Const stability checks v2
The const stability system has served us well ever since `const fn` were first stabilized. It's main feature is that it enforces *recursive* validity -- a stable const fn cannot internally make use of unstable const features without an explicit marker in the form of `#[rustc_allow_const_fn_unstable]`. This is done to make sure that we don't accidentally expose unstable const features on stable in a way that would be hard to take back. As part of this, it is enforced that a `#[rustc_const_stable]` can only call `#[rustc_const_stable]` functions. However, some problems have been coming up with increased usage:
- It is baffling that we have to mark private or even unstable functions as `#[rustc_const_stable]` when they are used as helpers in regular stable `const fn`, and often people will rather add `#[rustc_allow_const_fn_unstable]` instead which was not our intention.
- The system has several gaping holes: a private `const fn` without stability attributes whose inherited stability (walking up parent modules) is `#[stable]` is allowed to call *arbitrary* unstable const operations, but can itself be called from stable `const fn`. Similarly, `#[allow_internal_unstable]` on a macro completely bypasses the recursive nature of the check.
Fundamentally, the problem is that we have *three* disjoint categories of functions, and not enough attributes to distinguish them:
1. const-stable functions
2. private/unstable functions that are meant to be callable from const-stable functions
3. functions that can make use of unstable const features
Functions in the first two categories cannot use unstable const features and they can only call functions from the first two categories.
This PR implements the following system:
- `#[rustc_const_stable]` puts functions in the first category. It may only be applied to `#[stable]` functions.
- `#[rustc_const_unstable]` by default puts functions in the third category. The new attribute `#[rustc_const_stable_indirect]` can be added to such a function to move it into the second category.
- `const fn` without a const stability marker are in the second category if they are still unstable. They automatically inherit the feature gate for regular calls, it can now also be used for const-calls.
Also, all the holes mentioned above have been closed. There's still one potential hole that is hard to avoid, which is when MIR building automatically inserts calls to a particular function in stable functions -- which happens in the panic machinery. Those need to be manually marked `#[rustc_const_stable_indirect]` to be sure they follow recursive const stability. But that's a fairly rare and special case so IMO it's fine.
The net effect of this is that a `#[unstable]` or unmarked function can be constified simply by marking it as `const fn`, and it will then be const-callable from stable `const fn` and subject to recursive const stability requirements. If it is publicly reachable (which implies it cannot be unmarked), it will be const-unstable under the same feature gate. Only if the function ever becomes `#[stable]` does it need a `#[rustc_const_unstable]` or `#[rustc_const_stable]` marker to decide if this should also imply const-stability.
Adding `#[rustc_const_unstable]` is only needed for (a) functions that need to use unstable const lang features (including intrinsics), or (b) `#[stable]` functions that are not yet intended to be const-stable. Adding `#[rustc_const_stable]` is only needed for functions that are actually meant to be directly callable from stable const code. `#[rustc_const_stable_indirect]` is used to mark intrinsics as const-callable and for `#[rustc_const_unstable]` functions that are actually called from other, exposed-on-stable `const fn`. No other attributes are required.
Also see the updated dev-guide at https://github.com/rust-lang/rustc-dev-guide/pull/2098.
I think in the future we may want to tweak this further, so that in the hopefully common case where a public function's const-stability just exactly mirrors its regular stability, we never have to add any attribute. But right now, once the function is stable this requires `#[rustc_const_stable]`.
### Open question
There is one point I could see we might want to do differently, and that is putting `#[rustc_const_unstable]` functions (but not intrinsics) in category 2 by default, and requiring an extra attribute for `#[rustc_const_not_exposed_on_stable]` or so. This would require a bunch of extra annotations, but would have the advantage that turning a `#[rustc_const_unstable]` into `#[rustc_const_stable]` will never change the way the function is const-checked. Currently, we often discover in the const stabilization PR that a function needs some other unstable const things, and then we rush to quickly deal with that. In this alternative universe, we'd work towards getting rid of the `rustc_const_not_exposed_on_stable` before stabilization, and once that is done stabilization becomes a trivial matter. `#[rustc_const_stable_indirect]` would then only be used for intrinsics.
