Personal tools

The Monad.Reader/Issue2/FunWithLinearImplicitParameters

From HaskellWiki

Jump to: navigation, search
Abstract.' Haskell is widely believed to be a purely functional language. While this is certainly true for Haskell98, GHC's various extensions can interplay in unforeseen ways and make it possible to write side-effecting code. In this article, we take the next step of impure programming by implementing Filinski's
functions for a wide class of monads.


1 Introduction

The following sections provide a short introduction into the various concepts our implementation uses. The code presented here is no longer available as an attachment. It has however been successfully tested with ghc-6.2.2 and ghc-6.4.

1.1 Shift and Reset

control operators provide a way to manipulate delimited continuations, which are similar to the undelimited continuation the familiar
uses, but more powerful. There are more detailed descriptions available e.g. in [1] and [2]; moreover, [3] give a unifying treatment of various forms of other "subcontinuations". Instead of capturing an undelimited continuation as
only captures the subcontinuation/context up to the the next
, and reifies it into a function value. The result of the evaluation of the body then becomes the result of the
. For example in
reset (1 + shift (\k -> k 1 + k 2)) :: Int
the context of
k = \x -> x + 1
, so the expression evaluates to
k 1 + k 2 = 2 + 3 = 5
. The interpretation of
is very easy in the continuation monad.
-- An action in the continuation monad maps a continuation,
-- i.e the "rest" of the computation, to a final result of type r.
newtype Cont r a = Cont { runCont :: (a -> r) -> r }
instance Functor (Cont r) where {- ... -}
instance Monad (Cont r) where {- ... -}
-- NB. In the attached Article.hs file, these are called shiftC and resetC.
shift :: ((a -> r) -> Cont r r) -> Cont r a
shift e = Cont $ \k -> reset (e k)
reset :: Cont a a -> a
reset e = e `runCont` id
-- The above example written in monadic style
* Main> reset $ (1 +) `fmap` shift (\k -> return $ k 1 + k 2)
As we can see,
reset e
delimits all effects of
and returns a pure value;
lets us explicitly construct the mapping from continuations to final results, so it is very similar to the data constructor
. Therefore
give us full control over the underlying continuation monad and are thereby strictly more expressive than
, which is polymorphic in the answer type
. To treat the direct-style
safely in a typed setting, it is necessary to express the answer type of the underlying continuation monad in the types. The Hindley-Milner type system cannot express this, but luckily, Haskell allows type information to be hidden in contexts, which provides our approach with full static type safety as opposed to Filinski's implementation in SML.

1.2 Monadic Reflection

Monadic reflection [4] enables us to write monadic code in direct style.
"reflects" a monadic value into a first-class value of our language. The side effects can then be observed by "reifing" a value back into monadic form. For example,
> reify (reflect [0,2] + reflect [0,1]) :: [Int]


> liftM2 (+) [0,2] [0,1]
both yield the same result, namely
In order to understand how monadic reflection can be implemented, we combine the observation that
give us the full power over an underlying continuation monad with an arbitrary answer type with Wadler's observation [5] that every monad can be embedded in the continuation monad. So using a direct-style
, we can write arbitrary monadic code in direct style.

Explicitly (but hiding the wrapping necessary for the ContT monad transformer), Wadler's transformation is as follows

embed :: Monad m => m a -> (forall r. (a -> m r) -> m r)
embed m = \k -> k =<< m
project :: Monad m => (forall r. (a -> m r) -> m r) -> m a
project f = f return
project . embed === id
and the property of
constituting monad morphisms between the monad
and the monad
forall r. ContT m r a
can easily be checked. Translating these morphisms into direct style, we immediately arrive at Filinski's
reflect m = shift (\k -> k =<< m)
reify t = reset (return t)

Now let us have a closer look at the above example to see how it works operationally.

e = reify (reflect [0,2] + reflect [0,1])

Substituting the definitions, this becomes

e = reset (return (shift (\k -> k =<< [0,2]) + shift (\k -> k =<< [0,1])))

which simplifies to

e = reset [shift (\k -> k 0 ++ k 2) + shift (\k' -> k' 0 ++ k' 1)]
Assuming left to right evaluation, the result of this expression is
k 0 ++ k 2
is bound to the subcontinuation
k = \x -> reset [x + shift (\k' -> k' 0 ++ k' 1)]
Again, in this term,
is bound to
\y -> reset [x + y]
, so
is the function
k = \x -> [x + 0] ++ [x + 1] = \x -> [x,x+1]

