[`implied_bounds_in_impls`]: include (previously omitted) associated types in suggestion
Fixes#11435
It now includes associated types from the implied bound that were omitted in the second bound. Example:
```rs
fn f() -> impl Iterator<Item = u8> + ExactSizeIterator> {..}
```
Suggestion before this change:
```diff
- pub fn my_iter() -> impl Iterator<Item = u32> + ExactSizeIterator {
+ pub fn my_iter() -> impl ExactSizeIterator {
```
It didn't include `<Item = u32>` on `ExactSizeIterator`. Now, with this change, it does.
```diff
- pub fn my_iter() -> impl Iterator<Item = u32> + ExactSizeIterator {
+ pub fn my_iter() -> impl ExactSizeIterator<Item = u32> {
```
We also now extend the span to include not just possible `+` ahead of it, but also behind it (an example for this is in the linked issue as well).
**Note:** The overall diff is a bit noisy, because building up the suggestion involves quite a bit more logic now and I decided to extract that into its own function. For that reason, I split this PR up into two commits. The first commit contains the actual "logic" changes. Second commit just moves code around.
changelog: [`implied_bounds_in_impls`]: include (previously omitted) associated types in suggestion
changelog: [`implied_bounds_in_impls`]: include the `+` behind bound if it's the last bound
Rename incorrect_impls to non_canonical_impls, move them to warn by default
The wording/category of these feel too strong to me, I would expect most of the time it's linting the implementations aren't going to be *incorrect*, just unnecessary
changelog: rename `incorrect_clone_impl_on_copy_type` to [`non_canonical_clone_impl`]
changelog: rename `incorrect_partial_ord_impl_on_ord_type` to [`non_canonical_partial_ord_impl`]
changelog: Move [`non_canonical_clone_impl`], [`non_canonical_partial_ord_impl`] to suspicious
Preserve literals and range kinds in `manual_range_patterns`
Fixes#11461
Also enables linting when there are 3 or fewer alternatives if one of them is already a range pattern
changelog: none
[`slow_vector_initialization`]: use the source span of vec![] macro and fix another FP
Fixes#11408
<details>
<summary>Also fixes a FP when the vec initializer comes from a macro other than `vec![]`</summary>
```rs
macro_rules! x {
() => { vec![] }
}
fn f() {
let mut v = x!();
v.resize(10, 0);
}
```
This shouldn't warn. The `x!` macro might be doing other things, so just replacing `x!()` with `vec![0; 10]` is not always an option.
</details>
I added some test cases for macro expansions, however I don't think there's a way to write a test for that specific warning that appeared in the linked issue. As far as I understand, that happens when the rust-src rustup component isn't installed (so the stdlib source is unavailable) and the span points to the `vec![]` *expansion*, instead of the `vec![]` that the user wrote.
changelog: [`slow_vector_initialization`]: use the source span of `vec![]` macro
changelog: [`slow_vector_initialization`]: only warn on `vec![]` expansions and allow other macros
fix fp when [`undocumented_unsafe_blocks`] not able to detect comment on globally defined const/static variables
fixes: #11246
changelog: fix detection on global variables for [`undocumented_unsafe_blocks`]
skip `todo!()` in `never_loop`
As promised in #11450, here is an implementation which skips occurrences of the `todo!()` macro.
changelog: [`never_loop`]: skip loops containing `todo!()`
Don't pass extra generic arguments in `needless_borrow`
fixes#10253
Also switches to using `implements_trait` which does ICE when clippy's debug assertions are enabled.
changelog: None
[`implied_bounds_in_impls`]: don't ICE on default generic parameter and move to nursery
Fixes#11422
This fixes two ICEs ([1](https://github.com/rust-lang/rust-clippy/issues/11422#issue-1872351763), [2](https://play.rust-lang.org/?version=stable&mode=debug&edition=2021&gist=2901e6febb479d3bd2a74f8a5b8a9305)), and moves it to nursery for now, because this lint needs some improvements in its suggestion (see #11435, for one such example).
changelog: Moved [`implied_bounds_in_impls`] to nursery (Now allow-by-default)
[#11437](https://github.com/rust-lang/rust-clippy/pull/11437)
changelog: [`implied_bounds_in_impls`]: don't ICE on default generic parameter in supertrait clause
r? `@xFrednet` (since you reviewed my PR that added this lint, I figured it might make sense to have you review this as well since you have seen this code before. If you don't want to review this, sorry! Feel free to reroll then)
--------
As for the ICE, it's pretty complicated and very confusing imo, so I'm going to try to explain the idea here (partly for myself, too, because I've confused myself several times writing- and fixing this):
<details>
<summary>Expand</summary>
The general idea behind the lint is that, if we have this function:
```rs
fn f() -> impl PartialEq<i32> + PartialOrd<i32> { 0 }
```
We want to lint the `PartialEq` bound because it's unnecessary. That exact bound is already specified in `PartialOrd<i32>`'s supertrait clause:
```rs
trait PartialOrd<Rhs>: PartialEq<Rhs> {}
// PartialOrd<i32>: PartialEq<i32>
```
The way it does this is in two steps:
- Go through all of the bounds in the `impl Trait` return type and collect each of the trait's supertrait bounds into a vec. We also store the generic arguments for later.
- `PartialEq` has no supertraits, nothing to add.
- `PartialOrd` is defined as `trait PartialOrd: PartialEq`, so add `PartialEq` to the list, as well as the generic argument(s) `<i32>`
Once we are done, we have these entries in the vec: `[(PartialEq, [i32])]`
- Go through all the bounds again, and looking for those bounds that have their trait `DefId` in the implied bounds vec.
