diff options
author | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-04-19 00:47:55 +0000 |
---|---|---|
committer | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-04-19 00:47:55 +0000 |
commit | 26a029d407be480d791972afb5975cf62c9360a6 (patch) | |
tree | f435a8308119effd964b339f76abb83a57c29483 /third_party/rust/rayon/src/slice | |
parent | Initial commit. (diff) | |
download | firefox-26a029d407be480d791972afb5975cf62c9360a6.tar.xz firefox-26a029d407be480d791972afb5975cf62c9360a6.zip |
Adding upstream version 124.0.1.upstream/124.0.1
Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>
Diffstat (limited to 'third_party/rust/rayon/src/slice')
-rw-r--r-- | third_party/rust/rayon/src/slice/chunks.rs | 389 | ||||
-rw-r--r-- | third_party/rust/rayon/src/slice/mergesort.rs | 755 | ||||
-rw-r--r-- | third_party/rust/rayon/src/slice/mod.rs | 1041 | ||||
-rw-r--r-- | third_party/rust/rayon/src/slice/quicksort.rs | 897 | ||||
-rw-r--r-- | third_party/rust/rayon/src/slice/rchunks.rs | 386 | ||||
-rw-r--r-- | third_party/rust/rayon/src/slice/test.rs | 170 |
6 files changed, 3638 insertions, 0 deletions
diff --git a/third_party/rust/rayon/src/slice/chunks.rs b/third_party/rust/rayon/src/slice/chunks.rs new file mode 100644 index 0000000000..a3cc709536 --- /dev/null +++ b/third_party/rust/rayon/src/slice/chunks.rs @@ -0,0 +1,389 @@ +use crate::iter::plumbing::*; +use crate::iter::*; +use crate::math::div_round_up; +use std::cmp; + +/// Parallel iterator over immutable non-overlapping chunks of a slice +#[derive(Debug)] +pub struct Chunks<'data, T: Sync> { + chunk_size: usize, + slice: &'data [T], +} + +impl<'data, T: Sync> Chunks<'data, T> { + pub(super) fn new(chunk_size: usize, slice: &'data [T]) -> Self { + Self { chunk_size, slice } + } +} + +impl<'data, T: Sync> Clone for Chunks<'data, T> { + fn clone(&self) -> Self { + Chunks { ..*self } + } +} + +impl<'data, T: Sync + 'data> ParallelIterator for Chunks<'data, T> { + type Item = &'data [T]; + + fn drive_unindexed<C>(self, consumer: C) -> C::Result + where + C: UnindexedConsumer<Self::Item>, + { + bridge(self, consumer) + } + + fn opt_len(&self) -> Option<usize> { + Some(self.len()) + } +} + +impl<'data, T: Sync + 'data> IndexedParallelIterator for Chunks<'data, T> { + fn drive<C>(self, consumer: C) -> C::Result + where + C: Consumer<Self::Item>, + { + bridge(self, consumer) + } + + fn len(&self) -> usize { + div_round_up(self.slice.len(), self.chunk_size) + } + + fn with_producer<CB>(self, callback: CB) -> CB::Output + where + CB: ProducerCallback<Self::Item>, + { + callback.callback(ChunksProducer { + chunk_size: self.chunk_size, + slice: self.slice, + }) + } +} + +struct ChunksProducer<'data, T: Sync> { + chunk_size: usize, + slice: &'data [T], +} + +impl<'data, T: 'data + Sync> Producer for ChunksProducer<'data, T> { + type Item = &'data [T]; + type IntoIter = ::std::slice::Chunks<'data, T>; + + fn into_iter(self) -> Self::IntoIter { + self.slice.chunks(self.chunk_size) + } + + fn split_at(self, index: usize) -> (Self, Self) { + let elem_index = cmp::min(index * self.chunk_size, self.slice.len()); + let (left, right) = self.slice.split_at(elem_index); + ( + ChunksProducer { + chunk_size: self.chunk_size, + slice: left, + }, + ChunksProducer { + chunk_size: self.chunk_size, + slice: right, + }, + ) + } +} + +/// Parallel iterator over immutable non-overlapping chunks of a slice +#[derive(Debug)] +pub struct ChunksExact<'data, T: Sync> { + chunk_size: usize, + slice: &'data [T], + rem: &'data [T], +} + +impl<'data, T: Sync> ChunksExact<'data, T> { + pub(super) fn new(chunk_size: usize, slice: &'data [T]) -> Self { + let rem_len = slice.len() % chunk_size; + let len = slice.len() - rem_len; + let (slice, rem) = slice.split_at(len); + Self { + chunk_size, + slice, + rem, + } + } + + /// Return the remainder of the original slice that is not going to be + /// returned by the iterator. The returned slice has at most `chunk_size-1` + /// elements. + pub fn remainder(&self) -> &'data [T] { + self.rem + } +} + +impl<'data, T: Sync> Clone for ChunksExact<'data, T> { + fn clone(&self) -> Self { + ChunksExact { ..*self } + } +} + +impl<'data, T: Sync + 'data> ParallelIterator for ChunksExact<'data, T> { + type Item = &'data [T]; + + fn drive_unindexed<C>(self, consumer: C) -> C::Result + where + C: UnindexedConsumer<Self::Item>, + { + bridge(self, consumer) + } + + fn opt_len(&self) -> Option<usize> { + Some(self.len()) + } +} + +impl<'data, T: Sync + 'data> IndexedParallelIterator for ChunksExact<'data, T> { + fn drive<C>(self, consumer: C) -> C::Result + where + C: Consumer<Self::Item>, + { + bridge(self, consumer) + } + + fn len(&self) -> usize { + self.slice.len() / self.chunk_size + } + + fn with_producer<CB>(self, callback: CB) -> CB::Output + where + CB: ProducerCallback<Self::Item>, + { + callback.callback(ChunksExactProducer { + chunk_size: self.chunk_size, + slice: self.slice, + }) + } +} + +struct ChunksExactProducer<'data, T: Sync> { + chunk_size: usize, + slice: &'data [T], +} + +impl<'data, T: 'data + Sync> Producer for ChunksExactProducer<'data, T> { + type Item = &'data [T]; + type IntoIter = ::std::slice::ChunksExact<'data, T>; + + fn into_iter(self) -> Self::IntoIter { + self.slice.chunks_exact(self.chunk_size) + } + + fn split_at(self, index: usize) -> (Self, Self) { + let elem_index = index * self.chunk_size; + let (left, right) = self.slice.split_at(elem_index); + ( + ChunksExactProducer { + chunk_size: self.chunk_size, + slice: left, + }, + ChunksExactProducer { + chunk_size: self.chunk_size, + slice: right, + }, + ) + } +} + +/// Parallel iterator over mutable non-overlapping chunks of a slice +#[derive(Debug)] +pub struct ChunksMut<'data, T: Send> { + chunk_size: usize, + slice: &'data mut [T], +} + +impl<'data, T: Send> ChunksMut<'data, T> { + pub(super) fn new(chunk_size: usize, slice: &'data mut [T]) -> Self { + Self { chunk_size, slice } + } +} + +impl<'data, T: Send + 'data> ParallelIterator for ChunksMut<'data, T> { + type Item = &'data mut [T]; + + fn drive_unindexed<C>(self, consumer: C) -> C::Result + where + C: UnindexedConsumer<Self::Item>, + { + bridge(self, consumer) + } + + fn opt_len(&self) -> Option<usize> { + Some(self.len()) + } +} + +impl<'data, T: Send + 'data> IndexedParallelIterator for ChunksMut<'data, T> { + fn drive<C>(self, consumer: C) -> C::Result + where + C: Consumer<Self::Item>, + { + bridge(self, consumer) + } + + fn len(&self) -> usize { + div_round_up(self.slice.len(), self.chunk_size) + } + + fn with_producer<CB>(self, callback: CB) -> CB::Output + where + CB: ProducerCallback<Self::Item>, + { + callback.callback(ChunksMutProducer { + chunk_size: self.chunk_size, + slice: self.slice, + }) + } +} + +struct ChunksMutProducer<'data, T: Send> { + chunk_size: usize, + slice: &'data mut [T], +} + +impl<'data, T: 'data + Send> Producer for ChunksMutProducer<'data, T> { + type Item = &'data mut [T]; + type IntoIter = ::std::slice::ChunksMut<'data, T>; + + fn into_iter(self) -> Self::IntoIter { + self.slice.chunks_mut(self.chunk_size) + } + + fn split_at(self, index: usize) -> (Self, Self) { + let elem_index = cmp::min(index * self.chunk_size, self.slice.len()); + let (left, right) = self.slice.split_at_mut(elem_index); + ( + ChunksMutProducer { + chunk_size: self.chunk_size, + slice: left, + }, + ChunksMutProducer { + chunk_size: self.chunk_size, + slice: right, + }, + ) + } +} + +/// Parallel iterator over mutable non-overlapping chunks of a slice +#[derive(Debug)] +pub struct ChunksExactMut<'data, T: Send> { + chunk_size: usize, + slice: &'data mut [T], + rem: &'data mut [T], +} + +impl<'data, T: Send> ChunksExactMut<'data, T> { + pub(super) fn new(chunk_size: usize, slice: &'data mut [T]) -> Self { + let rem_len = slice.len() % chunk_size; + let len = slice.len() - rem_len; + let (slice, rem) = slice.split_at_mut(len); + Self { + chunk_size, + slice, + rem, + } + } + + /// Return the remainder of the original slice that is not going to be + /// returned by the iterator. The returned slice has at most `chunk_size-1` + /// elements. + /// + /// Note that this has to consume `self` to return the original lifetime of + /// the data, which prevents this from actually being used as a parallel + /// iterator since that also consumes. This method is provided for parity + /// with `std::iter::ChunksExactMut`, but consider calling `remainder()` or + /// `take_remainder()` as alternatives. + pub fn into_remainder(self) -> &'data mut [T] { + self.rem + } + + /// Return the remainder of the original slice that is not going to be + /// returned by the iterator. The returned slice has at most `chunk_size-1` + /// elements. + /// + /// Consider `take_remainder()` if you need access to the data with its + /// original lifetime, rather than borrowing through `&mut self` here. + pub fn remainder(&mut self) -> &mut [T] { + self.rem + } + + /// Return the remainder of the original slice that is not going to be + /// returned by the iterator. The returned slice has at most `chunk_size-1` + /// elements. Subsequent calls will return an empty slice. + pub fn take_remainder(&mut self) -> &'data mut [T] { + std::mem::take(&mut self.rem) + } +} + +impl<'data, T: Send + 'data> ParallelIterator for ChunksExactMut<'data, T> { + type Item = &'data mut [T]; + + fn drive_unindexed<C>(self, consumer: C) -> C::Result + where + C: UnindexedConsumer<Self::Item>, + { + bridge(self, consumer) + } + + fn opt_len(&self) -> Option<usize> { + Some(self.len()) + } +} + +impl<'data, T: Send + 'data> IndexedParallelIterator for ChunksExactMut<'data, T> { + fn drive<C>(self, consumer: C) -> C::Result + where + C: Consumer<Self::Item>, + { + bridge(self, consumer) + } + + fn len(&self) -> usize { + self.slice.len() / self.chunk_size + } + + fn with_producer<CB>(self, callback: CB) -> CB::Output + where + CB: ProducerCallback<Self::Item>, + { + callback.callback(ChunksExactMutProducer { + chunk_size: self.chunk_size, + slice: self.slice, + }) + } +} + +struct ChunksExactMutProducer<'data, T: Send> { + chunk_size: usize, + slice: &'data mut [T], +} + +impl<'data, T: 'data + Send> Producer for ChunksExactMutProducer<'data, T> { + type Item = &'data mut [T]; + type IntoIter = ::std::slice::ChunksExactMut<'data, T>; + + fn into_iter(self) -> Self::IntoIter { + self.slice.chunks_exact_mut(self.chunk_size) + } + + fn split_at(self, index: usize) -> (Self, Self) { + let elem_index = index * self.chunk_size; + let (left, right) = self.slice.split_at_mut(elem_index); + ( + ChunksExactMutProducer { + chunk_size: self.chunk_size, + slice: left, + }, + ChunksExactMutProducer { + chunk_size: self.chunk_size, + slice: right, + }, + ) + } +} diff --git a/third_party/rust/rayon/src/slice/mergesort.rs b/third_party/rust/rayon/src/slice/mergesort.rs new file mode 100644 index 0000000000..fec309d273 --- /dev/null +++ b/third_party/rust/rayon/src/slice/mergesort.rs @@ -0,0 +1,755 @@ +//! Parallel merge sort. +//! +//! This implementation is copied verbatim from `std::slice::sort` and then parallelized. +//! The only difference from the original is that the sequential `mergesort` returns +//! `MergesortResult` and leaves descending arrays intact. + +use crate::iter::*; +use crate::slice::ParallelSliceMut; +use crate::SendPtr; +use std::mem; +use std::mem::size_of; +use std::ptr; +use std::slice; + +unsafe fn get_and_increment<T>(ptr: &mut *mut T) -> *mut T { + let old = *ptr; + *ptr = ptr.offset(1); + old +} + +unsafe fn decrement_and_get<T>(ptr: &mut *mut T) -> *mut T { + *ptr = ptr.offset(-1); + *ptr +} + +/// When dropped, copies from `src` into `dest` a sequence of length `len`. +struct CopyOnDrop<T> { + src: *const T, + dest: *mut T, + len: usize, +} + +impl<T> Drop for CopyOnDrop<T> { + fn drop(&mut self) { + unsafe { + ptr::copy_nonoverlapping(self.src, self.dest, self.len); + } + } +} + +/// Inserts `v[0]` into pre-sorted sequence `v[1..]` so that whole `v[..]` becomes sorted. +/// +/// This is the integral subroutine of insertion sort. +fn insert_head<T, F>(v: &mut [T], is_less: &F) +where + F: Fn(&T, &T) -> bool, +{ + if v.len() >= 2 && is_less(&v[1], &v[0]) { + unsafe { + // There are three ways to implement insertion here: + // + // 1. Swap adjacent elements until the first one gets to its final destination. + // However, this way we copy data around more than is necessary. If elements are big + // structures (costly to copy), this method will be slow. + // + // 2. Iterate until the right place for the first element is found. Then shift the + // elements succeeding it to make room for it and finally place it into the + // remaining hole. This is a good method. + // + // 3. Copy the first element into a temporary variable. Iterate until the right place + // for it is found. As we go along, copy every traversed element into the slot + // preceding it. Finally, copy data from the temporary variable into the remaining + // hole. This method is very good. Benchmarks demonstrated slightly better + // performance than with the 2nd method. + // + // All methods were benchmarked, and the 3rd showed best results. So we chose that one. + let tmp = mem::ManuallyDrop::new(ptr::read(&v[0])); + + // Intermediate state of the insertion process is always tracked by `hole`, which + // serves two purposes: + // 1. Protects integrity of `v` from panics in `is_less`. + // 2. Fills the remaining hole in `v` in the end. + // + // Panic safety: + // + // If `is_less` panics at any point during the process, `hole` will get dropped and + // fill the hole in `v` with `tmp`, thus ensuring that `v` still holds every object it + // initially held exactly once. + let mut hole = InsertionHole { + src: &*tmp, + dest: &mut v[1], + }; + ptr::copy_nonoverlapping(&v[1], &mut v[0], 1); + + for i in 2..v.len() { + if !is_less(&v[i], &*tmp) { + break; + } + ptr::copy_nonoverlapping(&v[i], &mut v[i - 1], 1); + hole.dest = &mut v[i]; + } + // `hole` gets dropped and thus copies `tmp` into the remaining hole in `v`. + } + } + + // When dropped, copies from `src` into `dest`. + struct InsertionHole<T> { + src: *const T, + dest: *mut T, + } + + impl<T> Drop for InsertionHole<T> { + fn drop(&mut self) { + unsafe { + ptr::copy_nonoverlapping(self.src, self.dest, 1); + } + } + } +} + +/// Merges non-decreasing runs `v[..mid]` and `v[mid..]