Therefore, as we expected, the whole expression evaluates to

e = k 0 ++ k 2 = [0,1] ++ [2,3] = [0,1,2,3]

1.3 Implicit Parameters

Implicit parameters [6] are GHC-specific type system extension providing dynamically bound variables. They are passed in the same way as type class dictionaries, but unlike type class dictionaries, their value can be changed for a subexpression. The types of the implicit parameters a function expects appear in type contexts which now make sense at arbitrary argument positions.

addThree :: (?foo :: Int) => Int
addThree = 3 + ?foo
withFour :: ((?foo :: Int) => a) -> a
withFour x = let ?foo = 4 in x
* Main> withFour addThree

We see that implicit parameters act like a reader monad written in direct style. The commutativity of the reader monad ensures that the code still is referentially transparent (the monomorphic recursion issue aside that will be discussed below).

Linear implicit parameters [7] work very much like regular implicit parameters, but the type of the parameter is required to be an instance of the class
with the single method
split :: a -> (a,a)
. At each branching point of the computation, the parameter gets split, so that each value is used only once. However, as we shall later see, this linearity is not enforced in all circumstances, with higher order functions and a certain class of recursive functions being the notable exceptions.

Possible uses are random number distribution, fresh name generation (if you do

not mind the names becoming very long) or a direct-style [6]. In this article, they will be used to store a subcontinuation from an enclosing
. The syntax is exactly the same as in the implicit case with the
replaced by
. We give a small example illustrating their intended use.
import qualified System.Random as R
instance Splittable R.StdGen where split = R.split
randInts :: R.StdGen -> (Int, Int, Int)
randInts gen = let %gen = gen in (rand, rand, rand) where
 rand :: (%gen :: R.StdGen) => Int
 rand = fst $ R.random %gen
* Main> print . randInts =<< R.getStdGen

As in the implicit case, the semantics of linear implicit parameters can be described in terms of a "monad", which, however, does not obey the monad laws in any nontrivial case.

newtype Split r a = Split { runSplit :: r -> a }
instance Functor (Split r) where
 f `fmap` Split x = Split $ f . x
instance Splittable r => Monad (Split r) where
 return x = Split $ const x
 Split x >>= f = Split $ \s ->
 let (s1,s2) = split s in f (x s1) `runSplit` s2
toSplit :: ((%foo :: r) => a) -> Split r a
toSplit x = Split $ \r -> let %foo = r in x
fromSplit :: Split r a -> ((%foo :: r) => a)
fromSplit (Split f) = f %foo

The ability to freely transform between "monadic" and "implicit" style is often very helpful, e.g. to work around GHC's limitation that signature contexts in a mutually recursive group must all be identical.

1.4 Unsafe Operations

The code below uses two unsafe operations Cite error: Closing </ref> missing for <ref> tag

infixr 0 `deepSeq`, $!!
class DeepSeq a where
 deepSeq :: a -> b -> b
($!!) :: (DeepSeq a) => (a -> b) -> a -> b
f $!! x = x `deepSeq` f x
Not all types can be made an instance of
. In particular, functions with an infinite domain and
actions cannot be fully evaluated in a sensible way.

2 Implementation

This section discusses the implementation of the monadic reflection library. It safely be skipped, especially the first two subsections are very technical.

2.1 Basic Declarations

k :-> v
is just an abstract representation of a finite map from k to v, The type
will be used to store the context of the evaluation, so it should have the property that different sequences of applications of
to an
yield different values. A
stores a value of arbitrary type. The most interesting declaration is that of
. The field
saves the position of the current expression relative to the next enclosing reset,
is the expression this next enclosing
stores the subexpressions that already have been assigned a value, and
will be used for exception handling.
infixr 9 :->
lookup :: Ord k => (k :-> v) -> k -> Maybe v
insert :: Ord k => (k :-> v) -> k -> v -> k :-> v
empty :: k :-> v
leftPos :: Position -> Position
rightPos :: Position -> Position
initPos :: Position
type Facts = Position :-> Cell
data Cell = forall a. Cell a deriving Typeable
data Prompt r = Prompt {
 position :: Position,
 prompt :: Direct r r,
 facts :: Facts,
 promptID :: Unique
newPrompt :: Facts -> Direct r r -> Prompt r
instance Splittable (Prompt r) where
 split p = (p {position = leftPos pos},
 p {position = rightPos pos}) where
 pos = position p
type Direct r a = (%ans :: Prompt r) => a

