER CLASS REALNUMBER(REALVAL); REAL REALVAL;
BEGIN
PROCEDURE OUT; OUTFIX(REALVAL, 4, 10);
END REALNUMBER;
! SIMULA CANNOT RETURN FUNCTIONS - SIMULATE FUNCTIONS WITH CLASSES ;
CLASS ACCUMULATOR(ACC); REF(NUMBER) ACC;
BEGIN
PROCEDURE SWITCHTOREAL(Y); REAL Y;
BEGIN
REF(REALNUMBER) NEWACC;
NEWACC :- NEW REALNUMBER(ACC QUA INTEGERNUMBER.INTVAL);
NEWACC.REALVAL:= NEWACC.REALVAL + Y;
ACC :- NEWACC;
END SWITCHTOREAL;
REF(NUMBER) PROCEDURE ACCUMULATE(OTHERNUM); REF(NUMBER) OTHERNUM;
BEGIN
INSPECT ACC
WHEN INTEGERNUMBER DO
BEGIN
INSPECT OTHERNUM
WHEN INTEGERNUMBER DO
ACC QUA INTEGERNUMBER.INTVAL
:= ACC QUA INTEGERNUMBER.INTVAL + INTVAL
WHEN REALNUMBER DO
SWITCHTOREAL(REALVAL)
END
WHEN REALNUMBER DO
BEGIN
INSPECT OTHERNUM
WHEN INTEGERNUMBER DO
ACC QUA REALNUMBER.REALVAL
:= ACC QUA REALNUMBER.REALVAL + INTVAL
WHEN REALNUMBER DO
ACC QUA REALNUMBER.REALVAL
:= ACC QUA REALNUMBER.REALVAL + REALVAL
END;
ACCUMULATE :- ACC;
END ACCUMULATE;
PROCEDURE OUT; ACC.OUT;
END FOO;
REF(ACCUMULATOR) FOO;
FOO :- NEW ACCUMULATOR(NEW INTEGERNUMBER(1)); FOO.OUT; OUTIMAGE;
FOO.ACCUMULATE(NEW INTEGERNUMBER(5)); FOO.OUT; OUTIMAGE;
NEW ACCUMULATOR(NEW INTEGERNUMBER(3));
FOO.ACCUMULATE(NEW REALNUMBER(2.3)); FOO.OUT; OUTIMAGE;
END.
- Output:
1 6 8.3000
Smalltalk
Object subclass: AccumulatorFactory [
AccumulatorFactory class >> new: aNumber [
|r sum|
sum := aNumber.
r := [ :a |
sum := sum + a.
sum
].
^r
]
]
|x y|
x := AccumulatorFactory new: 1.
x value: 5.
y := AccumulatorFactory new: 3.
(x value: 2.3) displayNl.
"x inspect."
"de-comment the previous line to show that x is a block closure"
the above can also be done without a class to hold the block, simply by putting it into another block (i.e. an outer closure for the sum, returning an inner function which updates that sum):
|factory accu1 accu2|
factory := [:initial |
[
|sum|
sum := initial.
[:addend | sum := sum + addend].
] value.
].
accu1 := factory value:1.
accu1 value:5.
accu2 := factory value:10.
accu2 value:5.
(accu1 value:2.3) printCR. "-> 8.3 (a float)"
(accu2 value:0) printCR. "-> 15 (an integer)"
(accu2 value:22 factorial) printCR. "-> a large integer"
- Output:
8.3 15 1124000727777607680015
Standard ML
Deviations: An accumulator instance can take either integers or reals, but not both mixed (due to lack of runtime polymorphism).
fun accumulator (sum0:real) : real -> real = let
val sum = ref sum0
in
fn n => (
sum := !sum + n;
!sum)
end;
let
val x = accumulator 1.0
val _ = x 5.0
val _ = accumulator 3.0
in
print (Real.toString (x 2.3) ^ "\n")
end;
- Output:
8.3
Swift
func makeAccumulator(var sum: Double) -> Double -> Double {
return {
sum += $0
return sum
}
}
let x = makeAccumulator(1)
x(5)
let _ = makeAccumulator(3)
println(x(2.3))
- Output:
8.3
Tcl
This uses nested coroutines to manage the state, which for the outer coroutine is a counter used to generate unique instances of the inner coroutine, and for the inner coroutine it is the actual accumulator variable. Note that Tcl commands (including coroutines) are never nameless, but it is trivial to synthesize a name for them. It's possible to guarantee uniqueness of names, but just using a simple sequence generator gets 90% of the effect for 10% of the effort.