` using `buf` as temporary storage, and +/// stores the result into `v[..]`. +/// +/// # Safety +/// +/// The two slices must be non-empty and `mid` must be in bounds. Buffer `buf` must be long enough +/// to hold a copy of the shorter slice. Also, `T` must not be a zero-sized type. +unsafe fn merge<T, F>(v: &mut [T], mid: usize, buf: *mut T, is_less: &F) +where + F: Fn(&T, &T) -> bool, +{ + let len = v.len(); + let v = v.as_mut_ptr(); + let v_mid = v.add(mid); + let v_end = v.add(len); + + // The merge process first copies the shorter run into `buf`. Then it traces the newly copied + // run and the longer run forwards (or backwards), comparing their next unconsumed elements and + // copying the lesser (or greater) one into `v`. + // + // As soon as the shorter run is fully consumed, the process is done. If the longer run gets + // consumed first, then we must copy whatever is left of the shorter run into the remaining + // hole in `v`. + // + // Intermediate state of the process is always tracked by `hole`, which serves two purposes: + // 1. Protects integrity of `v` from panics in `is_less`. + // 2. Fills the remaining hole in `v` if the longer run gets consumed first. + // + // Panic safety: + // + // If `is_less` panics at any point during the process, `hole` will get dropped and fill the + // hole in `v` with the unconsumed range in `buf`, thus ensuring that `v` still holds every + // object it initially held exactly once. + let mut hole; + + if mid <= len - mid { + // The left run is shorter. + ptr::copy_nonoverlapping(v, buf, mid); + hole = MergeHole { + start: buf, + end: buf.add(mid), + dest: v, + }; + + // Initially, these pointers point to the beginnings of their arrays. + let left = &mut hole.start; + let mut right = v_mid; + let out = &mut hole.dest; + + while *left < hole.end && right < v_end { + // Consume the lesser side. + // If equal, prefer the left run to maintain stability. + let to_copy = if is_less(&*right, &**left) { + get_and_increment(&mut right) + } else { + get_and_increment(left) + }; + ptr::copy_nonoverlapping(to_copy, get_and_increment(out), 1); + } + } else { + // The right run is shorter. + ptr::copy_nonoverlapping(v_mid, buf, len - mid); + hole = MergeHole { + start: buf, + end: buf.add(len - mid), + dest: v_mid, + }; + + // Initially, these pointers point past the ends of their arrays. + let left = &mut hole.dest; + let right = &mut hole.end; + let mut out = v_end; + + while v < *left && buf < *right { + // Consume the greater side. + // If equal, prefer the right run to maintain stability. + let to_copy = if is_less(&*right.offset(-1), &*left.offset(-1)) { + decrement_and_get(left) + } else { + decrement_and_get(right) + }; + ptr::copy_nonoverlapping(to_copy, decrement_and_get(&mut out), 1); + } + } + // Finally, `hole` gets dropped. If the shorter run was not fully consumed, whatever remains of + // it will now be copied into the hole in `v`. + + // When dropped, copies the range `start..end` into `dest..`. + struct MergeHole<T> { + start: *mut T, + end: *mut T, + dest: *mut T, + } + + impl<T> Drop for MergeHole<T> { + fn drop(&mut self) { + // `T` is not a zero-sized type, so it's okay to divide by its size. + unsafe { + let len = self.end.offset_from(self.start) as usize; + ptr::copy_nonoverlapping(self.start, self.dest, len); + } + } + } +} + +/// The result of merge sort. +#[must_use] +#[derive(Clone, Copy, PartialEq, Eq)] +enum MergesortResult { + /// The slice has already been sorted. + NonDescending, + /// The slice has been descending and therefore it was left intact. + Descending, + /// The slice was sorted. + Sorted, +} + +/// A sorted run that starts at index `start` and is of length `len`. +#[derive(Clone, Copy)] +struct Run { + start: usize, + len: usize, +} + +/// Examines the stack of runs and identifies the next pair of runs to merge. More specifically, +/// if `Some(r)` is returned, that means `runs[r]` and `runs[r + 1]` must be merged next. If the +/// algorithm should continue building a new run instead, `None` is returned. +/// +/// TimSort is infamous for its buggy implementations, as described here: +/// http://envisage-project.eu/timsort-specification-and-verification/ +/// +/// The gist of the story is: we must enforce the invariants on the top four runs on the stack. +/// Enforcing them on just top three is not sufficient to ensure that the invariants will still +/// hold for *all* runs in the stack. +/// +/// This function correctly checks invariants for the top four runs. Additionally, if the top +/// run starts at index 0, it will always demand a merge operation until the stack is fully +/// collapsed, in order to complete the sort. +#[inline] +fn collapse(runs: &[Run]) -> Option<usize> { + let n = runs.len(); + + if n >= 2 + && (runs[n - 1].start == 0 + || runs[n - 2].len <= runs[n - 1].len + || (n >= 3 && runs[n - 3].len <= runs[n - 2].len + runs[n - 1].len) + || (n >= 4 && runs[n - 4].len <= runs[n - 3].len + runs[n - 2].len)) + { + if n >= 3 && runs[n - 3].len < runs[n - 1].len { + Some(n - 3) + } else { + Some(n - 2) + } + } else { + None + } +} + +/// Sorts a slice using merge sort, unless it is already in descending order. +/// +/// This function doesn't modify the slice if it is already non-descending or descending. +/// Otherwise, it sorts the slice into non-descending order. +/// +/// This merge sort borrows some (but not all) ideas from TimSort, which is described in detail +/// [here](https://github.com/python/cpython/blob/main/Objects/listsort.txt). +/// +/// The algorithm identifies strictly descending and non-descending subsequences, which are called +/// natural runs. There is a stack of pending runs yet to be merged. Each newly found run is pushed +/// onto the stack, and then some pairs of adjacent runs are merged until these two invariants are +/// satisfied: +/// +/// 1. for every `i` in `1..runs.len()`: `runs[i - 1].len > runs[i].len` +/// 2. for every `i` in `2..runs.len()`: `runs[i - 2].len > runs[i - 1].len + runs[i].len` +/// +/// The invariants ensure that the total running time is *O*(*n* \* log(*n*)) worst-case. +/// +/// # Safety +/// +/// The argument `buf` is used as a temporary buffer and must be at least as long as `v`. +unsafe fn mergesort<T, F>(v: &mut [T], buf: *mut T, is_less: &F) -> MergesortResult +where + T: Send, + F: Fn(&T, &T) -> bool + Sync, +{ + // Very short runs are extended using insertion sort to span at least this many elements. + const MIN_RUN: usize = 10; + + let len = v.len(); + + // In order to identify natural runs in `v`, we traverse it backwards. That might seem like a + // strange decision, but consider the fact that merges more often go in the opposite direction + // (forwards). According to benchmarks, merging forwards is slightly faster than merging + // backwards. To conclude, identifying runs by traversing backwards improves performance. + let mut runs = vec![]; + let mut end = len; + while end > 0 { + // Find the next natural run, and reverse it if it's strictly descending. + let mut start = end - 1; + + if start > 0 { + start -= 1; + + if is_less(v.get_unchecked(start + 1), v.get_unchecked(start)) { + while start > 0 && is_less(v.get_unchecked(start), v.get_unchecked(start - 1)) { + start -= 1; + } + + // If this descending run covers the whole slice, return immediately. + if start == 0 && end == len { + return MergesortResult::Descending; + } else { + v[start..end].reverse(); + } + } else { + while start > 0 && !is_less(v.get_unchecked(start), v.get_unchecked(start - 1)) { + start -= 1; + } + + // If this non-descending run covers the whole slice, return immediately. + if end - start == len { + return MergesortResult::NonDescending; + } + } + } + + // Insert some more elements into the run if it's too short. Insertion sort is faster than + // merge sort on short sequences, so this significantly improves performance. + while start > 0 && end - start < MIN_RUN { + start -= 1; + insert_head(&mut v[start..end], &is_less); + } + + // Push this run onto the stack. + runs.push(Run { + start, + len: end - start, + }); + end = start; + + // Merge some pairs of adjacent runs to satisfy the invariants. + while let Some(r) = collapse(&runs) { + let left = runs[r + 1]; + let right = runs[r]; + merge( + &mut v[left.start..right.start + right.len], + left.len, + buf, + &is_less, + ); + + runs[r] = Run { + start: left.start, + len: left.len + right.len, + }; + runs.remove(r + 1); + } + } + + // Finally, exactly one run must remain in the stack. + debug_assert!(runs.len() == 1 && runs[0].start == 0 && runs[0].len == len); + + // The original order of the slice was neither non-descending nor descending. + MergesortResult::Sorted +} + +//////////////////////////////////////////////////////////////////////////// +// Everything above this line is copied from `std::slice::sort` (with very minor tweaks). +// Everything below this line is parallelization. +//////////////////////////////////////////////////////////////////////////// + +/// Splits two sorted slices so that they can be merged in parallel. +/// +/// Returns two indices `(a, b)` so that slices `left[..a]` and `right[..b]` come before +/// `left[a..]` and `right[b..]`. +fn split_for_merge<T, F>(left: &[T], right: &[T], is_less: &F) -> (usize, usize) +where + F: Fn(&T, &T) -> bool, +{ + let left_len = left.len(); + let right_len = right.len(); + + if left_len >= right_len { + let left_mid = left_len / 2; + + // Find the first element in `right` that is greater than or equal to `left[left_mid]`. + let mut a = 0; + let mut b = right_len; + while a < b { + let m = a + (b - a) / 2; + if is_less(&right[m], &left[left_mid]) { + a = m + 1; + } else { + b = m; + } + } + + (left_mid, a) + } else { + let right_mid = right_len / 2; + + // Find the first element in `left` that is greater than `right[right_mid]`. + let mut a = 0; + let mut b = left_len; + while a < b { + let m = a + (b - a) / 2; + if is_less(&right[right_mid], &left[m]) { + b = m; + } else { + a = m + 1; + } + } + + (a, right_mid) + } +} + +/// Merges slices `left` and `right` in parallel and stores the result into `dest`. +/// +/// # Safety +/// +/// The `dest` pointer must have enough space to store the result. +/// +/// Even if `is_less` panics at any point during the merge process, this function will fully copy +/// all elements from `left` and `right` into `dest` (not necessarily in sorted order). +unsafe fn par_merge<T, F>(left: &mut [T], right: &mut [T], dest: *mut T, is_less: &F) +where + T: Send, + F: Fn(&T, &T) -> bool + Sync, +{ + // Slices whose lengths sum up to this value are merged sequentially. This number is slightly + // larger than `CHUNK_LENGTH`, and the reason is that merging is faster than merge sorting, so + // merging needs a bit coarser granularity in order to hide the overhead of Rayon's task + // scheduling. + const MAX_SEQUENTIAL: usize = 5000; + + let left_len = left.len(); + let right_len = right.len(); + + // Intermediate state of the merge process, which serves two purposes: + // 1. Protects integrity of `dest` from panics in `is_less`. + // 2. Copies the remaining elements as soon as one of the two sides is exhausted. + // + // Panic safety: + // + // If `is_less` panics at any point during the merge process, `s` will get dropped and copy the + // remaining parts of `left` and `right` into `dest`. + let mut s = State { + left_start: left.as_mut_ptr(), + left_end: left.as_mut_ptr().add(left_len), + right_start: right.as_mut_ptr(), + right_end: right.as_mut_ptr().add(right_len), + dest, + }; + + if left_len == 0 || right_len == 0 || left_len + right_len < MAX_SEQUENTIAL { + while s.left_start < s.left_end && s.right_start < s.right_end { + // Consume the lesser side. + // If equal, prefer the left run to maintain stability. + let to_copy = if is_less(&*s.right_start, &*s.left_start) { + get_and_increment(&mut s.right_start) + } else { + get_and_increment(&mut s.left_start) + }; + ptr::copy_nonoverlapping(to_copy, get_and_increment(&mut s.dest), 1); + } + } else { + // Function `split_for_merge` might panic. If that happens, `s` will get destructed and copy + // the whole `left` and `right` into `dest`. + let (left_mid, right_mid) = split_for_merge(left, right, is_less); + let (left_l, left_r) = left.split_at_mut(left_mid); + let (right_l, right_r) = right.split_at_mut(right_mid); + + // Prevent the destructor of `s` from running. Rayon will ensure that both calls to + // `par_merge` happen. If one of the two calls panics, they will ensure that elements still + // get copied into `dest_left` and `dest_right``. + mem::forget(s); + + // Wrap pointers in SendPtr so that they can be sent to another thread + // See the documentation of SendPtr for a full explanation + let dest_l = SendPtr(dest); + let dest_r = SendPtr(dest.add(left_l.len() + right_l.len())); + rayon_core::join( + move || par_merge(left_l, right_l, dest_l.get(), is_less), + move || par_merge(left_r, right_r, dest_r.get(), is_less), + ); + } + // Finally, `s` gets dropped if we used sequential merge, thus copying the remaining elements + // all at once. + + // When dropped, copies arrays `left_start..left_end` and `right_start..right_end` into `dest`, + // in that order. + struct State<T> { + left_start: *mut T, + left_end: *mut T, + right_start: *mut T, + right_end: *mut T, + dest: *mut T, + } + + impl<T> Drop for State<T> { + fn drop(&mut self) { + let size = size_of::<T>(); + let left_len = (self.left_end as usize - self.left_start as usize) / size; + let right_len = (self.right_end as usize - self.right_start as usize) / size; + + // Copy array `left`, followed by `right`. + unsafe { + ptr::copy_nonoverlapping(self.left_start, self.dest, left_len); + self.dest = self.dest.add(left_len); + ptr::copy_nonoverlapping(self.right_start, self.dest, right_len); + } + } + } +} + +/// Recursively merges pre-sorted chunks inside `v`. +/// +/// Chunks of `v` are stored in `chunks` as intervals (inclusive left and exclusive right bound). +/// Argument `buf` is an auxiliary buffer that will be used during the procedure. +/// If `into_buf` is true, the result will be stored into `buf`, otherwise it will be in `v`. +/// +/// # Safety +/// +/// The number of chunks must be positive and they must be adjacent: the right bound of each chunk +/// must equal the left bound of the following chunk. +/// +/// The buffer must be at least as long as `v`. +unsafe fn recurse<T, F>( + v: *mut T, + buf: *mut T, + chunks: &[(usize, usize)], + into_buf: bool, + is_less: &F, +) where + T: Send, + F: Fn(&T, &T) -> bool + Sync, +{ + let len = chunks.len(); + debug_assert!