Lazy functional programming with Husky

• Husky syntax
• Definitions
• Types
• Examples
• Functions
• Lambdas
• List comprehension
• Macros
• Special constructs
• Commands
• Gotchas
• Installation
• Author
• License

Husky is a lazy functional language similar to Haskell, but with a more conventional syntax. Husky implements a Hindley-Milner-style parametric polymorphic type system, higher-order functions, argument pattern matching for function specialization, currying, macros, list comprehension, and monads.

Husky uses NOR (normal order reduction), WHNF (weak head normal form), sharing (an important lazy evaluation optimization), and lazy constructors to create infinite "lazy" data structures.

For example, we can create an infinite list of integers, filter the even numbers, and only pick the first ten even numbers:

> take(10, filter(even, from(1))).
[2, 4, 6, 8, 10, 12, 14, 16, 18, 20] :: [num].

Function from(1) returns the infinite list [1, 2, 3, ...], even(n) returns true if n is even, and filter retains all even values in the list using even as a filter.

Husky is written in SWI-Prolog using Prolog unification for efficient beta normal-order reduction of the lambda A -> B applied to argument X:

apply(A -> B, X, B) :- !, (var(A) -> A = X; eval(X, V) -> A = V).

The apply operation takes a lambda A -> B to apply to the unevaluated X, then returns B when A is a variable. If A is not a variable, then pattern matching is used by evaluating X to V and unifying A with V.

Sharing is implemented internally by using an eval(X, V) functor that holds an (un)evaluated expression X and variable V. When X is evaluated to a value, variable V is set to that value, thereby sharing the value of V to all uses of V (which are of the form of eval(X, V)) in an expression.

This reduction method makes Husky run fast, asymptotically as fast as Haskell. Except that Husky does not optimize tail recursion, meaning that very deep recursive calls will eventually fail.

Type inference is performed by Prolog inference. Type rules are expressed as Prolog clauses for the Husky type inference operator ::. These clauses have a one-to-one correspondence to Post system rules that are often used to define polymorphic type systems with rules for type inference. For example, the polymorphic type of a lambda is defined as a Prolog rule:

(A -> B) :: Alpha -> Beta :- A :: Alpha, B :: Beta.

This rule infers the type of a lambda as Alpha -> Beta if argument A is of type Alpha and B is of type Beta.

The type rule for lambda application (:) uses Prolog unification with the "occurs check" to safely implement the Hindley-Milner-style parametric polymorphic type system of Husky:

F:A :: Beta :- !, A :: Alpha, F :: Gamma, unify_with_occurs_check(Gamma, (Alpha -> Beta))

It is that simple!

The Curry-Howard correspondence (the link between logic and computation) trivially follows by observing that the plain unoptimized beta reduction rule is the same as the type inference rule for application without the occurs check:

apply(F, A, B) :- eval(F, (A -> B)).
F:A :: Beta :- F :: (Alpha -> Beta), A :: Alpha.

In both rules A and its type Alpha are unified with the lambda argument to produce B and its type Beta.

Because Husky evaluation rules and type inference rules are defined in Prolog, the implementation is easy to understand and change, for example to add features, to change the language, and to experiment with type systems.

Husky syntax

Husky syntax follows the usual syntax of arithmetic expressions with functions. Functions are applied to one or more parenthesized arguments:

func(arg1, arg2, ...)

Values are Booleans true and false, numbers, strings (quoted with "), lists written with [ and ] brackets, and arbitrary data structures.

At the Husky prompt you can enter expressions which are evaluated and the answer is returned together with its inferred type. Global type declarations and function definitions entered at the command prompt are stored with the program. Input may span one or more lines and should end with a period (.).

> 1+2.
3 :: num

> x^2 where x := sin(1)-2.
1.3421894790419853 :: num

> true /\ false \/ true.
true :: bool

> "hello" // "world".
"helloworld" :: string

> [1,2,3] ++ [4,5].
[1, 2, 3, 4, 5] :: [num]

> (1+2, \ true).
(3, false) :: (num, bool)

> x->x+1.
$0->$0+1 :: num->num

> f where f(x) := x+1.
$0->$0+1 :: num->num

> map(f, [1,2,3]) where f(x) := x+1.
[2,3,4] :: [num]

> zip([1, 2, 3], [true, true, false]).
[(1, true), (2, true), (3, false)] :: [(num, bool)]

Type variables and formal arguments of functions and lambdas are internally anonymized as Prolog variables and displayed as $i where i is an integer.

