As I’ve marveled before, the Scala type system is extremely expressive.  In this post, I want to look at one particular aspect of the type system.  I want to look at a few of the ways to pass behaviors into functions.  Using classes, traits, function types, and a construct that looks like an ad hoc interface (and probably other ways I haven’t learned yet) you can exercise precise control over what units of behavior your functions will accept and use.

First, let’s define a zoo of classes that we will experiment on:

trait TestTrait {
  def func(s1: String): Unit

class TestClass1 {
  def func(s1: String): Unit = println("TestClass1: "+s1)

class TestClass2 {
  def func(s1: String): Int = { println("TestClass2: "+s1); 999; }

class TestClass3 extends TestTrait {
  def func(s1: String) = println("TestClass3: "+s1)

class TestClass4 {
  def func(a1: Any) = println("TestClass4: "+a1)

class TestClass5 {
  def otherFunc(s1: String) = println("TestClass5: "+s1)

class TestClass6 {
  def wrongFunc(list: List[Int]) = println("TestClass6: "+list)

There.  These are very simple, but I’ll just point out the salient points of each:

  • TestTrait – Basically, an interface with one method.
  • TestClass1 – One method, func, takes a String and returns Unit (void, to Java coders).
  • TestClass2 – Same as TestClass1, but returns Int.
  • TestClass3 – Same as TestClass1, but also extends TestTrait.
  • TestClass4 – Same as TestClass1, but takes an Any parameter.
  • TestClass5 – Same as TestClass1, but its function has a different name
  • TestClass6 – One method, different name, different parameter type, does not extend TestTrait.

Each of these (except TestTrait) defines some behavior that we might want to pass into a function.  That behavior is trivial in these classes, but I’m trying to keep the examples short.  Let’s now define several functions that might be able to use some of this behavior.

def test1(fn : (String) => Unit) = { fn("test1"); }

def test2(fn : (String) => Int)  = { fn("test2"); }

def test3(obj: { def func(x: String): Unit }) = { obj.func("test3") }

def test4(obj: TestClass1) = { obj.func("test4") }

def test5(obj: TestTrait)  = { obj.func("test5") }

def test6(obj: { def func(x: Any): Unit }) = { obj.func("test6") }

Now we just need to instantiate each one class and start passing parameters!  Create an instance of each class, TestClass1 to TestClass6, and name them tc1 to tc6.  The pass each instance’s function into test1 and test2, and pass the instance itself into test3 through test6.  Let’s do this in the Scala interpreter, rather than compiling.  Like so:

val tc1 = new TestClass1
val tc2 = new TestClass2
val tc3 = new TestClass3
val tc4 = new TestClass4
val tc5 = new TestClass5
val tc6 = new TestClass6







Did you run that?  It didn’t work, did it?  At least not all of it.  Now we we have something to discuss.  Here’s a chart of what did and didn’t work:

test1 test2 test3 test4 test5 test6

Taking these results one test function at a time:

test1(fn : (String) => Unit)

All but one of the test objects has at least one member function that can be passed in to this function.  It’s easy to see why this works for tc1, tc3, and tc5.  They each have a function that matches the type of parameter fn exactly.  It’s also easy to see why it doesn’t work for tc6.  TestClass6.wrongFunc takes a List[Int] parameter, which is unrelated to fn’s expected String parameter.

But what about tc2 and tc4?  TestClass2.func takes a String parameter, just like the test1’s fn parameter, but it returns an Int instead of Unit.  But it still works.  Instead of requiring a Unit return type for fn, Scala allows functions that return values.  It simply throws away any returned value and treats the passed in function as though it returned Unit.  Scala’s typesafety requirements in this case are very permissive!  This can be convenient, but be aware that you have very little control of what functions can be passed in when you use this kind of parameter type.

TestClass4.func returns Unit, but it takes a parameter of type Any.  Since test1 knows parameter fn as a function with a String parameter, test1 will only ever pass Strings into fn.  A String “is-a” Any so Scala says it’s ok to pass Strings into TestClass4.func.