(len > 0); + + // Base case of the algorithm. + // If only one chunk is remaining, there's no more work to split and merge. + if len == 1 { + if into_buf { + // Copy the chunk from `v` into `buf`. + let (start, end) = chunks[0]; + let src = v.add(start); + let dest = buf.add(start); + ptr::copy_nonoverlapping(src, dest, end - start); + } + return; + } + + // Split the chunks into two halves. + let (start, _) = chunks[0]; + let (mid, _) = chunks[len / 2]; + let (_, end) = chunks[len - 1]; + let (left, right) = chunks.split_at(len / 2); + + // After recursive calls finish we'll have to merge chunks `(start, mid)` and `(mid, end)` from + // `src` into `dest`. If the current invocation has to store the result into `buf`, we'll + // merge chunks from `v` into `buf`, and vice versa. + // + // Recursive calls flip `into_buf` at each level of recursion. More concretely, `par_merge` + // merges chunks from `buf` into `v` at the first level, from `v` into `buf` at the second + // level etc. + let (src, dest) = if into_buf { (v, buf) } else { (buf, v) }; + + // Panic safety: + // + // If `is_less` panics at any point during the recursive calls, the destructor of `guard` will + // be executed, thus copying everything from `src` into `dest`. This way we ensure that all + // chunks are in fact copied into `dest`, even if the merge process doesn't finish. + let guard = CopyOnDrop { + src: src.add(start), + dest: dest.add(start), + len: end - start, + }; + + // Wrap pointers in SendPtr so that they can be sent to another thread + // See the documentation of SendPtr for a full explanation + let v = SendPtr(v); + let buf = SendPtr(buf); + rayon_core::join( + move || recurse(v.get(), buf.get(), left, !into_buf, is_less), + move || recurse(v.get(), buf.get(), right, !into_buf, is_less), + ); + + // Everything went all right - recursive calls didn't panic. + // Forget the guard in order to prevent its destructor from running. + mem::forget(guard); + + // Merge chunks `(start, mid)` and `(mid, end)` from `src` into `dest`. + let src_left = slice::from_raw_parts_mut(src.add(start), mid - start); + let src_right = slice::from_raw_parts_mut(src.add(mid), end - mid); + par_merge(src_left, src_right, dest.add(start), is_less); +} + +/// Sorts `v` using merge sort in parallel. +/// +/// The algorithm is stable, allocates memory, and `O(n log n)` worst-case. +/// The allocated temporary buffer is of the same length as is `v`. +pub(super) fn par_mergesort<T, F>(v: &mut [T], is_less: F) +where + T: Send, + F: Fn(&T, &T) -> bool + Sync, +{ + // Slices of up to this length get sorted using insertion sort in order to avoid the cost of + // buffer allocation. + const MAX_INSERTION: usize = 20; + // The length of initial chunks. This number is as small as possible but so that the overhead + // of Rayon's task scheduling is still negligible. + const CHUNK_LENGTH: usize = 2000; + + // Sorting has no meaningful behavior on zero-sized types. + if size_of::<T>() == 0 { + return; + } + + let len = v.len(); + + // Short slices get sorted in-place via insertion sort to avoid allocations. + if len <= MAX_INSERTION { + if len >= 2 { + for i in (0..len - 1).rev() { + insert_head(&mut v[i..], &is_less); + } + } + return; + } + + // Allocate a buffer to use as scratch memory. We keep the length 0 so we can keep in it + // shallow copies of the contents of `v` without risking the dtors running on copies if + // `is_less` panics. + let mut buf = Vec::<T>::with_capacity(len); + let buf = buf.as_mut_ptr(); + + // If the slice is not longer than one chunk would be, do sequential merge sort and return. + if len <= CHUNK_LENGTH { + let res = unsafe { mergesort(v, buf, &is_less) }; + if res == MergesortResult::Descending { + v.reverse(); + } + return; + } + + // Split the slice into chunks and merge sort them in parallel. + // However, descending chunks will not be sorted - they will be simply left intact. + let mut iter = { + // Wrap pointer in SendPtr so that it can be sent to another thread + // See the documentation of SendPtr for a full explanation + let buf = SendPtr(buf); + let is_less = &is_less; + + v.par_chunks_mut(CHUNK_LENGTH) + .with_max_len(1) + .enumerate() + .map(move |(i, chunk)| { + let l = CHUNK_LENGTH * i; + let r = l + chunk.len(); + unsafe { + let buf = buf.get().add(l); + (l, r, mergesort(chunk, buf, is_less)) + } + }) + .collect::<Vec<_>>() + .into_iter() + .peekable() + }; + + // Now attempt to concatenate adjacent chunks that were left intact. + let mut chunks = Vec::with_capacity(iter.len()); + + while let Some((a, mut b, res)) = iter.next() { + // If this chunk was not modified by the sort procedure... + if res != MergesortResult::Sorted { + while let Some(&(x, y, r)) = iter.peek() { + // If the following chunk is of the same type and can be concatenated... + if r == res && (r == MergesortResult::Descending) == is_less(&v[x], &v[x - 1]) { + // Concatenate them. + b = y; + iter.next(); + } else { + break; + } + } + } + + // Descending chunks must be reversed. + if res == MergesortResult::Descending { + v[a..b].reverse(); + } + + chunks.push((a, b)); + } + + // All chunks are properly sorted. + // Now we just have to merge them together. + unsafe { + recurse(v.as_mut_ptr(), buf, &chunks, false, &is_less); + } +} + +#[cfg(test)] +mod tests { + use super::split_for_merge; + use rand::distributions::Uniform; + use rand::{thread_rng, Rng}; + + #[test] + fn test_split_for_merge() { + fn check(left: &[u32], right: &[u32]) { + let (l, r) = split_for_merge(left, right, &|&a, &b| a < b); + assert!(left[..l] + .iter() + .all(|&x| right[r..].iter().all(|&y| x <= y))); + assert!(right[..r].iter().all(|&x| left[l..].iter().all(|&y| x < y))); + } + + check(&[1, 2, 2, 2, 2, 3], &[1, 2, 2, 2, 2, 3]); + check(&[1, 2, 2, 2, 2, 3], &[]); + check(&[], &[1, 2, 2, 2, 2, 3]); + + let rng = &mut thread_rng(); + + for _ in 0..100 { + let limit: u32 = rng.gen_range(1..21); + let left_len: usize = rng.gen_range(0..20); + let right_len: usize = rng.gen_range(0..20); + + let mut left = rng + .sample_iter(&Uniform::new(0, limit)) + .take(left_len) + .collect::<Vec<_>>(); + let mut right = rng + .sample_iter(&Uniform::new(0, limit)) + .take(right_len) + .collect::<Vec<_>>(); + + left.sort(); + right.sort(); + check(&left, &right); + } + } +} diff --git a/third_party/rust/rayon/src/slice/mod.rs b/third_party/rust/rayon/src/slice/mod.rs new file mode 100644 index 0000000000..dab56deb32 --- /dev/null +++ b/third_party/rust/rayon/src/slice/mod.rs @@ -0,0 +1,1041 @@ +//! Parallel iterator types for [slices][std::slice] +//! +//! You will rarely need to interact with this module directly unless you need +//! to name one of the iterator types. +//! +//! [std::slice]: https://doc.rust-lang.org/stable/std/slice/ + +mod chunks; +mod mergesort; +mod quicksort; +mod rchunks; + +mod test; + +use self::mergesort::par_mergesort; +use self::quicksort::par_quicksort; +use crate::iter::plumbing::*; +use crate::iter::*; +use crate::split_producer::*; +use std::cmp; +use std::cmp::Ordering; +use std::fmt::{self, Debug}; +use std::mem; + +pub use self::chunks::{Chunks, ChunksExact, ChunksExactMut, ChunksMut}; +pub use self::rchunks::{RChunks, RChunksExact, RChunksExactMut, RChunksMut}; + +/// Parallel extensions for slices. +pub trait ParallelSlice<T: Sync> { + /// Returns a plain slice, which is used to implement the rest of the + /// parallel methods. + fn as_parallel_slice(&self) -> &[T]; + + /// Returns a parallel iterator over subslices separated by elements that + /// match the separator. + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// let smallest = [1, 2, 3, 0, 2, 4, 8, 0, 3, 6, 9] + /// .par_split(|i| *i == 0) + /// .map(|numbers| numbers.iter().min().unwrap()) + /// .min(); + /// assert_eq!(Some(&1), smallest); + /// ``` + fn par_split<P>(&self, separator: P) -> Split<'_, T, P> + where + P: Fn(&T) -> bool + Sync + Send, + { + Split { + slice: self.as_parallel_slice(), + separator, + } + } + + /// Returns a parallel iterator over all contiguous windows of length + /// `window_size`. The windows overlap. + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// let windows: Vec<_> = [1, 2, 3].par_windows(2).collect(); + /// assert_eq!(vec![[1, 2], [2, 3]], windows); + /// ``` + fn par_windows(&self, window_size: usize) -> Windows<'_, T> { + Windows { + window_size, + slice: self.as_parallel_slice(), + } + } + + /// Returns a parallel iterator over at most `chunk_size` elements of + /// `self` at a time. The chunks do not overlap. + /// + /// If the number of elements in the iterator is not divisible by + /// `chunk_size`, the last chunk may be shorter than `chunk_size`. All + /// other chunks will have that exact length. + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// let chunks: Vec<_> = [1, 2, 3, 4, 5].par_chunks(2).collect(); + /// assert_eq!(chunks, vec![&[1, 2][..], &[3, 4], &[5]]); + /// ``` + #[track_caller] + fn par_chunks(&self, chunk_size: usize) -> Chunks<'_, T> { + assert!(chunk_size != 0, "chunk_size must not be zero"); + Chunks::new(chunk_size, self.as_parallel_slice()) + } + + /// Returns a parallel iterator over `chunk_size` elements of + /// `self` at a time. The chunks do not overlap. + /// + /// If `chunk_size` does not divide the length of the slice, then the + /// last up to `chunk_size-1` elements will be omitted and can be + /// retrieved from the remainder function of the iterator. + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// let chunks: Vec<_> = [1, 2, 3, 4, 5].par_chunks_exact(2).collect(); + /// assert_eq!(chunks, vec![&[1, 2][..], &[3, 4]]); + /// ``` + #[track_caller] + fn par_chunks_exact(&self, chunk_size: usize) -> ChunksExact<'_, T> { + assert!(chunk_size != 0, "chunk_size must not be zero"); + ChunksExact::new(chunk_size, self.as_parallel_slice()) + } + + /// Returns a parallel iterator over at most `chunk_size` elements of `self` at a time, + /// starting at the end. The chunks do not overlap. + /// + /// If the number of elements in the iterator is not divisible by + /// `chunk_size`, the last chunk may be shorter than `chunk_size`. All + /// other chunks will have that exact length. + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// let chunks: Vec<_> = [1, 2, 3, 4, 5].par_rchunks(2).collect(); + /// assert_eq!(chunks, vec![&[4, 5][..], &[2, 3], &[1]]); + /// ``` + #[track_caller] + fn par_rchunks(&self, chunk_size: usize) -> RChunks<'_, T> { + assert!(chunk_size != 0, "chunk_size must not be zero"); + RChunks::new(chunk_size, self.as_parallel_slice()) + } + + /// Returns a parallel iterator over `chunk_size` elements of `self` at a time, + /// starting at the end. The chunks do not overlap. + /// + /// If `chunk_size` does not divide the length of the slice, then the + /// last up to `chunk_size-1` elements will be omitted and can be + /// retrieved from the remainder function of the iterator. + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// let chunks: Vec<_> = [1, 2, 3, 4, 5].par_rchunks_exact(2).collect(); + /// assert_eq!(chunks, vec![&[4, 5][..], &[2, 3]]); + /// ``` + #[track_caller] + fn par_rchunks_exact(&self, chunk_size: usize) -> RChunksExact<'_, T> { + assert!(chunk_size != 0, "chunk_size must not be zero"); + RChunksExact::new(chunk_size, self.as_parallel_slice()) + } +} + +impl<T: Sync> ParallelSlice<T> for [T] { + #[inline] + fn as_parallel_slice(&self) -> &[T] { + self + } +} + +/// Parallel extensions for mutable slices. +pub trait ParallelSliceMut<T: Send> { + /// Returns a plain mutable slice, which is used to implement the rest of + /// the parallel methods. + fn as_parallel_slice_mut(&mut self) -> &mut [T]; + + /// Returns a parallel iterator over mutable subslices separated by + /// elements that match the separator. + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// let mut array = [1, 2, 3, 0, 2, 4, 8, 0, 3, 6, 9]; + /// array.par_split_mut(|i| *i == 0) + /// .for_each(|slice| slice.reverse()); + /// assert_eq!(array, [3, 2, 1, 0, 8, 4, 2, 0, 9, 6, 3]); + /// ``` + fn par_split_mut<P>(&mut self, separator: P) -> SplitMut<'_, T, P> + where + P: Fn(&T) -> bool + Sync + Send, + { + SplitMut { + slice: self.as_parallel_slice_mut(), + separator, + } + } + + /// Returns a parallel iterator over at most `chunk_size` elements of + /// `self` at a time. The chunks are mutable and do not overlap. + /// + /// If the number of elements in the iterator is not divisible by + /// `chunk_size`, the last chunk may be shorter than `chunk_size`. All + /// other chunks will have that exact length. + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// let mut array = [1, 2, 3, 4, 5]; + /// array.par_chunks_mut(2) + /// .for_each(|slice| slice.reverse()); + /// assert_eq!(array, [2, 1, 4, 3, 5]); + /// ``` + #[track_caller] + fn par_chunks_mut(&mut self, chunk_size: usize) -> ChunksMut<'_, T> { + assert!(chunk_size != 0, "chunk_size must not be zero"); + ChunksMut::new(chunk_size, self.as_parallel_slice_mut()) + } + + /// Returns a parallel iterator over `chunk_size` elements of + /// `self` at a time. The chunks are mutable and do not overlap. + /// + /// If `chunk_size` does not divide the length of the slice, then the + /// last up to `chunk_size-1` elements will be omitted and can be + /// retrieved from the remainder function of the iterator. + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// let mut array = [1, 2, 3, 4, 5]; + /// array.par_chunks_exact_mut(3) + /// .for_each(|slice| slice.reverse()); + /// assert_eq!