Many functions and operators are predefined in Husky, either as built-ins or as prelude functions, see Functions.

In Husky, v->b is a lambda (abstraction) with argument v and body expression b. Applications of lambdas are written with a colon (:), for example (v->v+1):3 applies the lambda v->v+1 to the value 3 giving 3+1 which evaluates to 4.

When applying named functions the usual parenthesized form can be used, for example f(3) to apply f to the value 3. Also f:3 is permitted.

Currying of functions and operators is supported. For example, addition can be written in functional form +(x,y) and (+):x:y and even +(x):y. These are all different representations of the same expression. This makes Currying intuitive, for example +(1) is the increment function, *(2) doubles, and /(1) inverts (note the parameter placement order in this case!).

Lists in Husky are written with brackets, for example[] and [1,2]. The special form [x|xs] represents a list with head expression x and tail expression xs. The dot notation x.xs represents a list with head x and tail xs, where . is the list constructor function commonly referred to as "cons".

Definitions

A definition in Husky is written as LHS := RHS, where LHS is a name or a function name with parenthesized arguments and RHS is an expression.

For example:

one := 1.
inc(x) := x + 1.
hypotenuse(x, y) := sqrt(x^2 + y^2).

Definitions of functions can span more than one LHS := RHS pair in which the definitions should be separated with a semicolon (;). This allows you to define pattern arguments for function specialization:

fact(0) := 1;
fact(n) := n * fact(n - 1).

Multiple definitions of a function typically have constants and data structures as arguments for which the function has a specialized function body.

A wildcard formal argument _ is permitted and represents an unused argument. For example to define the const combinator function:

const(x, _) := x.

Types

Types are automatically inferred, like in Haskell. Types can also be explicitly associated with definitions using the :: operator. New named types and parametric types can be defined.

Husky has three built-in atomic types:

bool           Booleans
num            integers, rationals, and floats
string         strings

• the Boolean type has two values true and false;
• rationals are written with rdiv in the form numerator rdiv denominator;
• floats are written conventionally with +/- sign, a period, and exponent E;
• strings are sequences of characters enclosed in double quotes (");
• the nil constant denotes the so-called "bottom value" (an undefined value) and is polymorphic.

Husky has three built-in type constructors:

[alpha]          a list of elements of type alpha
(alpha, beta)    a tuple is the product of types alpha and beta
alpha -> beta    a function mapping elements of type alpha to type beta

New (parametric) data types and their constructors can be defined as follows:

MyType :: type.
Constructor1(arg1, arg2, ...) :: MyType(parm1, parm2, ...).
...
ConstructorN(arg1, arg2, ...) :: MyType(parm1, parm2, ...).

A type variable is a name, except bool, num, and string. Also symbols such as * and ** can be used as type variables. For example, a polymorphic list type can be written as [alpha] or [*] and the comparison operator = has type (=) :: * -> * -> bool.

Some examples of named types and parametric types:

one :: num.
one := 1.

inc :: num->num.
inc := +(1).

Day :: type.
Mon :: Day.
Tue :: Day.
Wed :: Day.
Thu :: Day.
Fri :: Day.
Sat :: Day.
Sun :: Day.
WeekDays :: [Day].
WeekDays := [Mon,Tue,Wed,Thu,Fri].
isWeekDay(Mon) := true;
isWeekDay(Tue) := true;
isWeekDay(Wed) := true;
isWeekDay(Thu) := true;
isWeekDay(Fri) := true;
isWeekDay(day) := false.
verify := all(map(isWeekDay, WeekDays)).

BinTree :: type.
Leaf :: BinTree(*).
Node(*, BinTree(*), BinTree(*)) :: BinTree(*).

Examples

> a := 2.
{ DEFINED a::num }

> f(x) := x+1.
{ DEFINED f::num->num }

> f(a).
3 :: num

% the same function f, using the colon syntax for application:
> f:x := (+):x:1.
{ DEFINED f::num->num }

% the same function f, using currying:
> f := +(1).
{ DEFINED f::num->num }

% function definitions are translated to lambdas,
% typing the function name returns its lambda expression:

> f(a) := a+1.
{ DEFINED f::num->num }
> f.
$0->$0+1 :: num->num

> plus(x,y) := x+y.
{ DEFINED plus::num->num->num }
> plus.
$0->$1->$0+$1 :: num->num->num

% a lambda can be specialized by multiple abstractions, each
% separated by a ';', e.g. the factorial function:
> fact(0) := 1;
  fact(n) := n*fact(n-1).
{ DEFINED fac::num->num }
> fact.
0->1;$0->$0*fac($0-1) :: num->num
> fact(6).
720 :: num

% constants can be defined to construct new data types

% declare a BinTree type:
> BinTree :: type.
{ DECLARED BinTree::type }

% define a Leaf constant:
> Leaf :: BinTree(*).
{ DECLARED Leaf::BinTree($0) }

% define a Node constructor:
> Node(*, BinTree(*), BinTree(*)) :: BinTree(*).
Node(num, BinTree(num), BinTree(num)) :: BinTree(num).