I think I like this idea, but might want to do it in a follow-up PR, as it will need a whole bunch of annotations in the standard library. Also, we probably want to convert all const intrinsics to the "new" form (`#[rustc_intrinsic]` instead of an `extern` block) before doing this to avoid having to deal with two different ways of declaring intrinsics.
Cc `@rust-lang/wg-const-eval` `@rust-lang/libs-api`
Part of https://github.com/rust-lang/rust/issues/129815 (but not finished since this is not yet sufficient to safely let us expose `const fn` from hashbrown)
Fixes https://github.com/rust-lang/rust/issues/131073 by making it so that const-stable functions are always stable
try-job: test-various
Fundamentally, we have *three* disjoint categories of functions:
1. const-stable functions
2. private/unstable functions that are meant to be callable from const-stable functions
3. functions that can make use of unstable const features
This PR implements the following system:
- `#[rustc_const_stable]` puts functions in the first category. It may only be applied to `#[stable]` functions.
- `#[rustc_const_unstable]` by default puts functions in the third category. The new attribute `#[rustc_const_stable_indirect]` can be added to such a function to move it into the second category.
- `const fn` without a const stability marker are in the second category if they are still unstable. They automatically inherit the feature gate for regular calls, it can now also be used for const-calls.
Also, several holes in recursive const stability checking are being closed.
There's still one potential hole that is hard to avoid, which is when MIR
building automatically inserts calls to a particular function in stable
functions -- which happens in the panic machinery. Those need to *not* be
`rustc_const_unstable` (or manually get a `rustc_const_stable_indirect`) to be
sure they follow recursive const stability. But that's a fairly rare and special
case so IMO it's fine.
The net effect of this is that a `#[unstable]` or unmarked function can be
constified simply by marking it as `const fn`, and it will then be
const-callable from stable `const fn` and subject to recursive const stability
requirements. If it is publicly reachable (which implies it cannot be unmarked),
it will be const-unstable under the same feature gate. Only if the function ever
becomes `#[stable]` does it need a `#[rustc_const_unstable]` or
`#[rustc_const_stable]` marker to decide if this should also imply
const-stability.
Adding `#[rustc_const_unstable]` is only needed for (a) functions that need to
use unstable const lang features (including intrinsics), or (b) `#[stable]`
functions that are not yet intended to be const-stable. Adding
`#[rustc_const_stable]` is only needed for functions that are actually meant to
be directly callable from stable const code. `#[rustc_const_stable_indirect]` is
used to mark intrinsics as const-callable and for `#[rustc_const_unstable]`
functions that are actually called from other, exposed-on-stable `const fn`. No
other attributes are required.
Rename Receiver -> LegacyReceiver
As part of the "arbitrary self types v2" project, we are going to replace the current `Receiver` trait with a new mechanism based on a new, different `Receiver` trait.
This PR renames the old trait to get it out the way. Naming is hard. Options considered included:
* HardCodedReceiver (because it should only be used for things in the standard library, and hence is sort-of hard coded)
* LegacyReceiver
* TargetLessReceiver
* OldReceiver
These are all bad names, but fortunately this will be temporary. Assuming the new mechanism proceeds to stabilization as intended, the legacy trait will be removed altogether.
Although we expect this trait to be used only in the standard library, we suspect it may be in use elsehwere, so we're landing this change separately to identify any surprising breakages.
It's known that this trait is used within the Rust for Linux project; a patch is in progress to remove their dependency.
This is a part of the arbitrary self types v2 project,
https://github.com/rust-lang/rfcs/pull/3519https://github.com/rust-lang/rust/issues/44874
r? `@wesleywiser`
Optimize `Rc<T>::default`
The missing piece of https://github.com/rust-lang/rust/pull/131460.
Also refactored `Arc<T>::default` by using a safe `NonNull::from(Box::leak(_))` to replace the unnecessarily unsafe call to `NonNull::new_unchecked(Box::into_raw(_))`. The remaining unsafety is coming from `[Rc|Arc]::from_inner`, which is safe from the construction of `[Rc|Arc]Inner`.
better default capacity for str::replace
Adds smarter capacity for str::replace in cases where we know that the output will be at least as long as the original string.