(array, [3, 2, 1, 4, 5]); + /// ``` + #[track_caller] + fn par_chunks_exact_mut(&mut self, chunk_size: usize) -> ChunksExactMut<'_, T> { + assert!(chunk_size != 0, "chunk_size must not be zero"); + ChunksExactMut::new(chunk_size, self.as_parallel_slice_mut()) + } + + /// Returns a parallel iterator over at most `chunk_size` elements of `self` at a time, + /// starting at the end. The chunks are mutable and do not overlap. + /// + /// If the number of elements in the iterator is not divisible by + /// `chunk_size`, the last chunk may be shorter than `chunk_size`. All + /// other chunks will have that exact length. + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// let mut array = [1, 2, 3, 4, 5]; + /// array.par_rchunks_mut(2) + /// .for_each(|slice| slice.reverse()); + /// assert_eq!(array, [1, 3, 2, 5, 4]); + /// ``` + #[track_caller] + fn par_rchunks_mut(&mut self, chunk_size: usize) -> RChunksMut<'_, T> { + assert!(chunk_size != 0, "chunk_size must not be zero"); + RChunksMut::new(chunk_size, self.as_parallel_slice_mut()) + } + + /// Returns a parallel iterator over `chunk_size` elements of `self` at a time, + /// starting at the end. The chunks are mutable and do not overlap. + /// + /// If `chunk_size` does not divide the length of the slice, then the + /// last up to `chunk_size-1` elements will be omitted and can be + /// retrieved from the remainder function of the iterator. + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// let mut array = [1, 2, 3, 4, 5]; + /// array.par_rchunks_exact_mut(3) + /// .for_each(|slice| slice.reverse()); + /// assert_eq!(array, [1, 2, 5, 4, 3]); + /// ``` + #[track_caller] + fn par_rchunks_exact_mut(&mut self, chunk_size: usize) -> RChunksExactMut<'_, T> { + assert!(chunk_size != 0, "chunk_size must not be zero"); + RChunksExactMut::new(chunk_size, self.as_parallel_slice_mut()) + } + + /// Sorts the slice in parallel. + /// + /// This sort is stable (i.e., does not reorder equal elements) and *O*(*n* \* log(*n*)) worst-case. + /// + /// When applicable, unstable sorting is preferred because it is generally faster than stable + /// sorting and it doesn't allocate auxiliary memory. + /// See [`par_sort_unstable`](#method.par_sort_unstable). + /// + /// # Current implementation + /// + /// The current algorithm is an adaptive merge sort inspired by + /// [timsort](https://en.wikipedia.org/wiki/Timsort). + /// It is designed to be very fast in cases where the slice is nearly sorted, or consists of + /// two or more sorted sequences concatenated one after another. + /// + /// Also, it allocates temporary storage the same size as `self`, but for very short slices a + /// non-allocating insertion sort is used instead. + /// + /// In order to sort the slice in parallel, the slice is first divided into smaller chunks and + /// all chunks are sorted in parallel. Then, adjacent chunks that together form non-descending + /// or descending runs are concatenated. Finally, the remaining chunks are merged together using + /// parallel subdivision of chunks and parallel merge operation. + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// + /// let mut v = [-5, 4, 1, -3, 2]; + /// + /// v.par_sort(); + /// assert_eq!(v, [-5, -3, 1, 2, 4]); + /// ``` + fn par_sort(&mut self) + where + T: Ord, + { + par_mergesort(self.as_parallel_slice_mut(), T::lt); + } + + /// Sorts the slice in parallel with a comparator function. + /// + /// This sort is stable (i.e., does not reorder equal elements) and *O*(*n* \* log(*n*)) worst-case. + /// + /// The comparator function must define a total ordering for the elements in the slice. If + /// the ordering is not total, the order of the elements is unspecified. An order is a + /// total order if it is (for all `a`, `b` and `c`): + /// + /// * total and antisymmetric: exactly one of `a < b`, `a == b` or `a > b` is true, and + /// * transitive, `a < b` and `b < c` implies `a < c`. The same must hold for both `==` and `>`. + /// + /// For example, while [`f64`] doesn't implement [`Ord`] because `NaN != NaN`, we can use + /// `partial_cmp` as our sort function when we know the slice doesn't contain a `NaN`. + /// + /// ``` + /// use rayon::prelude::*; + /// + /// let mut floats = [5f64, 4.0, 1.0, 3.0, 2.0]; + /// floats.par_sort_by(|a, b| a.partial_cmp(b).unwrap()); + /// assert_eq!(floats, [1.0, 2.0, 3.0, 4.0, 5.0]); + /// ``` + /// + /// When applicable, unstable sorting is preferred because it is generally faster than stable + /// sorting and it doesn't allocate auxiliary memory. + /// See [`par_sort_unstable_by`](#method.par_sort_unstable_by). + /// + /// # Current implementation + /// + /// The current algorithm is an adaptive merge sort inspired by + /// [timsort](https://en.wikipedia.org/wiki/Timsort). + /// It is designed to be very fast in cases where the slice is nearly sorted, or consists of + /// two or more sorted sequences concatenated one after another. + /// + /// Also, it allocates temporary storage the same size as `self`, but for very short slices a + /// non-allocating insertion sort is used instead. + /// + /// In order to sort the slice in parallel, the slice is first divided into smaller chunks and + /// all chunks are sorted in parallel. Then, adjacent chunks that together form non-descending + /// or descending runs are concatenated. Finally, the remaining chunks are merged together using + /// parallel subdivision of chunks and parallel merge operation. + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// + /// let mut v = [5, 4, 1, 3, 2]; + /// v.par_sort_by(|a, b| a.cmp(b)); + /// assert_eq!(v, [1, 2, 3, 4, 5]); + /// + /// // reverse sorting + /// v.par_sort_by(|a, b| b.cmp(a)); + /// assert_eq!(v, [5, 4, 3, 2, 1]); + /// ``` + fn par_sort_by<F>(&mut self, compare: F) + where + F: Fn(&T, &T) -> Ordering + Sync, + { + par_mergesort(self.as_parallel_slice_mut(), |a, b| { + compare(a, b) == Ordering::Less + }); + } + + /// Sorts the slice in parallel with a key extraction function. + /// + /// This sort is stable (i.e., does not reorder equal elements) and *O*(*m* \* *n* \* log(*n*)) + /// worst-case, where the key function is *O*(*m*). + /// + /// For expensive key functions (e.g. functions that are not simple property accesses or + /// basic operations), [`par_sort_by_cached_key`](#method.par_sort_by_cached_key) is likely to + /// be significantly faster, as it does not recompute element keys. + /// + /// When applicable, unstable sorting is preferred because it is generally faster than stable + /// sorting and it doesn't allocate auxiliary memory. + /// See [`par_sort_unstable_by_key`](#method.par_sort_unstable_by_key). + /// + /// # Current implementation + /// + /// The current algorithm is an adaptive merge sort inspired by + /// [timsort](https://en.wikipedia.org/wiki/Timsort). + /// It is designed to be very fast in cases where the slice is nearly sorted, or consists of + /// two or more sorted sequences concatenated one after another. + /// + /// Also, it allocates temporary storage the same size as `self`, but for very short slices a + /// non-allocating insertion sort is used instead. + /// + /// In order to sort the slice in parallel, the slice is first divided into smaller chunks and + /// all chunks are sorted in parallel. Then, adjacent chunks that together form non-descending + /// or descending runs are concatenated. Finally, the remaining chunks are merged together using + /// parallel subdivision of chunks and parallel merge operation. + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// + /// let mut v = [-5i32, 4, 1, -3, 2]; + /// + /// v.par_sort_by_key(|k| k.abs()); + /// assert_eq!(v, [1, 2, -3, 4, -5]); + /// ``` + fn par_sort_by_key<K, F>(&mut self, f: F) + where + K: Ord, + F: Fn(&T) -> K + Sync, + { + par_mergesort(self.as_parallel_slice_mut(), |a, b| f(a).lt(&f(b))); + } + + /// Sorts the slice in parallel with a key extraction function. + /// + /// During sorting, the key function is called at most once per element, by using + /// temporary storage to remember the results of key evaluation. + /// The key function is called in parallel, so the order of calls is completely unspecified. + /// + /// This sort is stable (i.e., does not reorder equal elements) and *O*(*m* \* *n* + *n* \* log(*n*)) + /// worst-case, where the key function is *O*(*m*). + /// + /// For simple key functions (e.g., functions that are property accesses or + /// basic operations), [`par_sort_by_key`](#method.par_sort_by_key) is likely to be + /// faster. + /// + /// # Current implementation + /// + /// The current algorithm is based on [pattern-defeating quicksort][pdqsort] by Orson Peters, + /// which combines the fast average case of randomized quicksort with the fast worst case of + /// heapsort, while achieving linear time on slices with certain patterns. It uses some + /// randomization to avoid degenerate cases, but with a fixed seed to always provide + /// deterministic behavior. + /// + /// In the worst case, the algorithm allocates temporary storage in a `Vec<(K, usize)>` the + /// length of the slice. + /// + /// All quicksorts work in two stages: partitioning into two halves followed by recursive + /// calls. The partitioning phase is sequential, but the two recursive calls are performed in + /// parallel. Finally, after sorting the cached keys, the item positions are updated sequentially. + /// + /// [pdqsort]: https://github.com/orlp/pdqsort + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// + /// let mut v = [-5i32, 4, 32, -3, 2]; + /// + /// v.par_sort_by_cached_key(|k| k.to_string()); + /// assert!(v == [-3, -5, 2, 32, 4]); + /// ``` + fn par_sort_by_cached_key<K, F>(&mut self, f: F) + where + F: Fn(&T) -> K + Sync, + K: Ord + Send, + { + let slice = self.as_parallel_slice_mut(); + let len = slice.len(); + if len < 2 { + return; + } + + // Helper macro for indexing our vector by the smallest possible type, to reduce allocation. + macro_rules! sort_by_key { + ($t:ty) => {{ + let mut indices: Vec<_> = slice + .par_iter_mut() + .enumerate() + .map(|(i, x)| (f(&*x), i as $t)) + .collect(); + // The elements of `indices` are unique, as they are indexed, so any sort will be + // stable with respect to the original slice. We use `sort_unstable` here because + // it requires less memory allocation. + indices.par_sort_unstable(); + for i in 0..len { + let mut index = indices[i].1; + while (index as usize) < i { + index = indices[index as usize].1; + } + indices[i].1 = index; + slice.swap(i, index as usize); + } + }}; + } + + let sz_u8 = mem::size_of::<(K, u8)>(); + let sz_u16 = mem::size_of::<(K, u16)>(); + let sz_u32 = mem::size_of::<(K, u32)>(); + let sz_usize = mem::size_of::<(K, usize)>(); + + if sz_u8 < sz_u16 && len <= (std::u8::MAX as usize) { + return sort_by_key!(u8); + } + if sz_u16 < sz_u32 && len <= (std::u16::MAX as usize) { + return sort_by_key!(u16); + } + if sz_u32 < sz_usize && len <= (std::u32::MAX as usize) { + return sort_by_key!(u32); + } + sort_by_key!(usize) + } + + /// Sorts the slice in parallel, but might not preserve the order of equal elements. + /// + /// This sort is unstable (i.e., may reorder equal elements), in-place + /// (i.e., does not allocate), and *O*(*n* \* log(*n*)) worst-case. + /// + /// # Current implementation + /// + /// The current algorithm is based on [pattern-defeating quicksort][pdqsort] by Orson Peters, + /// which combines the fast average case of randomized quicksort with the fast worst case of + /// heapsort, while achieving linear time on slices with certain patterns. It uses some + /// randomization to avoid degenerate cases, but with a fixed seed to always provide + /// deterministic behavior. + /// + /// It is typically faster than stable sorting, except in a few special cases, e.g., when the + /// slice consists of several concatenated sorted sequences. + /// + /// All quicksorts work in two stages: partitioning into two halves followed by recursive + /// calls. The partitioning phase is sequential, but the two recursive calls are performed in + /// parallel. + /// + /// [pdqsort]: https://github.com/orlp/pdqsort + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// + /// let mut v = [-5, 4, 1, -3, 2]; + /// + /// v.par_sort_unstable(); + /// assert_eq!(v, [-5, -3, 1, 2, 4]); + /// ``` + fn par_sort_unstable(&mut self) + where + T: Ord, + { + par_quicksort(self.as_parallel_slice_mut(), T::lt); + } + + /// Sorts the slice in parallel with a comparator function, but might not preserve the order of + /// equal elements. + /// + /// This sort is unstable (i.e., may reorder equal elements), in-place + /// (i.e., does not allocate), and *O*(*n* \* log(*n*)) worst-case. + /// + /// The comparator function must define a total ordering for the elements in the slice. If + /// the ordering is not total, the order of the elements is unspecified. An order is a + /// total order if it is (for all `a`, `b` and `c`): + /// + /// * total and antisymmetric: exactly one of `a < b`, `a == b` or `a > b` is true, and + /// * transitive, `a < b` and `b < c` implies `a < c`. The same must hold for both `==` and `>`. + /// + /// For example, while [`f64`] doesn't implement [`Ord`] because `NaN != NaN`, we can use + /// `partial_cmp` as our sort function when we know the slice doesn't contain a `NaN`. + /// + /// ``` + /// use rayon::prelude::*; + /// + /// let mut floats = [5f64, 4.0, 1.0, 3.0, 2.0]; + /// floats.par_sort_unstable_by(|a, b| a.partial_cmp(b).unwrap()); + /// assert_eq!