Fix `ErrorGuaranteed` unsoundness with stash/steal.
When you stash an error, the error count is incremented. You can then use the non-zero error count to get an `ErrorGuaranteed`. You can then steal the error, which decrements the error count. You can then cancel the error.
Example code:
```
fn unsound(dcx: &DiagCtxt) -> ErrorGuaranteed {
let sp = rustc_span::DUMMY_SP;
let k = rustc_errors::StashKey::Cycle;
dcx.struct_err("bogus").stash(sp, k); // increment error count on stash
let guar = dcx.has_errors().unwrap(); // ErrorGuaranteed from error count > 0
let err = dcx.steal_diagnostic(sp, k).unwrap(); // decrement error count on steal
err.cancel(); // cancel error
guar // ErrorGuaranteed with no error emitted!
}
```
This commit fixes the problem in the simplest way: by not counting stashed errors in `DiagCtxt::{err_count,has_errors}`.
However, just doing this without any other changes leads to over 40 ui test failures. Mostly because of uninteresting extra errors (many saying "type annotations needed" when type inference fails), and in a few cases, due to delayed bugs causing ICEs when no normal errors are printed.
To fix these, this commit adds `DiagCtxt::stashed_err_count`, and uses it in three places alongside `DiagCtxt::{has_errors,err_count}`. It's dodgy to rely on it, because unlike `DiagCtxt::err_count` it can go up and down. But it's needed to preserve existing behaviour, and at least the three places that need it are now obvious.
r? oli-obk
Fix more `ty::Error` ICEs in MIR passes
Fixes#120791 - Add a check for `ty::Error` in the `ByMove` coroutine pass
Fixes#120816 - Add a check for `ty::Error` in the MIR validator
Also a drive-by fix for a FIXME I had asked oli to add
r? oli-obk
A drive-by rewrite of `give_region_a_name()`
This drive-by rewrite makes the cache-updating nature of the method clearer, using the Entry API into the hash table for region names to capture the update-insert nature of the method. May be marginally more efficient since it only runtime-borrows and indexes the map once, but in this context the performance impact is almost certainly completely negligible.
Note that this commit should preserve all externally visible behaviour. Notably, it preserves the debug logging:
1. printing even in the case of a `None` for the new computed name, and
2. only printing on new values, begin silent on reused values
Invert diagnostic lints.
That is, change `diagnostic_outside_of_impl` and `untranslatable_diagnostic` from `allow` to `deny`, because more than half of the compiler has been converted to use translated diagnostics.
This commit removes more `deny` attributes than it adds `allow` attributes, which proves that this change is warranted.
r? ````@davidtwco````
Make privacy visitor use types more (instead of HIR)
r? ``@petrochenkov``
This is a prerequisite to normalizing projections, as otherwise we have too many invalid bound vars (hir_ty_to_ty is creating types that have bound vars, but no binder).
The commits are still chaotic, I'm gonna clean them up, but I just wanted to let you know about the general direction and wondering if we could land this before adding normalization, as normalization is where behavioral changes happen, and I'd like to keep that part as minimal as possible.
[context can be found on zulip](https://rust-lang.zulipchat.com/#narrow/stream/315482-t-compiler.2Fetc.2Fopaque-types/topic/weak.20type.20aliases.20and.20privacy)
Toggle assert_unsafe_precondition in codegen instead of expansion
The goal of this PR is to make some of the unsafe precondition checks in the standard library available in debug builds. Some UI tests are included to verify that it does that.
The diff is large, but most of it is blessing mir-opt tests and I've also split up this PR so it can be reviewed commit-by-commit.
This PR:
1. Adds a new intrinsic, `debug_assertions` which is lowered to a new MIR NullOp, and only to a constant after monomorphization
2. Rewrites `assume_unsafe_precondition` to check the new intrinsic, and be monomorphic.
3. Skips codegen of the `assume` intrinsic in unoptimized builds, because that was silly before but with these checks it's *very* silly
4. The checks with the most overhead are `ptr::read`/`ptr::write` and `NonNull::new_unchecked`. I've simply added `#[cfg(debug_assertions)]` to the checks for `ptr::read`/`ptr::write` because I was unable to come up with any (good) ideas for decreasing their impact. But for `NonNull::new_unchecked` I found that the majority of callers can use a different function, often a safe one.
