Strange behaviour related to type test of path-dependent types and divergence from Scala 2

[liufengyun]

Compiler version

3.5.0-RC1-bin-20240506-1cdf99f-NIGHTLY

Minimized code

trait CmdLineParser { outer =>

  trait Opt[+T] {
    val default: T
    val names: Set[String]
    val help: String
  }

  trait IntOpt extends Opt[Int] {
    val parser = outer                          //       <=== comment out this line, we get "true true"
  }
}

object FirstParser extends CmdLineParser {
  object OptMinSuccess extends IntOpt {
    val default = 100
    val names = Set("bla")
    val help = "bla"
  }

  val opts = List(OptMinSuccess)
}

object SecondParser extends CmdLineParser {
  object OptMinSuccess extends IntOpt {
    val default = 50
    val names = Set("bla")
    val help = "help"
  }
}


val a = SecondParser.OptMinSuccess.isInstanceOf[FirstParser.IntOpt]

println(a)

(SecondParser.OptMinSuccess: SecondParser.IntOpt)  match {
  case _: FirstParser.IntOpt => println("true")
  case _ => println("false")
}

Output

true
false

If we comment out the following line in trait IntOpt:

  trait IntOpt extends Opt[Int] {
    // val parser = outer                          //       <=== comment out this line, we get "true true"
  }

We get

true
true

In contrast, Scala 2 reports a compilation error for the pattern match in both versions:

pattern type is incompatible with expected type;
 found   : Playground.FirstParser.IntOpt
 required: Playground.SecondParser.IntOpt

Expectation

• We would expect the output to be always the same, ideally false false
• Change code in a trait should not change type test result

[odersky]

What does Scla 2 give for the isInstanceOf?

[odersky]

The problem that IntOpt with the member commented out does not get an outer pointer, and we cannot change that.

We could restrict the type system to prevent the match as Scala 2 does. But it could be easily circumvented by widening:

(SecondParser.OptMinSuccess: Any)  match {
  case _: FirstParser.IntOpt => println("true")
  case _ => println("false")
}

So I think it's best to verify that we have an outer pointer and if that's not the case we issue a warning that the full type cannot be tested at run time.

[liufengyun]

What does Scala 2 give for the isInstanceOf?

It returns true.

The problem that IntOpt with the member commented out does not get an outer pointer, and we cannot change that.

I was surprised to find my understanding of outers in Scala incorrect. Is this behavior specified somewhere? Or it's just an optimization that has become de facto standard? /cc: @sjrd

[odersky]

It would be quite inefficient if every inner trait needed an outer pointer. So it gets one only if the body references an outer member. I don't know whether that's specced.