Improve binary compatibility for inner classes that lose the outer pointer

[retronym]

$ cat sandbox/A_1.scala
class C {
  class Inner
}

object Test {
  def main(args: Array[String]): Unit = {
    val c = new C
    val i = new c.Inner
    (i: Any) match {
      case _: c.Inner =>
        println("matched `case _: c.Inner`")
    }
  }
}

$ cat sandbox/A_2.scala
class C {
  final class Inner
}

$ scalac sandbox/A_1.scala && scala Test && scalac sandbox/A_2.scala && scala Test
matched `case _: c.Inner`
java.lang.NoSuchMethodError: C$Inner.C$Inner$$$outer()LC;
	at Test$.main(A_1.scala:10)
	at Test.main(A_1.scala)

Inner classes marked final do not store the outer pointer unless it is captured. They still have the outer parameter in their constructor, so new c.Inner is binary compatible in the example above.

  // access flags 0x1
  public main([Ljava/lang/String;)V
    // parameter final  args
    NEW C
    DUP
    INVOKESPECIAL C.<init> ()V
    ASTORE 3
    NEW C$Inner
    DUP
    ALOAD 3
    INVOKESPECIAL C$Inner.<init> (LC;)V

But type tests in pattern matching and isInstanceOf that check prefixes have a binary fragility.

    ASTORE 4
    ALOAD 4
    ASTORE 5
    ALOAD 5
    INSTANCEOF C$Inner
    IFEQ L0
    ALOAD 5
    CHECKCAST C$Inner
    INVOKEVIRTUAL C$Inner.C$Inner$$$outer ()LC;
    ALOAD 3
    IF_ACMPNE L0
    GETSTATIC scala/Predef$.MODULE$ : Lscala/Predef$;
    LDC "matched `case _: c.Inner`"

The runtime prefix check is elided if it can the prefix of the scrutinee is statically the same as that of the tested type, hence the upcast to Any in this test case.

We could make this more binary compatible by still generating the $outer accessor in final inner classes (it would return null) and treating a null prefix in the scrutinee as a wildcard for the prefix check.