package require Tcl 8.6
# make the creation of coroutines without procedures simpler
proc coro {name arguments body args} {
coroutine $name apply [list $arguments $body] {*}$args
}
# Wrap the feeding of values in and out of a generator
proc coloop {var body} {
set val [info coroutine]
upvar 1 $var v
while 1 {
set v [yield $val]
if {$v eq "stop"} break
set val [uplevel 1 $body]
}
}
# The outer coroutine is the accumulator factory
# The inner coroutine is the particular accumulator
coro accumulator {} {
coloop n {
coro accumulator.[incr counter] n {
coloop i {
set n [expr {$n + $i}]
}
} $n
}
}
Sample usage (extra characters over Paul's example to show more clearly what is going on):
% set x [accumulator 1]
::accumulator.1
% $x 5
6
% accumulator 3
::accumulator.2
% puts ">>[$x 2.3]<<"
>>8.3<<
TXR
Verbose
(defun accumulate (sum)
(lambda (n)
(inc sum n)))
;; test
(for ((f (accumulate 0)) num)
((set num (iread : : nil)))
((format t "~s -> ~s\n" num [f num])))
(exit 0)
- Run:
$ txr accumulator-factory.tl 1 1 -> 1 2 2 -> 3 3 3 -> 6 400000000000000000000000000000000000000000000000000000000000000000000000 400000000000000000000000000000000000000000000000000000000000000000000000 -> 400000000000000000000000000000000000000000000000000000000000000000000006 5.3 5.3 -> 4e71 1e71 1e71 -> 5e71 [Ctrl-D][Enter] $
Sugared
(let ((f (let ((sum 0)) (do inc sum @1))))
(mapdo (do put-line `@1 -> @[f @1]`) (gun (iread : : nil))))
- Output:
$ echo "1 2 3 4.5" | txr accumulator-factory2.tl 1 -> 1 2 -> 3 3 -> 6 4.5 -> 10.5
Yield-based
Using the obtain
/yield
interface to delimited continuations, we can turn an imperative for loop into an accumulation function:
(defun accum ()
(for ((sum (yield-from accum)))
()
((inc sum (yield-from accum sum)))))
(let ((f (obtain (accum))))
(mapdo (do put-line `@1 -> @[f @1]`) (gun (iread : : nil))))
- Output:
$ echo "1 2 3 4.5" | txr accumulator-factory2.tl 1 -> 1 2 -> 3 3 -> 6 4.5 -> 10.5
OOP-based
OOP languages can use objects to simulate closures. In particular, function-objects which can be called as if they were functions, without any visible method being referenced. TXR Lisp supports functors as an expression of irony in language design. A structure object for which a method named lambda
is defined can be used as function. Arguments applied to the objects are applied to lambda, preceded by the object itself as the leftmost argument:
(defstruct (accum count) nil
(count 0))
(defmeth accum lambda (self delta)
(inc self.count delta))
;; Identical test code to Yield-Based and Sugared, except for
;; the construction of the function object bound to variable f.
(let ((f (new (accum 0))))
(mapdo (do put-line `@1 -> @[f @1]`) (gun (iread : : nil))))
Unicon
Strictly speaking, genAcc(n) returns a co-expression, not a function. However, the invocation syntax here is indistinguishable from calling a function.
procedure main()
a := genAcc(3)
b := genAcc(5)
write(" " ,center("a",5), " ", center("b", 5))
write("genAcc: ", right(a(4),5), " ", right(b(4), 5))
write("genAcc: ", right(a(2),5), " ", right(b(3),5))
write("genAcc: ", right(a(4.5),5)," ", right(b(1.3),5))
end
procedure genAcc(n) # The generator factory
return makeProc { while i := (n@&source)[1] do n +:= i }
end
procedure makeProc(A) # A Programmer-Defined Control Operation
return (@A[1],A[1])
end
Note: The co-expression calling sequence used is Unicon specific.
- Output:
a b genAcc: 7 9 genAcc: 9 12 genAcc: 13.5 13.3
UNIX Shell
Deviation from task: The accumulator factory returns a global function, which stores the sum in a global variable. Other code can modify the function or the variable, perhaps by accident.
The shell is a bad choice for this task. This example plays tricks with eval. The difficulty with eval is to put the quotation marks " and dollar signs $ in the correct place, and escape them with the correct number of backslashes \. One missing (or one extra) backslash can ruin the entire program.
#!/bin/sh
accumulator() {
# Define a global function named $1
# with a global variable named ${1}_sum.
eval "${1}_sum=\$2"
eval "$1() {
${1}_sum=\$(echo \"(\$${1}_sum) + (\$2)\" | bc)
eval \"\$1=\\\$${1}_sum\" # Provide the current sum.
