IncompatibleClassChangeError on inner class java.lang.Class

[nicolasstucki]

object Test {
  def main(args: Array[String]): Unit = {
    class Foo
    classOf[Foo].getDeclaringClass
  }
}

fails with

Exception in thread "main" java.lang.IncompatibleClassChangeError: Test and Test$Foo$1 disagree on InnerClasses attribute
	at java.lang.Class.getDeclaringClass0(Native Method)
	at java.lang.Class.getDeclaringClass(Class.java:1235)
	at Test$.main(Foo.scala:4)
	at Test.main(Foo.scala)

It works on scalac.

[smarter]

This is probably related to all the changes to the InnerClass table generation done by @lrytz in the scalac backend since the dotty backend was forked (September 2014), good luck figuring it out: https://github.com/scala/scala/blob/2a88024236cdab266cabe11414120a6af796a56f/src/compiler/scala/tools/nsc/backend/jvm/BTypes.scala#L333-L588

[farquet]

For reference, this issue is a blocker for generating GraalVM native images for code generated by the dotty compiler. GraalVM static analysis relies on those InnerClasses attributes to be correct to be able to work.

The scala compiler had a similar issue which got fixed in scala 2.12.7. scala/bug/issues/10487

[prolativ]

Not sure if this should be treated as a separate issue or as a part of this one but while trying to construct a more complex test case for this I found out that there is a discrepancy between scala 2 and 3 about how class files are generated.
In this case

object Test {
  class Quux

  def main(args: Array[String]): Unit = {
    class Foo {
      class Bar

      println(classOf[Bar].getDeclaringClass)

      def test() = {
        class Baz
        println(classOf[Baz].getDeclaringClass)
      }
    }

    println(classOf[Foo].getDeclaringClass)
    val foo = new Foo
    foo.test()
  }
}

as the result of transformations done in lambdaLift and flatten phases Baz turns into Test$Foo$1$Baz$1 in scala 2 but into Test$Baz$1 in scala 3. Is this expected/acceptable? Is there a particular reason why this got changed? Taking into account that in both cases Quux becomes Test$Quux and Foo becomes Test$Foo$1, the scheme used in scala 2 seems more regular although it produces class files with longer names which might cause problems in some extreme cases (e.g. when one uses a Linux with an encrypted disk partition).
@smarter @nicolasstucki any ideas?

[smarter]

Is this expected/acceptable?

Yes, those are just implementation details.

it produces class files with longer names which might cause problems in some extreme cases (e.g. when one uses a Linux with an encrypted disk partition).

That won't be an issue once #5812 is fixed.

[propensive]

In Scala 2, at least, I believe that very long names should have their "middle" replaced by a hash (MD5, I think), so they are never longer than a certain length. It's the "middle" so that the few characters at the start and end remain unencrypted and readable.

[farquet]

Awesome!
Looking forward to test dotty native image compatibility once 3.0.0-RC2 is out

[smarter]

If you wish to try it out early, the fix should be part of the latest nightly: https://repo1.maven.org/maven2/org/scala-lang/scala3-compiler_3.0.0-RC2/3.0.0-RC2-bin-20210310-4af1386-NIGHTLY/