Make "Assemble stage1 compiler" orders of magnitude faster
This used to take upwards of 5 seconds for me locally. I found that the culprit was copying the downloaded LLVM shared object:
```
[22:28:03] Install "/home/jnelson/rust-lang/rust/build/x86_64-unknown-linux-gnu/ci-llvm/lib/libLLVM-14-rust-1.62.0-nightly.so" to "/home/jnelson/rust-lang/rust/build/x86_64-unknown-linux-gnu/stage1/lib/rustlib/x86_64-unknown-linux-gnu/lib/libLLVM-14-rust-1.62.0-nightly.so"
[22:28:09] c Sysroot { compiler: Compiler { stage: 1, host: x86_64-unknown-linux-gnu(x86_64-unknown-linux-gnu) } }
```
It turned out that `install()` used full copies unconditionally. Change it to try using a hard-link before falling back to copying.
Fortify handing of where bounds on trait & trait alias definitions
Closes https://github.com/rust-lang/rust/issues/96664
Closes https://github.com/rust-lang/rust/issues/96665
Since https://github.com/rust-lang/rust/pull/93803, when listing all bounds and predicates we now need to account for the possible presence of predicates on any of the generic parameters. Both bugs were hidden by the special handling of bounds at the generic parameter declaration position.
Trait alias expansion used to confuse predicates on `Self` and where predicates.
Exiting too late when listing all the bounds caused a cycle error.
Optimize switch sources representation and usage
* Avoid constructing switch sources unless necessary - switch sources are used by backward analysis with a custom switch int edge effects, but are otherwise unnecessarily computed.
* Use sparse representation of switch sources to avoid quadratic space overhead.
Some subst cleanup
Two separate things here. Both changes are useful for some refactoring I'm doing to add an "EarlyBinder" newtype. (Part of chalkification).
1) Remove `subst_spanned` and just use `subst`. It wasn't used much anyways. In practice, I think we can probably get most of the info just from the actual error message. If not, outputting logs should do the trick. (The specific line probably wouldn't help much anyways).
2) Call `.subst()` before `replace_bound_vars_with_fresh_vars` and `erase_late_bound_regions` in three places that do the opposite. I think there might have been some time in the past that the order here matter for something, but this shouldn't be the case anymore. Conceptually, it makes more sense to the of the *early bound* vars on `fn`s as "outside" the late bound vars.
Actually fix ICE from #96583
PR #96746 fixed a very similar bug, so the same logic is used in a different place.
I originally concluded that the two issues (#96583 and #96738) were identical by comparing the backtrace, but I didn't look close enough.
Revert "Implement [OsStr]::join", which was merged without FCP.
This reverts commit 4fcbc53820, see https://github.com/rust-lang/rust/pull/96744. (I'm terribly sorry, and truly don't remember r+ing it, or even having seen it before yesterday, which is... genuinely very worrisome for me).
r? `@m-ou-se`
Warn on unused `#[doc(hidden)]` attributes on trait impl items
[Zulip conversation](https://rust-lang.zulipchat.com/#narrow/stream/266220-rustdoc/topic/.E2.9C.94.20Validy.20checks.20for.20.60.23.5Bdoc.28hidden.29.5D.60).
Whether an associated item in a trait impl is shown or hidden in the documentation entirely depends on the corresponding item in the trait declaration. Rustdoc completely ignores `#[doc(hidden)]` attributes on impl items. No error or warning is emitted:
```rust
pub trait Tr { fn f(); }
pub struct Ty;
impl Tr for Ty { #[doc(hidden)] fn f() {} }
// ^^^^^^^^^^^^^^ ignored by rustdoc and currently
// no error or warning issued
```
This may lead users to the wrong belief that the attribute has an effect. In fact, several such cases are found in the standard library (I've removed all of them in this PR).
There does not seem to exist any incentive to allow this in the future either: Impl'ing a trait for a type means the type *fully* conforms to its API. Users can add `#[doc(hidden)]` to the whole impl if they want to hide the implementation or add the attribute to the corresponding associated item in the trait declaration to hide the specific item. Hiding an implementation of an associated item does not make much sense: The associated item can still be found on the trait page.
This PR emits the warn-by-default lint `unused_attribute` for this case with a future-incompat warning.
