Question

I couldn’t find the answer to this in any other question. Suppose that I have an abstract superclass Abstract0 with two subclasses, Concrete1 and Concrete1. I want to be able to define in Abstract0 something like

def setOption(...): Self = {...}

where Self would be the concrete subtype. This would allow chaining calls to setOption like this:

val obj = new Concrete1.setOption(...).setOption(...)

and still get Concrete1 as the inferred type of obj.

What I don’t want is to define this:

abstract class Abstract0[T <: Abstract0[T]]

because it makes it harder for clients to handle this type. I tried various possibilities including an abstract type:

abstract class Abstract0 {
  type Self <: Abstract0
}

class Concrete1 extends Abstract0 {
  type Self = Concrete1
}

but then it is impossible to implement setOption, because this in Abstract0 does not have type Self. And using this: Self => also doesn’t work in Abstract0.

What solutions are there to this issue?

Answer 1

This is what this.type is for:

scala> abstract class Abstract0 {
     |   def setOption(j: Int): this.type
     | }
defined class Abstract0

scala> class Concrete0 extends Abstract0 {
     |   var i: Int = 0
     |   def setOption(j: Int) = {i = j; this}
     | }
defined class Concrete0

scala> (new Concrete0).setOption(1).setOption(1)
res72: Concrete0 = Concrete0@a50ea1

As you can see setOption returns the actual type used, not Abstract0. If Concrete0 had setOtherOption then (new Concrete0).setOption(1).setOtherOption(...) would work

UPDATE: To answer JPP’s followup question in the comment (how to return new instances:
The general approach described in the question is the right one (using abstract types). However, the creation of the new instances needs to be explicit for each subclass.

One approach is:

abstract class Abstract0 {
  type Self <: Abstract0

  var i = 0

  def copy(i: Int) : Self

  def setOption(j: Int): Self = copy(j)
}

class Concrete0(i: Int) extends Abstract0 {
  type Self = Concrete0
  def copy(i: Int) = new Concrete0(i)
}

Another one is to follow the builder pattern used in Scala’s collection library. That is, setOption receives an implicit builder parameter. This has the advantages that building the new instance can be done with more methods than just ‘copy’ and that complex builds can be done. E.g. a setSpecialOption can specify that the return instance must be SpecialConcrete.

Here’s an illustration of the solution:

trait Abstract0Builder[To] {
    def setOption(j: Int)
    def result: To
}

trait CanBuildAbstract0[From, To] {
  def apply(from: From): Abstract0Builder[To]
}


abstract class Abstract0 {
  type Self <: Abstract0

  def self = this.asInstanceOf[Self]

  def setOption[To <: Abstract0](j: Int)(implicit cbf: CanBuildAbstract0[Self, To]): To = {
    val builder = cbf(self)
    builder.setOption(j)
    builder.result
  }

}

class Concrete0(i: Int) extends Abstract0 {
  type Self = Concrete0
}

object Concrete0 {
    implicit def cbf = new CanBuildAbstract0[Concrete0, Concrete0] {
        def apply(from: Concrete0) = new Abstract0Builder[Concrete0] {
           var i = 0
           def setOption(j: Int) = i = j
           def result = new Concrete0(i)
        }
    }
}

object Main {
    def main(args: Array[String]) {
    val c = new Concrete0(0).setOption(1)
    println("c is " + c.getClass)
    }
}

UPDATE 2:
Replying to JPP’s second comment. In case of several levels of nesting, use a type parameter instead of type member and make Abstract0 into a trait:

trait Abstract0[+Self <: Abstract0[_]] {
  // ...
}

class Concrete0 extends Abstract0[Concrete0] {
  // ....
}

class RefinedConcrete0 extends Concrete0 with Abstract0[RefinedConcrete0] {
 // ....
}