Add command to report unresolved references

[darichey]

Adds rust-analyzer unresolved-references which reports unresolved references. This is useful for debugging and regression testing for both rust-analyzer and project generators like Buck's rust-project.

As discussed: https://rust-lang.zulipchat.com/#narrow/stream/185405-t-compiler.2Frust-analyzer/topic/Command.20to.20report.20unresolved.20references

[darichey]

This isn't quite working totally yet, but opening the PR to get some feedback and ask a couple questions. The approach is based on NameRefClass, but there are some cases this doesn't handle properly on its own. I think highlight.rs has all of the answers, but I'd appreciate some pointers :)

This is the output when running on rust-analyzer itself: https://gist.github.com/darichey/55e411f55b722d9cc1602e8f273c1a67

It does find some actual issues! e.g.,

rust-analyzer/crates/hir-expand/src/files.rs:199:71: text_range
rust-analyzer/crates/salsa/src/runtime.rs:468:45: LEN
rust-analyzer/crates/syntax/src/ted.rs:71:31: last_child_or_token
rust-analyzer/crates/hir-def/src/resolver.rs:1058:13: module

However, there's a lot of false positives. In particular, code within #[tracing::instrument] (and other proc macros, like Self inside a #[salsa::database]) gets reported as unresolved. I'm not sure why that is.

Also, as reported in #10935, proc macro attributes are unresolved. Should we copy the hack for highlighting in #10937 or keep them since they are technically unresolved?

[darichey]

I've fixed the handling of macros and decided to ignore attributes. This is the new output when running on rust-analyzer: https://gist.github.com/darichey/2f74b0e45f5c1ff18d4ea78039e2c831

All of those are correctly-identified unresolved references except the last one:

rust-analyzer/crates/cfg/src/tests.rs

Line 248 in 9b72445

let cfg = CfgExpr::arbitrary(&mut u).unwrap();

I assume it has something to do with the conditional deriving of Arbitrary, but I wasn't able to repro in isolation.
Other than that issue, I think this is ready for review.

[Veykril]

Had a quick look but I am also not sure why that is. Either way commited some small cleansups and fixes (maybe that fixes things though I doubt it)

[darichey]

Figured it out! We need to set all_targets: true in the CargoConfig. Without it, we don't get artifact info for dev-dependencies when building build dependencies, so we can't expand proc macros from them.

For now, I've switched to creating the CargoConfig via Config, the same way scip does. This ensures we get the same defaults rust-analyzer does in normal usage. As a follow up, it would be good to extract the common workspace loading logic from scip, lsif, diagnostics, and unresolved_references to make sure they all do the right thing (it looks like lsif and diagnostics would have the same problem).

With that, there are no more false positives in rust-analyzer: https://gist.github.com/darichey/10a7b13961216e680f0d6b1085f79ab5

I will run it on our monorepo to check for any more false positives, but I think this is a good start we can iterate on!

[Veykril]

@bors r+

[bors]

📌 Commit e602e01 has been approved by Veykril

It is now in the queue for this repository.

[bors]

⌛ Testing commit e602e01 with merge 6c56df8...

[bors]

☀️ Test successful - checks-actions
Approved by: Veykril
Pushing 6c56df8 to master...