7426: Create all `ModuleId`s through a `DefMap` method r=jonas-schievink a=jonas-schievink
`ModuleId` needs to be able to represent blocks, and only the
associated `DefMap` will know how to construct that `ModuleId`
bors r+
Co-authored-by: Jonas Schievink <jonasschievink@gmail.com>
7359: ItemTree: store a mapping from blocks to inner items r=jonas-schievink a=jonas-schievink
To do name resolution within block expressions, we need to know which inner items are located inside each block expression. This adds such a mapping to `ItemTree`, replacing the previous one, which was seemingly unused other than to access all the inner items.
This also assigns `AstId`s to block expressions, which is needed to store the mapping in salsa.
Co-authored-by: Jonas Schievink <jonasschievink@gmail.com>
7336: Rename `CrateDefMap` to `DefMap` r=matklad a=jonas-schievink
I propose handling local items by computing a `DefMap` for every block expression, using the regular (early) name resolution algorithm. The result of that will be a `DefMap` that has a reference to the parent `DefMap`, which is either the one computed for the containing block expression, or the crate's root `DefMap`. Name resolution will fall back to a name in the parent `DefMap` if it cannot be resolved in the inner block.
The `DefMap`s computed for block expressions will go through a separate query that can be garbage-collected much more aggressively, since these `DefMap`s should be cheap to compute and are never part of a crate's public API.
The first step towards that is to make `CrateDefMap` not specific to crates anymore, hence this rename (if this plans sounds reasonable).
cc https://github.com/rust-analyzer/rust-analyzer/issues/7325 and https://github.com/rust-analyzer/rust-analyzer/issues/1165
Co-authored-by: Jonas Schievink <jonasschievink@gmail.com>
7276: Remove map module from la-arena public API r=lnicola a=arzg
It’s unlikely that more items will be added to the module, so it’s simpler for users if `ArenaMap` is re-exported and the module made private.
This doesn’t compile for the same reason that #7275 doesn’t:
> This pull request doesn’t compile because dependencies on la-arena go through crates.io, so existing dependencies on the crate are referencing an old version. As such, this PR will only compile once a new la-arena version has been published.
Co-authored-by: Aramis Razzaghipour <aramisnoah@gmail.com>
7110: Deduplicate macros when offering completion r=matklad a=AdnoC
Closes https://github.com/rust-analyzer/rust-analyzer/issues/7081
When iterating over the names within the `hir_def::resolver::Scope` for a module, track what macros are in the `hir_def::item_scope::ItemScope::legacy_macros` collection for the module. When iterating over names from the prelude, do not proccess the name if it had been in the `legacy_macros` collection.
This is implemented with a `FxHashSet` in the `Scope::process_names` function that is populated when iterating over `legacy_macros` and checked when iterating over the prelude.
Alternative implementation could instead query the `legacy_macros` `FxHashMap` directly when processing names in the prelude.
Also, I'd like to add a test for this, but I'm not sure where it could be added.
Co-authored-by: AdnoC <adam.r.cutler@gmail.com>
7145: Proper handling $crate Take 2 [DO NOT MERGE] r=edwin0cheng a=edwin0cheng
Similar to previous PR (#7133) , but improved the following things :
1. Instead of storing the whole `ExpansionInfo`, we store a similar but stripped version `HygieneInfo`.
2. Instread of storing the `SyntaxNode` (because every token we are interested are IDENT), we store the `TextRange` only.
