3 Fun Tricks with Call-By-Name in Scala

Call-by-Name (CBN) is one of those Scala features that had more confused programmers than happy programmers. It’s deceptively powerful, often abused, most of the time misused, almost never employed to its full potential, and it has a terrible name.

The genesis

I’m not going to use 3-dollar words to describe what CBN does. Instead, here’s the thing: if you write a normal function

def byValueFunction(x: Int) = ...

then when you call it

byValueFunction(2 + 3)

the argument is evaluated before the function is invoked. So before the function call, 2 + 3 becomes 5 and you actually call byValueFunction(5). This has been the case since the beginning of time with most programming languages. However, you can also mentally conceive a mechanism where you could pass 2 + 3 just as it is, and let the function evaluate it whenever it wants, if it needs to.

def byNameFunction(x: => Int) = ...

The arrow sign before the Int type means “I’ll take an Int expression literally and I’ll take care to evaluate it if I need to”. So when you call

byNameFunction(2 + 3)

what you’re passing is not the value of 2 + 3, but the expression 2 + 3, literally, as it is. The function is now responsible for evaluating the expression by simply using the argument name, which is, I suspect, where the terrible and very confusing “call-by-name” name came from. In the creators’ defense, I can’t find a simple, one-two-punch name for this either. Call-by-expression? Call-literally? Phony call? DETISS (don’t evaluate till I say so)?

Anyway. I think you get the picture in a few sentences rather than from half a book on referential transparency.

Trick 1 - reevaluation

So this call-by-name mechanism is actually very powerful and allows greater expressiveness in Scala. The first example I want to show is that when you pass a by-name argument, you can evaluate it as many times as you like. “Why would you want that? 2 + 3 evaluates to 5 at every hour of the day.”

Yes, but not quite everything does.

def byValueFunction(x: Long): Unit = {
    println(x)
    println(x)
}

def byNameFunction(x: => Long): Unit = {
    println(x)
    println(x)
}

I’ve defined two functions: one with a classical “by-value” argument, one with the fancy “by-name” argument. Both have the same implementation. Now let me run the following:

byValueFunction(System.nanoTime())

and I get

180615587360888
180615587360888

which is understandable, since nanoTime was evaluated to 180615587360888 before the call. However, look at this:

byNameFunction(System.nanoTime())

which gets me

180615853269641
180615853294140

that is, two different results. You can probably guess why: because the expression is passed literally and it’s being used twice, then it’s evaluated twice, at different moments.

There is great power - and great responsibility - in being able to reevaluate your arguments as you see fit. But more on that, another time.

Trick 2 - manageable infinity

Call-by-name used in conjunction with lazy evaluation - which is another beast - allows us to manage infinite data structures in Scala:

abstract class LazyList[+T] {
    def head: T
    def tail: LazyList[T]
}

case object Empty extends LazyList[Nothing] {
    override def head = throw new NoSuchElementException
    override def tail = throw new NoSuchElementException
}

class NonEmpty[+T](h: => T, t: => LazyList[T]) extends LazyList[T] {
    override lazy val head = h
    override lazy val tail = t
}

The magic happens in the NonEmpty class. because we’re using by-name arguments, that means the arguments are not evaluated until used. Because the only places where we evaluate the arguments are the initialization of lazy values - which also trigger on first use - it means that neither the arguments, nor the internal fields are evaluated on construction. Even better - if the fields are ever used, they’re evaluated once and then reused, so no more System.nanoTime shenanigans if you’re afraid of them. If you’re brave, you can also use something like

class NonEmpty[+T](h: => T, t: => LazyList[T], amber: => Long = System.nanoTime()) extends LazyList[T] { ... }

to forever hold - with some handling inside - the moment when this instance was “collapsed”, or all its members evaluated. Trapped in amber.

Trick 3 - hold the door

The third and most powerful aspect of CBN is that it prevents the computation of the argument so that the expression can be handled in some other way. For example, handled safely in case it threw something:

val anAttempt: Try[Int] = Try(throw new NullPointerException)

If not for CBN, the Try wrapper would have been impossible because the argument would not have any choice but to be evaluated first, and blow up in our face. Scala makes it more beautiful by combining that with the curly brace syntax for single-argument functions, which makes Try look like it belonged to the language itself:

val anAttempt: Try[Int] = Try {
  // something that can blow up
  throw new NullPointerException
}

Same with a Future - if not for CBN, you couldn’t pass an expression to be evaluated on another thread:

val aFuture = Future {
    // some nasty computation
    42
}

Epilogue

I’d be really interested to see if anyone had a better name for call-by-name, if anyone has an idea - comment below.