Pattern matching on unchecked generic types (with isInstanceOf guard) does not work correctly

Compiler version

scala 3.1.1

Minimized code

trait Fruit
case class Apple() extends Fruit
case class Orange() extends Fruit

case class Box[C](fruit: C) extends Fruit

val apple = Box(fruit = Apple())
val orange = Box(fruit = Orange())


val result = List(apple, orange).map {
  case appleBox: Box[Apple] @unchecked if appleBox.fruit.isInstanceOf[Apple] => //contains apple
    "apple"
  case _ => 
    "orange"
}

assert(result == List("apple", "orange"))

Output

result == List("apple", "apple")

Expectation

Only Box with Apple inside should be matched as “apple”. It works for 2.13.x scala version. It can be workarounded with

case appleBox: Box[_] @unchecked if appleBox.fruit.isInstanceOf[Apple] =>
  val a = appleBox.asInstanceOf[Box[Apple]]

[som-snytt]

I think the Scala 3 behavior is correct, but on the isInstanceOf, it should warn,

fruitless type test

with the correct pun on fruit.

[sjrd]

1. Behave like Scala 2 and don’t optimize isInstanceOf checks, at least not if the scrutinee has an @unchecked type.

If it is only “if the scrutinee has an @unchecked type”, it means that

case appleBox: Box[Apple] @unchecked if appleBox.fruit.isInstanceOf[Apple] =>
  1
case _ =>
  2

is not equivalent to

case appleBox: Box[Apple] @unchecked =>
  if appleBox.fruit.isInstanceOf[Apple] then
    1
  else
    2
case _ =>
  2

anymore. Is that acceptable?

In other words, it’s not really a question of optimizing or not. It’s a question of changing the specified semantics of types inside the guard of a pattern match that has @unchecked. Do you want to put that in the spec? We would have to say something like, in the guard of a case, we don’t trust the types beyond their erasure. What does that even mean in terms of type system? Can we still remotely reason about soundness in that situation? What about GADT soundness in this context? There is so much we don’t know once we “relax” run-time semantics like that. Previously, I could explain the meaning of @unchecked in one sentenced (see my previous comment); now I need to make sweeping changes to my understanding of the type system.