2.2 Shift and Reset

first saves the
and checks if this
has already been assigned a value using the
dictionary. If so, it just returns that value, otherwise, the outer
should return the value of
applied to the subcontinuation from the
to the
. The subcontinuation we pass to
creates a new copy of the
on every invocation, updates the
dictionary with the additional information that instead of the current
, the value
should be returned, and finally executes the
computation of the enclosing
. In order to pass the result of
up to the next
, we use exception handling, the unique ID of the
ensures that it is handled at the right place; the value, although known to be of type
is put in a
because we do not know whether
is an instance of the class
. Now all
has to do is evaluate the expression with a fresh
, and return the thrown value instead if an exception is caught. This gets a little more complicated because we need to be able to handle the effects of nested
shift :: ((a -> r) -> Direct r r) -> Direct r a
shift f :: Direct r a =
 let ans :: Prompt r
 ans = %ans
 in case lookup (facts ans) (position ans) of
 Just (Cell a) -> unsafeCoerce# a
 Nothing -> throwDyn . (,) (promptID ans) . Cell . f $ \x ->
 let %ans = newPrompt
 (insert (facts ans) (position ans) (Cell x))
 (prompt ans)
 in prompt ans
reset :: DeepSeq r => Direct r r -> r
reset e :: r = let %ans = newPrompt empty res in res where
 res :: Direct r r
 res = unsafePerformIO $ do
 let catchEsc e' = evaluate (id $!! e') `catchDyn`
 \err@(i, Cell result) ->
 if i == promptID %ans
 then catchEsc $ unsafeCoerce# result
 else throwDyn err
 catchEsc e
It is interesting to observe that in case of the error monad, this code uses the
monad's exception handling mechanism to propagate the error. Finally, we need to check the unsafe features are used in a safe way as described above. The
calls are always coercing to type
and it is clear that always the same
is in scope which we are ensuring using the
i == promptID
is only used for a "pure exception handling", which destroys purity, but still satisfies the weaker condition that the behavior does not depend on the outside world, which is essential here, as we rely on the property that a computation performs exactly the same steps when rerun.

2.3 Reflection and Reification

With working
functions, we can now turn to monadic reflection primitives. We first consider the case of the continuation monad.

2.3.1 Reflecting the Cont Monad

reflectCont :: Cont r a -> Direct r a
reflectCont (Cont f) = shift f
reifyCont :: DeepSeq r => Direct r a -> Cont r a
reifyCont e = Cont $ \k -> reset (k e)
As an example, we lift the function
to direct-style.
callCC' :: DeepSeq r => ((a -> b) -> Direct r a) -> Direct r a
callCC' f = reflectCont $ callCC $ \c -> reifyCont $ f $ reflectCont . c
However, the
operation can be implemented much more nicely using only two
s, as in
callCC' :: ((forall b. a -> b) -> Direct r a) -> Direct r a
callCC' f = shift $ \k -> k $ f (\x -> shift $ \_ -> k x)

In both versions, the expression

reset (callCC' (\k x -> k (x+)) 5) :: Int
correctly evaluates to
. It is a nice exercise to do this in Haskell's continuation monad; but be warned that it is a little harder than the above direct-style version.

2.3.2 Reflecting Arbitrary Monads

Now, implementing
is easier than in Filinski's implementation in SML, because the stronger static guarantees of our
functions eliminate the need for unsafe coercion functions.
-- Type alias for more concise type signatures of direct-style code.
type Monadic m a = forall r. Direct (m r) a
reflect :: Monad m => m a -> Monadic m a
reflect m = shift (\k -> k =<< m)
reify :: (DeepSeq (m a), Monad m) => Monadic m a -> m a
reify t = reset (return t)

3 Interface

For quick reference, we repeat the type signatures of the most important library functions.

type Direct r a = (%ans :: Prompt r) => a
shift :: ((a -> r) -> Direct r r) -> Direct r a
reset :: DeepSeq r => Direct r r -> r
type Monadic m a = forall r. Direct (m r) a
reflect :: Monad m => m a -> Monadic m a
reify :: (DeepSeq (m a), Monad m) => Monadic m a -> m a

4 Resolving Ambiguities

The use of linear implicit parameters comes with a few surprises. The GHC manual [7] even writes

So the semantics of the program depends on whether or not foo has a type signature. Yikes! You may say that this is a good reason to dislike linear implicit parameters and you'd be right. That is why they are an experimental feature.

However, most of the problems can be circumvented quite easily, and the property that the meaning of a program can depend on the signatures given is actually a good thing.

