Allow package private classes to change binary API

[dwijnand]

Refs #53

[SethTisue]

is there substantial copy-and-paste here as you'd feared earlier today, or did it end up being fresh code?

if this PR opens up a whole new space of possibilities for MiMa features, as we discussed, then some coordinated doc updates (maybe even just a paragraph or two in the README?) might be appropriate to include?

[dwijnand]

I stripped scalac 2.10's UnPickler, which (I realised after submitting) won't work against scalac 2.11/2.12.

This adds a reasonable start for extracting more out of the pickler info (as in you need to get it to pull out more things for you) but it doesn't, for instance, even close #53 because it only does the class, not its methods...

[lrytz]

I'm not sure we should go in that direction (starting to read scala signatures).

Somewhat related: it's really difficult to create a correct mapping between the symbol table and the classfile level (name-mangled method names, anonymous classes, etc). This was the major source of issues for the Scala inliner before 2.12.

I think MiMa should only work on the bytecode. However, we could think of new channels to pass information from the compiler to MiMa. For example we could have make the Scala compiler emit annotations on methods that were made public by the compiler. We could also have an annotation in the standard library that people could use for classes/members that MiMa should ignore (an alternative to the mima config - maybe easier to use/manage?).

Disclaimer: I don't know enough about how MiMa works in detail, I might be missing something! What kind of exceptions does it currently have, for example to not fail for anonymous function classes, or foo$anonfun methods?

[dwijnand]

I'm keen to have a solution for current versions of Scala. If we want to add a more robust, or in other ways better, solution in future Scala, then we can migrate future MiMa to that. But I'm guessing reading scala signatures is the best we can do right now (outside of just not ship the features/enhancements).

Of course, I'm happy to find a happy compromise on how it's implemented (e.g how much to re-use or not re-use out of scalac, and what to duplicate in MiMa, how to handle scalac differences, etc).

If it's easier we can jump in a call to discuss, once I'm back in the office.

[dwijnand]

This change is a hacky rush, needs fixing to land, and there's some concern.

I don't have the bandwidth to land it.