(floats, [1.0, 2.0, 3.0, 4.0, 5.0]); + /// ``` + /// + /// # Current implementation + /// + /// The current algorithm is based on [pattern-defeating quicksort][pdqsort] by Orson Peters, + /// which combines the fast average case of randomized quicksort with the fast worst case of + /// heapsort, while achieving linear time on slices with certain patterns. It uses some + /// randomization to avoid degenerate cases, but with a fixed seed to always provide + /// deterministic behavior. + /// + /// It is typically faster than stable sorting, except in a few special cases, e.g., when the + /// slice consists of several concatenated sorted sequences. + /// + /// All quicksorts work in two stages: partitioning into two halves followed by recursive + /// calls. The partitioning phase is sequential, but the two recursive calls are performed in + /// parallel. + /// + /// [pdqsort]: https://github.com/orlp/pdqsort + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// + /// let mut v = [5, 4, 1, 3, 2]; + /// v.par_sort_unstable_by(|a, b| a.cmp(b)); + /// assert_eq!(v, [1, 2, 3, 4, 5]); + /// + /// // reverse sorting + /// v.par_sort_unstable_by(|a, b| b.cmp(a)); + /// assert_eq!(v, [5, 4, 3, 2, 1]); + /// ``` + fn par_sort_unstable_by<F>(&mut self, compare: F) + where + F: Fn(&T, &T) -> Ordering + Sync, + { + par_quicksort(self.as_parallel_slice_mut(), |a, b| { + compare(a, b) == Ordering::Less + }); + } + + /// Sorts the slice in parallel with a key extraction function, but might not preserve the order + /// of equal elements. + /// + /// This sort is unstable (i.e., may reorder equal elements), in-place + /// (i.e., does not allocate), and *O*(m \* *n* \* log(*n*)) worst-case, + /// where the key function is *O*(*m*). + /// + /// # Current implementation + /// + /// The current algorithm is based on [pattern-defeating quicksort][pdqsort] by Orson Peters, + /// which combines the fast average case of randomized quicksort with the fast worst case of + /// heapsort, while achieving linear time on slices with certain patterns. It uses some + /// randomization to avoid degenerate cases, but with a fixed seed to always provide + /// deterministic behavior. + /// + /// Due to its key calling strategy, `par_sort_unstable_by_key` is likely to be slower than + /// [`par_sort_by_cached_key`](#method.par_sort_by_cached_key) in cases where the key function + /// is expensive. + /// + /// All quicksorts work in two stages: partitioning into two halves followed by recursive + /// calls. The partitioning phase is sequential, but the two recursive calls are performed in + /// parallel. + /// + /// [pdqsort]: https://github.com/orlp/pdqsort + /// + /// # Examples + /// + /// ``` + /// use rayon::prelude::*; + /// + /// let mut v = [-5i32, 4, 1, -3, 2]; + /// + /// v.par_sort_unstable_by_key(|k| k.abs()); + /// assert_eq!(v, [1, 2, -3, 4, -5]); + /// ``` + fn par_sort_unstable_by_key<K, F>(&mut self, f: F) + where + K: Ord, + F: Fn(&T) -> K + Sync, + { + par_quicksort(self.as_parallel_slice_mut(), |a, b| f(a).lt(&f(b))); + } +} + +impl<T: Send> ParallelSliceMut<T> for [T] { + #[inline] + fn as_parallel_slice_mut(&mut self) -> &mut [T] { + self + } +} + +impl<'data, T: Sync + 'data> IntoParallelIterator for &'data [T] { + type Item = &'data T; + type Iter = Iter<'data, T>; + + fn into_par_iter(self) -> Self::Iter { + Iter { slice: self } + } +} + +impl<'data, T: Send + 'data> IntoParallelIterator for &'data mut [T] { + type Item = &'data mut T; + type Iter = IterMut<'data, T>; + + fn into_par_iter(self) -> Self::Iter { + IterMut { slice: self } + } +} + +/// Parallel iterator over immutable items in a slice +#[derive(Debug)] +pub struct Iter<'data, T: Sync> { + slice: &'data [T], +} + +impl<'data, T: Sync> Clone for Iter<'data, T> { + fn clone(&self) -> Self { + Iter { ..*self } + } +} + +impl<'data, T: Sync + 'data> ParallelIterator for Iter<'data, T> { + type Item = &'data T; + + fn drive_unindexed<C>(self, consumer: C) -> C::Result + where + C: UnindexedConsumer<Self::Item>, + { + bridge(self, consumer) + } + + fn opt_len(&self) -> Option<usize> { + Some(self.len()) + } +} + +impl<'data, T: Sync + 'data> IndexedParallelIterator for Iter<'data, T> { + fn drive<C>(self, consumer: C) -> C::Result + where + C: Consumer<Self::Item>, + { + bridge(self, consumer) + } + + fn len(&self) -> usize { + self.slice.len() + } + + fn with_producer<CB>(self, callback: CB) -> CB::Output + where + CB: ProducerCallback<Self::Item>, + { + callback.callback(IterProducer { slice: self.slice }) + } +} + +struct IterProducer<'data, T: Sync> { + slice: &'data [T], +} + +impl<'data, T: 'data + Sync> Producer for IterProducer<'data, T> { + type Item = &'data T; + type IntoIter = ::std::slice::Iter<'data, T>; + + fn into_iter(self) -> Self::IntoIter { + self.slice.iter() + } + + fn split_at(self, index: usize) -> (Self, Self) { + let (left, right) = self.slice.split_at(index); + (IterProducer { slice: left }, IterProducer { slice: right }) + } +} + +/// Parallel iterator over immutable overlapping windows of a slice +#[derive(Debug)] +pub struct Windows<'data, T: Sync> { + window_size: usize, + slice: &'data [T], +} + +impl<'data, T: Sync> Clone for Windows<'data, T> { + fn clone(&self) -> Self { + Windows { ..*self } + } +} + +impl<'data, T: Sync + 'data> ParallelIterator for Windows<'data, T> { + type Item = &'data [T]; + + fn drive_unindexed<C>(self, consumer: C) -> C::Result + where + C: UnindexedConsumer<Self::Item>, + { + bridge(self, consumer) + } + + fn opt_len(&self) -> Option<usize> { + Some(self.len()) + } +} + +impl<'data, T: Sync + 'data> IndexedParallelIterator for Windows<'data, T> { + fn drive<C>(self, consumer: C) -> C::Result + where + C: Consumer<Self::Item>, + { + bridge(self, consumer) + } + + fn len(&self) -> usize { + assert!(self.window_size >= 1); + self.slice.len().saturating_sub(self.window_size - 1) + } + + fn with_producer<CB>(self, callback: CB) -> CB::Output + where + CB: ProducerCallback<Self::Item>, + { + callback.callback(WindowsProducer { + window_size: self.window_size, + slice: self.slice, + }) + } +} + +struct WindowsProducer<'data, T: Sync> { + window_size: usize, + slice: &'data [T], +} + +impl<'data, T: 'data + Sync> Producer for WindowsProducer<'data, T> { + type Item = &'data [T]; + type IntoIter = ::std::slice::Windows<'data, T>; + + fn into_iter(self) -> Self::IntoIter { + self.slice.windows(self.window_size) + } + + fn split_at(self, index: usize) -> (Self, Self) { + let left_index = cmp::min(self.slice.len(), index + (self.window_size - 1)); + let left = &self.slice[..left_index]; + let right = &self.slice[index..]; + ( + WindowsProducer { + window_size: self.window_size, + slice: left, + }, + WindowsProducer { + window_size: self.window_size, + slice: right, + }, + ) + } +} + +/// Parallel iterator over mutable items in a slice +#[derive(Debug)] +pub struct IterMut<'data, T: Send> { + slice: &'data mut [T], +} + +impl<'data, T: Send + 'data> ParallelIterator for IterMut<'data, T> { + type Item = &'data mut T; + + fn drive_unindexed<C>(self, consumer: C) -> C::Result + where + C: UnindexedConsumer<Self::Item>, + { + bridge(self, consumer) + } + + fn opt_len(&self) -> Option<usize> { + Some(self.len()) + } +} + +impl<'data, T: Send + 'data> IndexedParallelIterator for IterMut<'data, T> { + fn drive<C>(self, consumer: C) -> C::Result + where + C: Consumer<Self::Item>, + { + bridge(self, consumer) + } + + fn len(&self) -> usize { + self.slice.len() + } + + fn with_producer<CB>(self, callback: CB) -> CB::Output + where + CB: ProducerCallback<Self::Item>, + { + callback.callback(IterMutProducer { slice: self.slice }) + } +} + +struct IterMutProducer<'data, T: Send> { + slice: &'data mut [T], +} + +impl<'data, T: 'data + Send> Producer for IterMutProducer<'data, T> { + type Item = &'data mut T; + type IntoIter = ::std::slice::IterMut<'data, T>; + + fn into_iter(self) -> Self::IntoIter { + self.slice.iter_mut() + } + + fn split_at(self, index: usize) -> (Self, Self) { + let (left, right) = self.slice.split_at_mut(index); + ( + IterMutProducer { slice: left }, + IterMutProducer { slice: right }, + ) + } +} + +/// Parallel iterator over slices separated by a predicate +pub struct Split<'data, T, P> { + slice: &'data [T], + separator: P, +} + +impl<'data, T, P: Clone> Clone for Split<'data, T, P> { + fn clone(&self) -> Self { + Split { + separator: self.separator.clone(), + ..*self + } + } +} + +impl<'data, T: Debug, P> Debug for Split<'data, T, P> { + fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { + f.debug_struct("Split").field("slice", &self.slice).finish() + } +} + +impl<'data, T, P> ParallelIterator for Split<'data, T, P> +where + P: Fn(&T) -> bool + Sync + Send, + T: Sync, +{ + type Item = &'data [T]; + + fn drive_unindexed<C>(self, consumer: C) -> C::Result + where + C: UnindexedConsumer<Self::Item>, + { + let producer = SplitProducer::new(self.slice, &self.separator); + bridge_unindexed(producer, consumer) + } +} + +/// Implement support for `SplitProducer`. +impl<'data, T, P> Fissile<P> for &'data [T] +where + P: Fn(&T) -> bool, +{ + fn length(&self) -> usize { + self.len() + } + + fn midpoint(&self, end: usize) -> usize { + end / 2 + } + + fn find(&self, separator: &P, start: usize, end: usize) -> Option<usize> { + self[start..end].iter().position(separator) + } + + fn rfind(&self, separator: &P, end: usize) -> Option<usize> { + self[..end].iter().rposition(separator) + } + + fn split_once(self, index: usize) -> (Self, Self) { + let (left, right) = self.split_at(index); + (left, &right[1..]) // skip the separator + } + + fn fold_splits<F>(self, separator: &P, folder: F, skip_last: bool) -> F + where + F: Folder<Self>, + Self: Send, + { + let mut split = self.split(separator); + if skip_last { + split.next_back(); + } + folder.consume_iter(split) + } +} + +/// Parallel iterator over mutable slices separated by a predicate +pub struct SplitMut<'data, T, P> { + slice: &'data mut [T], + separator: P, +} + +impl<'data, T: Debug, P> Debug for SplitMut<'data, T, P> { + fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { + f.debug_struct("SplitMut") + .field("slice", &self.slice) + .finish() + } +} + +impl<'data, T, P> ParallelIterator for SplitMut<'data, T, P> +where + P: Fn(&T) -> bool + Sync + Send, + T: Send, +{ + type Item = &'data mut [T]; + + fn drive_unindexed<C>(self, consumer: C) -> C::Result + where + C: UnindexedConsumer<Self::Item>, + { + let producer = SplitProducer::new(self.slice, &self.separator); + bridge_unindexed(producer, consumer) + } +} + +/// Implement support for `SplitProducer`. +impl<'data, T, P> Fissile<P> for &'data mut [T] +where + P: Fn(&T) -> bool, +{ + fn length(&self) -> usize { + self.len() + } + + fn midpoint(&self, end: usize) -> usize { + end / 2 + } + + fn find(&self, separator: &P, start: usize, end: usize) -> Option<usize> { + self[start..end].iter().position(separator) + } + + fn rfind(&self, separator: &P, end: usize) -> Option<usize> { + self[..end].iter().rposition(separator) + } + + fn split_once(self, index: usize) -> (Self, Self) { + let (left, right) = self.split_at_mut(index); + (left, &mut right[1..]) // skip the separator + } + + fn fold_splits<F>(self, separator: &P, folder: F, skip_last: bool) -> F + where + F: Folder<Self>, + Self: Send, + { + let mut split = self.split_mut(separator); + if skip_last { + split.next_back(); + } + folder.consume_iter(split) + } +} diff --git a/third_party/rust/rayon/src/slice/quicksort.rs b/third_party/rust/rayon/src/slice/quicksort.rs new file mode 100644 index 0000000000..d71079ec3b --- /dev/null +++ b/third_party/rust/rayon/src/slice/quicksort.rs @@ -0,0 +1,897 @@ +//! Parallel quicksort. +//! +//! This implementation is copied verbatim from `std::slice::sort_unstable` and then parallelized. +//! The only difference from the original is that calls to `recurse` are executed in parallel using +//! `rayon_core::join`. + +use std::cmp; +use std::mem::{self, MaybeUninit}; +use std::ptr; + +/// When dropped, copies from `src` into `dest`. +struct CopyOnDrop<T> { + src: *const T, + dest: *mut T, +} + +impl<T> Drop for CopyOnDrop<T> { + fn drop(&mut self) { + // SAFETY: This is a helper class. + // Please refer to its usage for correctness. + // Namely, one must be sure that `src` and `dst` does not overlap as required by `ptr::copy_nonoverlapping`. + unsafe { + ptr::copy_nonoverlapping(self.src, self.dest, 1); + } + } +} + +/// Shifts the first element to the right until it encounters a greater or equal element. +fn shift_head<T, F>(v: &mut [T], is_less: &F) +where + F: Fn(&T, &T) -> bool, +{ + let len = v.len(); + // SAFETY: The unsafe operations below involves indexing without a bounds check (by offsetting a + // pointer) and copying memory (`ptr::copy_nonoverlapping`). + // + // a. Indexing: + // 1. We checked the size of the array to >=2. + // 2. All the indexing that we will do is always between {0 <= index < len} at most. + // + // b. Memory copying + // 1. We are obtaining pointers to references which are guaranteed to be valid. + // 2. They cannot overlap because we obtain pointers to difference indices of the slice. + // Namely, `i` and `i-1`. + // 3. If the slice is properly aligned, the elements are properly aligned. + // It is the caller's responsibility to make sure the slice is properly aligned. + // + // See comments below for further detail. + unsafe { + // If the first two elements are out-of-order... + if len >= 2 && is_less(v.get_unchecked(1), v.get_unchecked(0)) { + // Read the first element into a stack-allocated variable. If a following comparison + // operation panics, `hole` will get dropped and automatically write the element back + // into the slice. + let tmp = mem::ManuallyDrop::new(ptr::read(v.get_unchecked(0))); + let v = v.as_mut_ptr(); + let mut hole = CopyOnDrop { + src: &*tmp, + dest: v.add(1), + }; + ptr::copy_nonoverlapping(v.add(1), v.add(0), 1); + + for i in 2..len { + if !is_less(&*v.add(i), &*tmp) { + break; + } + + // Move `i`-th element one place to the left, thus shifting the hole to the right. + ptr::copy_nonoverlapping(v.add(i), v.add(i - 1), 1); + hole.dest = v.add(i); + } + // `hole` gets dropped and thus copies `tmp` into the remaining hole in `v`. + } + } +} + +/// Shifts the last element to the left until it encounters a smaller or equal element. +fn shift_tail<T, F>(v: &mut [T], is_less: &F) +where + F: Fn(&T, &T) -> bool, +{ + let len = v.len(); + // SAFETY: The unsafe operations below involves indexing without a bound check (by offsetting a + // pointer) and