- `PartialEq` is in that vec. However, that is not enough, because the trait is generic. If the user wrote `impl PartialEq<String> + PartialOrd<i32>`, then `PartialOrd` clearly doesn't imply `PartialEq`. Which means, we also need to check that the generic parameters match. This is why we also collected the generic arguments in `PartialOrd<i32>`. This process of checking generic arguments is pretty complicated and is also where the two ICEs happened.
The way it checks that the generic arguments match is by comparing the generic parameters in the super trait clause:
```rs
trait PartialOrd<Rhs>: PartialEq<Rhs> {}
// ^^^^^^^^^^^^^^
```
...this needs to match...
```rs
fn f() -> impl PartialEq<i32> + ...
// ^^^^^^^^^^^^^^
```
In the compiler, the `Rhs` generic parameter is its own type and we cannot just compare it to `i32`. We need to "substitute" it.
Internally, `Rhs` is represented as `Rhs#1` (the number next to # represents the type parameter index. They start at 0, but 0 is "reserved" for the implicit `Self` generic parameter).
How do we go from `Rhs#1` to `i32`? Well, we know that all the generic parameters had to be substituted in the `impl ... + PartialOrd<i32>` type. So we subtract 1 from the type parameter index, giving us 0 (`Self` is not specified in that list of arguments). We use that as the index into the generic argument list `<i32>`. That's `i32`. Now we know that the supertrait clause looks like `: PartialEq<i32>`.
Then, we can compare that to what the user actually wrote on the bound that we think is being implied: `impl PartialEq<i32> + ...`.
Now to the actual bug: this whole logic doesn't take into account *default* generic parameters. Actually, `PartialOrd` is defined like this:
```rs
trait PartialOrd<Rhs = Self>: PartialEq<Rhs> {}
```
If we now have a function like this:
```rs
fn f() -> impl PartialOrd + PartialEq {}
```
that logic breaks apart... We look at the supertrait predicate `: PartialEq<Rhs>` (`Rhs` is `Rhs#1`), then take the first argument in the generic argument list `PartialEq<..>` to resolve the `Rhs`, but at this point we crash because there *is no* generic argument.
The index 0 is out of bounds. If this happens (and we even get to linting here, which could only happen if it passes typeck), it must mean that that generic parameter has a default type that is not required to be specified.
This PR changes the logic such that if we have a type parameter index that is out of bounds, it looks at the definition of the trait and check that there exists a default type that we can use instead.
So, we see `<Rhs = Self>`, and use `Self` for substitution, and end up with this predicate: `: PartialEq<Self>`. No crash this time.
</details>
`never_loop` catches `loop { panic!() }`
* Depends on: #11447
This is an outgrowth of #11447 which I felt would best be done as a separate PR because it yields significant new results.
This uses typecheck results to determine divergence, meaning we can now detect cases like `loop { std::process::abort() }` or `loop { panic!() }`. A downside is that `loop { unimplemented!() }` is also being linted, which is arguably a false positive. I'm not really sure how to check this from HIR though, and it seems best to leave this epicycle for a later PR.
changelog: [`never_loop`]: Now lints on `loop { panic!() }` and similar constructs
Fix span when linting `explicit_auto_deref` immediately after `needless_borrow`
fixes#11366
changelog: `explicit_auto_deref`: Fix span when linting immediately after `needless_borrow`
Add config flag for reborrows in explicit_iter_loop
This PR adds a config flag for enforcing explicit into iter lint for reborrowed values. The config flag, `enforce_iter_loop_reborrow`, can be added to clippy.toml files to enable the linting behaviour. By default the reborrow lint is disabled.
fixes: #11074
changelog: [`explicit_iter_loop`]: add config flag `enforce_iter_loop_reborrow` to disable reborrow linting by default
new lint: `iter_out_of_bounds`
Closes#11345
The original idea in the linked issue seemed to be just about arrays afaict, but I extended this to catch some other iterator sources such as `iter::once` or `iter::empty`.
I'm not entirely sure if this name makes a lot of sense now that it's not just about arrays anymore (specifically, not sure if you can call `.take(1)` on an `iter::Empty` to be "out of bounds"?).
changelog: [`iter_out_of_bounds`]: new lint
[`unnecessary_unwrap`]: lint on `.as_ref().unwrap()`
Closes#11371
This turned out to be a little more code than I originally thought, because the lint also makes sure to not lint if the user tries to mutate the option:
```rs
if option.is_some() {
option = None;
option.unwrap(); // don't lint here
}
```
... which means that even if we taught this lint to recognize `.as_mut()`, it would *still* not lint because that would count as a mutation. So we need to allow `.as_mut()` calls but reject other kinds of mutations.
Unfortunately it doesn't look like this is possible with `is_potentially_mutated` (seeing what kind of mutation happened).
This replaces it with a custom little visitor that does basically what it did before, but also allows `.as_mut()`.
changelog: [`unnecessary_unwrap`]: lint on `.as_ref().unwrap()`
skip float_cmp check if lhs is a custom type
*Please write a short comment explaining your change (or "none" for internal only changes)*
changelog: [`float_cmp`]: allow float eq comparison when lhs is a custom type that implements PartialEq<f32/f64>
If the lhs of a comparison is not float, it means there is a user implemented PartialEq, and the caller is invoking that custom version of `==`, instead of the default floating point equal comparison.
People may wrap f32 with a struct (say `MyF32`) and implement its PartialEq that will do the `is_close()` check, so that `MyF32` can be compared with either f32 or `MyF32`.
