WIP: Module name resolution

[matklad]

work towards #231

[aochagavia]

Nice work! I didn't have time to tackle this earlier this week, but I enjoyed reading the PR.

[matklad]

Some rough perf numbers:

INFO [ra_analysis::descriptors::module::nameres] item_map: 149.49299ms
INFO [ra_analysis::descriptors::module::nameres] item_map: 56.36107ms
INFO [ra_analysis::descriptors::module::nameres] item_map: 64.422592ms
INFO [ra_analysis::descriptors::module::nameres] item_map: 46.608787ms
INFO [ra_analysis::descriptors::module::nameres] item_map: 66.772202ms
INFO [ra_analysis::descriptors::module::nameres] item_map: 67.200399ms
INFO [ra_analysis::descriptors::module::nameres] item_map: 71.504211ms
INFO [ra_analysis::descriptors::module::nameres] item_map: 55.688357ms
INFO [ra_analysis::descriptors::module::nameres] item_map: 53.92913ms
INFO [ra_analysis::descriptors::module::nameres] item_map: 78.610024ms
INFO [ra_analysis::descriptors::module::nameres] item_map: 52.666113ms

This is for rust-lang/rust repo. Each item_map is roughly a simplified from scratch resolve of all uses. These numbers are pretty existing: there are good enough for completion as is, and there's a ton of room for improvement. Specifically:

• currently, we basically treat all of rust-lang/rust as a single crate (a single SourceRoot, more specifically). Ideally, we should create a SourceRoot per cargo package. That would allow us to have per-crate invalidation of imports.

• currently, we store LocalSyntaxPtrs in the data we compute. This is bad, because they change with every edit. Ideally, we should recompute ItemMap only when the user adds/removes items. This can be achieved using a smarter version of LocalSyntaxPtr: instead of storing a pair of offsets, we will store something like "the second struct with name Foo", which should be stable across most edits.

With these two optimizations implemented, I think that such top-down approach to name resolution would be speedy enough even with proper support for globs and macros (:crossed_fingers: )

[aochagavia]

Nice work! Are you planning to add tests as well? This seems quite tricky to get right...

[matklad]

Nice work! Are you planning to add tests as well? This seems quite tricky to get right...

Not in this PR: this is covered by the completion tests atm.

But yeah, we definitely should have nameres specific unit-tests. There's one in this PR for the start. We also should add a test which checks that we don't recompute item_map if user types inside function body.

[matklad]

The main problem is that we've got an ICE...

[matklad]

bors r+

[bors]

Build succeeded

• continuous-integration/travis-ci/push