copying memory (`ptr::copy_nonoverlapping`). + // + // a. Indexing: + // 1. We checked the size of the array to >= 2. + // 2. All the indexing that we will do is always between `0 <= index < len-1` at most. + // + // b. Memory copying + // 1. We are obtaining pointers to references which are guaranteed to be valid. + // 2. They cannot overlap because we obtain pointers to difference indices of the slice. + // Namely, `i` and `i+1`. + // 3. If the slice is properly aligned, the elements are properly aligned. + // It is the caller's responsibility to make sure the slice is properly aligned. + // + // See comments below for further detail. + unsafe { + // If the last two elements are out-of-order... + if len >= 2 && is_less(v.get_unchecked(len - 1), v.get_unchecked(len - 2)) { + // Read the last element into a stack-allocated variable. If a following comparison + // operation panics, `hole` will get dropped and automatically write the element back + // into the slice. + let tmp = mem::ManuallyDrop::new(ptr::read(v.get_unchecked(len - 1))); + let v = v.as_mut_ptr(); + let mut hole = CopyOnDrop { + src: &*tmp, + dest: v.add(len - 2), + }; + ptr::copy_nonoverlapping(v.add(len - 2), v.add(len - 1), 1); + + for i in (0..len - 2).rev() { + if !is_less(&*tmp, &*v.add(i)) { + break; + } + + // Move `i`-th element one place to the right, thus shifting the hole to the left. + ptr::copy_nonoverlapping(v.add(i), v.add(i + 1), 1); + hole.dest = v.add(i); + } + // `hole` gets dropped and thus copies `tmp` into the remaining hole in `v`. + } + } +} + +/// Partially sorts a slice by shifting several out-of-order elements around. +/// +/// Returns `true` if the slice is sorted at the end. This function is *O*(*n*) worst-case. +#[cold] +fn partial_insertion_sort<T, F>(v: &mut [T], is_less: &F) -> bool +where + F: Fn(&T, &T) -> bool, +{ + // Maximum number of adjacent out-of-order pairs that will get shifted. + const MAX_STEPS: usize = 5; + // If the slice is shorter than this, don't shift any elements. + const SHORTEST_SHIFTING: usize = 50; + + let len = v.len(); + let mut i = 1; + + for _ in 0..MAX_STEPS { + // SAFETY: We already explicitly did the bound checking with `i < len`. + // All our subsequent indexing is only in the range `0 <= index < len` + unsafe { + // Find the next pair of adjacent out-of-order elements. + while i < len && !is_less(v.get_unchecked(i), v.get_unchecked(i - 1)) { + i += 1; + } + } + + // Are we done? + if i == len { + return true; + } + + // Don't shift elements on short arrays, that has a performance cost. + if len < SHORTEST_SHIFTING { + return false; + } + + // Swap the found pair of elements. This puts them in correct order. + v.swap(i - 1, i); + + // Shift the smaller element to the left. + shift_tail(&mut v[..i], is_less); + // Shift the greater element to the right. + shift_head(&mut v[i..], is_less); + } + + // Didn't manage to sort the slice in the limited number of steps. + false +} + +/// Sorts a slice using insertion sort, which is *O*(*n*^2) worst-case. +fn insertion_sort<T, F>(v: &mut [T], is_less: &F) +where + F: Fn(&T, &T) -> bool, +{ + for i in 1..v.len() { + shift_tail(&mut v[..i + 1], is_less); + } +} + +/// Sorts `v` using heapsort, which guarantees *O*(*n* \* log(*n*)) worst-case. +#[cold] +fn heapsort<T, F>(v: &mut [T], is_less: &F) +where + F: Fn(&T, &T) -> bool, +{ + // This binary heap respects the invariant `parent >= child`. + let sift_down = |v: &mut [T], mut node| { + loop { + // Children of `node`. + let mut child = 2 * node + 1; + if child >= v.len() { + break; + } + + // Choose the greater child. + if child + 1 < v.len() && is_less(&v[child], &v[child + 1]) { + child += 1; + } + + // Stop if the invariant holds at `node`. + if !is_less(&v[node], &v[child]) { + break; + } + + // Swap `node` with the greater child, move one step down, and continue sifting. + v.swap(node, child); + node = child; + } + }; + + // Build the heap in linear time. + for i in (0..v.len() / 2).rev() { + sift_down(v, i); + } + + // Pop maximal elements from the heap. + for i in (1..v.len()).rev() { + v.swap(0, i); + sift_down(&mut v[..i], 0); + } +} + +/// Partitions `v` into elements smaller than `pivot`, followed by elements greater than or equal +/// to `pivot`. +/// +/// Returns the number of elements smaller than `pivot`. +/// +/// Partitioning is performed block-by-block in order to minimize the cost of branching operations. +/// This idea is presented in the [BlockQuicksort][pdf] paper. +/// +/// [pdf]: https://drops.dagstuhl.de/opus/volltexte/2016/6389/pdf/LIPIcs-ESA-2016-38.pdf +fn partition_in_blocks<T, F>(v: &mut [T], pivot: &T, is_less: &F) -> usize +where + F: Fn(&T, &T) -> bool, +{ + // Number of elements in a typical block. + const BLOCK: usize = 128; + + // The partitioning algorithm repeats the following steps until completion: + // + // 1. Trace a block from the left side to identify elements greater than or equal to the pivot. + // 2. Trace a block from the right side to identify elements smaller than the pivot. + // 3. Exchange the identified elements between the left and right side. + // + // We keep the following variables for a block of elements: + // + // 1. `block` - Number of elements in the block. + // 2. `start` - Start pointer into the `offsets` array. + // 3. `end` - End pointer into the `offsets` array. + // 4. `offsets - Indices of out-of-order elements within the block. + + // The current block on the left side (from `l` to `l.add(block_l)`). + let mut l = v.as_mut_ptr(); + let mut block_l = BLOCK; + let mut start_l = ptr::null_mut(); + let mut end_l = ptr::null_mut(); + let mut offsets_l = [MaybeUninit::<u8>::uninit(); BLOCK]; + + // The current block on the right side (from `r.sub(block_r)` to `r`). + // SAFETY: The documentation for .add() specifically mention that `vec.as_ptr().add(vec.len())` is always safe` + let mut r = unsafe { l.add(v.len()) }; + let mut block_r = BLOCK; + let mut start_r = ptr::null_mut(); + let mut end_r = ptr::null_mut(); + let mut offsets_r = [MaybeUninit::<u8>::uninit(); BLOCK]; + + // FIXME: When we get VLAs, try creating one array of length `min(v.len(), 2 * BLOCK)` rather + // than two fixed-size arrays of length `BLOCK`. VLAs might be more cache-efficient. + + // Returns the number of elements between pointers `l` (inclusive) and `r` (exclusive). + fn width<T>(l: *mut T, r: *mut T) -> usize { + assert!(mem::size_of::<T>() > 0); + // FIXME: this should *likely* use `offset_from`, but more + // investigation is needed (including running tests in miri). + // TODO unstable: (r.addr() - l.addr()) / mem::size_of::<T>() + (r as usize - l as usize) / mem::size_of::<T>() + } + + loop { + // We are done with partitioning block-by-block when `l` and `r` get very close. Then we do + // some patch-up work in order to partition the remaining elements in between. + let is_done = width(l, r) <= 2 * BLOCK; + + if is_done { + // Number of remaining elements (still not compared to the pivot). + let mut rem = width(l, r); + if start_l < end_l || start_r < end_r { + rem -= BLOCK; + } + + // Adjust block sizes so that the left and right block don't overlap, but get perfectly + // aligned to cover the whole remaining gap. + if start_l < end_l { + block_r = rem; + } else if start_r < end_r { + block_l = rem; + } else { + // There were the same number of elements to switch on both blocks during the last + // iteration, so there are no remaining elements on either block. Cover the remaining + // items with roughly equally-sized blocks. + block_l = rem / 2; + block_r = rem - block_l; + } + debug_assert!(block_l <= BLOCK && block_r <= BLOCK); + debug_assert!(width(l, r) == block_l + block_r); + } + + if start_l == end_l { + // Trace `block_l` elements from the left side. + // TODO unstable: start_l = MaybeUninit::slice_as_mut_ptr(&mut offsets_l); + start_l = offsets_l.as_mut_ptr() as *mut u8; + end_l = start_l; + let mut elem = l; + + for i in 0..block_l { + // SAFETY: The unsafety operations below involve the usage of the `offset`. + // According to the conditions required by the function, we satisfy them because: + // 1. `offsets_l` is stack-allocated, and thus considered separate allocated object. + // 2. The function `is_less` returns a `bool`. + // Casting a `bool` will never overflow `isize`. + // 3. We have guaranteed that `block_l` will be `<= BLOCK`. + // Plus, `end_l` was initially set to the begin pointer of `offsets_` which was declared on the stack. + // Thus, we know that even in the worst case (all invocations of `is_less` returns false) we will only be at most 1 byte pass the end. + // Another unsafety operation here is dereferencing `elem`. + // However, `elem` was initially the begin pointer to the slice which is always valid. + unsafe { + // Branchless comparison. + *end_l = i as u8; + end_l = end_l.offset(!is_less(&*elem, pivot) as isize); + elem = elem.offset(1); + } + } + } + + if start_r == end_r { + // Trace `block_r` elements from the right side. + // TODO unstable: start_r = MaybeUninit::slice_as_mut_ptr(&mut offsets_r); + start_r = offsets_r.as_mut_ptr() as *mut u8; + end_r = start_r; + let mut elem = r; + + for i in 0..block_r { + // SAFETY: The unsafety operations below involve the usage of the `offset`. + // According to the conditions required by the function, we satisfy them because: + // 1. `offsets_r` is stack-allocated, and thus considered separate allocated object. + // 2. The function `is_less` returns a `bool`. + // Casting a `bool` will never overflow `isize`. + // 3. We have guaranteed that `block_r` will be `<= BLOCK`. + // Plus, `end_r` was initially set to the begin pointer of `offsets_` which was declared on the stack. + // Thus, we know that even in the worst case (all invocations of `is_less` returns true) we will only be at most 1 byte pass the end. + // Another unsafety operation here is dereferencing `elem`. + // However, `elem` was initially `1 * sizeof(T)` past the end and we decrement it by `1 * sizeof(T)` before accessing it. + // Plus, `block_r` was asserted to be less than `BLOCK` and `elem` will therefore at most be pointing to the beginning of the slice. + unsafe { + // Branchless comparison. + elem = elem.offset(-1); + *end_r = i as u8; + end_r = end_r.offset(is_less(&*elem, pivot) as isize); + } + } + } + + // Number of out-of-order elements to swap between the left and right side. + let count = cmp::min(width(start_l, end_l), width(start_r, end_r)); + + if count > 0 { + macro_rules! left { + () => { + l.offset(*start_l as isize) + }; + } + macro_rules! right { + () => { + r.offset(-(*start_r as isize) - 1) + }; + } + + // Instead of swapping one pair at the time, it is more efficient to perform a cyclic + // permutation. This is not strictly equivalent to swapping, but produces a similar + // result using fewer memory operations. + + // SAFETY: The use of `ptr::read` is valid because there is at least one element in + // both `offsets_l` and `offsets_r`, so `left!` is a valid pointer to read from. + // + // The uses of `left!` involve calls to `offset` on `l`, which points to the + // beginning of `v`. All the offsets pointed-to by `start_l` are at most `block_l`, so + // these `offset` calls are safe as all reads are within the block. The same argument + // applies for the uses of `right!`. + // + // The calls to `start_l.offset` are valid because there are at most `count-1` of them, + // plus the final one at the end of the unsafe block, where `count` is the minimum number + // of collected offsets in `offsets_l` and `offsets_r`, so there is no risk of there not + // being enough elements. The same reasoning applies to the calls to `start_r.offset`. + // + // The calls to `copy_nonoverlapping` are safe because `left!` and `right!` are guaranteed + // not to overlap, and are valid because of the reasoning above. + unsafe { + let tmp = ptr::read(left!()); + ptr::copy_nonoverlapping(right!(), left!(), 1); + + for _ in 1..count { + start_l = start_l.offset(1); + ptr::copy_nonoverlapping(left!(), right!(), 1); + start_r = start_r.offset(1); + ptr::copy_nonoverlapping(right!(), left!(), 1); + } + + ptr::copy_nonoverlapping(&tmp, right!(), 1); + mem::forget(tmp); + start_l = start_l.offset(1); + start_r = start_r.offset(1); + } + } + + if start_l == end_l { + // All out-of-order elements in the left block were moved. Move to the next block. + + // block-width-guarantee + // SAFETY: if `!is_done` then the slice width is guaranteed to be at least `2*BLOCK` wide. There + // are at most `BLOCK` elements in `offsets_l` because of its size, so the `offset` operation is + // safe. Otherwise, the debug assertions in the `is_done` case guarantee that + // `width(l, r) == block_l + block_r`, namely, that the block sizes have been adjusted to account + // for the smaller number of remaining elements. + l = unsafe { l.add(block_l) }; + } + + if start_r == end_r { + // All out-of-order elements in the right block were moved. Move to the previous block. + + // SAFETY: Same argument as [block-width-guarantee]. Either this is a full block `2*BLOCK`-wide, + // or `block_r` has been adjusted for the last handful of elements. + r = unsafe { r.offset(-(block_r as isize)) }; + } + + if is_done { + break; + } + } + + // All that remains now is at most one block (either the left or the right) with out-of-order + // elements that need to be moved. Such remaining elements can be simply shifted to the end + // within their block. + + if start_l < end_l { + // The left block remains. + // Move its remaining out-of-order elements to the far right. + debug_assert_eq!