NLL and Borrowck dataflow disagree on the meaning of a location

[nikomatsakis]

In the NLL regions, if a region contains a location L, that means that the region contains the entry point to that location. In the borrow dataflow, however, when we apply statement effects, they take place on the exit from a location. The end result is that kills for borrows are applied slightly too late. Consider something like:

L1: P = &X;
L2: use(P); // last use
L3: X = 1;

The region for P would be inferred to {L1, L2}. Therefore, the kill of L2 needs to include the borrow on line 1. However, what we actually do is to insert the kill after L3.

I noticed this because I was trying to stop emitting EndRegion statements when using NLL. However, this caused false errors to be omitted. For now, there is a FIXME in the code.

[nikomatsakis]

Sample test case that was failing: src/test/ui/nll/ty-outlives/projection-implied-bound.rs

[nikomatsakis]

After some review of the code and discussion with @pnkfelix, we proposed generalizing the dataflow API slightly to handle this. The idea is to have:

• statement_effect(statement) means the effect a statement has
• terminator_effect(terminator) means the effect of the terminator up to -- but not including! -- the branches (also kind of as today)
• edge_effect(terminator, successor index) means the effect of a particular outgoing edge
  • today we have this, but only for one edge -- the "non-unwind" edge of a call

[arielb1]

In addition to adding edge effects, another possible fix would be to make the "out of scope" effect a "pre-statement" effect, that is evaluated by FlowAtLocation for the location.