[nll] optimize places_conflict... somehow

[nikomatsakis]

places_conflict currently accounts for 8% of MIR borrowck time. Most of that time, based on perf annotate results, seems to be spent in and around the SmallVec operations (e.g., push and into_iter). The tricky part here is that the Place data structures we are using are setup as a kind of tree, so that x.a.b is sort of like this:

• Projection of b from:
  • Projection of a from:
    • Local variable x

But we want to iterate "bottom up", so x, then a, then b. This doesn't lend itself to a simple iterator so we currently build up a SmallVec of results.

It might be a win instead to build up a (stack-allocated) linked list by recursively "unrolling" the paths for both sides. Or perhaps dong other experiments, such as sharing a Vec buffer (that seemed a bit tricky from a type perspective when I looked into it).

Longer term, I'd like to move away from using Place here (and perhaps change how Place is setup in MIR) but it seems worth trying to tackle this in a targeted way.

[nikomatsakis]

I'm going to take a stab at doing this, since I think my description will be nigh incomprehensible to anyone else.