Imported From: https://issues.scala-lang.org/browse/SI-4819?orig=1
Reporter: @gkossakowski
Affected Versions: 2.9.1

@gkossakowski said:
I gave it another look and it seems like the root cause is the same as #4820. To keep story short, constructor of test.Test$Y$ refers to A parameter because it takes Test instance as it's argument <init>(Ltest/Test;)V.

Now, test.Test$Y$ is not declared as inner class in InnerClasses attribute so from ecj's point of view A is not in scope. I guess that fix for #4820 will fix this issue too.

@paulp said:
There are several bugs, I'm finding, but with your tests to draw the lines for me the resolutions are making themselves apparent. The major remaining unknown is whether everything will reconcile nicely once we throw in the inner class issues of yesteryear.

Commit Message Bot (anonymous) said:
(grek in r25639) Fix various InnerClasses bugs.

This commit fixes two major problems:

1. InnerClasses table missed entries
   that would close the chain between
   nested and top-level class.

2. In some situations, classes
   corresponding to objects would be
   not be reported in the InnerClasses
   table.

For details it's the best to check #4819, #4820 and
#4983.

First problem mentioned above was straightforward to
fix so I won't be going into details.

The second one deserves more attention. From now, classes
corresponding to objects are properly reported as inner
classes. Also, members (classes, objects) of objects are
reported as inner classes of classes corresponding to
objects.

There's one caveat though: top level objects get two
classes (regular and mirror). Members of top-level
objects are declared as inner classes of mirror class
and not regular one. The reason for that is to allow
importing them from Java. For example:

object A { class B }

will be compiled into following classes: A, A$, A$B.
If we declared A$B as inner class of A$ (regular class
for objects) then it would be impossible to import
B using "import A.B" or "import A$.B" constructs. The
reason for that is that Java compiler seems to blindly
put dollars instead of looking at InnerClasses attribute.

Since non-top-level objects don't have a mirror class
it's impossible to use the same solution. Thus, in case
like this:

object A { object B { class C } }

it's impossible to import C from Java. That's the tradeoff
for fixing other (more serious) problems. It's never been
possible to do that in a clean way so we are not making
situation worse.

As a nice consequence of this change, we get better way to
refer to inner members of top-level objects. It's been
reflected in one of test-cases that is updated by this
change.

Fixes #4789 #4819 #4820 #4983 and possibly some
other tickets related to reflection.

Review by extempore, dragos.