- In `emit_producing_error_guaranteed`, only allow `Level::Error`.
- In `emit_diagnostic`, only produce `ErrorGuaranteed` for `Level` and
`DelayedBug`. (Not `Bug` or `Fatal`. They don't need it, because the
relevant `emit` methods abort.)
- Add/update various comments.
`main_args` calls `from_matches`, which does lots of initialization. If
anything goes wrong, `from_matches` emits an error message and returns
`Err(1)` (or `Err(3)`). `main_args` then turns the `Err(1)` into
`Err(ErrorGuaranteed)`, because that's what `catch_with_exit_code`
requires on error. But `catch_with_exit_code` doesn't do anything with
the `ErrorGuaranteed`, it just exits with `EXIT_FAILURE`.
We can avoid the creation of the `ErrorGuaranteed` (which requires
an undesirable `unchecked_claim_error_was_emitted` call), by changing
`from_matches` to instead eagerly abort if anything goes wrong. The
behaviour from the user's point of view is the same: an early abort with
an `EXIT_FAILURE` exit code.
And we can also simplify `from_matches` to return an `Option` instead of
a `Result`:
- Old `Err(0)` case --> `None`
- Old `Err(_)` case --> fatal error.
This requires similar changes to `ScrapeExamplesOptions::new` and
`load_call_locations`.
By making non-unicode arguments a fatal error instead of a warning, we
don't need to handle what comes after, which avoids the need for an
`unchecked_claim_error_was_emitted` call.
When `catch_fatal_errors` catches a `FatalErrorMarker`, it returns an
`ErrorGuaranteed` that is conjured out of thin air with
`unchecked_claim_error_was_emitted`. But that `ErrorGuaranteed` is never
used.
This commit changes it to instead conjure a `FatalError` out of thin
air. (A non-deprecated action!) This makes more sense because
`FatalError` and `FatalErrorMarker` are a natural pairing -- a
`FatalErrorMarker` is created by calling `FatalError::raise`, so this is
effectively getting back the original `FatalError`.
This requires a tiny change in `catch_with_exit_code`. The old result of
the `catch_fatal_errors` call there was
`Result<Result<(), ErrorGuaranteed>, ErrorGuaranteed>` which could be
`flatten`ed into `Result<(), ErrorGuaranteed>`. The new result of the
`catch_fatal_errors` calls is
`Result<Result<(), ErrorGuaranteed>, FatalError>`, which can't be
`flatten`ed but is still easily matched for the success case.
Rollup of 12 pull requests
Successful merges:
- #120520 (Some cleanups around diagnostic levels.)
- #120575 (Simplify codegen diagnostic handling)
- #120597 (Suggest `[tail @ ..]` on `[..tail]` and `[...tail]` where `tail` is unresolved)
- #120602 (rustc_monomorphize: fix outdated comment in partition)
- #120609 (hir: Stop keeping prefixes for most of `use` list stems)
- #120631 (Emit a diagnostic for invalid target options)
- #120632 (For E0223, suggest associated functions that are similar to the path)
- #120670 (cleanup effect var handling)
- #120673 (rustc_metadata: fix typo)
- #120683 (miri: fix ICE with symbolic alignment check on extern static)
- #120690 (Remove b-naber from the compiler review rotation)
- #120713 (Make async closures test use async bound modifier)
r? `@ghost`
`@rustbot` modify labels: rollup
Remove b-naber from the compiler review rotation
They have been inactive for over two months and occasionally S-waiting-on-review PRs didn't get reviewed for a while. I've stolen a few of them.
I've discussed this with `@b-naber` in private and they agreed to leave the review rotation for now.
Thank you very much for all the helpful PR reviews you did in the past! <3
r? compiler
hir: Stop keeping prefixes for most of `use` list stems
And make sure all other imports have non-empty resolution lists.
Addresses one of FIXMEs in https://github.com/rust-lang/rust/pull/120206.
rustc_monomorphize: fix outdated comment in partition
`max_cgu_count` was removed in 51821515b3, but not comment (usage in `merge_codegen_units` was removed earlier).
r? `@nnethercote`
Suggest `[tail @ ..]` on `[..tail]` and `[...tail]` where `tail` is unresolved
Fixes#120591.
~~Will conflict with #120570~~ (rebased).
r? estebank or compiler
Some cleanups around diagnostic levels.
Plus some refactoring in and around diagnostic levels and emission. Details in the individual commit logs.
r? ````@oli-obk````
Rework support for async closures; allow them to return futures that borrow from the closure's captures
This PR implements a new lowering for async closures via `TyKind::CoroutineClosure` which handles the curious relationship between the closure and the coroutine that it returns.
I wrote up a bunch in [this hackmd](https://hackmd.io/`@compiler-errors/S1HvqQxca)` which will be copied to the dev guide after this PR lands, and hopefully left sufficient comments in the source code explaining why this change is as large as it is.
This also necessitates that they begin implementing the `AsyncFn`-family of traits, rather than the `Fn`-family of traits -- if you need `Fn` implementations, you should probably use the non-sugar `|| async {}` syntax instead.