4.1 Recursive Functions

Indeed, omitting a type signature can sometimes result in a different behavior. Consider the following code, where
shift (\k -> k n)
should behave identically.
-- Without the explicit signature for k GHC does not infer a
-- sufficiently general type.
down 0 = []
down (n+1) = shift (\(k::Int -> [Int]) -> k n): down n
* Main> reset (down 4)
[3,3,3,3] -- wrong!
GHC considers the function
to be monomorphically recursive, but in fact the recursive call to
should be in a different context (with the implicit parameter bound to a different value), so
should actually be polymorphically recursive. This is semantically different and ensures the linearity. We can persuade GHC to treat it correctly by giving the function an explicit signature.
down' :: Int -> Direct [Int] [Int]
{- ... -}
* Main> reset (down' 4)
[3,2,1,0] -- right!
Furthermore, we have to watch out for a GHC bug [8] that appears to happen when expressions with differently polymorphic linear implicit parameter constraints are unified. In the above example, this occurs when
's explicit type signature is dropped and the signature of
is not generalized to
Int -> Direct r [Int]

4.2 Higher order functions

Implicit parameters are particularly tricky when functions using implicit parameters are passed to higher order functions. Consider the following example.

-- The prelude definition of the function map
map :: (a -> b) -> [a] -> [b]
map _ [] = []
map f (x:xs) = f x : map f xs
foo :: [[Int]]
foo = reify (map f [1,2,3]) where
 f :: Int -> Monadic [] Int
 f x = reflect [-x,x]
* Main> foo
[[-1,-1,-1],[1,1,1]] -- wrong!
The first surprise is that this code type checks at all: The type of the function
Int -> Monadic [] Int
but in order to be passed to
, the function
must have the different type
Monadic [] (Int -> Int)
. GHC pushes contexts at covariant argument positions as far to the left as possible using a technique called for-all-hoisting #ref6 6, which is of course sensible for type class constraints and implicit parameters, but destroys the linearity, which seems bad even in the motivating examples of random number or fresh name generation, and is only OK in the !QuickCheck example. So we always have to watch out for effectful functions that are passed as parameters, but at least we can copy the implementation of the higher order functions we want to use.
map' :: (a -> Direct r b) -> [a] -> Direct r [b]
{- Implementation as above -}
foo = reify (map' f [1,2,3]) where {- ... -}
* Main> foo
[1,2,3]] -- right!

4.3 The Monomorphism Restriction

What should the expression

reify (let x = reflect [0,1] in [x,x+2,x+4])
evaluate to? Two possibilities come to mind: Either we choose a value for the variable
first, and then evaluate the lists
or we view
as the reflected list
and the choice whether
stands for
is made whenever
it is evaluated. It is immediately clear how both variants can be achieved in monadic style.
* Main> do x <- [0,1]; return [x,x+2,x+4]
* Main> let x = [0,1] in sequence [x,(+2) `fmap` x, (+4) `fmap` x]

In direct style, this is even easier, but the meaning of our code now depends on the type signature.

* Main> reify (let x :: Int; x = reflect [0,1] in [x,x+2,x+4])
* Main> reify (let x :: Monadic [] Int; x = reflect [0,1] in [x,x+2,x+4])

It is important that we give a real type signature:

x :: Int = reflect [0,1]
does not make any difference! This is a nice and very natural way to describe both situations, but the answer to the question which one GHC chooses when no signature is given is less satisfactory: It depends on the status of the flag
. With the monomorphism "restriction" #ref11 11 turned on,
must have a monomorphic type, so the first situation applies, without the restriction
gets the most general type which leads to the second behavior. In my opinion, it would be nice if there were a flag that, in order to give the programmer a chance to disambiguate his code, causes a warning to be emitted whenever the monomorphism restriction kicks in; a similar warning has been proven useful to detect numeric defaulting.

5 Examples

We now present some examples reflecting the

5.1 Lazy Evaluation

The use of monads in Haskell models an impure language with call-by-value semantics. This is not surprising as one motivation for the use of monads is the need to do IO. For IO, evaluation order is important and call-by-value makes evaluation order easier to reason about. For the
monad this certainly the right decision, and if desired, the
function can be used to execute
operations lazily.

But such a lazy monadic behavior would be practical for other monads, too: The list monad is very susceptible to space leaks and unnecessary recomputation. The reflected list monad, however, is often closer to the desired behavior, as the following examples suggest.