(width(l, r), block_l); + while start_l < end_l { + // remaining-elements-safety + // SAFETY: while the loop condition holds there are still elements in `offsets_l`, so it + // is safe to point `end_l` to the previous element. + // + // The `ptr::swap` is safe if both its arguments are valid for reads and writes: + // - Per the debug assert above, the distance between `l` and `r` is `block_l` + // elements, so there can be at most `block_l` remaining offsets between `start_l` + // and `end_l`. This means `r` will be moved at most `block_l` steps back, which + // makes the `r.offset` calls valid (at that point `l == r`). + // - `offsets_l` contains valid offsets into `v` collected during the partitioning of + // the last block, so the `l.offset` calls are valid. + unsafe { + end_l = end_l.offset(-1); + ptr::swap(l.offset(*end_l as isize), r.offset(-1)); + r = r.offset(-1); + } + } + width(v.as_mut_ptr(), r) + } else if start_r < end_r { + // The right block remains. + // Move its remaining out-of-order elements to the far left. + debug_assert_eq!(width(l, r), block_r); + while start_r < end_r { + // SAFETY: See the reasoning in [remaining-elements-safety]. + unsafe { + end_r = end_r.offset(-1); + ptr::swap(l, r.offset(-(*end_r as isize) - 1)); + l = l.offset(1); + } + } + width(v.as_mut_ptr(), l) + } else { + // Nothing else to do, we're done. + width(v.as_mut_ptr(), l) + } +} + +/// Partitions `v` into elements smaller than `v[pivot]`, followed by elements greater than or +/// equal to `v[pivot]`. +/// +/// Returns a tuple of: +/// +/// 1. Number of elements smaller than `v[pivot]`. +/// 2. True if `v` was already partitioned. +fn partition<T, F>(v: &mut [T], pivot: usize, is_less: &F) -> (usize, bool) +where + F: Fn(&T, &T) -> bool, +{ + let (mid, was_partitioned) = { + // Place the pivot at the beginning of slice. + v.swap(0, pivot); + let (pivot, v) = v.split_at_mut(1); + let pivot = &mut pivot[0]; + + // Read the pivot into a stack-allocated variable for efficiency. If a following comparison + // operation panics, the pivot will be automatically written back into the slice. + + // SAFETY: `pivot` is a reference to the first element of `v`, so `ptr::read` is safe. + let tmp = mem::ManuallyDrop::new(unsafe { ptr::read(pivot) }); + let _pivot_guard = CopyOnDrop { + src: &*tmp, + dest: pivot, + }; + let pivot = &*tmp; + + // Find the first pair of out-of-order elements. + let mut l = 0; + let mut r = v.len(); + + // SAFETY: The unsafety below involves indexing an array. + // For the first one: We already do the bounds checking here with `l < r`. + // For the second one: We initially have `l == 0` and `r == v.len()` and we checked that `l < r` at every indexing operation. + // From here we know that `r` must be at least `r == l` which was shown to be valid from the first one. + unsafe { + // Find the first element greater than or equal to the pivot. + while l < r && is_less(v.get_unchecked(l), pivot) { + l += 1; + } + + // Find the last element smaller that the pivot. + while l < r && !is_less(v.get_unchecked(r - 1), pivot) { + r -= 1; + } + } + + ( + l + partition_in_blocks(&mut v[l..r], pivot, is_less), + l >= r, + ) + + // `_pivot_guard` goes out of scope and writes the pivot (which is a stack-allocated + // variable) back into the slice where it originally was. This step is critical in ensuring + // safety! + }; + + // Place the pivot between the two partitions. + v.swap(0, mid); + + (mid, was_partitioned) +} + +/// Partitions `v` into elements equal to `v[pivot]` followed by elements greater than `v[pivot]`. +/// +/// Returns the number of elements equal to the pivot. It is assumed that `v` does not contain +/// elements smaller than the pivot. +fn partition_equal<T, F>(v: &mut [T], pivot: usize, is_less: &F) -> usize +where + F: Fn(&T, &T) -> bool, +{ + // Place the pivot at the beginning of slice. + v.swap(0, pivot); + let (pivot, v) = v.split_at_mut(1); + let pivot = &mut pivot[0]; + + // Read the pivot into a stack-allocated variable for efficiency. If a following comparison + // operation panics, the pivot will be automatically written back into the slice. + // SAFETY: The pointer here is valid because it is obtained from a reference to a slice. + let tmp = mem::ManuallyDrop::new(unsafe { ptr::read(pivot) }); + let _pivot_guard = CopyOnDrop { + src: &*tmp, + dest: pivot, + }; + let pivot = &*tmp; + + // Now partition the slice. + let mut l = 0; + let mut r = v.len(); + loop { + // SAFETY: The unsafety below involves indexing an array. + // For the first one: We already do the bounds checking here with `l < r`. + // For the second one: We initially have `l == 0` and `r == v.len()` and we checked that `l < r` at every indexing operation. + // From here we know that `r` must be at least `r == l` which was shown to be valid from the first one. + unsafe { + // Find the first element greater than the pivot. + while l < r && !is_less(pivot, v.get_unchecked(l)) { + l += 1; + } + + // Find the last element equal to the pivot. + while l < r && is_less(pivot, v.get_unchecked(r - 1)) { + r -= 1; + } + + // Are we done? + if l >= r { + break; + } + + // Swap the found pair of out-of-order elements. + r -= 1; + let ptr = v.as_mut_ptr(); + ptr::swap(ptr.add(l), ptr.add(r)); + l += 1; + } + } + + // We found `l` elements equal to the pivot. Add 1 to account for the pivot itself. + l + 1 + + // `_pivot_guard` goes out of scope and writes the pivot (which is a stack-allocated variable) + // back into the slice where it originally was. This step is critical in ensuring safety! +} + +/// Scatters some elements around in an attempt to break patterns that might cause imbalanced +/// partitions in quicksort. +#[cold] +fn break_patterns<T>(v: &mut [T]) { + let len = v.len(); + if len >= 8 { + // Pseudorandom number generator from the "Xorshift RNGs" paper by George Marsaglia. + let mut random = len as u32; + let mut gen_u32 = || { + random ^= random << 13; + random ^= random >> 17; + random ^= random << 5; + random + }; + let mut gen_usize = || { + if usize::BITS <= 32 { + gen_u32() as usize + } else { + (((gen_u32() as u64) << 32) | (gen_u32() as u64)) as usize + } + }; + + // Take random numbers modulo this number. + // The number fits into `usize` because `len` is not greater than `isize::MAX`. + let modulus = len.next_power_of_two(); + + // Some pivot candidates will be in the nearby of this index. Let's randomize them. + let pos = len / 4 * 2; + + for i in 0..3 { + // Generate a random number modulo `len`. However, in order to avoid costly operations + // we first take it modulo a power of two, and then decrease by `len` until it fits + // into the range `[0, len - 1]`. + let mut other = gen_usize() & (modulus - 1); + + // `other` is guaranteed to be less than `2 * len`. + if other >= len { + other -= len; + } + + v.swap(pos - 1 + i, other); + } + } +} + +/// Chooses a pivot in `v` and returns the index and `true` if the slice is likely already sorted. +/// +/// Elements in `v` might be reordered in the process. +fn choose_pivot<T, F>(v: &mut [T], is_less: &F) -> (usize, bool) +where + F: Fn(&T, &T) -> bool, +{ + // Minimum length to choose the median-of-medians method. + // Shorter slices use the simple median-of-three method. + const SHORTEST_MEDIAN_OF_MEDIANS: usize = 50; + // Maximum number of swaps that can be performed in this function. + const MAX_SWAPS: usize = 4 * 3; + + let len = v.len(); + + // Three indices near which we are going to choose a pivot. + #[allow(clippy::identity_op)] + let mut a = len / 4 * 1; + let mut b = len / 4 * 2; + let mut c = len / 4 * 3; + + // Counts the total number of swaps we are about to perform while sorting indices. + let mut swaps = 0; + + if len >= 8 { + // Swaps indices so that `v[a] <= v[b]`. + // SAFETY: `len >= 8` so there are at least two elements in the neighborhoods of + // `a`, `b` and `c`. This means the three calls to `sort_adjacent` result in + // corresponding calls to `sort3` with valid 3-item neighborhoods around each + // pointer, which in turn means the calls to `sort2` are done with valid + // references. Thus the `v.get_unchecked` calls are safe, as is the `ptr::swap` + // call. + let mut sort2 = |a: &mut usize, b: &mut usize| unsafe { + if is_less(v.get_unchecked(*b), v.get_unchecked(*a)) { + ptr::swap(a, b); + swaps += 1; + } + }; + + // Swaps indices so that `v[a] <= v[b] <= v[c]`. + let mut sort3 = |a: &mut usize, b: &mut usize, c: &mut usize| { + sort2(a, b); + sort2(b, c); + sort2(a, b); + }; + + if len >= SHORTEST_MEDIAN_OF_MEDIANS { + // Finds the median of `v[a - 1], v[a], v[a + 1]` and stores the index into `a`. + let mut sort_adjacent = |a: &mut usize| { + let tmp = *a; + sort3(&mut (tmp - 1), a, &mut (tmp + 1)); + }; + + // Find medians in the neighborhoods of `a`, `b`, and `c`. + sort_adjacent(&mut a); + sort_adjacent(&mut b); + sort_adjacent(&mut c); + } + + // Find the median among `a`, `b`, and `c`. + sort3(&mut a, &mut b, &mut c); + } + + if swaps < MAX_SWAPS { + (b, swaps == 0) + } else { + // The maximum number of swaps was performed. Chances are the slice is descending or mostly + // descending, so reversing will probably help sort it faster. + v.reverse(); + (len - 1 - b, true) + } +} + +/// Sorts `v` recursively. +/// +/// If the slice had a predecessor in the original array, it is specified as `pred`. +/// +/// `limit` is the number of allowed imbalanced partitions before switching to `heapsort`. If zero, +/// this function will immediately switch to heapsort. +fn recurse<'a, T, F>(mut v: &'a mut [T], is_less: &F, mut pred: Option<&'a mut T>, mut limit: u32) +where + T: Send, + F: Fn(&T, &T) -> bool + Sync, +{ + // Slices of up to this length get sorted using insertion sort. + const MAX_INSERTION: usize = 20; + // If both partitions are up to this length, we continue sequentially. This number is as small + // as possible but so that the overhead of Rayon's task scheduling is still negligible. + const MAX_SEQUENTIAL: usize = 2000; + + // True if the last partitioning was reasonably balanced. + let mut was_balanced = true; + // True if the last partitioning didn't shuffle elements (the slice was already partitioned). + let mut was_partitioned = true; + + loop { + let len = v.len(); + + // Very short slices get sorted using insertion sort. + if len <= MAX_INSERTION { + insertion_sort(v, is_less); + return; + } + + // If too many bad pivot choices were made, simply fall back to heapsort in order to + // guarantee `O(n * log(n))` worst-case. + if limit == 0 { + heapsort(v, is_less); + return; + } + + // If the last partitioning was imbalanced, try breaking patterns in the slice by shuffling + // some elements around. Hopefully we'll choose a better pivot this time. + if !was_balanced { + break_patterns(v); + limit -= 1; + } + + // Choose a pivot and try guessing whether the slice is already sorted. + let (pivot, likely_sorted) = choose_pivot(v, is_less); + + // If the last partitioning was decently balanced and didn't shuffle elements, and if pivot + // selection predicts the slice is likely already sorted... + if was_balanced && was_partitioned && likely_sorted { + // Try identifying several out-of-order elements and shifting them to correct + // positions. If the slice ends up being completely sorted, we're done. + if partial_insertion_sort(v, is_less) { + return; + } + } + + // If the chosen pivot is equal to the predecessor, then it's the smallest element in the + // slice. Partition the slice into elements equal to and elements greater than the pivot. + // This case is usually hit when the slice contains many duplicate elements. + if let Some(ref p) = pred { + if !is_less(p, &v[pivot]) { + let mid = partition_equal(v, pivot, is_less); + + // Continue sorting elements greater than the pivot. + v = &mut v[mid..]; + continue; + } + } + + // Partition the slice. + let (mid, was_p) = partition(v, pivot, is_less); + was_balanced = cmp::min(mid, len - mid) >= len / 8; + was_partitioned = was_p; + + // Split the slice into `left`, `pivot`, and `right`. + let (left, right) = v.split_at_mut(mid); + let (pivot, right) = right.split_at_mut(1); + let pivot = &mut pivot[0]; + + if cmp::max(left.len(), right.len()) <= MAX_SEQUENTIAL { + // Recurse into the shorter side only in order to minimize the total number of recursive + // calls and consume less stack space. Then just continue with the longer side (this is + // akin to tail recursion). + if left.len() < right.len() { + recurse(left, is_less, pred, limit); + v = right; + pred = Some(pivot); + } else { + recurse(right, is_less, Some(pivot), limit); + v = left; + } + } else { + // Sort the left and right half in parallel. + rayon_core::join( + || recurse(left, is_less, pred, limit), + || recurse(right, is_less, Some(pivot), limit), + ); + break; + } + } +} + +/// Sorts `v` using pattern-defeating quicksort in parallel. +/// +/// The algorithm is unstable, in-place, and *O*(*n* \* log(*n*)) worst-case. +pub(super) fn par_quicksort<T, F>(v: &mut [T], is_less: F) +where + T: Send, + F: Fn(&T, &T) -> bool + Sync, +{ + // Sorting has no meaningful behavior on zero-sized types. + if mem::size_of::<T>() == 0 { + return; + } + + // Limit the number of imbalanced partitions to `floor(log2(len)) + 1`. + let limit = usize::BITS - v.len().leading_zeros(); + + recurse(v, &is_less, None, limit); +} + +#[cfg(test)] +mod tests { + use super::heapsort; + use rand::distributions::Uniform; + use rand::{thread_rng, Rng}; + + #[test] + fn test_heapsort() { + let rng = &mut thread_rng(); + + for len in (0..25).chain(500..501) { + for &modulus in &[5, 10, 100] { + let dist = Uniform::new(0, modulus); + for _ in 0..100 { + let v: Vec<i32> = rng.sample_iter(&dist).take(len).collect(); + + // Test heapsort using `<` operator. + let mut tmp = v.clone(); + heapsort(&mut tmp, &|a, b| a < b); + assert!(tmp.windows(2).all(|w| w[0] <= w[1])); + + // Test heapsort using `>` operator. + let mut tmp = v.clone(); + heapsort(&mut tmp, &|a, b| a > b); + assert!(tmp.windows(2).all(|w| w[0] >= w[1])); + } + } + } + + // Sort using a completely random comparison function. + // This will reorder the elements *somehow*, but won't panic. + let mut v: Vec<_> = (0..100).collect(); + heapsort(&mut v, &|_, _| thread_rng().gen()); + heapsort(&mut v, &|a, b| a < b); + + for (i, &entry) in v.iter().enumerate() { + assert_eq!