As part of the "arbitrary self types v2" project, we are going to
replace the current `Receiver` trait with a new mechanism based on a
new, different `Receiver` trait.
This PR renames the old trait to get it out the way. Naming is hard.
Options considered included:
* HardCodedReceiver (because it should only be used for things in the
standard library, and hence is sort-of hard coded)
* LegacyReceiver
* TargetLessReceiver
* OldReceiver
These are all bad names, but fortunately this will be temporary.
Assuming the new mechanism proceeds to stabilization as intended, the
legacy trait will be removed altogether.
Although we expect this trait to be used only in the standard library,
we suspect it may be in use elsehwere, so we're landing this change
separately to identify any surprising breakages.
It's known that this trait is used within the Rust for Linux project; a
patch is in progress to remove their dependency.
This is a part of the arbitrary self types v2 project,
https://github.com/rust-lang/rfcs/pull/3519https://github.com/rust-lang/rust/issues/44874
r? @wesleywiser
Fix predicate signatures in retain_mut docs
This is my first PR here so let me know if I'm doing anything wrong.
The docs for `retain_mut` in `LinkedList` and `VecDeque` say the predicate takes `&e`, but it should be `&mut e` to match the actual signature. `Vec` [has it documented](https://doc.rust-lang.org/std/vec/struct.Vec.html#method.retain_mut) correctly already.
optimize str.replace
Adds a fast path for str.replace for the ascii to ascii case. This allows for autovectorizing the code. Also should this instead be done with specialization? This way we could remove one branch. I think it is the kind of branch that is easy to predict though.
Benchmark for the fast path (replace all "a" with "b" in the rust wikipedia article, using criterion) :
| N | Speedup | Time New (ns) | Time Old (ns) |
|----------|---------|---------------|---------------|
| 2 | 2.03 | 13.567 | 27.576 |
| 8 | 1.73 | 17.478 | 30.259 |
| 11 | 2.46 | 18.296 | 45.055 |
| 16 | 2.71 | 17.181 | 46.526 |
| 37 | 4.43 | 18.526 | 81.997 |
| 64 | 8.54 | 18.670 | 159.470 |
| 200 | 9.82 | 29.634 | 291.010 |
| 2000 | 24.34 | 81.114 | 1974.300 |
| 20000 | 30.61 | 598.520 | 18318.000 |
| 1000000 | 29.31 | 33458.000 | 980540.000 |
Implemented `FromStr` for `CString` and `TryFrom<CString>` for `String`
The motivation of this change is making it possible to use `CString` in generic methods with `FromStr` and `TryInto<String>` trait bounds. The same traits are already implemented for `OsString` which is an FFI type too.
remove const_cow_is_borrowed feature gate
The two functions guarded by this are still unstable, and there's no reason to require a separate feature gate for their const-ness -- we can just have `cow_is_borrowed` cover both kinds of stability.
Cc #65143
Port sort-research-rs test suite to Rust stdlib tests
This PR is a followup to https://github.com/rust-lang/rust/pull/124032. It replaces the tests that test the various sort functions in the standard library with a test-suite developed as part of https://github.com/Voultapher/sort-research-rs. The current tests suffer a couple of problems:
- They don't cover important real world patterns that the implementations take advantage of and execute special code for.
- The input lengths tested miss out on code paths. For example, important safety property tests never reach the quicksort part of the implementation.
- The miri side is often limited to `len <= 20` which means it very thoroughly tests the insertion sort, which accounts for 19 out of 1.5k LoC.
- They are split into to core and alloc, causing code duplication and uneven coverage.
- ~~The randomness is tied to a caller location, wasting the space exploration capabilities of randomized testing.~~ The randomness is not repeatable, as it relies on `std:#️⃣:RandomState::new().build_hasher()`.
Most of these issues existed before https://github.com/rust-lang/rust/pull/124032, but they are intensified by it. One thing that is new and requires additional testing, is that the new sort implementations specialize based on type properties. For example `Freeze` and non `Freeze` execute different code paths.
