Properly pass down immutability info for thread-locals.

[EdSchouten]

For thread-locals we call into cat_rvalue_node() to create a CMT
(Category, Mutability, Type) that always has McDeclared. This is
incorrect for thread-locals that don't have the 'mut' keyword; we should
use McImmutable there.

Extend cat_rvalue_node() to have an additional mutability parameter. Fix
up all the callers to make use of that function. Also extend one of the
existing unit tests to cover this.

Fixes: #47053

[rust-highfive]

r? @michaelwoerister

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

[eddyb]

r? @nikomatsakis

This is only needed because of how #[thread_local] is hackily handled in the old borrowck.
Does the MIR borrowck treat references to #[thread_local] statics as to function-local temporaries?

[nikomatsakis]

This is indeed fixed in MIR borrowck, as you can see from this example. It may still be worth fixing in AST borrowck, if it's easy -- we're still going to need some version of the mem-cat / expr-use-visitor code for upvar inference, though I hope we'll be able to greatly simplify it.

[nikomatsakis]

This is only needed because of how #[thread_local] is hackily handled in the old borrowck.

OK, I think I see what you are saying. I don't think I realized how this was being handled. Basically I guess thread-locals are being converted into rvalues so that borrowck will assign a more limited lifetime? That does indeed feel like rather a hack.

[nikomatsakis]

Maybe I should take a look at what it might mean to integrate more "properly" into borrowck? It feels like that would likely be very easy to do (i.e., maybe even a smaller diff than what is here).
Or @EdSchouten if you feel like digging you can do some explring on your own. =)

I think the main thing is that this line has to change:

rust/src/librustc_borrowck/borrowck/gather_loans/lifetime.rs

Lines 83 to 85 in da569fa

	Categorization::StaticItem => {
	Ok(())
	}

This basically says that a loan of static content can have any lifetime. But in the case of a static item tagged with #[thread_local], we would want to check that the loan lifetime does not exceed current function (mostly, it should be a ReScope, but we might have to be careful around closures).

[EdSchouten]

Hi @nikomatsakis! That does sound like a decent approach. That said, I've only been using Rust for the last four weeks. I was already quite happy to come up with this patch. I suspect it would be wiser for you to take a stab at this. Would that be all right?

[nikomatsakis]

Hmm, in that case, I might be inclined to let this "fix itself" as we transition to the MIR-based borrowck.

[EdSchouten]

That sounds reasonable to me. Before closing this pull request: do we want to add this as a unit test for the new borrowck?

[nikomatsakis]

@EdSchouten yep, I've been tracking such bugs in #47366 -- I'll add #47053 to that list

[nikomatsakis]

@EdSchouten maybe repurpose this PR for that purpose? #47366 has a bit of directions for how to go about it.

[EdSchouten]

I've just reworked this PR to only add testing coverage for NLL. Does this look all right?

[nikomatsakis]

@bors r+ rollup

[bors]

📌 Commit e47cc69 has been approved by nikomatsakis

[nikomatsakis]

@EdSchouten looks great =)