3. Because of 2, we now can put it in Salsa.
4. And most important improvement: Instead of computing the whole frames every single time, we compute it recursively through salsa: (Such that in the best scenario, we only need to compute the first layer of frame)
```rust
let def_site = db.hygiene_frame(info.def.file_id);
let call_site = db.hygiene_frame(info.arg.file_id);
HygieneFrame { expansion: Some(info), local_inner, krate, call_site, def_site }
```
The overall speed compared to previous PR is much faster (65s vs 45s) :
```
[WITH old PR]
Database loaded 644.86ms, 284mi
Crates in this dir: 36
Total modules found: 576
Total declarations: 11153
Total functions: 8715
Item Collection: 15.78s, 91562mi
Total expressions: 240721
Expressions of unknown type: 2635 (1%)
Expressions of partially unknown type: 2064 (0%)
Type mismatches: 865
Inference: 49.84s, 250747mi
Total: 65.62s, 342310mi
rust-analyzer -q analysis-stats . 66.72s user 0.57s system 99% cpu 1:07.40 total
[WITH this PR]
Database loaded 665.83ms, 284mi
Crates in this dir: 36
Total modules found: 577
Total declarations: 11188
Total functions: 8743
Item Collection: 15.28s, 84919mi
Total expressions: 241229
Expressions of unknown type: 2637 (1%)
Expressions of partially unknown type: 2064 (0%)
Type mismatches: 868
Inference: 30.15s, 135293mi
Total: 45.43s, 220213mi
rust-analyzer -q analysis-stats . 46.26s user 0.74s system 99% cpu 47.294 total
```
*HOWEVER*, it is still a perf regression (35s vs 45s):
```
[WITHOUT this PR]
Database loaded 657.42ms, 284mi
Crates in this dir: 36
Total modules found: 577
Total declarations: 11177
Total functions: 8735
Item Collection: 12.87s, 72407mi
Total expressions: 239380
Expressions of unknown type: 2643 (1%)
Expressions of partially unknown type: 2064 (0%)
Type mismatches: 868
Inference: 22.88s, 97889mi
Total: 35.74s, 170297mi
rust-analyzer -q analysis-stats . 36.71s user 0.63s system 99% cpu 37.498 total
```
Co-authored-by: Edwin Cheng <edwin0cheng@gmail.com>
7140: Store trait associated items in fst r=matklad a=SomeoneToIgnore
Store imported traits' associated function/methods and constants into `ImportMap.fst` and pefrorm the imports search on them.
This is a first step towards trait autoimport during completion functionality, the way I see it, after this PR, only a few major things are left to be done:
* store all traits' assoc items into fst, not only the ones in scope, as we do now. Any code pointers on how to do this are welcome 😄
* adjust a few modules in completions crate (`dot.rs`, `qualified_path.rs` at least) to query the import map, reusing the `import_assets` logic heavily
==
With the current import and autoimport implementations, it looks like for a single query, we're either interested in either associated items lookup or in all other `fst` contents lookup, but never both simultaneously.
I would rather not split `fst` in two but add another `Query` parameter to separate those, but let me know if you have any ideas.
Co-authored-by: Kirill Bulatov <mail4score@gmail.com>
1) the set of attributes for all variants of an enum, and
2) the set of attributes for all fields of a variant.
This avoids the current n^2 behavior when rendering completion for variants, which
prevents completion for enums with large numbers of variants.
7133: Proper handling $crate and local_inner_macros r=jonas-schievink a=edwin0cheng
This PR introduces `HygineFrames` to store the macro definition/call site hierarchy in hyginee and when resolving `local_inner_macros` and `$crate`, we use the token to look up the corresponding frame and return the correct value.
See also: https://rustc-dev-guide.rust-lang.org/macro-expansion.html#hygiene-and-hierarchies
fixe #6890 and #6788
r? @jonas-schievink
Co-authored-by: Edwin Cheng <edwin0cheng@gmail.com>
Added the error message to the doc for the UnresolvedProcMacro
diagnostic, explaining that either enabling the procMacro setting
or disabling this diagnostic should make the warnings go away.
7010: Update ungrammar for const block patterns r=matklad a=Veykril
Fixes#6848
Adds const blocks and const block patterns to the AST and parses them.
Blocked on https://github.com/rust-analyzer/ungrammar/pull/17/, will merge that PR there once this one gets the OK so I can remove the local ungrammar dependency path and fix the Cargo.lock.
Co-authored-by: Lukas Wirth <lukastw97@gmail.com>
6921: Higher-ranked trait bounds for where clauses r=flodiebold a=Veykril
There is a slight problem with this which is also noted in a FIXME now but `LifetimeParameters` of these ForLifetime where clauses allocate the lifetimes in the corresponding arena as if they were lifetimes of the item itself and not just the clause they belong to. I wasn't entirely sure what I could do about this but given nothing really uses lifetimes like that currently I figured it might be fine? Open to suggestions for that problem.