Effectively there are three dimensions that matter:
- Input type
- Input length
- Input pattern
The ported test-suite tests various properties along all three dimensions, greatly improving test coverage. It side-steps the miri issue by preferring sampled approaches. For example the test that checks if after a panic the set of elements is still the original one, doesn't do so for every single possible panic opportunity but rather it picks one at random, and performs this test across a range of input length, which varies the panic point across them. This allows regular execution to easily test inputs of length 10k, and miri execution up to 100 which covers significantly more code. The randomness used is tied to a fixed - but random per process execution - seed. This allows for fully repeatable tests and fuzzer like exploration across multiple runs.
Structure wise, the tests are previously found in the core integration tests for `sort_unstable` and alloc unit tests for `sort`. The new test-suite was developed to be a purely black-box approach, which makes integration testing the better place, because it can't accidentally rely on internal access. Because unwinding support is required the tests can't be in core, even if the implementation is, so they are now part of the alloc integration tests. Are there architectures that can only build and test core and not alloc? If so, do such platforms require sort testing? For what it's worth the current implementation state passes miri `--target mips64-unknown-linux-gnuabi64` which is big endian.
The test-suite also contains tests for properties that were and are given by the current and previous implementations, and likely relied upon by users but weren't tested. For example `self_cmp` tests that the two parameters `a` and `b` passed into the comparison function are never references to the same object, which if the user is sorting for example a `&mut [Mutex<i32>]` could lead to a deadlock.
Instead of using the hashed caller location as rand seed, it uses seconds since unix epoch / 10, which given timestamps in the CI should be reasonably easy to reproduce, but also allows fuzzer like space exploration.
---
Test run-time changes:
Setup:
```
Linux 6.10
rustc 1.83.0-nightly (f79a912d9 2024-09-18)
AMD Ryzen 9 5900X 12-Core Processor (Zen 3 micro-architecture)
CPU boost enabled.
```
master: e9df22f
Before core integration tests:
```
$ LD_LIBRARY_PATH=build/x86_64-unknown-linux-gnu/stage0-std/x86_64-unknown-linux-gnu/release/deps/ hyperfine build/x86_64-unknown-linux-gnu/stage0-std/x86_64-unknown-linux-gnu/release/deps/coretests-219cbd0308a49e2f
Time (mean ± σ): 869.6 ms ± 21.1 ms [User: 1327.6 ms, System: 95.1 ms]
Range (min … max): 845.4 ms … 917.0 ms 10 runs
# MIRIFLAGS="-Zmiri-disable-isolation" to get real time
$ MIRIFLAGS="-Zmiri-disable-isolation" ./x.py miri library/core
finished in 738.44s
```
After core integration tests:
```
$ LD_LIBRARY_PATH=build/x86_64-unknown-linux-gnu/stage0-std/x86_64-unknown-linux-gnu/release/deps/ hyperfine build/x86_64-unknown-linux-gnu/stage0-std/x86_64-unknown-linux-gnu/release/deps/coretests-219cbd0308a49e2f
Time (mean ± σ): 865.1 ms ± 14.7 ms [User: 1283.5 ms, System: 88.4 ms]
Range (min … max): 836.2 ms … 885.7 ms 10 runs
$ MIRIFLAGS="-Zmiri-disable-isolation" ./x.py miri library/core
finished in 752.35s
```
Before alloc unit tests:
```
LD_LIBRARY_PATH=build/x86_64-unknown-linux-gnu/stage0-std/x86_64-unknown-linux-gnu/release/deps/ hyperfine build/x86_64-unknown-linux-gnu/stage0-std/x86_64-unknown-linux-gnu/release/deps/alloc-19c15e6e8565aa54
Time (mean ± σ): 295.0 ms ± 9.9 ms [User: 719.6 ms, System: 35.3 ms]
Range (min … max): 284.9 ms … 319.3 ms 10 runs
