Internally change regions to be covariant

[lcnr]

THIS ISSUE IS ONLY ABOUT AN IMPLEMENTATION DETAIL OF RUSTC, IT DOES NOT AFFECT THE RUST LANGUAGE.

Right now regions in rustc are considered to be contravariant wherever they are used:

rust/compiler/rustc_middle/src/ty/relate.rs

Lines 447 to 454 in cdd7afe

	let region_bound = relation.with_cause(Cause::ExistentialRegionBound, |relation| {
	relation.relate_with_variance(
	ty::Contravariant,
	ty::VarianceDiagInfo::default(),
	a_region,
	b_region,
	)
	})?;

rust/compiler/rustc_middle/src/ty/relate.rs

Lines 492 to 497 in cdd7afe

	let r = relation.relate_with_variance(
	ty::Contravariant,
	ty::VarianceDiagInfo::default(),
	a_r,
	b_r,
	)?;

This is confusing and different from outlives, as 'a subtype 'b currently means 'b outlives 'a. Whenever we switch between "outlives" and "subtyping" for regions, we have to swap the order, e.g.

rust/compiler/rustc_infer/src/infer/outlives/env.rs

Lines 137 to 148 in cdd7afe

	OutlivesBound::RegionSubParam(r_a, param_b) => {
	self.region_bound_pairs
	.insert(ty::OutlivesPredicate(GenericKind::Param(param_b), r_a));
	}
	OutlivesBound::RegionSubProjection(r_a, projection_b) => {
	self.region_bound_pairs
	.insert(ty::OutlivesPredicate(GenericKind::Projection(projection_b), r_a));
	}
	OutlivesBound::RegionSubOpaque(r_a, def_id, substs) => {
	self.region_bound_pairs
	.insert(ty::OutlivesPredicate(GenericKind::Opaque(def_id, substs), r_a));
	}

We've previously discussed this in https://rust-lang.zulipchat.com/#narrow/stream/144729-t-types/topic/relating.20regions.20variance where we ended up agreeing that the variance should be changed to covariance for regions.

Doing so is non-trivial as my first attempt at this ended up missing at least one place resulting in hard to diagnose errors in core. I still want to see this fixed and would be up to review or mentor any work here. This requires some familiarity with the compiler (some knowledge about borrowck and/or the lexical regionck would be helpful) to correctly update relevant comments when doing this change.

edit: we also have to update the crate_variances query for this to work. probably that was the issue I encountered

[aliemjay]

@rustbot claim

[lcnr]

see the discussion starting in #107339 (comment)

we still want to remove the current meaning of Sub inside of the region solver where 'a <= 'b if 'b: 'a