`@rustbot` label T-compiler T-rustdoc A-lint
Improve floating point documentation
This is my attempt to improve/solve https://github.com/rust-lang/rust/issues/95468 and https://github.com/rust-lang/rust/issues/73328 .
Added/refined explanations:
- Refine the "NaN as a special value" top level explanation of f32
- Refine `const NAN` docstring: add an explanation about there being multitude of NaN bitpatterns and disclaimer about the portability/stability guarantees.
- Refine `fn is_sign_positive` and `fn is_sign_negative` docstrings: add disclaimer about the sign bit of NaNs.
- Refine `fn min` and `fn max` docstrings: explain the semantics and their relationship to the standard and libm better.
- Refine `fn trunc` docstrings: explain the semantics slightly more.
- Refine `fn powi` docstrings: add disclaimer that the rounding behaviour might be different from `powf`.
- Refine `fn copysign` docstrings: add disclaimer about payloads of NaNs.
- Refine `minimum` and `maximum`: add disclaimer that "propagating NaN" doesn't mean that propagating the NaN bit patterns is guaranteed.
- Refine `max` and `min` docstrings: add "ignoring NaN" to bring the one-row explanation to parity with `minimum` and `maximum`.
Cosmetic changes:
- Reword `NaN` and `NAN` as plain "NaN", unless they refer to the specific `const NAN`.
- Reword "a number" to `self` in function docstrings to clarify.
- Remove "Returns NAN if the number is NAN" from `abs`, as this is told to be the default behavior in the top explanation.
optimize `promote_consts` by caching the results of `validate_local`
From the FIXME in the impl of `promote_consts`. Early return the `validate_local` should save some compile time.
`qualif_local` is similar to this, but requires futher changing because there are different types of qualif checks. If this PR is effective, I will do it as well.
Remove `#[rustc_deprecated]`
This removes `#[rustc_deprecated]` and introduces diagnostics to help users to the right direction (that being `#[deprecated]`). All uses of `#[rustc_deprecated]` have been converted. CI is expected to fail initially; this requires #95958, which includes converting `stdarch`.
I plan on following up in a short while (maybe a bootstrap cycle?) removing the diagnostics, as they're only intended to be short-term.
This used to take upwards of 5 seconds for me locally. I found that the culprit was copying the downloaded LLVM shared object:
```
[22:28:03] Install "/home/jnelson/rust-lang/rust/build/x86_64-unknown-linux-gnu/ci-llvm/lib/libLLVM-14-rust-1.62.0-nightly.so" to "/home/jnelson/rust-lang/rust/build/x86_64-unknown-linux-gnu/stage1/lib/rustlib/x86_64-unknown-linux-gnu/lib/libLLVM-14-rust-1.62.0-nightly.so"
[22:28:09] c Sysroot { compiler: Compiler { stage: 1, host: x86_64-unknown-linux-gnu(x86_64-unknown-linux-gnu) } }
```
It turned out that `install()` used full copies unconditionally. Change it to try using a hard-link before falling back to copying.
Support tool lints with the `#[expect]` attribute (RFC 2383)
This PR fixes the ICE https://github.com/rust-lang/rust/issues/94953 by making the assert for converted expectation IDs conditional.
Additionally, it moves the lint expectation check into a separate query to support rustdoc and other tools. On the way, I've also added some tests to ensure that the attribute works for Clippy and rustdoc lints.
The number of changes comes from the long test file. This may look like a monster PR, this may smell like a monster PR and this may be a monster PR, but it's a harmless monster. 🦕
---
Closes: https://github.com/rust-lang/rust/issues/94953
cc: https://github.com/rust-lang/rust/issues/85549
r? `@wesleywiser`
cc: `@rust-lang/rustdoc`
Switch sources are used by backward analysis with a custom switch int
edge effects, but are otherwise unnecessarily computed.
Delay the computation until we know that switch sources are indeed
required and avoid the computation otherwise.
Further elaborate the lack of guarantees from `Hasher`
I realized that I got too excited in #94598 by adding new methods, and forgot to do the documentation to really answer the core question in #94026.
This PR just has that doc update.
r? `@Amanieu`
Add more diagnostic items
This just adds a handful diagnostic items I noticed were missing.
Would it be worth doing this for all of the remaining types? I'm willing to do it if it'd be helpful.