% try it out by building a small tree (see examples/btree.sky):
> Node(3, Node(1, Leaf, Leaf), Node(4, Leaf, Leaf)).
Node(3, Node(1, Leaf, Leaf), Node(4, Leaf, Leaf)) :: BinTree(num)

Functions

Built-in operators and functions are defined in prelude.sky:

- x         unary minus (can also use 'neg' for unary minus, e.g. for currying)
x + y       addition
x - y       subtraction
x * y       multiplication
x / y       division
x rdiv y    rational division with "infinite" precision
x div y     integer division
x mod y     modulo
x ^ y       power
min(x,y)    minimum
max(x,y)    maximum
x = y       equal (structural comparison)
x <> y      not equal (structural comparison)
x < y       less (num and string)
x <= y      less or equal (num and string)
x > y       greater (num and string)
x >= y      greater or equal (num and string)
\ x         logical not
x /\ y      logical and
x \/ y      logical or
gcd(x, y)   GCD
lcm(x, y)   LCM
fac(x)      factorial
fib(x)      Fibonacci
id(x)       identity (returns x)
abs(x)      absolute value
sin(x)      sine
cos(x)      cosine
tan(x)      tangent
asin(x)     arc sine
acos(x)     arc cosine
atan(x)     arc tangent
exp(x)      e^x
log(x)      natural logarithm (nil when x<0)
sqrt(x)     root (nil when x<0)
x // y      string concatenation
(x,y)       tuple (note: tuple constructors are strict, not lazy)
x.xs        a list with head x and tail xs (list constructors are lazy)
[x|xs]      same as above: a list with head x and tail xs
# xs        list length
xs ? n      element of list xs at index n >= 1
hd(xs)      head element of list xs (nil when x=[])
tl(xs)      tail element of list xs (nil when x=[])
init(xs)    init of xs without the last element
last(xs)    last element of xs
xs ++ ys    list concatenation
xs \\ ys    list subtraction
x is_in xs  list membership check (positive test)
x not_in xs list membership check (negative test)
fst((x,y))  first (=x)
snd((x,y))  second (=y)
even(x)     true when x is even
odd(x)      true when x is odd

num(s)      convert string s to number (nil when s is not numeric)
str(x)      convert number x to string
s2ascii(s)  convert string s to list of ASCII codes
ascii2s(xs) convert list xs of ASCII codes to string

read        prompts user and returns input string
write(x)    display x (returns x)
writeln(x)  display x '\n' (returns x)
getc        returns single input char in ASCII
putc(x)     display character of ASCII code x (returns x)

map         map(f, xs) maps f onto xs [f(x1), f(x2), ..., f(xk)]
filter      filter(p, xs) yields elements of xs that fulfill p
foldl       foldl(f, x, xs) = f(x, ... f(f(f(x1, x2), x3), x4), ...)
foldr       foldr(f, x, xs) = f(... f(xk-2, f(xk-1, f(xk, x))) ... )
foldl1      foldl1(f, xs) foldl over non-empty list
foldr1      foldr1(f, xs) foldr over non-empty list
scanl       scanl left scan
scanr       scanr right scan
all         all(p, xs) = true if all elements in xs fulfill p
any         any(p, xs) = true if any element in xs fulfills p
concat      concat list of lists
zip         zip(xs, ys) zips two lists
unzip       unzip(xys) unzips list of tuples
zipwith     zipwith(f, xs, ys) zips two lists using operator f
until       until(p, f, x) applies f to x until p(x) holds
iterate     iterate(f, x) = [x, f(x), f(f(x)), f(f(f(x))), ... ]
from        from(n) = [n, n+1, n+2, ...]
repeat      repeat(x) produces [x,x,x,...]
replicate   replicate(n, x) generates list of n x's
cycle       cycle(xs) generates cyclic list xs ++ xs ++ ...
drop        drop(n, xs) drops n elements from xs
take        take(n, xs) takes n elements from xs
dropwhile   dropwhile(p, xs) drops first elements fulfilling p from xs
takewhile   takewhile(p, xs) takes first elements fulfilling p from xs
reverse     reverse list
f @ g       function composition ("f after g")
twiddle     twiddle operands twiddle(f:x, y) means f(y, x)
flip        flip(f, x, y) flips operands to apply f(y, x)
curry       curry function curry(f, x, y) := f((x, y))
uncurry     uncurry function uncurry(f, (x, y)) := f(x, y)