$ MIRIFLAGS="-Zmiri-disable-isolation" ./x.py miri library/alloc
finished in 322.75s
```
After alloc unit tests:
```
LD_LIBRARY_PATH=build/x86_64-unknown-linux-gnu/stage0-std/x86_64-unknown-linux-gnu/release/deps/ hyperfine build/x86_64-unknown-linux-gnu/stage0-std/x86_64-unknown-linux-gnu/release/deps/alloc-19c15e6e8565aa54
Time (mean ± σ): 97.4 ms ± 4.1 ms [User: 297.7 ms, System: 28.6 ms]
Range (min … max): 92.3 ms … 109.2 ms 27 runs
$ MIRIFLAGS="-Zmiri-disable-isolation" ./x.py miri library/alloc
finished in 309.18s
```
Before alloc integration tests:
```
$ LD_LIBRARY_PATH=build/x86_64-unknown-linux-gnu/stage0-std/x86_64-unknown-linux-gnu/release/deps/ hyperfine build/x86_64-unknown-linux-gnu/stage0-std/x86_64-unknown-linux-gnu/release/deps/alloctests-439e7300c61a8046
Time (mean ± σ): 103.2 ms ± 1.7 ms [User: 135.7 ms, System: 39.4 ms]
Range (min … max): 99.7 ms … 107.3 ms 28 runs
$ MIRIFLAGS="-Zmiri-disable-isolation" ./x.py miri library/alloc
finished in 231.35s
```
After alloc integration tests:
```
$ LD_LIBRARY_PATH=build/x86_64-unknown-linux-gnu/stage0-std/x86_64-unknown-linux-gnu/release/deps/ hyperfine build/x86_64-unknown-linux-gnu/stage0-std/x86_64-unknown-linux-gnu/release/deps/alloctests-439e7300c61a8046
Time (mean ± σ): 379.8 ms ± 4.7 ms [User: 4620.5 ms, System: 1157.2 ms]
Range (min … max): 373.6 ms … 386.9 ms 10 runs
$ MIRIFLAGS="-Zmiri-disable-isolation" ./x.py miri library/alloc
finished in 449.24s
```
In my opinion the results don't change iterative library development or CI execution in meaningful ways. For example currently the library doc-tests take ~66s and incremental compilation takes 10+ seconds. However I only have limited knowledge of the various local development workflows that exist, and might be missing one that is significantly impacted by this change.
The `Box<T: Default>` impl currently calls `T::default()` before allocating
the `Box`.
Most `Default` impls are trivial, which should in theory allow
LLVM to construct `T: Default` directly in the `Box` allocation when calling
`<Box<T>>::default()`.
However, the allocation may fail, which necessitates calling `T's` destructor if it has one.
If the destructor is non-trivial, then LLVM has a hard time proving that it's
sound to elide, which makes it construct `T` on the stack first, and then copy it into the allocation.
Create an uninit `Box` first, and then write `T::default` into it, so that LLVM now only needs to prove
that the `T::default` can't panic, which should be trivial for most `Default` impls.
liballoc: introduce String, Vec const-slicing
This change `const`-qualifies many methods on `Vec` and `String`, notably `as_slice`, `as_str`, `len`. These changes are made behind the unstable feature flag `const_vec_string_slice`.
## Motivation
This is to support simultaneous variance over ownership and constness. I have an enum type that may contain either `String` or `&str`, and I want to produce a `&str` from it in a possibly-`const` context.
```rust
enum StrOrString<'s> {
Str(&'s str),
String(String),
}
impl<'s> StrOrString<'s> {
const fn as_str(&self) -> &str {
match self {
// In a const-context, I really only expect to see this variant, but I can't switch the implementation
// in some mode like #[cfg(const)] -- there has to be a single body
Self::Str(s) => s,
// so this is a problem, since it's not `const`
Self::String(s) => s.as_str(),
}
}
}
```
Currently `String` and `Vec` don't support this, but can without functional changes. Similar logic applies for `len`, `capacity`, `is_empty`.
## Changes
The essential thing enabling this change is that `Unique::as_ptr` is `const`. This lets us convert `RawVec::ptr` -> `Vec::as_ptr` -> `Vec::as_slice` -> `String::as_str`.
