Aleksey Kladov [Tue, 22 Jun 2021 18:22:36 +0000 (21:22 +0300)]
internal: remove one more accidentally quadratic code-path
Definition::visibility was implemented in a rather roundabout way -- by
asking the parent module about the effective visibility.
This is problematic for a couple of reasons:
* first, it doesn't work for local items
* second, asking module about visibility of a child is a linear
operation (that's a problem in itself, tracked in #9378)
Instead, lets ask the declared visibility directly, we have all the code
for it, and need only to actually us it.
Aleksey Kladov [Mon, 21 Jun 2021 17:14:38 +0000 (20:14 +0300)]
feature: massively improve performance for large files
This story begins in #8384, where we added a smart test for our syntax
highting, which run the algorithm on synthetic files of varying length
in order to guesstimate if the complexity is O(N^2) or O(N)-ish.
The test turned out to be pretty effective, and flagged #9031 as a
change that makes syntax highlighting accidentally quadratic. There was
much rejoicing, for the time being.
Then, lnicola asked an ominous question[1]: "Are we sure that the time
is linear right now?"
Of course it turned out that our sophisticated non-linearity detector
*was* broken, and that our syntax highlighting *was* quadratic.
Investigating that, many brave hearts dug deeper and deeper into the
guts of rust-analyzer, only to get lost in a maze of traits delegating
to traits delegating to macros.
Eventually, matklad managed to peel off all layers of abstraction one by
one, until almost nothing was left. In fact, the issue was discovered in
the very foundation of the rust-analyzer -- in the syntax trees.
Worse, it was not a new problem, but rather a well-know, well-understood
and event (almost) well-fixed (!) performance bug.
The problem lies within `SyntaxNodePtr` type -- a light-weight "address"
of a node in a syntax tree [3]. Such pointers are used by rust-analyzer all
other the place to record relationships between IR nodes and the
original syntax.
Internally, the pointer to a syntax node is represented by node's range.
To "dereference" the pointer, you traverse the syntax tree from the
root, looking for the node with the right range. The inner loop of this
search is finding a node's child whose range contains the specified
range. This inner loop was implemented by naive linear search over all
the children. For wide trees, dereferencing a single `SyntaxNodePtr` was
linear. The problem with wide trees though is that they contain a lot of
nodes! And dereferencing pointers to all the nodes is quadratic in the
size of the file!
The solution to this problem is to speed up the children search --
rather than doing a linear lookup, we can use binary search to locate
the child with the desired interval.
Doing this optimization was one of the motivations (or rather, side
effects) of #6857. That's why `rowan` grew the useful
`child_or_token_at_range` method which does exactly this binary search.
But looks like we've never actually switch to this method? Oups.
Lesson learned: do not leave broken windows in the fundamental infra.
Otherwise, you'll have to repeatedly re-investigate the issue, by
digging from the top of the Everest down to the foundation!
bors[bot] [Mon, 21 Jun 2021 13:41:27 +0000 (13:41 +0000)]
Merge #9227
9227: Add a config setting to disable the 'test' cfg in specified crates r=matklad a=lf-
If you are opening libcore from rust-lang/rust as opposed to e.g.
goto definition from some other crate which would use the sysroot
instance of libcore, a `#![cfg(not(test))]` would previously have made
all the code excluded from the module tree, breaking the editor
experience.
Core does not need to ever be edited with `#[cfg(test)]` enabled,
as the tests are in another crate.
This PR puts in a slight hack that checks for the crate name "core" and
turns off `#[cfg(test)]` for that crate.
bors[bot] [Sun, 20 Jun 2021 17:38:27 +0000 (17:38 +0000)]
Merge #9346
9346: Refactor / clean up hir_ty tests r=flodiebold a=flodiebold
Notable changes:
- unify `check_types` and `check_mismatches` into `check`, which supports both kinds of annotations (`check_types` still exists because I didn't want to change all the annotations, but uses the same implementation)
- because of that, `check_types` now fails on any type mismatches; also annotations always have to hit the exact range
- there's also `check_no_mismatches` for when we explicitly just want to check that there are no type mismatches without giving any annotations (`check` will fail without annotations)
- test annotations can now be overlapping (they point to the nearest line that has actual code in that range):
```
// ^^^^ annotation
// ^^^^^^^^^ another annotation
```
Jade [Sat, 12 Jun 2021 12:16:22 +0000 (05:16 -0700)]
Fix libcore not being included in rust-lang/rust module tree
If you are opening libcore from rust-lang/rust as opposed to e.g.
goto definition from some other crate which would use the sysroot
instance of libcore, a `#![cfg(not(test))]` would previously have made
all the code excluded from the module tree, breaking the editor
experience.
This puts in a slight hack that checks for the crate name "core" and
turns off `#[cfg(test)]`.
bors[bot] [Fri, 18 Jun 2021 16:47:58 +0000 (16:47 +0000)]
Merge #9321
9321: Inline generics in const and function trait completions r=Veykril a=RDambrosio016
This PR does a couple of things:
- moves path_transform from ide_assists to ide_db to be shared by both assists and completions
- when completing a const or a function for a trait, it will "inline" any generics in those associated items instead
of leaving the generic's name. For example:
```rust
trait Foo<T> {
const BAR: T;
fn foo() -> T;
}
struct Bar;
impl Foo<u32> for Bar {
// autocompletes to this
fn foo() -> u32;
// and not this (old)
fn foo() -> T;
// also works for associated consts and where clauses
const BAR: u32 = /* */
}
```
Currently this does not work for const generics, because `PathTransform` does not seem to account for them. If this should work on const generics too, `PathTransform` will need to be changed. However, it is uncommon to implement a trait only for a single const value, so this isnt a huge concern.