-- Lazy repeat, Prelude.repeat would allow the side effect
-- of the argument to take place only once
repeat' :: Direct r a -> Direct r [a]
repeat' x = x:repeat' x
* Main> take 3 `fmap` sequence (repeat [1,2::Int])
<< Does not terminate. >>
* Main> reify (take 3 $ repeat' (reflect [1,2::Int]))
* Main> fst `fmap` liftM2 (,) [1,2::Int] [3,4::Int]
* Main> reify (fst (reflect [1,2::Int], reflect [3,4::Int]))
* Main> reify (fst $!! (reflect [1,2::Int], reflect [3,4::Int]))

The last expression shows that we can easily revert to the eager version by adding appropriate strictness annotations.

5.2 Filtering Permutations

As a typical problem where the lazy behavior of our implementation is advantageous, we consider a small combinatorial example: Find all permutations of


such that all the sums of the initial sequences of the permutations are primes.

-- NB. This section's example code can be found in the files Perms.*.
-- _very_ simple primality test.
isPrime :: Int -> Bool
isPrime n = n >= 2 && all (\k -> n `mod` k /= 0)
 (takeWhile (\k -> k*k <= n) $ 2:[3,5..])
-- check if all the initial sums are primes.
goodPerm :: [Int] -> Bool
goodPerm xs = all isPrime (scanl1 (+) xs)
If we want to solve the problem in Haskell, we need to make a big compromise: Either we take the easy road and generate a list of the permutations and then
the good ones, which is unfortunately very slow because all permutations must be checked even if it already turns out after inspecting a few list elements that no permutation starting this way can have the property.

Alternatively, we can hand-optimize the algorithm by performing the construction of the permutation step-wise and interleaving the primality checks appropriately. In our example, this is not really hard and the list monad is a great help, but it feels low-level, error-prone and lacks modularity. We would like the declarativity of the first approach while retaining the speed improvements the lazy checking provides.

So, should we to switch to another language? An obvious candidate is curry [9], a lazily evaluated hybrid functional-logic language with a very Haskell-like syntax and feel. Curry allows nondeterministic functions to be written by simply declaring the function multiple times; however, the nondeterminacy cannot be expressed on the type level. Using monadic reflection, we can do something very similar as follows.

-- nondeterministic choice
(?) :: DeepSeq a => Monadic [] a -> Monadic [] a -> Monadic [] a
x ? y = reflect (reify x `mplus` reify y)
-- nondeterministically select a permutation
permute :: [Int] -> Monadic [] [Int]
permute [] = []
permute xs = y: permute ys where
 y::Int; ys::[Int]
 (y,ys) = select xs
select :: [Int] -> Monadic [] (Int,[Int])
select [] = reflect []
select (x:xs) = (x,xs) ? second (x:) (select xs) where
 -- a special case of Control.Arrow.second
 second f (x,y) = (x,f y)

Now we only need to ensure that the computation fails when the permutation does not have the desired property.

solve :: Int -> Monadic [] [Int]
solve n = if goodPerm xs then xs else reflect [] where
 xs :: [Int]
 xs = permute $ map (2^) [0..n-1]
* Main> reify (solve 17)
The relative performance of the different approaches is not surprising: The manual Haskell solution (GHC) is the fastest, the Curry solution (Muenster Curry) is about six times slower while the solution using monadic reflection is another four times slower (and gets slightly worse for larger values of
), since a lot of recomputation is implied by the way
are implemented. Finally, the naÔve solution would probably take years to finish.

6 Further Ideas

This section discusses some further directions in which the ideas of this article might be extended.

6.1 Denotational Semantics

The relationship between laziness and direct-style continuation effects, despite often following the intuition, needs some further clarification. For that purpose, I wrote two interpreters of a simple untyped combinator language, which use a continuation-like monad and the monadic reflection library, respectively. They can be checked for coincidence using !QuickCheck tests generating type-checking expressions for the language. The monad the interpreter is built upon is an
monad augmented with continuations of answer type
using the
newtype Eval s a
 = Eval { runEval :: ContT Int (ST s) a }
 deriving (Functor, Monad)

The interpreter maps the source language's expressions into the following universal type.

type U s = Eval s (Ref s `Either` U' s)
data U' s
 = Int { runInt :: Int }
 | Fun { runFun :: U s -> U s }
 | List { runList :: Maybe (U s, U s) }
newtype Ref s = Ref { unRef :: STRef s (U' s `Either` U s) }
So an
U s
is either a reference or a value of type
U' s
; references either point to a thunk of type
U s
or to an evaluated value of type
U' s
. Laziness is provided by two functions of the following types.
-- Delays a computation
delay :: U s -> U s
-- Force evaluation of a reference to a normal form.
force :: U s -> Eval s (U' s)

Details can be found in the [attachment:Reflection.tar.gz tarball] provided with this article. The distribution also contains two interpreters for a strict version of the language, which can be more straightforwardly implemented using the plain continuation monad and, in case of the direct-style interpreter, some strictness annotations.

