Crate metadata nondefid args

[cramertj]

Builds on #41766
cc #41417

I started working on queryifying the crate metadata functions that take CrateNum as an argument. However, a lot of these showed up in trans, which doesn't currently have access to tcx. I plumbed it through, but after doing so I'm not sure how to set it up in driver, which makes me think this wasn't the right thing to do 😆 I'm imagining there's a reason that trans relies upon Session instead of TyCtxt. If that's the case, how should I proceed?

r? @nikomatsakis

[nikomatsakis]

@cramertj

a lot of these showed up in trans, which doesn't currently have access to tcx

wait, what? Trans definitely has access to a tcx.

[nikomatsakis]

@cramertj can you give a more specific pointer to the part of trans you meant?

[cramertj]

@nikomatsakis Yeah, that's what I would've expected haha. But
here, here, here, and here it looks like I only had access to &Session, when I needed TyCtxt. I'm clearly misunderstanding something here.

[cramertj]

@nikomatsakis and I just discussed this on IRC. In the LLVM passes there's no tcx right now. This is intentional, as some of the tcx arenas are dropped before the LLVM phase to limit peak memory usage. For now, I'm going to change this PR to leave those methods as-is, tag with a FIXME, and open an issue tracking this problem.

[bors]

☔ The latest upstream changes (presumably #41709) made this pull request unmergeable. Please resolve the merge conflicts.

[eddyb]

You can just use MetaDataByCrateNum above, the functions are used when the variants don't match the key exactly.

[nikomatsakis]

I don't think we should have DepNode with a cratenum in it, though, at least not yet, as it won't get remapped properly when being loaded back up.

[cramertj]

@nikomatsakis what would you suggest instead?

[eddyb]

There's something in MIR const qualif doing this check, so that could be fully replaced by this (right now users of this query have to take into account the fact that it only works cross-crate).

[eddyb]

@cramertj I think the thing to do here is to query the information ahead of time and maybe even not have &Session around - that is, post-translation the backend should be as detached as possible from rustc.

[cramertj]

@eddyb What's the right way to do that so as not to mess up dependency tracking? Or do we even care about dependencies post-translation?

[eddyb]

@cramertj Post-translation should be fully deterministic in terms of dependencies.
Treat is as "creating a build plan" that post-translation obeys. Unless you want to hash its results, it's fine.

[alexcrichton]

ping @cramertj just wanted to keep this on your radar, do you think you'll have a chance to rebase?

[cramertj]

@alexcrichton This is based on top of another PR of mine that requires a test on Windows, and I haven't been able to get LLVM to build properly on my Windows machine. I'll close this for now and open a new PR when I have that sorted out.