Continuation-passing style (CPS)

Do you want to program in style? Well, you probably already do if you write programs at all. The most common style of programming is called ‘direct style’. In this style functions are called with some arguments and return some values (or nothing). When you want to combine functions, you can do so by just calling them individually or inside other functions. You can also exploit the composition operator.

So is there any other way to go about returning from a function? How can a function not just return a value that it has computed? How can you use such a function at all? A function written in continuation-passing style returns a continuation instead. So what is this ‘continuation’ thing? Let’s look at it’s type first:

type C r a = a -> r

This is just a function from type a to type r. What makes it so special? Because it can be used as a higher-order function. Moreover, it gains importance when used in the continuation monad. If you don’t know what a monad is, it doesn’t matter. Here is its type:

type Cont r a = (a -> r) -> r

where r stands for return value and a for intermediate value. The monad type is a higher-order function from a function that takes a value of type r and returns a value of type r to a value of type r. You could also say it is a function that takes a function and returns the result of that function. This sounds to me like it does nothing at all! Well it does do something, namely apply a function to a value of type a.

I suggest to now think really hard about functions that have this Cont type. We know it takes a function as its argument and applies something to it. Here is a simple example:

foo :: (Int -> r) -> r
foo c = c 42

foo' :: Cont r Int
foo' c = c 42

foo'' :: C r Int -> r
foo'' = \ c -> c 42

I have written the function with three different notations, they do exactly the same thing. In this blog post I will stick mostly with the first way of writing these functions. The types will get a bit messy later on, but I want to be as explicit as possible. These three functions take an argument c, that represents the current continuation and apply it to the integer value 42.

How do we get our number back? By simply applying foo to the identity function like so:

fortyTwo :: Int
fortyTwo = foo id

fortyTwo' :: Int
fortyTwo' = foo (\ fortyTwo -> fortyTwo)

We call the identity function with the value 42; returning 42. Can we use other functions with foo? Yes, we can. Can we use all functions with foo? No, we can’t. We can only use functions that support integers as their first argument. Luckily, there are many of these functions:

fortyThree :: Int
fortyThree = foo (+1)

fortyTwoStr :: String
fortyTwoStr = foo show

So far we have called foo with functions written in direct style, causing it to return a value. However, we can call it with other functions written in CPS. You may now ask yourself: “Can I call foo with itself?”. Yes, you can, but we have to make a slight modification. Remember what foo does: apply the continuation to the value 42. If foo is that continuation, this is analogous to applying foo to 42. We can use a lambda abstraction to deal with this:

foofoo :: (Int -> r) -> r
foofoo = foo (\ fortyTwo -> foo)

foofoo' :: (Int -> r) -> r
foofoo' = (\ fortyTwo -> (\ c -> c 42)) 42

and create an immensely useless function, because exactly nothing has happened and we get our original function back. You can also come to the conclusion that foo can not be applied to foo by the reasoning that foo does not support integers as its first argument, but functions from integers to r. It would look like the illegal expression (\ c -> c 42) 42.

How to play the CPS game

A continuation can be seen as a todo. This part of our code may be completed later. In direct style this occurs frequently. When writing a program a certain function can be left undefined or with a placeholder: to be completed later. It may be called by some other function. This is what our c is: a placeholder for someone to complete later. We completed our foo function by calling it with the id function.

The person writing the implementation of the placeholder needs to only consider the intermediate value a passed to the continuation. She can remain agnostic about the rest of the original function. If she is lazy herself, she can even insert some continuations (todos) in her code. She then passes on the code to someone else, and this person continues… You get the idea. We can keep using continuations inside our continuations. CPS is just a long list of todo‘s.

When writing in CPS we will have two new rules:

• Functions take an extra continuation argument; c in our case.
• Functions are applied to variables or other functions (i.e., lambda expressions).

This has the consequence that the innermost part of an expression must be evaluated first. Making the order of evaluation explicit.

We will encounter two main usages of continuations:

• Apply the continuation to something.
• Apply something to the continuation.

If you want you may think about the possible types of these two somethings. The first one restricts the type of the continuation, as it needs to accept the thing it is applied to. In the foo example this is an integer, making the intermediate value an integer as well. The second one restricts the type of the something applied to the continuation, namely it needs to be a function that accepts continuations as arguments. When applying a function to a continuation, we are actually applying the continuation to the result of that function. It is like performing the function first and after it has finished just continuing on.

We will show some idioms of CPS. They will make more sense once you have read the entirety of this blog post.

• bind to result of continuation
• bind to continuation
• modify continuation

We saw an example of this with our foo function, where we bound the value 42. More generally we can bind the result of any continuation.

bindResult :: ((a -> r) -> r) -> (a -> r) -> r
bindResult m c = m (\ a -> c a)

This can be ead as: perform computation m and bind a to the result. What it does is pass the intermediate result to the current continuation. This becomes useful later, when we want to do some computation first and then continue with the result of that computation. We can also bind continuations to other functions. Like so:

bindContinuation :: ((a -> r) -> (a -> r) -> (a -> r) -> r) -> (a -> r) -> r
bindContinuation k c = k c c c

The type of modification function looks like a combination of the two other idioms:

withc :: ((b -> r) -> r) -> ((a -> r) -> b -> r) -> (a -> r) -> r
withc m f c = m (f c)

Chains that bind

With this knowledge we can now start chaining functions like no other. Let’s first define a useful function which helps us do this:

unit :: a -> (a -> r) -> r
unit a c = c a

fortyTwo'' :: Int
fortyTwo'' = unit 42 id

With this function we can now make a continuation from any value of type a. Multiple unit‘s can be chained together.

