Normalize parameter environment with its own substitution

[edwardw]

Otherwise, the substitution is carried out from outside perspective,
which would be against the very purpose of ParameterEnvironment.

Closes #21750
Closes #22066
Closes #22077

[rust-highfive]

r? @pcwalton

(rust_highfive has picked a reviewer for you, use r? to override)

[edwardw]

r? @nikomatsakis

[nikomatsakis]

Nit: this should not need move

[nikomatsakis]

So I did some digging here. I think what's going wrong is that we wind up creating a region variable during environment normalization and then we don't replace it with its inferred variable. This is a pretty big oversight on my part! I'm trying to figure out the correct fix here, but I don't think that applying the substitutions in this way is right. The best thing would be to fixup the parameter environment normalization to replace the region variables with an inferred result, but my naive attempt to do so revealed that there were not enough constraints getting generated to yield reasonable results. I need to dig further to understand why not and what needs to be done to fix it.

[edwardw]

No more dubious "ambient substs", but otherwise the patch employes the same approach. So this is still FYI, I guess.

[edwardw]

r? @nikomatsakis
I think I've figured it out.

[nikomatsakis]

So I've got a rather different patch. The branch is here:

https://github.com/nikomatsakis/rust/tree/issue-21750-normalization-with-regions

It's a bit rough atm but it also fixes #21750/bootstraps etc. Basically what I think is going wrong with #21750 is that I forgot that although we fully resolve the type variables in the param env, we AREN'T replace the region variables. That is quite the no-no, because those region variables are created in a temporary inference context that we then throw away, leading to ICEs and other nonsense. The problem is that just resolving the region variables fails, because the results are inferred to be empty regions. This is due to the way that variance and associated types interact at present. The problem is that the Self parameter is contravariant in most cases, and hence if Self is &'a T, when we project out we wind up with a region variable '0 and the constraint that 'a : '0 -- basically we know that '0 cannot be larger than 'a, but we have no lower bound on '0, and hence we infer it to an empty region.

I think that the current variance rules are sound, but they are not perhaps super helpful. This has bitten me before. In that patch I've modified the variance inference so that any trait which defines associated types is effectively invariant with respect to its inputs. This solves the problem. (It also fits well with the subtyping rules for projected types, which require an exact match on the trait reference being projected out of.) The patch still needs to be cleaned up to handle errors better. I also would like to draw up some examples showing why I think this variance treatment for assoc types is superior.

[edwardw]

Ah, thanks for the very detailed explanation. Feel free to use test cases from this patch that you think useful.