I had to move the `Deref` implementations into `as_{str,slice}` because `Deref` isn't `#[const_trait]`, but I would expect this change to be invisible up to inlining. I moved the `DerefMut` implementations as well for uniformity.
This change `const`-qualifies many methods on Vec and String, notably
`as_slice`, `as_str`, `len`. These changes are made behind the unstable
feature flag `const_vec_string_slice` with the following tracking issue:
https://github.com/rust-lang/rust/issues/129041
Mark some more types as having insignificant dtor
These were caught by https://github.com/rust-lang/rust/pull/129864#issuecomment-2376658407, which is implementing a lint for some changes in drop order for temporaries in tail expressions.
Specifically, the destructors of `CString` and the bitpacked repr for `std::io::Error` are insignificant insofar as they don't have side-effects on things like locking or synchronization; they just free memory.
See some discussion on #89144 for what makes a drop impl "significant"
This commit is a followup to https://github.com/rust-lang/rust/pull/124032. It
replaces the tests that test the various sort functions in the standard library
with a test-suite developed as part of
https://github.com/Voultapher/sort-research-rs. The current tests suffer a
couple of problems:
- They don't cover important real world patterns that the implementations take
advantage of and execute special code for.
- The input lengths tested miss out on code paths. For example, important safety
property tests never reach the quicksort part of the implementation.
- The miri side is often limited to `len <= 20` which means it very thoroughly
tests the insertion sort, which accounts for 19 out of 1.5k LoC.
- They are split into to core and alloc, causing code duplication and uneven
coverage.
- The randomness is not repeatable, as it
relies on `std:#️⃣:RandomState::new().build_hasher()`.
Most of these issues existed before
https://github.com/rust-lang/rust/pull/124032, but they are intensified by it.
One thing that is new and requires additional testing, is that the new sort
implementations specialize based on type properties. For example `Freeze` and
non `Freeze` execute different code paths.
Effectively there are three dimensions that matter:
- Input type
- Input length
- Input pattern
The ported test-suite tests various properties along all three dimensions,
greatly improving test coverage. It side-steps the miri issue by preferring
sampled approaches. For example the test that checks if after a panic the set of
elements is still the original one, doesn't do so for every single possible
panic opportunity but rather it picks one at random, and performs this test
across a range of input length, which varies the panic point across them. This
allows regular execution to easily test inputs of length 10k, and miri execution
up to 100 which covers significantly more code. The randomness used is tied to a
fixed - but random per process execution - seed. This allows for fully
repeatable tests and fuzzer like exploration across multiple runs.
Structure wise, the tests are previously found in the core integration tests for
`sort_unstable` and alloc unit tests for `sort`. The new test-suite was
developed to be a purely black-box approach, which makes integration testing the
better place, because it can't accidentally rely on internal access. Because
unwinding support is required the tests can't be in core, even if the
implementation is, so they are now part of the alloc integration tests. Are
there architectures that can only build and test core and not alloc? If so, do
such platforms require sort testing? For what it's worth the current
implementation state passes miri `--target mips64-unknown-linux-gnuabi64` which
is big endian.
The test-suite also contains tests for properties that were and are given by the
current and previous implementations, and likely relied upon by users but
weren't tested. For example `self_cmp` tests that the two parameters `a` and `b`
passed into the comparison function are never references to the same object,
which if the user is sorting for example a `&mut [Mutex<i32>]` could lead to a
deadlock.
Instead of using the hashed caller location as rand seed, it uses seconds since
unix epoch / 10, which given timestamps in the CI should be reasonably easy to
reproduce, but also allows fuzzer like space exploration.
Improve autovectorization of to_lowercase / to_uppercase functions
Refactor the code in the `convert_while_ascii` helper function to make it more suitable for auto-vectorization and also process the full ascii prefix of the string. The generic case conversion logic will only be invoked starting from the first non-ascii character.
The runtime on a microbenchmark with a small ascii-only input decreases from ~55ns to ~18ns per iteration. The new implementation also reduces the amount of unsafe code and encapsulates all unsafe inside the helper function.
Fixes#123712
