Scala LanguageContinuations Library

Introduction

Continuation passing style is a form of control flow that involves passing to functions the rest of the computation as a "continuation" argument. The function in question later invokes that continuation to continue program execution. One way to think of a continuation is as a closure. The Scala continuations library brings delimited continuations in the form of the primitives shift/reset to the language.

continuations library: https://github.com/scala/scala-continuations

Syntax

  • reset { ... } // Continuations extend up to the end of the enclosing reset block
  • shift { ... } // Create a continuation stating from after the call, passing it to the closure
  • A @cpsParam[B, C] // A computation that requires a function A => B to create a value of C
  • @cps[A] // Alias for @cpsParam[A, A]
  • @suspendable // Alias for @cpsParam[Unit, Unit]

Remarks

shift and reset are primitive control flow structures, like Int.+ is a primitive operation and Long is a primitive type. They are more primitive than either in that delimited continuations can actually be used to construct almost all control flow structures. They are not very useful "out-of-the-box", but they truly shine when they are used in libraries to create rich APIs.

Continuations and monads are also closely linked. Continuations can be made into the continuation monad, and monads are continuations because their flatMap operation takes a continuation as parameter.

Callbacks are Continutations

// Takes a callback and executes it with the read value
def readFile(path: String)(callback: Try[String] => Unit): Unit = ???

readFile(path) { _.flatMap { file1 =>
  readFile(path2) { _.foreach { file2 =>
    processFiles(file1, file2)
  }}
}}

The function argument to readFile is a continuation, in that readFile invokes it to continue program execution after it has done its job.

In order to rein in what can easily become callback hell, we use the continuations library.

reset { // Reset is a delimiter for continuations.
  for { // Since the callback hell is relegated to continuation library machinery.
        // a for-comprehension can be used
    file1 <- shift(readFile(path1)) // shift has type (((A => B) => C) => A)
    // We use it as (((Try[String] => Unit) => Unit) => Try[String])
    // It takes all the code that occurs after it is called, up to the end of reset, and
    // makes it into a closure of type (A => B).
    // The reason this works is that shift is actually faking its return type.
    // It only pretends to return A.
    // It actually passes that closure into its function parameter (readFile(path1) here),
    // And that function calls what it thinks is a normal callback with an A.
    // And through compiler magic shift "injects" that A into its own callsite.
    // So if readFile calls its callback with parameter Success("OK"),
    // the shift is replaced with that value and the code is executed until the end of reset,
    // and the return value of that is what the callback in readFile returns.
    // If readFile called its callback twice, then the shift would run this code twice too.
    // Since readFile returns Unit though, the type of the entire reset expression is Unit
    //
    // Think of shift as shifting all the code after it into a closure,
    // and reset as resetting all those shifts and ending the closures.
    file2 <- shift(readFile(path2))
  } processFiles(file1, file2)
}

// After compilation, shift and reset are transformed back into closures
// The for comprehension first desugars to:
reset {
  shift(readFile(path1)).flatMap { file1 => shift(readFile(path2)).foreach { file2 => processFiles(file1, file2) } }
}
// And then the callbacks are restored via CPS transformation
readFile(path1) { _.flatMap { file1 => // We see how shift moves the code after it into a closure
  readFile(path2) { _.foreach { file2 =>
    processFiles(file1, file2)
  }}
}}  // And we see how reset closes all those closures
// And it looks just like the old version!

Creating Functions That Take Continuations

If shift is called outside of a delimiting reset block, it can be used to create functions that themselves create continuations inside a reset block. It is important to note that shift's type is not just (((A => B) => C) => A), it is actually (((A => B) => C) => (A @cpsParam[B, C])). That annotation marks where CPS transformations are needed. Functions that call shift without reset have their return type "infected" with that annotation.

Inside a reset block, a value of A @cpsParam[B, C] seems to have a value of A, though really it's just pretending. The continuation that is needed to complete the computation has type A => B, so the code following a method that returns this type must return B. C is the "real" return type, and after CPS transformation the function call has the type C.

Now, the example, taken from the Scaladoc of the library

val sessions = new HashMap[UUID, Int=>Unit]
def ask(prompt: String): Int @suspendable = // alias for @cpsParam[Unit, Unit]. @cps[Unit] is also an alias. (@cps[A] = @cpsParam[A,A])
  shift {
    k: (Int => Unit) => {
      println(prompt)
      val id = uuidGen
      sessions += id -> k
    }
  }

def go(): Unit = reset {
  println("Welcome!")
  val first = ask("Please give me a number") // Uses CPS just like shift
  val second = ask("Please enter another number")
  printf("The sum of your numbers is: %d\n", first + second)
}

Here, ask will store the continuation into a map, and later some other code can retrieve that "session" and pass in the result of the query to the user. In this way, go can actually be using an asynchronous library while its code looks like normal imperative code.