gistfile1.scala

object Demo {
  // A couple of type classes with type members ...
  trait Foo[T] {
    type A
  }

  object Foo {
    implicit val fooIS = new Foo[Int] { type A = String }
  }

  trait Bar[T] {
    type B
    val value: B
  }

  object Bar {
    implicit val barSB = new Bar[String] {
      type B = Boolean
      val value = true
    }
  }

  // What we want to write ...
  //
  //  def run[T](t: T)(implicit foo: Foo[T], bar: Bar[foo.A]): bar.B = bar.value
  //
  // or maybe ...
  //
  //  def run[T](t: T)(implicit foo: Foo[T])(implicit bar: Bar[foo.A]): bar.B = bar.value
  //
  // but can't ... in the first case the compiler complains about a dependent type (foo.A)
  // appearing in the same parameter block as its prefix (foo); in the second the compiler
  // chokes on the multiple implicit parameter blocks.

  // But we can encode the above with the help of singleton types ...

  // SingletonOf[T, U] represents an implicit value of type T narrowed to its
  // singleton type U.
  case class SingletonOf[T, U](value: U)
  object SingletonOf {
    implicit def mkSingletonOf[T <: AnyRef](implicit t: T): SingletonOf[T, t.type] = SingletonOf(t)
  }

  // The implicit resolution of SingletonOf[Foo[T], fooT] will result in the type
  // fooT being inferred as the singleton type of the in-scope Foo[T] value.
  // We then rely on the equivalence between,
  //
  //   foo.A
  //
  // and,
  //
  //   foo.type#A
  //
  // to rewrite the problematic dependently chained parameter block to a form
  // that scalac is happy to digest ...
  def run[T, fooT <: { type A }](t: T)
    (implicit sFoo: SingletonOf[Foo[T], fooT], bar: Bar[fooT#A]): bar.B = bar.value

  val value = run(23)
  assert(value: Boolean)
}