This all goes back to variance.  In Scala, functions are contravariant with respect to parameters and covariant with respect to return type.  If you’re not familiar with variance, this means that if function f2’s parameters are simple supertypes (simple, meaning a supertype or the same type) of function f1’s parameters, and if f2’s return type is a simple subtype of f1’s return type then f2 is a subtype of f1 and can be used anywhere f1’s type is required.  The makes sense.  If a function takes type X, then you can pass in any value of type X or a subtype of X.  If a function returns type Y, then the result can be treated as Y or any supertype of Y.  Because Unit is a subtype of Int (and of all reference types) and String is a simple supertype of String, the function type “(String) => Int” is a subtype of “(String) => Unit” and can be used wherever a “(String) => Unit” is required.

test2(fn : (String) => Int)

This one only works for tc2.  TestClass2.func matches the type of parameter fn exactly.  But none of the other functions can be used here.  This is because none of the other functions take a parameter that is a simple subtype of String and return a simple supertype of Int, none of the others will work here.

A quick note about the use of functions in this way:  If you have an instance and pass a member function of that instance as a parameter, the function still has access to that instance’s data.  Moreover, if the passed function changes the state of the object of which it is a member, then function that receives the function as a parameter can mutate the object.  I don’t think I’m being clear.  Here’s some code to demonstrate what I mean.

class TestClass(name: String) {
  var memberStr: String = ""
  def func(str: String) = { memberStr += str }
  override def toString = name + ": " + memberStr

val testA = new TestClass("A")
val testB = new TestClass("B")

def test(fn : (String) => Unit, num: Int): Unit = {
  if (num > 0) { fn("X"); test(fn,num-1); }

test(testA.func, 5)
test(testB.func, 2)

println(testA + ", " + testB)

See?  Function test just takes a function as a parameter (and an Int).  It does NOT take a TestClass as a parameter, but it’s still able to alter those two instances of TestClass by way of the mutator member function TestClass.func.  So be aware that when you declare a function parameter like this, you never know whether you’re getting a bare function, or a member function.

test3(obj: { def func(x: String): Unit })

Here’s a construct that’s much more strict than the function types used in test1 and test2.  The test3 function take an ad-hoc interface as a parameter.  It’s not a named interface (a Trait), so the parameter’s type doesn’t have to be declared as extending anything.  The parameter is accepted if it has a member function of the given name and type.

Now, in this case, remember, the parameter isn’t the function, it’s the object.  So the permissive behavior that Scala allows in test1 and test2 because of the variance behavior of function types does not apply here.  An object of type A has a function funcA, and object of type B has a function funcB whose type is a subtype of funcA, that does NOT make B a subtype of A.  The names and types of the members declared in the ad-hoc interface must match the parameter exactly.  The parameter can have additional members, but it must at least provide all the members listed in the ad-hoc interface.

In this case, TestClass1 and TestClass3 are the only classes that have functions named func that take a String parameter and return Unit.  The other classes have either the wrong function name, parameter type, or return type.

test4(obj: TestClass1)

This is an easy one.  You know what’s happening here.  Function test4 states explicitly that it requires an object of type TestClass1.  Only a TestClass1 or a subtype of TestClass1 will do.

test5(obj: TestTrait)

Function test5 expects a TestTrait parameter.  Any object that has the trait TestTrait can be passed in.  In this case, TestClass3 is the only class that extends TestTrait.  We could have declared TestCase1 as extending TestTrait, but we didn’t.  Therefor, it doesn’t matter that TestClass1 has a function with the right name, the right parameter types, and the right return type.  It isn’t explicitly defined as extending TestTrait, so it isn’t allowed.

test6(obj: { def func(x: Any): Unit })

This is really just a more extreme example of the phenomenon demonstrated in test3.  This is another ad-hoc interface.  That member function is the supertype of ALL functions that have one parameter and return a reference type.  ALL of them.  You see why?  Any is the superclass of every type and Unit is a subclass of every reference type.  Function types are contravariate with respect to their parameters and covariant with respect to their return types.  But, of course, only our instance of TestClass4 is accepted because this parameter doesn’t have a function type, but an ad-hoc interface type.  It is not the case that all types are covariant with respect to the types of their member functions.  TestClass4 is the only one that has a function with the right name, parameter type, and return type.

Don’t forget to subscribe to my RSS feed, or follow this blog on Twitter.
Copyright © 2008 Matthew Jason Malone