Exploring inner modules for binary incompatibilities. Fix #127

[dotta]

The former implementation of PackageInfo#accessibleClasses expects that for any
module Foo compiled by scalac, both Foo.class and Foo$.class classfiles
are produced. Because Foo$.class and Foo.class are essentially mirrors, and
users never access directly Foo$.class, MiMa only traverses the Foo.class. This
works well for top-level modules, but it breaks for inner modules, because no
mirror class is created for inner modules (only the module class is created).

The fix consist in treating inner modules differently, hence making sure inner
modules are part of the set returned by PackageInfo#accessibleClasses.

[dotta]

@adriaanm @lrytz @retronym Could you guys have a quick look?

[retronym]

Because Foo$.class and Foo.class are essentially mirrors, and
users never access directly Foo$.class, MiMa only traverses the Foo.class

Could you expand on this with an example? It seems incorrect to me. scalac adds static forwarders to the class O for a subset of the members of O$.

For example:

⚡ printf "class Fooey { def foo = 42 }; class O extends Fooey; object O { def foo(x: Any) = x }" > /tmp/test.scala; scalac /tmp/test.scala; javap -classpath . O 'O$'
Compiled from "test.scala"
public class O extends Fooey {
  public O();
}
Compiled from "test.scala"
public final class O$ {
  public static final O$ MODULE$;
  public static {};
  public java.lang.Object foo(java.lang.Object);
}

O$.foo(Object) does not have a forwarders in O because O inherits a member with the same name.

[retronym]

It might be more robust to check the InnerClasses attribute of the module class.

⚡ printf 'object N { class O }; class C { object O }' > /tmp/test.scala && scalac /tmp/test.scala && javap -classpath . -v 'C$O$' 'N$O' 2>&1 | grep --after=1 InnerClasses:
InnerClasses:
     public #17= #2 of #16; //O$=class C$O$ of class C
--
InnerClasses:
     public static #14= #2 of #13; //O=class N$O of class N

[dotta]

Hi Jason, thanks a lot for the feedback!

Because Foo$.class and Foo.class are essentially mirrors, and
users never access directly Foo$.class, MiMa only traverses the Foo.class

Could you expand on this with an example? It seems incorrect to me. scalac adds static forwarders to the class O for a subset of the members of O$.

You are absolutely right. And, in fact, my statement "MiMa only traverses the Foo.class" is not false: MiMa does traverse both Foo.class and Foo$.class exactly because of what you said. Hence, the fix in this PR isn't good. As you pointed out in your review, the classfile attribute InnerClasses should be used instead.

Unfortunately, I don't have more spare cycles to dedicate to this at the moment. I'm closing this PR.