Sync portable-simd to remove autosplats
This PR syncs portable-simd in up to a8385522ad in order to address the type inference breakages documented on nightly in https://github.com/rust-lang/rust/issues/90904 by removing the vector + scalar binary operations (called "autosplats", "broadcasting", or "rank promotion", depending on who you ask) that allow `{scalar} + &'_ {scalar}` to fail in some cases, because it becomes possible the programmer may have meant `{scalar} + &'_ {vector}`.
A few quality-of-life improvements make their way in as well:
- Lane counts can now go to 64, as LLVM seems to have fixed their miscompilation for those.
- `{i,u}8x64` to `__m512i` is now available.
- a bunch of `#[must_use]` notes appear throughout the module.
- Some implementations, mostly instances of `impl core::ops::{Op}<Simd> for Simd` that aren't `{vector} + {vector}` (e.g. `{vector} + &'_ {vector}`), leverage some generics and `where` bounds now to make them easier to understand by reducing a dozen implementations into one (and make it possible for people to open the docs on less burly devices).
- And some internal-only improvements.
None of these changes should affect a beta backport, only actual users of `core::simd` (and most aren't even visible in the programmatic sense), though I can extract an even more minimal changeset for beta if necessary. It seemed simpler to just keep moving forward.
Make `array::{try_from_fn, try_map}` and `Iterator::try_find` generic over `Try`
Fixes#85115
This only updates unstable functions.
`array::try_map` didn't actually exist before; this adds it under the still-open tracking issue #79711 from the old PR #79713.
Tracking issue for the new trait: #91285
This would also solve the return type question in for the proposed `Iterator::try_reduce` in #87054
disable tests in Miri that take too long
Comparing slices of length `usize::MAX` diverges in Miri. In fact these tests even diverge in rustc unless `-O` is passed. I tried this code to check that:
```rust
#![feature(slice_take)]
const EMPTY_MAX: &'static [()] = &[(); usize::MAX];
fn main() {
let mut slice: &[_] = &[(); usize::MAX];
println!("1");
assert_eq!(Some(&[] as _), slice.take(usize::MAX..));
println!("2");
let remaining: &[_] = EMPTY_MAX;
println!("3");
assert_eq!(remaining, slice);
println!("4");
}
```
So, disable these tests in Miri for now.
Fixes 85115
This only updates unstable functions.
`array::try_map` didn't actually exist before, despite the tracking issue 79711 still being open from the old PR 79713.
Fix Iterator::advance_by contract inconsistency
The `advance_by(n)` docs state that in the error case `Err(k)` that k is always less than n.
It also states that `advance_by(0)` may return `Err(0)` to indicate an exhausted iterator.
These statements are inconsistent.
Since only one implementation (Skip) actually made use of that I changed it to return Ok(()) in that case too.
While adding some tests I also found a bug in `Take::advance_back_by`.
Methods that were only blocked on `const_panic` have been stabilized.
The remaining methods of `duration_consts_2` are all related to floats,
and as such have been placed behind the `duration_consts_float` feature
gate.
The `advance_by(n)` docs state that in the error case `Err(k)` that k is always less than n.
It also states that `advance_by(0)` may return `Err(0)` to indicate an exhausted iterator.
These statements are inconsistent.
Since only one implementation (Skip) actually made use of that I changed it to return Ok(()) in that case too.
While adding some tests I also found a bug in `Take::advance_back_by`.
Stabilize `const_raw_ptr_deref` for `*const T`
This stabilizes dereferencing immutable raw pointers in const contexts.
It does not stabilize `*mut T` dereferencing. This is behind the
same feature gate as mutable references.
closes https://github.com/rust-lang/rust/issues/51911
pub use core::simd;
A portable abstraction over SIMD has been a major pursuit in recent years for several programming languages. In Rust, `std::arch` offers explicit SIMD acceleration via compiler intrinsics, but it does so at the cost of having to individually maintain each and every single such API, and is almost completely `unsafe` to use. `core::simd` offers safe abstractions that are resolved to the appropriate SIMD instructions by LLVM during compilation, including scalar instructions if that is all that is available.
`core::simd` is enabled by the `#![portable_simd]` nightly feature tracked in https://github.com/rust-lang/rust/issues/86656 and is introduced here by pulling in the https://github.com/rust-lang/portable-simd repository as a subtree. We built the repository out-of-tree to allow faster compilation and a stochastic test suite backed by the proptest crate to verify that different targets, features, and optimizations produce the same result, so that using this library does not introduce any surprises. As these tests are technically non-deterministic, and thus can introduce overly interesting Heisenbugs if included in the rustc CI, they are visible in the commit history of the subtree but do nothing here. Some tests **are** introduced via the documentation, but these use deterministic asserts.
There are multiple unsolved problems with the library at the current moment, including a want for better documentation, technical issues with LLVM scalarizing and lowering to libm, room for improvement for the APIs, and so far I have not added the necessary plumbing for allowing the more experimental or libm-dependent APIs to be used. However, I thought it would be prudent to open this for review in its current condition, as it is both usable and it is likely I am going to learn something else needs to be fixed when bors tries this out.
