Implement friendly front-end format

[lqd]

Implement the friendlier front-end format described in issue #26 — destined to help create isolated tests without needing rustc NLL facts, but still generating the same data structure.

Here's how it looks:

// program description
universal_regions { 'a, 'b, 'c }

// block description
block B0 {
    // effects are by default emitted on the Mid point
    invalidates(L0);

    // but we can still emit some at the Start point if needed
    // by separating the Start point effects from the Mid point effects with a "/"
    invalidates(L1) / kill(L2);

    goto B1;
}
block B1 {
    use('a), outlives('a: 'b), borrow_region_at('b, L1);
}

[nikomatsakis]

This looks good, not quite there — do you think we should land and iterate in tree? Or would you prefer to hack on the PR a bit more?

[nikomatsakis]

Hmm, so for region_live_at, we have to do more — we need to actually compute liveness I think, but I realize that to do that we need to also have def('a), of course (and maybe we should make both def and use accept multiple lifetimes). We could also try to capture the idea of variables with types, though that is creeping towards MIR and I think I had hoped to do that as a second pass.

In other words, as an example:

bb0 {
  def('a); // statement 0
  ; // statement 1
  goto bb1;
}

bb1 {
  ; // statement 0
  use('a); // statement 1
}

This ought to make 'a live at all the following points:

• start(bb0/1),
• mid(bb0/1),
• start(bb1/0),
• mid(bb1/0),
• start(bb1/1),
• mid(bb1/1).

The idea would be: some variable X whose type includes 'a is used at mid(bb1/1). So, 'a is live at that point, because the current value of X matters -- this flows backwards until we reach mid(bb0/0), at which point X was assigned, so its current value doesn't matter.

[nikomatsakis]

We can of course use datafrog to do this computation =) you can imagine us emitting use and def facts and setting up a rule:

region_live_at(R, P) :- use(R, P).
region_live_at(R, P) :-
  use(R, Q),
  cfg_edge(P, Q),
  !def(R, P).

[nikomatsakis]

We also could potentially refactor the compiler to emit use and def facts and not region_live_at facts.

[nikomatsakis]

I would rather factor Effect into nested enums so that this case never happens.

[lqd]

Alright I'll fix this and reopen this later tonight