}"
}
accumulator x 1
x r 5
accumulator y 3
x r 2.3
echo $r
y r -3000
echo $r
- Output:
$ sh accumulator.sh 8.3 -2997
es
A better shell for this task is es, because it has lexical variables and closures. @ i {code}
is a lambda with parameter i, and fn accumulator n {code}
is sugar for fn-accumulator = @ n {code}
.
fn accumulator n {
result @ i {
n = `{echo $n + $i | bc}
result $n
}
}
fn-x = <={accumulator 1}
x 5
fn-y = <={accumulator 3}
echo <={x 2.3}
echo <={y -3000}
Ursalang
Ursalang has only a single number type.
let fac = fn(n) {
fn(i) {
n := n + i
}
}
let x = fac(1)
x(5)
fac(3)
print(x(2.3))
VBScript
I'm not entirely convinced that this is actually doing what is asked. A VBScript guru I'm not. The answer's right, though.
- Implementation
class accumulator
dim A
public default function acc(x)
A = A + x
acc = A
end function
public property get accum
accum = A
end property
end class
- Invocation
dim a
set a = new accumulator
x = a( 1 )
a 5
dim b
set b = new accumulator
b 3
wscript.echo a(2.3)
- Output:
8.3
Wart
def (accumulator n)
(fn() ++n)
Example usage:
a <- (accumulator 3) (a) => 4 (a) => 5 b <- (accumulator 23) (b) => 24 (a) => 6
Wren
var accumulator = Fn.new { |acc| Fn.new { |n| acc = acc + n } }
var x = accumulator.call(1)
x.call(5)
accumulator.call(3)
System.print(x.call(2.3))
- Output:
8.3
x86 Assembly
32 bit
The accumulator function that is returned uses a section of its instruction space to store the accumulated sum. This way it works without having to keep track of any addresses external to the function itself. The only deviation from the spec is that this only works with integer values. With some extra work, floating point numbers can be incorporated, but outputting would be trickier.
; Accumulator factory
; Returns a function that returns the sum of all numbers ever passed in
; Build:
; nasm -felf32 af.asm
; ld -m elf32_i386 af.o -o af
global _start
section .text
_start:
mov eax, 0x2D ; sys_brk(unsigned long brk)
xor ebx, ebx ; Returns current break on an error
int 0x80 ; syscall
push eax ; Save the initial program break
push 2 ; Get an accumulator initialized to 2
call factory
mov [acc1], eax ; Save the pointer in acc1
push 5 ; Get an accumulator initialized to 5
call factory
mov [acc2], eax ; Save the pointer in acc2
push 4 ; Call acc1 with 4
lea eax, [acc1]
call [eax]
push 4 ; Call acc2 with 4
lea eax, [acc2]
call [eax]
push -9 ; Call acc1 with -9
lea eax, [acc1]
call [eax]
push 13 ; Call acc1 with 13
lea eax, [acc1]
call [eax]
push eax ; Print the number, should be 10
call print_num
push -5 ; Call acc2 with -5
lea eax, [acc2]
call [eax]
push eax ; Print the number, should be 4
call print_num
mov eax, 0x2D ; Reset the program break
pop ebx
int 0x80
mov eax, 0x01 ; sys_exit(int error)
xor ebx, ebx ; error = 0 (success)
int 0x80
; int (*function)(int) factory (int n)
; Returns a pointer to a function that returns the sum of all numbers passed
; in to it, including the initial parameter n;
factory:
push ebp ; Create stack frame
mov ebp, esp
push ebx
push edi
push esi
mov eax, 0x2D ; Allocate memory for the accumulator
xor ebx, ebx
int 0x80
push eax ; Save the current program break
mov ebx, .acc_end ; Calculate the new program break
sub ebx, .acc
push ebx ; Save the length
add ebx, eax
mov eax, 0x2D
int 0x80
pop ecx ; Copy the accumulator code into memory
pop eax ; Set the returned address
mov edi, eax
mov esi, .acc
rep movsb
lea edi, [eax + 10] ; Copy the parameter to initialize accumulator
lea esi, [ebp + 8]
movsd
pop esi ; Tear down stack frame
pop edi
pop ebx
mov esp, ebp
pop ebp
ret 4 ; Return and remove parameter from stack
.acc: ; Start of the returned accumulator
push ebp
mov ebp, esp
push edi
push esi
call .acc_skip ; Jumps over storage, pushing address to stack
dd 0 ; The accumulator storage (32 bits)
.acc_skip:
pop esi ; Retrieve the accumulator using address on stack
lodsd
add eax, [ebp + 8] ; Add the parameter
lea edi, [esi - 4]
stosd ; Save the new value
pop esi
pop edi
mov esp, ebp
pop ebp
ret 4
.acc_end: ; End of accumulator
; void print_num (int n)
; Prints a positive integer and a newline
print_num:
push ebp
mov ebp, esp
mov eax, [ebp + 8] ; Get the number
lea ecx, [output + 10] ; Put a newline at the end
mov BYTE [ecx], 0x0A
mov ebx, 10 ; Divisor
.loop:
dec ecx ; Move backwards in string
xor edx, edx
div ebx
add edx, 0x30 ; Store ASCII digit
mov [ecx], dl
cmp eax, 0 ; Loop until all digits removed
jnz .loop
mov eax, 0x04 ; sys_write(int fd, char *buf, int len)
mov ebx, 0x01 ; stdout
lea edx, [output + 11] ; Calulate length
sub edx, ecx
int 0x80
mov esp, ebp
pop ebp
ret 4
section .bss
acc1: ; Variable that stores the first accumulator
resd 1
acc2: ; Variable that stores the second accumulator
resd 1
output: ; Holds the output buffer
resb 11
Output
10 4
XLISP
There are probably other ways of doing it, but this is one way.