Co-authored-by: Lukas Wirth <lukastw97@gmail.com>
ItemTree is per-file, so there is no unique crate associated with it.
This means that it cannot know the active CfgOptions and thus couldn't
handle `cfg_attr`.
Prepare it for `cfg_attr`s by avoiding accessing attributes.
6896: Node-ify lifetimes r=jonas-schievink a=Veykril
Let's see if this passes the tests 🤞
Depends on https://github.com/rust-analyzer/ungrammar/pull/15
Co-authored-by: Jonas Schievink <jonasschievink@gmail.com>
Co-authored-by: Jonas Schievink <jonas.schievink@ferrous-systems.com>
Co-authored-by: Lukas Wirth <lukastw97@gmail.com>
6897: Basic support for macros 2.0 r=jonas-schievink a=jonas-schievink
This adds support for (built-in-only) macros 2.0, and removes some hacks used for builtin derives, which are declared via macros 2.0 in libcore.
First steps for https://github.com/rust-analyzer/rust-analyzer/issues/2248.
Blocked on https://github.com/rust-analyzer/ungrammar/pull/16.
Co-authored-by: Jonas Schievink <jonasschievink@gmail.com>
Co-authored-by: Jonas Schievink <jonas.schievink@ferrous-systems.com>
6818: Add Lifetimes to the HIR r=matklad a=Veykril
This doesn't handle resolve yet as I don't know yet how that will be used. I'll get to that once I start moving the lifetime reference PR to the hir.
This also adds a new `hir` name type for lifetimes and labels, `hir::LifetimeName`.
Co-authored-by: Lukas Wirth <lukastw97@gmail.com>
6771: Properly attach inner attributes in Attrs::new r=matklad a=Veykril
Properly attach inner and outer attributes to the things they actually belong to in the HIR. ~~I can add some tests for this if wanted once I know where to put them/how to test for this.~~ Put some tests into `hover.rs`.
So the following snippet
```rust
mod foo {
//! Hello
}
```
now shows `Hello` on hover 🎉Fixes#2148
Co-authored-by: Lukas Wirth <lukastw97@gmail.com>
6698: Attach macro expansion errors to the right file r=jonas-schievink a=jonas-schievink
Previously it attached them to the result of the macro expansion (or, if no result was produced, to the file containing the invocation). Always use the file containing the invocation.
This doesn't seem to have any observable difference, but seems better in theory.
bors r+
Co-authored-by: Jonas Schievink <jonasschievink@gmail.com>
6435: Test Fixture ExplicitRoot + ModuleResolutionRelativePathOutsideRoot. r=matklad a=rickvanprim
Updates `module_resolution_relative_path_outside_root` test to check valid paths outside of the root, by moving the root to a subpath so that paths outside of it are possible. If this would be more appropriate as a new test, or if the original check for an invalid path should be left, I'm happy to update.
Co-authored-by: James Leitch <rickvanprim@gmail.com>
It's very useful when `pub` is equivalent to "this is crate's public
API", let's enforce this!
Ideally, we should enforce it for local `cargo test`, and only during
CI, but that needs https://github.com/rust-lang/cargo/issues/5034.
6324: Improve #[cfg] diagnostics r=jonas-schievink a=jonas-schievink
Unfortunately I ran into https://github.com/rust-analyzer/rust-analyzer/issues/4058 while testing this on https://github.com/nrf-rs/nrf-hal/, so I didn't see much of it in action yet, but it does seem to work.
Co-authored-by: Jonas Schievink <jonas.schievink@ferrous-systems.com>
Co-authored-by: Jonas Schievink <jonasschievink@gmail.com>
6307: Add whitelist of safe intrinsics r=frazar a=frazar
This PR should fix#5996, where intrinsic operations where all marked as unsafe.
I'm rather new to this codebase, so I might be doing something *very* wrong. Please forgive me!
In particular, I'm not sure how to "check that we are in extern `rust-intrinsics`" as mentioned [in this comment](https://github.com/rust-analyzer/rust-analyzer/issues/5996#issuecomment-709234802).
