Separate compilation issue because PrivateParamAccessor are only made non-private in TreeTransformer, not DenotTransformer

[smarter]

A.scala:

class A(val member: Int)

class SubA(member: Int) extends A(member)

B.scala:

class B(member: Int) extends SubA(member)

If you try to compile these two files together everything works fine, but if you first compile A.scala and then B.scala you get:

Exception in thread "main" java.lang.AssertionError: assertion failed: private method member in class SubA in try/exp/A.scala accessed from method member in class B in try/exp/B.scala
        at scala.Predef$.assert(Predef.scala:165)
        at dotty.tools.dotc.transform.ExpandPrivate.ensurePrivateAccessible(ExpandPrivate.scala:84)

The issue is that ExpandPrivate only transforms PrivateParamAccesor in trees:

  override def transformDefDef(tree: DefDef)(implicit ctx: Context, info: TransformerInfo) = {
    val sym = tree.symbol
    tree.rhs match {
      case Apply(sel @ Select(_: Super, _), _)
      if sym.is(PrivateParamAccessor) && sel.symbol.is(ParamAccessor) && sym.name == sel.symbol.name =>
        sym.ensureNotPrivate.installAfter(thisTransform)
      case _ =>
        if (isVCPrivateParamAccessor(sym))
          sym.ensureNotPrivate.installAfter(thisTransform)
    }
    tree
  }

So when we compile A.scala and B.scala together, the denotation for SubA#member gets transformed as expected, but when we do separate compilation, we see SubA#member as private, so ExpandPrivate explodes.
Note: see the comment at the top of ExpandPrivate if you're wondering why private param forwarders need to be made non-private.

I'm unsure how to fix this since it seems that we need access to the tree to decide whether or not the param forwarder should be made non-private. Maybe we need an extra flag to signal that a param forwarder forwards to an inherited param accessor with the same name?

This issue is blocking the Super Bootstrap

[DarkDimius]

A possible fix would be not make paramaccessors private, but name-mangle them instead while keeping them public.

[smarter]

@DarkDimius I'm not sure I understand what you mean. This is what the ensureNotPrivate in transformDefDef does already. Do you mean that we could make paramaccessors never private, even if they don't forward to a same-named param accessor? I agree that this would fix the issue.

[odersky]

Yes, I think making them not private to start with is the way to go.

[smarter]

@odersky OK, should EnsureNotPrivate be changed to always make them public, or should this be done before?