Lambdas

A lamba (abstraction) is written as v->b. A lambda expression may have several lambda alternatives for function specialization. Alternatives are represented by lambdas separated by semicolons (;). Each lambda should specify a specialized argument for selection, such as constants and partial data structures.

(v->b)                  a lambda abstraction
(v->b; w->c)            a lambda with two specializations

Argument v and w may be a name, a constant, or a constructor with named parameters. The body of a lambda is just an expression.

Lambdas should be parenthesized when used in expressions, because the -> and ; operators have a low precedence.

Function application is performed with the conventional syntax of argument parameterization:

> f := (x->x+1).
{ DEFINED f::num->num }
> f(3).
4 :: num
> iszero := (0->true; n->false).
{ DEFINED iszero::num->bool }

However, applying a lambda directly to an argument requires the : application operator:

> (x->x+1):3.
4 :: num

Functions and functions with specializations are automatically translated to the corresponding lambda abstractions when stored with the program:

> f(x) := x+1.
{ DEFINED f::num->num }
> f.
($0->$0+1) :: (num->num)
> iszero(0) := true;
|: iszero(n) := false.
{ DEFINED iszero::num->bool }
> iszero.
(0->true;$0->false) :: (num->bool)

List comprehension

A list comprehension has the following form:

[ expr | qual; qual; ...; qual ]

where the qualifiers are generators of the form x <- xs or guards over the free variables. Guards over one or more generator variables must be placed to the right of the generator, i.e. variable scoping applies to the right.

For example:

[ x^2 | x <- 1..10; even(x) ].

[ (a,b) | a <- 1..3; b <- 1..a ].

[ (a,b) | a <- 1..5; odd(a); b <- 1..a; a mod b = 0 ].

Macros

Macros are defined with == to serve as shorthands for expressions and types:

StartValue   == 0.
Address      == num.
MachineState == Address->num.
NumTree      == BinTree(num).

Macro expansion is structural: names, functions, lists, and other syntactic structures can be defined as macros with parameters, i.e. macros support pattern matching for macro specialization. Parameter names in macros are unified (viz. Prolog unification) to structurally match macro definitions without evaluating the arguments.

For example:

> macro(x, x) == "equal".
> macro(x, y) == "not equal".
> macro(1, 1).
"equal" :: string
> macro(1, 2).
"not equal" :: string
> macro(1, 1+0).
"not equal" :: string

In fact, list comprehension syntax is entirely implemented by recursive macro expansion into concat, map, and filter (defined in prelude.sky):

[ f | x <- xs; p; y <- ys; r ] == concat(map(x -> [ f | y <- ys; r ], filter(x -> p, xs))).
[ f | x <- xs; p; y <- ys    ] == concat(map(x -> [ f | y <- ys    ], filter(x -> p, xs))).

[ f | x <- xs; y <- ys; r ] == concat(map(x -> [ f | y <- ys; r ], xs)).
[ f | x <- xs; y <- ys    ] == concat(map(x -> [ f | y <- ys    ], xs)).

[ f | x <- xs; p; q; r ] == [ f | x <- xs; p /\ q; r ].
[ f | x <- xs; p; q    ] == [ f | x <- xs; p /\ q    ].

[ x | x <- xs; p ] == filter(x -> p, xs).
[ f | x <- xs; p ] == map(x -> f, filter(x -> p, xs)).

[ x | x <- xs ] == xs.
[ f | x <- xs ] == map(x -> f, xs).

The monad do operator (defined in monads.sky) is a macro:

do(x := y; z) == do(z) where x := y.
do(x <- y; z) == y >>= (x -> do(z)).
do(z) == z.

Macros do not evaluate arguments! Only syntactical structures can be pattern matched. New syntax can be introduced by declaring prefix, infix, and postfix operators, see Commands.