Co-authored-by: Francesco Zardi <frazar@users.noreply.github.com>
Declaration names sounds like a name of declaration -- something you
can use for analysis. It empathically isn't, and is just a label
displayed in various UI. It's important not to confuse the two, least
we accidentally mix semantics with UI (I believe, there's already a
case of this in the FamousDefs at least).
6130: Items case quick fix (snake_case / UPPER_SNAKE_CASE / CamelCase) r=matklad a=popzxc
Resolves#4598.
After a third try, it finally works. Boy, it appeared tougher than it seemed.
Initially I thought like "Ha, `rustc` already tells us where idents are named incorrectly. It shouldn't be that hard, should it?".
Well, the problems with the information provided by `rustc` appeared shortly:
- `rustc` warnings are `flycheck` warnings, which are slightly aside from our diagnostics with fixes.
When we map flycheck diagnostic to LSP, we can convert it into a fix, but only if it's marked as `Applicability::MachineApplicable`.
Name case fix is marked `Applicability::MaybeIncorrect`, and for a reason: it only suggest to rename symbol under cursor, without tracking any references.
- Warning spawned by `rustc` are identified by string labels rather than enum. It means that if one day the diagnostic will be renamed in `rustc`, `rust-analyzer` code will still compile, but won't find the required diagnostic by name anymore. If by chance this will happen when some unlucky guy will decide to create their first pull request, they'll be confused by suddenly failing tests (likely) not related to their changes.
- Even if we'll try to build fixes atop of `rustc` warnings, we'll have to do it in the `rust_analyzer::diagnostics::to_proto` module, which is far less
convenient for that matter than `ide` crate.
That's why I decided that it's worth a separate `rust-analyzer` diagnostic, which will implement `DiagnosticWithFix` trait.
After that, I discovered that currently `hir_ty::diagnostics` only check `DefWithBody` types, like function bodies. I had to add support for diagnostics
which look at any `ModuleDef`.
And of course, since I'd added a lot of new functionality, it required extensive testing.
That explains why the diff is so big for a (looking) relatively small feature.
I hope that this PR doesn't only add a small feature, but also creates a base for building another features.
## Example:
P.S. My eyes were bleeding when I had to write the code for the example...
6135: when generating new function, focus on return type instead of body r=matklad a=bnjjj
I made a little change when we use the assist to generate a new function, instead of focusing on the function body, it will focus on return type
Co-authored-by: Igor Aleksanov <popzxc@yandex.ru>
Co-authored-by: Benjamin Coenen <5719034+bnjjj@users.noreply.github.com>
6199: Fix `mut self` not emitting mutable binding on `self` use r=matklad a=Veykril
Prior to this, when `self` in a function is taken by value and bound mutably, its use inside of the method body won't be marked `mutably`.
Fixes#5461
Co-authored-by: Lukas Wirth <lukastw97@gmail.com>
5917: Add a command to open docs for the symbol under the cursor r=matklad a=zacps
#### Todo
- [ ] Decide if there should be a default keybind or context menu entry
- [x] Figure out how to get the documentation path for methods and other non-top-level defs
- [x] Design the protocol extension. In future we'll probably want parameters for local/remote documentation URLs, so that should maybe be done in this PR?
- [x] Code organisation
- [x] Tests
Co-authored-by: Zac Pullar-Strecker <zacmps@gmail.com>
Previously, "find all references" on a variant field wouldn't find any
references outside the defining module. This is because variant fields
were incorrectly assumed to be private, like struct fields without
explicit visibility, but they actually inherit the enum's visibility.
Currently a method only has defaultness if it is a provided trait
method, but this will change when specialisation is available and may
need to become a concept known to hir.
I opted to go for a 'fewest changes' approach given specialisation is
still under development.
6124: Better normalized crate name usage r=jonas-schievink a=SomeoneToIgnore
Closes https://github.com/rust-analyzer/rust-analyzer/issues/5343
Closes https://github.com/rust-analyzer/rust-analyzer/issues/5932
Uses normalized name for code snippets (to be able to test the fix), hover messages and documentation rewrite links (are there any tests for those?).
