Some more small salsa memory improvements
This does limit our lru limits to 2^16 but if you want to set them higher than that you might as well not set them at all. Also makes `LRU` opt-in per query now, allowing us to drop all the unnecessary LRU stuff for most queries
feature: teach rust-analyzer to discover `linked_projects`
This PR's been a long-time coming, but like the title says, it introduces server-side project discovery and removes the extension hooks I previously introduced. I don't think this PR is ready to land, but here are the things I'm feeling squishy about:
- I don't think I like the idea of introducing the `cargo-metadata` command-but-for-everything-else in the `flycheck` module, but the progress reporting infrastructure was too convenient to pass up. Happy to move it elsewhere.
Here are the things I _know_ I need to change:
- For progress reporting, I'm extracting from a `serde_json::Value` that corresponds to `tracing_subsciber::fmt::Layer`'s JSON output. I'd like to make this a bit more structured/documented than the current nonsense I wrote.
- The progress reporting currently hardcodes "Buck"; it should be deriving that from the previously mentioned more-structured-output.
- This doesn't handle *reloading* when a corresponding buildfile is changed. It should be doing that.
<details>
<summary>Anyway, here's a video of rust-analyzer discovering a Buck target.</summary>
https://github.com/rust-lang/rust-analyzer/assets/2067774/be6cd9b9-2c9a-402d-847f-05f860a91df1
</details>
feat: Add incorrect case diagnostics for enum variant fields and all variables/params
Updates the incorrect case diagnostic to check:
1. Fields of enum variants. Example:
```rust
enum Foo {
Variant { nonSnake: u8 }
}
```
2. All variable bindings, instead of just let bindings and certain match arm patters. Examples:
```rust
match 1 { nonSnake => () }
match 1 { nonSnake @ 1 => () }
match 1 { nonSnake1 @ nonSnake2 => () } // slightly cursed, but these both introduce new
// bindings that are bound to the same value.
const ONE: i32 = 1;
match 1 { nonSnake @ ONE } // ONE is ignored since it is not a binding
match Some(1) { Some(nonSnake) => () }
struct Foo { field: u8 }
match (Foo { field: 1 } ) {
Foo { field: nonSnake } => ();
}
struct Foo { nonSnake: u8 } // diagnostic here, at definition
match (Foo { nonSnake: 1 } ) { // no diagnostic here...
Foo { nonSnake } => (); // ...or here, since these are not where the name is introduced
}
for nonSnake in [] {}
struct Foo(u8);
for Foo(nonSnake) in [] {}
```
3. All parameter bindings, instead of just top-level binding identifiers. Examples:
```rust
fn func(nonSnake: u8) {} // worked before
struct Foo { field: u8 }
fn func(Foo { field: nonSnake }: Foo) {} // now get diagnostic for nonSnake
```
This is accomplished by changing the way binding identifier patterns are filtered:
- Previously, all binding idents were skipped, except a few classes of "good" binding locations that were checked.
- Now, all binding idents are checked, except field shorthands which are skipped.
Moving from a whitelist to a blacklist potentially makes the analysis more brittle:
If new pattern types are added in the future where ident pats don't introduce new names, then they may incorrectly create diagnostics.
But the benefit of the blacklist approach is simplicity: I think a whitelist approach would need to recursively visit patterns to collect renaming candidates?