(entry, i); + } + } +} diff --git a/third_party/rust/rayon/src/slice/rchunks.rs b/third_party/rust/rayon/src/slice/rchunks.rs new file mode 100644 index 0000000000..63e765d77e --- /dev/null +++ b/third_party/rust/rayon/src/slice/rchunks.rs @@ -0,0 +1,386 @@ +use crate::iter::plumbing::*; +use crate::iter::*; +use crate::math::div_round_up; + +/// Parallel iterator over immutable non-overlapping chunks of a slice, starting at the end. +#[derive(Debug)] +pub struct RChunks<'data, T: Sync> { + chunk_size: usize, + slice: &'data [T], +} + +impl<'data, T: Sync> RChunks<'data, T> { + pub(super) fn new(chunk_size: usize, slice: &'data [T]) -> Self { + Self { chunk_size, slice } + } +} + +impl<'data, T: Sync> Clone for RChunks<'data, T> { + fn clone(&self) -> Self { + RChunks { ..*self } + } +} + +impl<'data, T: Sync + 'data> ParallelIterator for RChunks<'data, T> { + type Item = &'data [T]; + + fn drive_unindexed<C>(self, consumer: C) -> C::Result + where + C: UnindexedConsumer<Self::Item>, + { + bridge(self, consumer) + } + + fn opt_len(&self) -> Option<usize> { + Some(self.len()) + } +} + +impl<'data, T: Sync + 'data> IndexedParallelIterator for RChunks<'data, T> { + fn drive<C>(self, consumer: C) -> C::Result + where + C: Consumer<Self::Item>, + { + bridge(self, consumer) + } + + fn len(&self) -> usize { + div_round_up(self.slice.len(), self.chunk_size) + } + + fn with_producer<CB>(self, callback: CB) -> CB::Output + where + CB: ProducerCallback<Self::Item>, + { + callback.callback(RChunksProducer { + chunk_size: self.chunk_size, + slice: self.slice, + }) + } +} + +struct RChunksProducer<'data, T: Sync> { + chunk_size: usize, + slice: &'data [T], +} + +impl<'data, T: 'data + Sync> Producer for RChunksProducer<'data, T> { + type Item = &'data [T]; + type IntoIter = ::std::slice::RChunks<'data, T>; + + fn into_iter(self) -> Self::IntoIter { + self.slice.rchunks(self.chunk_size) + } + + fn split_at(self, index: usize) -> (Self, Self) { + let elem_index = self.slice.len().saturating_sub(index * self.chunk_size); + let (left, right) = self.slice.split_at(elem_index); + ( + RChunksProducer { + chunk_size: self.chunk_size, + slice: right, + }, + RChunksProducer { + chunk_size: self.chunk_size, + slice: left, + }, + ) + } +} + +/// Parallel iterator over immutable non-overlapping chunks of a slice, starting at the end. +#[derive(Debug)] +pub struct RChunksExact<'data, T: Sync> { + chunk_size: usize, + slice: &'data [T], + rem: &'data [T], +} + +impl<'data, T: Sync> RChunksExact<'data, T> { + pub(super) fn new(chunk_size: usize, slice: &'data [T]) -> Self { + let rem_len = slice.len() % chunk_size; + let (rem, slice) = slice.split_at(rem_len); + Self { + chunk_size, + slice, + rem, + } + } + + /// Return the remainder of the original slice that is not going to be + /// returned by the iterator. The returned slice has at most `chunk_size-1` + /// elements. + pub fn remainder(&self) -> &'data [T] { + self.rem + } +} + +impl<'data, T: Sync> Clone for RChunksExact<'data, T> { + fn clone(&self) -> Self { + RChunksExact { ..*self } + } +} + +impl<'data, T: Sync + 'data> ParallelIterator for RChunksExact<'data, T> { + type Item = &'data [T]; + + fn drive_unindexed<C>(self, consumer: C) -> C::Result + where + C: UnindexedConsumer<Self::Item>, + { + bridge(self, consumer) + } + + fn opt_len(&self) -> Option<usize> { + Some(self.len()) + } +} + +impl<'data, T: Sync + 'data> IndexedParallelIterator for RChunksExact<'data, T> { + fn drive<C>(self, consumer: C) -> C::Result + where + C: Consumer<Self::Item>, + { + bridge(self, consumer) + } + + fn len(&self) -> usize { + self.slice.len() / self.chunk_size + } + + fn with_producer<CB>(self, callback: CB) -> CB::Output + where + CB: ProducerCallback<Self::Item>, + { + callback.callback(RChunksExactProducer { + chunk_size: self.chunk_size, + slice: self.slice, + }) + } +} + +struct RChunksExactProducer<'data, T: Sync> { + chunk_size: usize, + slice: &'data [T], +} + +impl<'data, T: 'data + Sync> Producer for RChunksExactProducer<'data, T> { + type Item = &'data [T]; + type IntoIter = ::std::slice::RChunksExact<'data, T>; + + fn into_iter(self) -> Self::IntoIter { + self.slice.rchunks_exact(self.chunk_size) + } + + fn split_at(self, index: usize) -> (Self, Self) { + let elem_index = self.slice.len() - index * self.chunk_size; + let (left, right) = self.slice.split_at(elem_index); + ( + RChunksExactProducer { + chunk_size: self.chunk_size, + slice: right, + }, + RChunksExactProducer { + chunk_size: self.chunk_size, + slice: left, + }, + ) + } +} + +/// Parallel iterator over mutable non-overlapping chunks of a slice, starting at the end. +#[derive(Debug)] +pub struct RChunksMut<'data, T: Send> { + chunk_size: usize, + slice: &'data mut [T], +} + +impl<'data, T: Send> RChunksMut<'data, T> { + pub(super) fn new(chunk_size: usize, slice: &'data mut [T]) -> Self { + Self { chunk_size, slice } + } +} + +impl<'data, T: Send + 'data> ParallelIterator for RChunksMut<'data, T> { + type Item = &'data mut [T]; + + fn drive_unindexed<C>(self, consumer: C) -> C::Result + where + C: UnindexedConsumer<Self::Item>, + { + bridge(self, consumer) + } + + fn opt_len(&self) -> Option<usize> { + Some(self.len()) + } +} + +impl<'data, T: Send + 'data> IndexedParallelIterator for RChunksMut<'data, T> { + fn drive<C>(self, consumer: C) -> C::Result + where + C: Consumer<Self::Item>, + { + bridge(self, consumer) + } + + fn len(&self) -> usize { + div_round_up(self.slice.len(), self.chunk_size) + } + + fn with_producer<CB>(self, callback: CB) -> CB::Output + where + CB: ProducerCallback<Self::Item>, + { + callback.callback(RChunksMutProducer { + chunk_size: self.chunk_size, + slice: self.slice, + }) + } +} + +struct RChunksMutProducer<'data, T: Send> { + chunk_size: usize, + slice: &'data mut [T], +} + +impl<'data, T: 'data + Send> Producer for RChunksMutProducer<'data, T> { + type Item = &'data mut [T]; + type IntoIter = ::std::slice::RChunksMut<'data, T>; + + fn into_iter(self) -> Self::IntoIter { + self.slice.rchunks_mut(self.chunk_size) + } + + fn split_at(self, index: usize) -> (Self, Self) { + let elem_index = self.slice.len().saturating_sub(index * self.chunk_size); + let (left, right) = self.slice.split_at_mut(elem_index); + ( + RChunksMutProducer { + chunk_size: self.chunk_size, + slice: right, + }, + RChunksMutProducer { + chunk_size: self.chunk_size, + slice: left, + }, + ) + } +} + +/// Parallel iterator over mutable non-overlapping chunks of a slice, starting at the end. +#[derive(Debug)] +pub struct RChunksExactMut<'data, T: Send> { + chunk_size: usize, + slice: &'data mut [T], + rem: &'data mut [T], +} + +impl<'data, T: Send> RChunksExactMut<'data, T> { + pub(super) fn new(chunk_size: usize, slice: &'data mut [T]) -> Self { + let rem_len = slice.len() % chunk_size; + let (rem, slice) = slice.split_at_mut(rem_len); + Self { + chunk_size, + slice, + rem, + } + } + + /// Return the remainder of the original slice that is not going to be + /// returned by the iterator. The returned slice has at most `chunk_size-1` + /// elements. + /// + /// Note that this has to consume `self` to return the original lifetime of + /// the data, which prevents this from actually being used as a parallel + /// iterator since that also consumes. This method is provided for parity + /// with `std::iter::RChunksExactMut`, but consider calling `remainder()` or + /// `take_remainder()` as alternatives. + pub fn into_remainder(self) -> &'data mut [T] { + self.rem + } + + /// Return the remainder of the original slice that is not going to be + /// returned by the iterator. The returned slice has at most `chunk_size-1` + /// elements. + /// + /// Consider `take_remainder()` if you need access to the data with its + /// original lifetime, rather than borrowing through `&mut self` here. + pub fn remainder(&mut self) -> &mut [T] { + self.rem + } + + /// Return the remainder of the original slice that is not going to be + /// returned by the iterator. The returned slice has at most `chunk_size-1` + /// elements. Subsequent calls will return an empty slice. + pub fn take_remainder(&mut self) -> &'data mut [T] { + std::mem::take(&mut self.rem) + } +} + +impl<'data, T: Send + 'data> ParallelIterator for RChunksExactMut<'data, T> { + type Item = &'data mut [T]; + + fn drive_unindexed<C>(self, consumer: C) -> C::Result + where + C: UnindexedConsumer<Self::Item>, + { + bridge(self, consumer) + } + + fn opt_len(&self) -> Option<usize> { + Some(self.len()) + } +} + +impl<'data, T: Send + 'data> IndexedParallelIterator for RChunksExactMut<'data, T> { + fn drive<C>(self, consumer: C) -> C::Result + where + C: Consumer<Self::Item>, + { + bridge(self, consumer) + } + + fn len(&self) -> usize { + self.slice.len() / self.chunk_size + } + + fn with_producer<CB>(self, callback: CB) -> CB::Output + where + CB: ProducerCallback<Self::Item>, + { + callback.callback(RChunksExactMutProducer { + chunk_size: self.chunk_size, + slice: self.slice, + }) + } +} + +struct RChunksExactMutProducer<'data, T: Send> { + chunk_size: usize, + slice: &'data mut [T], +} + +impl<'data, T: 'data + Send> Producer for RChunksExactMutProducer<'data, T> { + type Item = &'data mut [T]; + type IntoIter = ::std::slice::RChunksExactMut<'data, T>; + + fn into_iter(self) -> Self::IntoIter { + self.slice.rchunks_exact_mut(self.chunk_size) + } + + fn split_at(self, index: usize) -> (Self, Self) { + let elem_index = self.slice.len() - index * self.chunk_size; + let (left, right) = self.slice.split_at_mut(elem_index); + ( + RChunksExactMutProducer { + chunk_size: self.chunk_size, + slice: right, + }, + RChunksExactMutProducer { + chunk_size: self.chunk_size, + slice: left, + }, + ) + } +} diff --git a/third_party/rust/rayon/src/slice/test.rs b/third_party/rust/rayon/src/slice/test.rs new file mode 100644 index 0000000000..f74ca0f743 --- /dev/null +++ b/third_party/rust/rayon/src/slice/test.rs @@ -0,0 +1,170 @@ +#![cfg(test)] + +use crate::prelude::*; +use rand::distributions::Uniform; +use rand::seq::SliceRandom; +use rand::{thread_rng, Rng}; +use std::cmp::Ordering::{Equal, Greater, Less}; + +macro_rules! sort { + ($f:ident, $name:ident) => { + #[test] + fn $name() { + let rng = &mut thread_rng(); + + for len in (0..25).chain(500..501) { + for &modulus in &[5, 10, 100] { + let dist = Uniform::new(0, modulus); + for _ in 0..100 { + let v: Vec<i32> = rng.sample_iter(&dist).take(len).collect(); + + // Test sort using `<` operator. + let mut tmp = v.clone(); + tmp.$f(|a, b| a.cmp(b)); + assert!(tmp.windows(2).all(|w| w[0] <= w[1])); + + // Test sort using `>` operator. + let mut tmp = v.clone(); + tmp.$f(|a, b| b.cmp(a)); + assert!(tmp.windows(2).all(|w| w[0] >= w[1])); + } + } + } + + // Test sort with many duplicates. + for &len in &[1_000, 10_000, 100_000] { + for &modulus in &[5, 10, 100, 10_000] { + let dist = Uniform::new(0, modulus); + let mut v: Vec<i32> = rng.sample_iter(&dist).take(len).collect(); + + v.$f(|a, b| a.cmp(b)); + assert!(v.windows(2).all(|w| w[0] <= w[1])); + } + } + + // Test sort with many pre-sorted runs. + for &len in &[1_000, 10_000, 100_000] { + let len_dist = Uniform::new(0, len); + for &modulus in &[5, 10, 1000, 50_000] { + let dist = Uniform::new(0, modulus); + let mut v: Vec<i32> = rng.sample_iter(&dist).take(len).collect(); + + v.sort(); + v.reverse(); + + for _ in 0..5 { + let a = rng.sample(&len_dist); + let b = rng.sample(&len_dist); + if a < b { + v[a..b].reverse(); + } else { + v.swap(a, b); + } + } + + v.$f(|a, b| a.cmp(b)); + assert!(v.windows(2).all(|w| w[0] <= w[1])); + } + } + + // Sort using a completely random comparison function. + // This will reorder the elements *somehow*, but won't panic. + let mut v: Vec<_> = (0..100).collect(); + v.$f(|_, _| *[Less, Equal, Greater].choose(&mut thread_rng()).unwrap()); + v.$f(|a, b| a.cmp(b)); + for i in 0..v.len() { + assert_eq!(v[i], i); + } + + // Should not panic. + [0i32; 0].$f(|a, b| a.cmp(b)); + [(); 10].$f(|a, b| a.cmp(b)); + [(); 100].$f(|a, b| a.cmp(b)); + + let mut v = [0xDEAD_BEEFu64]; + v.$f(|a, b| a.cmp(b)); + assert!(v == [0xDEAD_BEEF]); + } + }; +} + +sort!(par_sort_by, test_par_sort); +sort!(par_sort_unstable_by, test_par_sort_unstable); + +#[test] +fn test_par_sort_stability() { + for len in (2..25).chain(500..510).chain(50_000..50_010) { + for _ in 0..10 { + let mut counts = [0; 10]; + + // Create a vector like [(6, 1), (5, 1), (6, 2), ...], + // where the first item of each tuple is random, but + // the second item represents which occurrence of that + // number this element is, i.e. the second elements + // will occur in sorted order. + let mut rng = thread_rng(); + let mut v: Vec<_> = (0..len) + .map(|_| { + let n: usize = rng.gen_range(0..10); + counts[n] += 1; + (n, counts[n]) + }) + .collect(); + + // Only sort on the first element, so an unstable sort + // may mix up the counts. + v.par_sort_by(|&(a, _), &(b, _)| a.cmp(&b)); + + // This comparison includes the count (the second item + // of the tuple), so elements with equal first items + // will need to be ordered with increasing + // counts... i.e. exactly asserting that this sort is + // stable. + assert!(v.windows(2).all(|w| w[0] <= w[1])); + } + } +} + +#[test] +fn test_par_chunks_exact_remainder() { + let v: &[i32] = &[0, 1, 2, 3, 4]; + let c = v.par_chunks_exact(2); + assert_eq!(c.remainder(), &[4]); + assert_eq!(c.len(), 2); +} + +#[test] +fn test_par_chunks_exact_mut_remainder() { + let v: &mut [i32] = &mut [0, 1, 2, 3, 4]; + let mut c = v.par_chunks_exact_mut(2); + assert_eq!(c.remainder(), &[4]); + assert_eq!(c.len(), 2); + assert_eq!(c.into_remainder(), &[4]); + + let mut c = v.par_chunks_exact_mut(2); + assert_eq!(c.take_remainder(), &[4]); + assert_eq!(c.take_remainder(), &[]); + assert_eq!(c.len(), 2); +} + +#[test] +fn test_par_rchunks_exact_remainder() { + let v: &[i32] = &[0, 1, 2, 3, 4]; + let c = v.par_rchunks_exact(2); + assert_eq!(c.remainder(), &[0]); + assert_eq!(c.len(), 2); +} + +#[test] +fn test_par_rchunks_exact_mut_remainder() { + let v: &mut [i32] = &mut [0, 1, 2, 3, 4]; + let mut c = v.par_rchunks_exact_mut(2); + assert_eq!(c.remainder(), &[0]); + assert_eq!(c.len(), 2); + assert_eq!(c.into_remainder(), &[0]); + + let mut c = v.par_rchunks_exact_mut(2); + assert_eq!(c.take_remainder(), &[0]); + assert_eq!(c.take_remainder(), &[]); + assert_eq!(c.len(), 2); +} |