Revisit the "Location::All" hack

[nikomatsakis]

As part of the nll-facts borrowck work, I added the idea of outlives relation that occur at all points. I did my best to explain it here:

https://github.com/rust-lang/rust/blob/0cd465087d800f6e1183ed3e8a71262039245902/src/librustc_mir/borrow_check/nll/type_check/mod.rs#L627-L657

I handled this by adding the relevant outlives relationships to all points. But, looking at the outlives.facts I saw that these relations account for a LOT of noise. Of the ~500K tuples in that file, about half are due to this. =) Moreover, these tuples will mess up any optimizations like #20. It's clear we need a better approach, but I'm not sure what.

That said, manually removing those tuples and re-running the analysis shows that they only have a small effect on overall performance (from ~40s to ~35s).

[nikomatsakis]

In my branch nll-alias-analysis-no-loc-call, I take a different approach that removes Location::All. However, I suspect it will have some "not obviously desirable" effects.

In particular, if you "reassign" a variable that has a type annotation, you get "fresh regions":

fn foo(mut x: &'static u32) {
  let y = 22;
  x = &y; // OK
  // ... here, `x` is not considered `'static` ..
}

This falls out from trying to be smart about examples like:

let mut p = &x;
let q = p;
p = &y;

I'm inclined to think we should run with it for now, but we'll have to revisit this issue prior to stabilization.

[nikomatsakis]

Preliminary PR posted in rustc to change how our input facts are organized:

rust-lang/rust#50938