The major types are
- `core::simd::Simd<T, N>`
- `core::simd::Mask<T, N>`
There is also the `LaneCount` struct, which, together with the SimdElement and SupportedLaneCount traits, limit the implementation's maximum support to vectors we know will actually compile and provide supporting logic for bitmasks. I'm hoping to simplify at least some of these out of the way as the compiler and library evolve.
These tests just verify some basic APIs of core::simd function, and
guarantees that attempting to access the wrong things doesn't work.
The majority of tests are stochastic, and so remain upstream, but
a few deterministic tests arrive in the subtree as doc tests.
This stabilizes dereferencing immutable raw pointers in const contexts.
It does not stabilize `*mut T` dereferencing. This is placed behind the
`const_raw_mut_ptr_deref` feature gate.
Add #[must_use] to expensive computations
The unifying theme for this commit is weak, admittedly. I put together a list of "expensive" functions when I originally proposed this whole effort, but nobody's cared about that criterion. Still, it's a decent way to bite off a not-too-big chunk of work.
Given the grab bag nature of this commit, the messages I used vary quite a bit. I'm open to wording changes.
For some reason clippy flagged four `BTreeSet` methods but didn't say boo about equivalent ones on `HashSet`. I stared at them for a while but I can't figure out the difference so I added the `HashSet` ones in.
```rust
// Flagged by clippy.
alloc::collections::btree_set::BTreeSet<T> fn difference<'a>(&'a self, other: &'a BTreeSet<T>) -> Difference<'a, T>;
alloc::collections::btree_set::BTreeSet<T> fn symmetric_difference<'a>(&'a self, other: &'a BTreeSet<T>) -> SymmetricDifference<'a, T>
alloc::collections::btree_set::BTreeSet<T> fn intersection<'a>(&'a self, other: &'a BTreeSet<T>) -> Intersection<'a, T>;
alloc::collections::btree_set::BTreeSet<T> fn union<'a>(&'a self, other: &'a BTreeSet<T>) -> Union<'a, T>;
// Ignored by clippy, but not by me.
std::collections::HashSet<T, S> fn difference<'a>(&'a self, other: &'a HashSet<T, S>) -> Difference<'a, T, S>;
std::collections::HashSet<T, S> fn symmetric_difference<'a>(&'a self, other: &'a HashSet<T, S>) -> SymmetricDifference<'a, T, S>
std::collections::HashSet<T, S> fn intersection<'a>(&'a self, other: &'a HashSet<T, S>) -> Intersection<'a, T, S>;
std::collections::HashSet<T, S> fn union<'a>(&'a self, other: &'a HashSet<T, S>) -> Union<'a, T, S>;
```
Parent issue: #89692
r? ```@joshtriplett```
Implement split_array and split_array_mut
This implements `[T]::split_array::<const N>() -> (&[T; N], &[T])` and `[T; N]::split_array::<const M>() -> (&[T; M], &[T])` and their mutable equivalents. These are another few “missing” array implementations now that const generics are a thing, similar to #74373, #75026, etc. Fixes#74674.
This implements `[T; N]::split_array` returning an array and a slice. Ultimately, this is probably not what we want, we would want the second return value to be an array of length N-M, which will likely be possible with future const generics enhancements. We need to implement the array method now though, to immediately shadow the slice method. This way, when the slice methods get stabilized, calling them on an array will not be automatic through coercion, so we won't have trouble stabilizing the array methods later (cf. into_iter debacle).
An unchecked version of `[T]::split_array` could also be added as in #76014. This would not be needed for `[T; N]::split_array` as that can be compile-time checked. Edit: actually, since split_at_unchecked is internal-only it could be changed to be split_array-only.
Make more `From` impls `const` (libcore)
Adding `const` to `From` implementations in the core. `rustc_const_unstable` attribute is not added to unstable implementations.
Tracking issue: #88674
<details>
<summary>Done</summary><div>
- `T` from `T`
- `T` from `!`
- `Option<T>` from `T`
- `Option<&T>` from `&Option<T>`
- `Option<&mut T>` from `&mut Option<T>`
- `Cell<T>` from `T`
- `RefCell<T>` from `T`
- `UnsafeCell<T>` from `T`
- `OnceCell<T>` from `T`
- `Poll<T>` from `T`
- `u32` from `char`
- `u64` from `char`
- `u128` from `char`
- `char` from `u8`
- `AtomicBool` from `bool`
- `AtomicPtr<T>` from `*mut T`
- `AtomicI(bits)` from `i(bits)`
- `AtomicU(bits)` from `u(bits)`
- `i(bits)` from `NonZeroI(bits)`
- `u(bits)` from `NonZeroU(bits)`
- `NonNull<T>` from `Unique<T>`
- `NonNull<T>` from `&T`
- `NonNull<T>` from `&mut T`
- `Unique<T>` from `&mut T`
- `Infallible` from `!`
- `TryIntError` from `!`
- `TryIntError` from `Infallible`
- `TryFromSliceError` from `Infallible`
- `FromResidual for Option<T>`
</div></details>
<details>
<summary>Remaining</summary><dev>
- `NonZero` from `NonZero`
These can't be made const at this time because these use Into::into.
https://github.com/rust-lang/rust/blob/master/library/core/src/convert/num.rs#L393
- `std`, `alloc`
There may still be many implementations that can be made `const`.
