Summary:
This implements most/all of the difficult parts of Diviner on top of Phabricator instead of as standalone components. See T988. In particular, here are the things I want to fix:
**Performance** The Diviner parser works in two stages. The first stage breaks source files into "Atoms". The second stage renders atoms into a display format (e.g., HTML). Diviner currently has a good caching story on the first step of the pipeline, but zero caching in the second step. This means it's very slow, even for a fairly small project like Phabricator. We must re-render every piece of documentation every time, instead of only changed documentation. Most of this diff concerns itself with addressing this problem. There's a fairly large explanatory comment about it, but the trickiest part is that when an atom changes, other atoms (defined in other places) may also change -- for example, if `class B extends A`, editing A should dirty B, even if B is in an entirely different file. We perform analysis in two stages to propagate these changes: first detecting direct changes, then detecting indirect changes. This isn't completely implemented -- we need to propagate 'extends' through more levels -- but I believe it's structurally correct and good enough until we actually document classes.
**Inheritance** Diviner currently has a very weak story on inheritance. I want to inherit a lot more metas/docs. If an interface documents a method, we should just pull that documentation in to every implementation by default (implementations can still override it if they want). It can be shown in grey or something, but it should be desirable and correct to omit documentation of a method implementation when you are implementing a parent. Similarly, I want to pull in inherited methods and @tasks and such. This diff sets up for that, by formalizing "extends" relationships between atoms.
**Overspecialization** Diviner currently specializes atoms (FileAtom, FunctionAtom, ClassAtom, etc.). This is pretty much not useful, because Atomizers (which produce the atoms) need to be highly specialized, and Renderers/Publishers (which consume the atoms) also need to be highly specialized. Nothing interesting actually lives in the atom specializations, and we don't benefit from having them -- it just costs us generality in storage/caches for them. In the new code, I've used a single Atom class to represent any type of atom.
**URIs** We have fairly hideous URIs right now, which are very cumbersome For in-app doc links, I want to provide nice URIs ("/h/notfications" or similar) which are stable redirects, and probably add remarkup for it: !{notifications} or similar. This diff isn't related to that since it's too premature.
**Search** Once we have a database generation target, we can index the documentation.
**Design** Chad has some nice mocks.
Test Plan: Ran `bin/diviner generate`, `bin/diviner generate --clean`. Saw appropriate graph propagation after edits. This diff doesn't do anything very useful yet.
Reviewers: btrahan, vrana
Reviewed By: btrahan
CC: aran
Maniphest Tasks: T988
Differential Revision: https://secure.phabricator.com/D4340