Also renamed the field to better resemble the semantics.
Co-authored-by: Kirill Bulatov <mail4score@gmail.com>
6139: Make find_path_prefixed configurable r=matklad a=Veykril
This makes `find_path_prefixed` more configurable allowing one to choose whether it always returns absolute paths, self-prefixed paths or to ignore local imports when building the path.
The config names are just thrown in here, taking better names if they exist :)
This should fix#6131 as well?
Co-authored-by: Lukas Wirth <lukastw97@gmail.com>
6019: Remove make::path_from_text r=matklad a=Veykril
This removes the `make::path_from_text` function, which according to a note should've been private. I removed it since it didn't really serve a purpose as it was simply wrapping `make::ast_from_text`.
Co-authored-by: Lukas Wirth <lukastw97@gmail.com>
6033: Make name resolution resolve proc macros instead of relying purely on the build system r=matklad a=jonas-schievink
This makes name resolution look at proc-macro declaration attributes like `#[proc_macro_derive]` and defines the right proc macro in the macro namespace, fixing unresolved custom derives like `thiserror::Error` (which can cause false positives, now that we emit diagnostics for unresolved imports).
This works even when proc-macro support is turned off, in which case we fall back to a dummy expander that always returns an error. IMO this is the right way to handle at least the name resolution part of proc. macros, while the *expansion* itself should rely on the build system to build and provide the macro DLL. It does mean that they may go out of sync, but we can provide diagnostics if that happens (something like "could not find macro X in crate Y – ensure that all files of crate Y are saved").
I think it is valuable to be able to reason about proc macros even when we can't expand them, since proc macro expansion can break between Rust releases or users might not want to turn it on for performance reasons. It allows us to provide better diagnostics on any proc macro invocation we're not expanding (like a weak warning that informs the user that proc macro support is turned off, or that it has been disabled because the server crashed).
Fixes https://github.com/rust-analyzer/rust-analyzer/issues/5763
Co-authored-by: Jonas Schievink <jonas.schievink@ferrous-systems.com>
6085: Mark unresolved imports diagnostic as experimental r=jonas-schievink a=jonas-schievink
It causes a lot of false positives for people. We collected all of the known ones during the last week.
Co-authored-by: Jonas Schievink <jonas.schievink@ferrous-systems.com>
6016: Emit diagnostics for unresolved imports and extern crates r=jonas-schievink a=jonas-schievink
AFAIK, we don't have any major bugs in name resolution that would cause a lot of false positives here (except procedural attribute macro support and some rare issues around `#[path]` on module files), so these are *not* marked as experimental diagnostics right now.
I noticed that diagnostics in a file sometimes don't get displayed after opening, but require some edit to be performed. This seems like a preexisting issue though.
Co-authored-by: Jonas Schievink <jonas.schievink@ferrous-systems.com>
5971: Implement async blocks r=flodiebold a=oxalica
Fix#4018
@flodiebold already gave a generic guide in the issue. Here's some concern about implementation detail:
- Chalk doesn't support generator type yet.
- Adding generator type as a brand new type (ctor) can be complex and need to *re-introduced* builtin impls. (Like how we implement closures before native closure support of chalk, which is already removed in #5401 )
- The output type of async block should be known after type inference of the whole body.
- We cannot directly get the type from source like return-positon-impl-trait. But we still need to provide trait bounds when chalk asking for `opaque_ty_data`.
- During the inference, the output type of async block can be temporary unknown and participate the later inference.
`let a = async { None }; let _: i32 = a.await.unwrap();`
So in this PR, the type of async blocks is inferred as an opaque type parameterized by the `Future::Output` type it should be, like what we do with closure type.
And it really works now.
Well, I still have some questions:
- The bounds `AsyncBlockImplType<T>: Future<Output = T>` is currently generated in `opaque_ty_data`. I'm not sure if we should put this code here.
- Type of async block is now rendered as `impl Future<Output = OutputType>`. Do we need to special display to hint that it's a async block? Note that closure type has its special format, instead of `impl Fn(..) -> ..` or function type.
Co-authored-by: oxalica <oxalicc@pm.me>