unitChain :: a -> (a -> r) -> r
unitChain a = (unit a) unit unit

Note that unit does not need a lambda abstraction to chain it together, because it takes a first argument of type a. We can also create an at first weird looking expression:

uu :: ((a -> (a -> r1) -> r1) -> r2) -> r2
uu = unit unit

this can be interpreted as a function that takes a function and applies unit to it. We apply id to get our unit function back. You can do the same thing with unitChain.

uuid :: a -> (a -> r) -> r
uuid = uu id

We can use the intermediate value of a CPS function and keep passing it on. What we can not do is use all of the previous intermediate values. For this reason we define bind to paste continuations together. We saw something similar when defining foofoo. What will be the type of this function? It will take a continuation, a function from the intermediate value of that continuation to a new continuation, and return this new continuation

bind :: ((a -> r) -> r) -> (a -> ((b -> r) -> r)) -> ((b -> r) -> r)
bind m f c = m (\ a -> f a c)

This function is a bit magical, but can be translated to English. Perform computation m and bind a to the result; apply f to the result a and apply that to the continuation c. The main difference in this function

is that the contination c is not applied to something, but f is applied to c. We do this to pass our current continuation. Consider functions that just return their passed continuation:

continueCont :: ((a -> r) -> r) -> (a -> r) -> r
continueCont m c = m c

continueCont' :: ((a -> r) -> r) -> (a -> r) -> r
continueCont' m c = m (\ a -> c a)

this can also be implemented with the id function. Here we see the resemblance with our bind function immediately.

Implementation example

I will give some functions to illustrate unit and bind. Since we are in the habit of making useless functions…

csqrd :: Floating a => a -> (a -> r) -> r
csqrd a c = c (a * a)

csqrt :: Floating a => a -> (a -> r) -> r
csqrt a c = c (sqrt a)

bar :: Floating a => a -> (a -> r) -> r
bar a = csqrd a csqrt

bar' :: Floating a => a -> (a -> r) -> r
bar' a = csqrt a csqrd

Now let’s say we want both the square and the square root

bothResults :: Floating a => a -> ((a, a) -> r) -> r
bothResults a = csqrt a `bind` \ x ->
                csqrd a `bind` \ y ->
                unit (x, y)

Here bind binds our intermediate results and unit helps us return the result. We can not use a normal tuple as final expression here? (Why?)

Control Flow

The examples in the previous section work, but are mostly clearer in direct style. What is something you can do in continuation style that you can not do in direct style? We can use CPS for most control flow statements! A few examples:

• jmp
• break
• continue
• try, throw, catch
• while
• if then else
• yield
• threading
• lazy evaluation

We may now pause and think about the reason why these things are possible. How can we just go to any part in our function from our function? This is where we must realize the power of higher-order functions. Calling functions with other functions gives us the possibility to interpret the passed functions as continuations (and call then).

Forward jumps

We will implement the cjmp, lbl pair to skip certain parts of our code. This is a forward jump. Backwards jumps are also possible, but a bit more elaborate. Let’s first think about the type of the conditional jump instruction. It should have a condition, which can be represented by a Bool, a continuation to jump to, and the current continuation, like all CPS function.

cjmp :: Bool -> a -> (a -> r) -> (a -> r) -> r
cjmp True  a c' _ = c' a
cjmp False a _  c = c  a

The True branch ignores the current continuation c and jumps to c'. The other branch does the same thing the other way around. Notice the similarity with the unit function.

The lbl function should take a label and return a continuation. What is a label? A label is function from two continuations to a result. We get the following:

lbl :: ((a -> r) -> (a -> r) -> r) -> (a -> r) -> r
lbl k c = k c c

With our fancy new function we can now do some flow control:

flow :: (String -> r) -> r
flow c = lbl (\ label ->
              unit "hello" `bind` \ _ ->
              unit "no jump"
           ) c

this returns "jump" when applied to id.

TODO – Continuation assignment

We will implement the function callcc, which stands for call with current continuation. It provides us with a function that ignores its continuation. We see the use of all idioms coming together.

callcc :: ((a -> (b -> r) -> r) -> (a -> r) -> r) -> (a -> r) -> r
callcc h c = h (\ a _ -> c a) c

fix :: (a -> a) -> a
fix f = let x = f x in x -- Y-combinator

getcc :: ((((a -> r) -> r) -> r) -> r)
getcc = callcc (unit . fix)

getcc' :: a -> (((a, a -> ((b -> r) -> r)) -> r) -> r)
getcc' x0 = callcc (\ c -> let f x = c (x, f) in unit (x0, f))

Conclusion

We have shown that CPS is a powerful tool for creating (sometimes hard to understand) functions with extra flow control. We have shown a couple of implementation examples. The takeway is that higher order functions have a lot of power to them. Monads and CPS are actually interchangable, if CPS is expressed as a monad like the one at the start of this post. This is exactly what we did in this post without calling attention to it. We defined the unit and bind function; the monad constructor can be emulated by defining a CPS function. Together they gives us a triple constituting a monad.

References

Wadler, Philip. “The essence of functional programming.” Proceedings of the 19th ACM SIGPLAN-SIGACT symposium on Principles of programming languages. 1992.

http://hackage.haskell.org/package/mtl-2.2.2/docs/Control-Monad-Cont.html

https://wiki.haskell.org/Continuation

https://en.wikipedia.org/wiki/Continuation-passing_style

https://apfelmus.nfshost.com/articles/operational-monad.html

http://www.haskellforall.com/2012/12/the-continuation-monad.html