Special constructs

if b then x else y      a conditional expression

x where y := z          each y in expression x is replaced by expression z

fix(f)                  the fixed-point Y-combinator (fix:f = f:(fix:f))

(let defs in val)       the 'let' construct (use parenthesis!)

case val of (const1 -> val1; const2 -> val2; ... constk -> valk)
                        the 'case' construct (translated to lambdas)

For example:

test := (let f := ^(2); start := 1; end := 10 in map(f, start..end)).

case n of (1 -> "one"; 2 -> "two"; 3 -> "three")

Commands

listing.                lists the contents of the loaded program(s)
bye.                    close session
reset.                  soft reset
trace.                  trace on
notrace.                trace off
profile.                collect execution profile stats
noprofile.              profile off
save.                   create standalone Husky executable
load "file".            load file
save "file".            save listing to file
:- Goal.                execute Prolog Goal
remove name.            remove name from definitions
prefix (op).            declare prefix operator op
postfix (op).           declare postfix operator op
infix (op).             declare non-associative infix operator op
infixl (op).            declare left associative infix operator op
infixr (op).            declare right associative infix operator op
num prefix (op).        declare prefix operator op with precedence num
num postfix (op).       declare postfix operator op with precedence num
num infix (op).         declare non-associative infix operator op w. prec. num
num infixl (op).        declare left associative infix operator op w. prec. num
num infixr (op).        declare right associative infix operator op w. prec. num

Gotchas

• An expression or command must be terminated by a period (.).

• There cannot be a space after the list cons dot (.) operator in x.xs, otherwise the system considers the input line has ended. As an alternative use [x|xs], for example if xs is a compound expression.

• There must be a space between different operators, otherwise the system is not able to separate the operators and parse the expression. For example:

  x\/\y           should be written         x\/ \y
  x+#l            should be written         x+ #l
  x*-y            should be written         x* -y
  x<-y            should be written         x< -y
  

• Parenthesization follows the Prolog syntax, so f(x,y) works, but f(x)(y) causes a syntax error. Instead, use colons, e.g. f:x:y or a mixed form such as f(x):y.

• The scope of the where construct is limited to its left-hand side. Therefore, it does not support recursive definitions:

  BAD:
  ... f(x) ... where f(n) := ... f(n-1) ...
  

  use the fix operator instead:

  GOOD:
  ... fix(f, x) ... where f(ff, n) := ... ff(n-1) ...
  

• When using tuples in where, be aware that tuple constructors are strict:

  x+1 where (x,y) := (1,write("abc");
  

  which displays "abc" even though y is not used, because tuple members are evaluated.

• Patterns in function arguments may consist of (parameterized) constants, empty list [], x.xs non-empty list with head x and tail xs, and tuples (x,y). These should not be nested:

  BAD:
  f((x,y).xs) := ...
  
  GOOD:
  f(p.xs) := ... where (x,y) := p.
  
  BAD:
  f(BinTree(BinTree(left, right), rest)) := ...
  
  GOOD:
  f(BinTree(tree, rest)) := ...
          where left := f(tree)
          where right := g(tree)
          where f(BinTree(l, r)) := l
          where g(BinTree(l, r)) := r.
  

Installation

Install SWI-Prolog 8.2.1 or greater from https://www.swi-prolog.org.

Then clone Husky with git:

$ git clone https://github.com/Genivia/husky

Run swipl then load Husky and run the husky interpreter:

$ swipl
?- [husky].
:- husky.

After starting Husky, load any one of the example .sky files with:

> load "examples/<filename>".

To delete programs from memory and reset Husky to system defaults:

> reset.

To create a stand-alone executable, run swipl and enter the following commands:

$ swipl
:- [husky].
:- husky.
> save.
> bye.

This creates the `Husky' executable with the prelude functions preloaded. You can also load one or more programs and then save the executable with the programs included.

Husky source files:

README.md             this file
husky.pl              the Husky interpreter
prelude.sky           built-in definitions

Husky program examples:

palindrome.sky        palindrome
btree.sky             binary trees
clientserver.sky      client-server stream example
combinators.sky       SKI combinator calculus
cr.sky                Chains of Recurrences algebra
hamming.sky           the Hamming problem
hangman.sky           a simple hangman game
matrix.sky            matrix and vector operations
monads.sky            monads a-la Haskell
primes.sky            primes
qsort.sky             quicksort
queens.sky            the queens problem
semantics.sky         denotational semantics example

Author

Husky is created by Prof. Robert A. van Engelen, engelen@acm.org.

License

Open source GPL (GNU public license).