In the `issue-53548` test added in this commit, the `Box<dyn Trait>`
type is expanded to `Box<dyn Trait + 'static>`, but the generator
"witness" that results is `for<'r> { Box<dyn Trait + 'r> }`. The WF
code was encountering an ICE (when debug-assertions were enabled) and
an unexpected compilation error (without debug-asserions) when trying
to process this `'r` region bound. In particular, to be WF, the region
bound must meet the requirements of the trait, and hence we got
`for<'r> { 'r: 'static }`. This would ICE because the `Binder`
constructor we were using was assering that no higher-ranked regions
were involved (because the WF code is supposed to skip those). The
error (if debug-asserions were disabled) came because we obviously
cannot prove that `'r: 'static` for any region `'r`. Pursuant with
our "lazy WF" strategy for higher-ranked regions, the fix is not to
require that `for<'r> { 'r: 'static }` holds (this is also analogous
to what we would do for higher-ranked regions appearing within the
trait in other positions).
Warning period for detecting nested impl trait
Here is some proposed code for making a warning period for the new checking of nested impl trait.
It undoes some of the corrective effects of PR #57730, by using boolean flags to track parts of the analysis that were previously skipped prior to PRs #57730 and #57981 landing.
Cc #57979
Filter away test annotations from UI test output
If you worked with UI tests for some time you could notice one issue affecting their readability and also readability of diffs when the tests change.
Look at the output of this test.
```rust
fn main() {
let 1 = 2; //~ ERROR refutable pattern in local binding
}
```
```
error[E0005]: refutable pattern in local binding: `-2147483648i32..=0i32` not covered
--> src/main.rs:2:9
|
2 | let 1 = 2; //~ ERROR refutable pattern in local binding
| ^ pattern `-2147483648i32..=0i32` not covered
error: aborting due to previous error
For more information about this error, try `rustc --explain E0005`.
```
You can see that the "refutable pattern in local binding" is duplicated.
One instance is the actual error, and the second instance is the expected error annotation.
This annotation is useful in the test input, but in the output it clutters the text and makes it harder to see what text refers to actual errors and what is just comments, especially if there are many errors in a single test file.
@estebank [reported](https://github.com/rust-lang/rust/pull/57379#discussion_r245523361) using the next trick to avoid the clutter
```rust
fn main() {
let 1 = 2;
//~^ ERROR refutable pattern in local binding
}
```
```
error[E0005]: refutable pattern in local binding: `-2147483648i32..=0i32` not covered
--> src/main.rs:2:9
|
2 | let 1 = 2;
| ^ pattern `-2147483648i32..=0i32` not covered
error: aborting due to previous error
For more information about this error, try `rustc --explain E0005`.
```
, i.e. using `//~^` and placing the annotation one line below will remove the annotation from the output.
However, this doesn't always works (consider errors with multi-line spans), and shouldn't be necessary in general!
`compiletest` could automatically filter away its own annotations from the output instead.
This is exactly what this PR does.
r? @davidtwco
Instead of a sticky-boolean flag that would downgrade errors to
warnings during further recursion into the type (which is overly broad
because we were not missing errors at arbitrarily deep levels), this
instead tracks state closer to what the original bug actually was.
In particular, the actual original bug was that we were failing to
record the existence of an outer `impl Trait` solely when it occurred
as an *immediate child* during the walk of the child types in
`visit_generic_args`.
Therefore, the correct way to precisely model when that bug would
manifest itself (and thus downgrade the error-to-warning accordingly)
is to track when those outer `impl Trait` cases were previously
unrecorded.
That's what this code does, by storing a flag with the recorded outer
`impl Trait` indicating at which point in the compiler's control flow
it had been stored.
I will note that this commit passes the current test suite. A
follow-up commit will also include tests illustrating the cases that
this commit gets right (and were handled incorrectly by the previous
sticky boolean).
Update RLS and Clippy due to #56732 (rustc_interface crate)
Closes#59060.
In addition to plain submodule bumps, this also contains update to rls-rustc. The in-tree, from the RLS monorepo, version is used instead of the crates.io one (@nrc I think we might stop publishing `rls-rustc` altogether, right? It's only there to work around passing `-Zsave-analysis` to stable `rustc` and meant to be used only by RLS, IIRC).
@Zoxc also due to how we need to access the expanded AST still from the RLS side in order to pass save analysis data in-memory, I delayed the AST drop after the `after_analysis` callback if the `-Zsave-analysis` is passed.
It'd be also good if you could take a look at the changes inside the `rls` and `rls-rustc`: 6a1b5a9cfd...6840dd69af. The `rls-rustc` is based on your [PR](https://github.com/rust-dev-tools/rls-rustc/pull/11) but I also had to change some bits in the RLS itself.
r? @Zoxc / @Manishearth
Make migrate mode work at item level granularity
Migrate mode now works entirely at the item level rather than the body level,
ensuring that we don't lose any errors in contained closures.
Closes#58776
r? @pnkfelix
Type checking associated constants can require trait bounds, but an empty
parameter environment was provided to the trait solver. Providing an
appropriate parameter environment seems to fix#54822 and also make one of the
cases in src/test/ui/nll/trait-associated-constant.rs that should compile
successfully do so. It also (slightly) improves the error message in
src/test/ui/associated-const/associated-const-generic-obligations.rs
CI: Trim some tests from i686-gnu
This removes some tests from the i686-gnu job. This job clocks in at 2hr 56min, and removing these should cut about 10 to 15 minutes, giving a little more breathing room. I suspect these don't need to be tested on every platform.