6.2 A Lightweight Notation for Monads

Haskell's do-notation is often criticized being too verbose, especially for commutative monads; and the process of transforming pure functions into monadic style because some (possibly deeply nested) function needs some effects is tedious and error-prone.

GHC already has special support for the (commutative) reader monad, through implicit parameters. This special rÙle of the reader monad might be justified by additional properties this monad has, for example that there are isomorphisms of type
m (a -> b) -> a -> m b
m (a, b) -> (m a, m b)
whose inverses are given by
\f x -> f `ap` return x
liftM2 (,)
, respectively. Also, special tools [10] are being developed that automatically transform a function from direct into monadic style, but this process requires arbitrary decisions where to apply effects, e.g. it is unclear if a function of type
Int -> Bool
should be monadified to a function of type
Monad m => m Int -> m Bool
Monad m => Int -> m Bool
, as both make sense in different circumstances.

As we showed in this article, Haskell's type system is almost ready to express these differences on the type level; the only remaining problem is that forall-hoisting [7] changes the meaning of expressions. On the other hand, because of the interaction with laziness, keeping the semantics of the library described in this article would result in a rather complicated translation, as we saw in the last section. In order to get rid of this obscurity, one might imagine a type-directed translation which translates (pseudo-code)

reflect :: m a -> (<m> => a)
reify :: Monad m => (<m> => a) -> m a
foo :: <[]> => Int
foo = reflect [0,2] + reflect [0,1]
bar :: [Int]
bar = reify foo

more strictly into

foo :: [Int]
foo = (+) `fmap` [0,2] `ap` [0,1]
bar :: [Int]
bar = foo

However, this contradicts Haskell's philosophy to make invocation of effects as explicit as possible, and would probably be considered an "underkill". Moreover, it would require a decent solution to the monomorphism restriction problem.

7 Conclusion

Do not take this too seriously: Our code heavily relies on unsafe and experimental features; time and space usage are increased by the suboptimal encoding of continuations and the recomputations; and the number of supported monads is limited by the
requirement. However, we provided a framework with strong static guarantees in which it is easy to experiment with the unfamiliar
operators, and we learned that GHC Haskell's type system goes well beyond Hindley-Milner and it is almost ready for an impure language where effects are declared explicitly on the type level. More importantly, it is great fun to abuse just about every unsafe feature of (GHC) Haskell, to create an impure sublanguage with monadic effects.

8 Acknowledgments

I would like to thank the GHC team for this great compiler with its many fascinating extensions.

I also want to thank Peter Eriksen, Cale Gibbard and Don Stewart for proof-reading the article and their valuable suggestions, as well as Brandon Moore and Autrijus Tang for their advice on the references.

9 References

  1. Olivier Danvy and Andrzej Filinski. "A Functional Abstraction of Typed Contexts". DIKU. DIKU Rapport 89/12. July 1989. Available online:
  2. Chung-chieh Shan. "Shift to Control". 2004 Scheme Workshop. September 2004. Available online:
  3. R. Kent Dybvig, Simon Peyton-Jones, and Amr Sabry. "A Monadic Framework for Subcontinuations". February 2005. Available online:
  4. Andrzej Filinski. Representing monads. In Conference Record of POPL '94: 21st ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, Portland, Oregon, pages 446--457. Available online:
  5. Philip Wadler. "The essence of functional programming". Invited talk, 19'th Symposium on Principles of Programming Languages, ACM Press. January 1992. Available online:
  6. 6.0 6.1 Koen Claessen and John Hughes. "!QuickCheck: An Automatic Testing Tool for Haskell".
  7. 7.0 7.1 7.2 The GHC Team. "The Glorious Glasgow Haskell Compilation System User's Guide, Version 6.4". BR Linear Implicit Parameters: BR Implicit Parameters: BR Forall-Hoisting:
  8. Thomas J‰ger "Linear implicit parameters: linearity not enforced". Mailing list post.
  9. Michael Hanus [editor] "Curry. An Integrated Functional Logic Language". Available online:
  10. "Monadification as a Refactoring".