Yes, this PR slows down the compile time of some programs. But in our benchmark suite it's never more than 1% icount, and the average icount change in debug-full programs is 0.22%. I think that is acceptable for such an improvement in developer experience.
https://github.com/rust-lang/rust/issues/120539#issuecomment-1922687101
When you stash an error, the error count is incremented. You can then
use the non-zero error count to get an `ErrorGuaranteed`. You can then
steal the error, which decrements the error count. You can then cancel
the error.
Example code:
```
fn unsound(dcx: &DiagCtxt) -> ErrorGuaranteed {
let sp = rustc_span::DUMMY_SP;
let k = rustc_errors::StashKey::Cycle;
dcx.struct_err("bogus").stash(sp, k); // increment error count on stash
let guar = dcx.has_errors().unwrap(); // ErrorGuaranteed from error count > 0
let err = dcx.steal_diagnostic(sp, k).unwrap(); // decrement error count on steal
err.cancel(); // cancel error
guar // ErrorGuaranteed with no error emitted!
}
```
This commit fixes the problem in the simplest way: by not counting
stashed errors in `DiagCtxt::{err_count,has_errors}`.
However, just doing this without any other changes leads to over 40 ui
test failures. Mostly because of uninteresting extra errors (many saying
"type annotations needed" when type inference fails), and in a few
cases, due to delayed bugs causing ICEs when no normal errors are
printed.
To fix these, this commit adds `DiagCtxt::stashed_err_count`, and uses
it in three places alongside `DiagCtxt::{has_errors,err_count}`. It's
dodgy to rely on it, because unlike `DiagCtxt::err_count` it can go up
and down. But it's needed to preserve existing behaviour, and at least
the three places that need it are now obvious.
Fix mir pass ICE in the presence of other errors
fixes#120779
it is impossible to add a ui test for this, because it only reproduces in build-fail, but a test that also has errors in check-fail mode can't be made build-fail 🙃
I would have to add a run-make test or sth, which is overkill for such a tiny thing imo.
Make `min_exhaustive_patterns` match `exhaustive_patterns` better
Split off from https://github.com/rust-lang/rust/pull/120742.
There remained two edge cases where `min_exhaustive_patterns` wasn't behaving like `exhaustive_patterns`. This fixes them, and tests the feature in a bunch more cases. I essentially went through all uses of `exhaustive_patterns` to see which ones would be interesting to compare between the two features.
r? `@compiler-errors`
No need to take `ImplTraitContext` by ref
We used to mutate `ImplTraitContext`, so it used to be `&mut` mutable ref. Then I think it used to have non-`Copy` data in it, so we took it by `&` ref. Now, none of that remains, so just copy it around.
Remove unused args from functions
`#[instrument]` suppresses the unused arguments from a function, *and* suppresses unused methods too! This PR removes things which are only used via `#[instrument]` calls, and fixes some other errors (privacy?) that I will comment inline.
It's possible that some of these arguments were being passed in for the purposes of being instrumented, but I am unconvinced by most of them.
Introduce `enter_forall` to supercede `instantiate_binder_with_placeholders`
r? `@lcnr`
Long term we'd like to experiment with decrementing the universe count after "exiting" binders so that we do not end up creating infer vars in non-root universes even when they logically reside in the root universe. The fact that we dont do this currently results in a number of issues in the new trait solver where we consider goals to be ambiguous because otherwise it would require lowering the universe of an infer var. i.e. the goal `?x.0 eq <T as Trait<?y.1>>::Assoc` where the alias is rigid would not be able to instantiate `?x` with the alias as there would be a universe error.
This PR is the first-ish sort of step towards being able to implement this as eventually we would want to decrement the universe in `enter_forall`. Unfortunately its Difficult to actually implement decrementing universes nicely so this is a separate step which moves us closer to the long term goal ✨
Rollup of 9 pull requests
Successful merges:
- #119592 (resolve: Unload speculatively resolved crates before freezing cstore)
- #120103 (Make it so that async-fn-in-trait is compatible with a concrete future in implementation)
- #120206 (hir: Make sure all `HirId`s have corresponding HIR `Node`s)
- #120214 (match lowering: consistently lower bindings deepest-first)
- #120688 (GVN: also turn moves into copies with projections)
- #120702 (docs: also check the inline stmt during redundant link check)
- #120727 (exhaustiveness: Prefer "`0..MAX` not covered" to "`_` not covered")
- #120734 (Add `SubdiagnosticMessageOp` as a trait alias.)
- #120739 (improve pretty printing for associated items in trait objects)
r? `@ghost`
`@rustbot` modify labels: rollup