Notably this PR does not yet implement `async Fn()` syntax sugar for bounds, but I expect to add those soon (**edit:** #120392). For now, users must use `AsyncFn()` traits directly, which necessitates adding the `async_fn_traits` feature gate as well. I will add this as a follow-up very soon.
r? oli-obk
This is based on top of #120322, but that PR is minimal.
Actually abort in -Zpanic-abort-tests
When a test fails in panic=abort, it can be useful to have a debugger or other tooling hook into the `abort()` call for debugging. Doing this some other way would require it to hard code details of Rust's panic machinery.
There's no reason we couldn't have done this in the first place; using a single exit code for "success" or "failed" was just simpler. Now we are aware of the special exit codes for posix and windows platforms, logging a special error if an unrecognized code is used on those platforms, and falling back to just "failure" on other platforms.
This continues to account for `#[should_panic]` inside the test process itself, so there's no risk of misinterpreting a random call to `abort()` as an expected panic. Any exit code besides `TR_OK` is logged as a test failure.
As an added benefit, this would allow us to support panic=immediate_abort (but not `#[should_panic]`), without noise about unexpected exit codes when a test fails.
Introduce support for `async` bound modifier on `Fn*` traits
Adds `async` to the list of `TraitBoundModifiers`, which instructs AST lowering to map the trait to an async flavor of the trait. For now, this is only supported for `Fn*` to `AsyncFn*`, and I expect that this manual mapping via lang items will be replaced with a better system in the future.
The motivation for adding these bounds is to separate the users of async closures from the exact trait desugaring of their callable bounds. Instead of users needing to be concerned with the `AsyncFn` trait, they should be able to write `async Fn()` and it will desugar to whatever underlying trait we decide is best for the lowering of async closures.
Note: rustfmt support can be done in the rustfmt repo after a subtree sync.
pattern_analysis: Gracefully abort on type incompatibility
This leaves the option for a consumer of the crate to return `Err` instead of panicking on type error. rust-analyzer could use that (e.g. https://github.com/rust-lang/rust-analyzer/issues/15808).
Since the only use of `TypeCx::bug` is in `Constructor::is_covered_by`, it is tempting to return `false` instead of `Err()`, but that would cause "non-exhaustive match" false positives.
r? `@compiler-errors`
Rollup of 8 pull requests
Successful merges:
- #113833 (`std::error::Error` -> Trait Implementations: lifetimes consistency improvement)
- #115386 (PartialEq, PartialOrd: update and synchronize handling of transitive chains)
- #116284 (make matching on NaN a hard error, and remove the rest of illegal_floating_point_literal_pattern)
- #118960 (Add LocalWaker and ContextBuilder types to core, and LocalWake trait to alloc.)
- #120384 (Use `<T, U>` for array/slice equality `impl`s)
- #120518 (riscv only supports split_debuginfo=off for now)
- #120657 (Remove unused struct)
- #120661 (target: default to the medium code model on LoongArch targets)
r? `@ghost`
`@rustbot` modify labels: rollup
All the other `emit`/`emit_diagnostic` methods were recently made
consuming (e.g. #119606), but this one wasn't. But it makes sense to.
Much of this is straightforward, and lots of `clone` calls are avoided.
There are a couple of tricky bits.
- `Emitter::primary_span_formatted` no longer takes a `Diagnostic` and
returns a pair. Instead it takes the two fields from `Diagnostic` that
it used (`span` and `suggestions`) as `&mut`, and modifies them. This
is necessary to avoid the cloning of `diag.children` in two emitters.
- `from_errors_diagnostic` is rearranged so various uses of `diag` occur
before the consuming `emit_diagnostic` call.
target: default to the medium code model on LoongArch targets
The Rust LoongArch targets have been using the default LLVM code model so far, which is "small" in LLVM-speak and "normal" in LoongArch-speak. As [described][1] in the "Code Model" section of LoongArch ELF psABI spec v20231219, one can only make function calls as far as ±128MiB with the "normal" code model; this is insufficient for very large software containing Rust components that needs to be linked into the big text section, such as Chromium.
Because:
* we do not want to ask users to recompile std if they are to build such software,
* objects compiled with larger code models can be linked with those with smaller code models without problems, and
* the "medium" code model is comparable to the "small"/"normal" one performance-wise (same data access pattern; each function call becomes 2-insn long and indirect, but this may be relaxed back into the direct 1-insn form in a future LLVM version), but is able to perform function calls within ±128GiB,
it is better to just switch the targets to the "medium" code model, which is also "medium" in LLVM-speak.
[1]: https://github.com/loongson/la-abi-specs/blob/v2.30/laelf.adoc#code-models
riscv only supports split_debuginfo=off for now
Disable packed/unpacked options for riscv linux/android. Other riscv targets already only have the off option.
The packed/unpacked options might be supported in the future. See upstream issue for more details:
https://github.com/llvm/llvm-project/issues/56642Fixes#110224