</div></details>
Automatic exponential formatting in Debug
Context: See [this comment from the libs team](https://github.com/rust-lang/rfcs/pull/2729#issuecomment-853454204)
---
Makes `"{:?}"` switch to exponential for floats based on magnitude. The libs team suggested exploring this idea in the discussion thread for RFC rust-lang/rfcs#2729. (**note:** this is **not** an implementation of the RFC; it is an implementation of one of the alternatives)
Thresholds chosen were 1e-4 and 1e16. Justification described [here](https://github.com/rust-lang/rfcs/pull/2729#issuecomment-864482954).
**This will require a crater run.**
---
As mentioned in the commit message of 8731d4dfb4, this behavior will not apply when a precision is supplied, because I wanted to preserve the following existing and useful behavior of `{:.PREC?}` (which recursively applies `{:.PREC}` to floats in a struct):
```rust
assert_eq!(
format!("{:.2?}", [100.0, 0.000004]),
"[100.00, 0.00]",
)
```
I looked around and am not sure where there are any tests that actually use this in the test suite, though?
All things considered, I'm surprised that this change did not seem to break even a single existing test in `x.py test --stage 2`. (even when I tried a smaller threshold of 1e6)
The unifying theme for this commit is weak, admittedly. I put together a
list of "expensive" functions when I originally proposed this whole
effort, but nobody's cared about that criterion. Still, it's a decent
way to bite off a not-too-big chunk of work.
Given the grab bag nature of this commit, the messages I used vary quite
a bit.
Speedup int log10 branchless
This is achieved with a branchless bit-twiddling implementation of the case x < 100_000, and using this as building block.
Benchmark on an Intel i7-8700K (Coffee Lake):
```
name old ns/iter new ns/iter diff ns/iter diff % speedup
num::int_log::u8_log10_predictable 165 169 4 2.42% x 0.98
num::int_log::u8_log10_random 438 423 -15 -3.42% x 1.04
num::int_log::u8_log10_random_small 438 423 -15 -3.42% x 1.04
num::int_log::u16_log10_predictable 633 417 -216 -34.12% x 1.52
num::int_log::u16_log10_random 908 471 -437 -48.13% x 1.93
num::int_log::u16_log10_random_small 945 471 -474 -50.16% x 2.01
num::int_log::u32_log10_predictable 1,496 1,340 -156 -10.43% x 1.12
num::int_log::u32_log10_random 1,076 873 -203 -18.87% x 1.23
num::int_log::u32_log10_random_small 1,145 874 -271 -23.67% x 1.31
num::int_log::u64_log10_predictable 4,005 3,171 -834 -20.82% x 1.26
num::int_log::u64_log10_random 1,247 1,021 -226 -18.12% x 1.22
num::int_log::u64_log10_random_small 1,265 921 -344 -27.19% x 1.37
num::int_log::u128_log10_predictable 39,667 39,579 -88 -0.22% x 1.00
num::int_log::u128_log10_random 6,456 6,696 240 3.72% x 0.96
num::int_log::u128_log10_random_small 4,108 3,903 -205 -4.99% x 1.05
```
Benchmark on an M1 Mac Mini:
```
name old ns/iter new ns/iter diff ns/iter diff % speedup
num::int_log::u8_log10_predictable 143 130 -13 -9.09% x 1.10
num::int_log::u8_log10_random 375 325 -50 -13.33% x 1.15
num::int_log::u8_log10_random_small 376 325 -51 -13.56% x 1.16
num::int_log::u16_log10_predictable 500 322 -178 -35.60% x 1.55
num::int_log::u16_log10_random 794 405 -389 -48.99% x 1.96
num::int_log::u16_log10_random_small 1,035 405 -630 -60.87% x 2.56
num::int_log::u32_log10_predictable 1,144 894 -250 -21.85% x 1.28
num::int_log::u32_log10_random 832 786 -46 -5.53% x 1.06
num::int_log::u32_log10_random_small 832 787 -45 -5.41% x 1.06
num::int_log::u64_log10_predictable 2,681 2,057 -624 -23.27% x 1.30
num::int_log::u64_log10_random 1,015 806 -209 -20.59% x 1.26
num::int_log::u64_log10_random_small 1,004 795 -209 -20.82% x 1.26
num::int_log::u128_log10_predictable 56,825 56,526 -299 -0.53% x 1.01
num::int_log::u128_log10_random 9,056 8,861 -195 -2.15% x 1.02
num::int_log::u128_log10_random_small 1,528 1,527 -1 -0.07% x 1.00
```
The 128 bit case remains ridiculously slow because llvm fails to optimize division by a constant 128-bit value to multiplications. This could be worked around but it seems preferable to fix this in llvm.
From u32 up, table lookup (like suggested [here](https://github.com/rust-lang/rust/issues/70887#issuecomment-881099813)) is still faster, but requires a hardware `leading_zeros` to be viable, and might clog up the cache.
