Regression in playframework/playframework for access to private static class public members

[WojciechMazur]

Based on Open CB failure in playframework/playframework - builds logs

Compiler version

3.6.0-RC1-bin-20240915-ad8c21a-NIGHTLY
Bisect points to e7d479f / #21362

Minimized code

import java.util.List;

public class RoutingDsl {

  final List<Route> routes = null;

  static class Method {}
  private static class Route {
    final Method actionMethod = null;
  }
}

@main def Test =
  val x: RoutingDsl = ???
  x.routes.forEach: route =>
    println(route.actionMethod) // error

Output

Compiling project (Scala 3.6.0-RC1-bin-20240915-ad8c21a-NIGHTLY, JVM (17))
[error] ./test.scala:4:13
[error] Unable to emit reference to value actionMethod in class Route, class Route is not accessible in the top-level definitions in package <empty>
[error]     println(route.actionMethod)
[error]             ^^^^^

Expectation

Should compile. This code does compile in Scala 3.5.0 and lower. It also compiles in Scala 2.12/2.13

[bracevac]

IMHO, this is the correct behavior for private static access in Java. E.g., we'll get a similar error message with an equivalent Java client program:

public class Main {
    public static void main(String[] args) {
        RoutingDsl x = null;
        for(RoutingDsl.Route route : x.routes) {
                System.out.println(route.actionMethod);
        }
    }
}

which results in

Main.java:4: error: Route has private access in RoutingDsl
        for(RoutingDsl.Route route : x.routes) {
                      ^
1 error

So it looks like we got it wrong in Scala 3 until #21362 which is to say, that playframework would have to fix that on their end.

Edit: The following is closer to the Scala version, but my point remains the same:

public class Main {
    public static void main(String[] args) {
        RoutingDsl x = null;
        x.routes.forEach(route -> System.out.println(route.actionMethod));
    }
}

[lrytz]

@bracevac I think one could argue both ways.

If the Route class is explicitly referenced in the Scala source, all Scala versions issue a "cannot be accessed" error.

However, in the example source code there's no reference to Route. The class is public in bytecode and the generated GETFIELD RoutingDsl$Route.actionMethod bytecode is valid. So the error message "Unable to emit reference to value actionMethod" is not really accurate.

I haven't looked in detail how Scala 2 makes the difference between the two cases, it could even be accidental / a bug.

To make an argument for your side, replacing RoutingDsl.Route route by var route in the equivalent Java program still causes an error in javac.

[dwijnand]

The class is public in bytecode and the generated GETFIELD RoutingDsl$Route.actionMethod bytecode is valid.

I wonder why javac emits the class as public. So it can be instantiated by RoutingDsl?

[lrytz]

class is public in bytecode

I was wrong here, the nested class has default access (package protected). Scala 2 is still correct, GETFIELD RoutingDsl$Route.actionMethod is OK within the same package. Moving the RoutingDsl into a different package, Scala 2 gives errors on compilation ("cannot emit reference").

Dale and me tried a simpler example:

// package p
public class A {
  public static B b() { return new B(); }
  private static class B {
    public String bar() { return "l"; }
  }
}

object T {
  def main(args: Array[String]): Unit = {
    println(/* p. */ A.b.bar)
  }
}

Again, no reference to the private B in the Scala file, but this time Scala 2 doesn't compile it ("Unable to emit reference"). So Scala 2 is inconsistent.

Scala 3.5 compiles the example even when moving A.java to a different package, which is clearly a bug, things then fail at runtime (IllegalAccessError).

Overall, I htink enforcing access as current main does is fine.