(defun accumulator (x)
(lambda (n)
(setq x (+ n x))
x ) )
Test it in a REPL:
[1] (define f (accumulator 1)) F [2] (define g (accumulator 3)) G [3] (f 5) 6 [4] (g 1.7) 4.7 [5] (f 9) 15
Yabasic
sub foo$(n)
local f$
f$ = "f" + str$(int(ran(1000000)))
compile("sub " + f$ + "(n): static acum : acum = acum + n : return acum : end sub")
execute(f$, n)
return f$
end sub
x$ = foo$(1)
execute(x$, 5)
foo$(3)
print execute(x$, 2.3)
Yorick
Yorick cannot dynamically create new functions. Instead, the accum function can be called in two ways: directly, in which case its first argument is numerical; or through a closure, where its first argument is implicitly an object and the second is the user-provided argument. This example uses closures and group objects, which require Yorick 2.2 or later.
func accum(data, n) {
if(!is_obj(data))
return closure(accum, save(total=data));
save, data, total=data.total + n;
return data.total;
}
Example of use (interactive session):
> x = accum(1) > x(5) 6 > y = accum(3) > x(2.3) 8.3 > y(2.3) 5.3
zkl
fcn foo(n){ fcn(n,acc){ acc.set(n+acc.value).value }.fp1(Ref(n)) }
A strong reference (Ref) is used as the accumulator, a Ref acts like a one element list. The Ref is bound to the new functions second parameter with the .fp1 method.
x:=foo(1) //--> partially applied function x(5) //-->6 (int) y:=foo(3) //-->new PFA x(2.3).println() 8.3 x(2) //-->10 (int) y(2) //-->5 (int)
The output switches between int and float based on the most recent input: With addition, the first operand casts the second: int + int|float --> int and float + int|float --> float. If the desire is to make the behavior "once float, always float", a 0 or 0.0 can be used to start the sum and stashed in a another bit of state.
- Programming Tasks
- Solutions by Programming Task
- MUMPS/Omit
- 11l
- 8th
- ABAP
- ActionScript
- Ada
- Aikido
- Aime
- ALGOL 68
- AppleScript
- Argile
- Astro
- BBC BASIC
- Bracmat
- Brat
- BQN
- C
- C sharp
- C++
- Ceylon
- Clay
- Clojure
- CoffeeScript
- Common Lisp
- Crystal
- D
- Dart
- Déjà Vu
- Delphi
- System.SysUtils
- System.Variants
- E
- EchoLisp
- Elena
- Elixir
- EMal
- Erlang
- ERRE
- F Sharp
- Factor
- Fantom
- Forth
- Fortran
- FreeBASIC
- Go
- Golo
- Groovy
- Haskell
- Icon
- Unicon
- Io
- J
- Java
- JavaScript
- Jsish
- Julia
- Kotlin
- Lambdatalk
- Lang
- LFE
- Lua
- M2000 Interpreter
- Maple
- Mathematica
- Wolfram Language
- Mercury
- MiniScript
- Nemerle
- NewLisp
- NGS
- Nim
- Nit
- Objeck
- Objective-C
- OCaml
- Octave
- Oforth
- OoRexx
- OxygenBasic
- Oz
- PARI/GP
- PascalABC.NET
- Perl
- Phix
- PHP
- PicoLisp
- Pony
- PostScript
- PowerShell
- Prolog
- Python
- Quackery
- R
- Racket
- Raku
- REBOL
- Retro
- REXX
- Ring
- RPL
- Ruby
- Rust
- Scala
- Scheme
- Sidef
- Simula
- Smalltalk
- Standard ML
- Swift
- Tcl
- TXR
- UNIX Shell
- Es
- Ursalang
- VBScript
- Wart
- Wren
- X86 Assembly
- XLISP
- Yabasic
- Yorick
- Zkl
- ARM Assembly/Omit
- Commodore BASIC/Omit
- C/Omit
- ML/I/Omit
- Scratch/Omit
- Pages with too many expensive parser function calls