| 1 |
//! Parallel quicksort. |
| 2 |
//! |
| 3 |
//! This implementation is copied verbatim from `std::slice::sort_unstable` and then parallelized. |
| 4 |
//! The only difference from the original is that calls to `recurse` are executed in parallel using |
| 5 |
//! `rayon_core::join`. |
| 6 |
|
| 7 |
use std::marker::PhantomData; |
| 8 |
use std::mem::{self, MaybeUninit}; |
| 9 |
use std::ptr; |
| 10 |
|
| 11 |
/// When dropped, copies from `src` into `dest`. |
| 12 |
#[must_use] |
| 13 |
struct CopyOnDrop<'a, T> { |
| 14 |
src: *const T, |
| 15 |
dest: *mut T, |
| 16 |
/// `src` is often a local pointer here, make sure we have appropriate |
| 17 |
/// PhantomData so that dropck can protect us. |
| 18 |
marker: PhantomData<&'a mut T>, |
| 19 |
} |
| 20 |
|
| 21 |
impl<'a, T> CopyOnDrop<'a, T> { |
| 22 |
/// Construct from a source pointer and a destination |
| 23 |
/// Assumes dest lives longer than src, since there is no easy way to |
| 24 |
/// copy down lifetime information from another pointer |
| 25 |
unsafe fn new(src: &'a T, dest: *mut T) -> Self { |
| 26 |
CopyOnDrop { |
| 27 |
src, |
| 28 |
dest, |
| 29 |
marker: PhantomData, |
| 30 |
} |
| 31 |
} |
| 32 |
} |
| 33 |
|
| 34 |
impl<T> Drop for CopyOnDrop<'_, T> { |
| 35 |
fn drop(&mut self) { |
| 36 |
// SAFETY: This is a helper class. |
| 37 |
// Please refer to its usage for correctness. |
| 38 |
// Namely, one must be sure that `src` and `dst` does not overlap as required by `ptr::copy_nonoverlapping`. |
| 39 |
unsafe { |
| 40 |
ptr::copy_nonoverlapping(self.src, self.dest, 1); |
| 41 |
} |
| 42 |
} |
| 43 |
} |
| 44 |
|
| 45 |
/// Shifts the first element to the right until it encounters a greater or equal element. |
| 46 |
fn shift_head<T, F>(v: &mut [T], is_less: &F) |
| 47 |
where |
| 48 |
F: Fn(&T, &T) -> bool, |
| 49 |
{ |
| 50 |
let len = v.len(); |
| 51 |
// SAFETY: The unsafe operations below involves indexing without a bounds check (by offsetting a |
| 52 |
// pointer) and copying memory (`ptr::copy_nonoverlapping`). |
| 53 |
// |
| 54 |
// a. Indexing: |
| 55 |
// 1. We checked the size of the array to >=2. |
| 56 |
// 2. All the indexing that we will do is always between {0 <= index < len} at most. |
| 57 |
// |
| 58 |
// b. Memory copying |
| 59 |
// 1. We are obtaining pointers to references which are guaranteed to be valid. |
| 60 |
// 2. They cannot overlap because we obtain pointers to difference indices of the slice. |
| 61 |
// Namely, `i` and `i-1`. |
| 62 |
// 3. If the slice is properly aligned, the elements are properly aligned. |
| 63 |
// It is the caller's responsibility to make sure the slice is properly aligned. |
| 64 |
// |
| 65 |
// See comments below for further detail. |
| 66 |
unsafe { |
| 67 |
// If the first two elements are out-of-order... |
| 68 |
if len >= 2 && is_less(v.get_unchecked(1), v.get_unchecked(0)) { |
| 69 |
// Read the first element into a stack-allocated variable. If a following comparison |
| 70 |
// operation panics, `hole` will get dropped and automatically write the element back |
| 71 |
// into the slice. |
| 72 |
let tmp = mem::ManuallyDrop::new(ptr::read(v.get_unchecked(0))); |
| 73 |
let v = v.as_mut_ptr(); |
| 74 |
let mut hole = CopyOnDrop::new(&*tmp, v.add(1)); |
| 75 |
ptr::copy_nonoverlapping(v.add(1), v.add(0), 1); |
| 76 |
|
| 77 |
for i in 2..len { |
| 78 |
if !is_less(&*v.add(i), &*tmp) { |
| 79 |
break; |
| 80 |
} |
| 81 |
|
| 82 |
// Move `i`-th element one place to the left, thus shifting the hole to the right. |
| 83 |
ptr::copy_nonoverlapping(v.add(i), v.add(i - 1), 1); |
| 84 |
hole.dest = v.add(i); |
| 85 |
} |
| 86 |
// `hole` gets dropped and thus copies `tmp` into the remaining hole in `v`. |
| 87 |
} |
| 88 |
} |
| 89 |
} |
| 90 |
|
| 91 |
/// Shifts the last element to the left until it encounters a smaller or equal element. |
| 92 |
fn shift_tail<T, F>(v: &mut [T], is_less: &F) |
| 93 |
where |
| 94 |
F: Fn(&T, &T) -> bool, |
| 95 |
{ |
| 96 |
let len = v.len(); |
| 97 |
// SAFETY: The unsafe operations below involves indexing without a bound check (by offsetting a |
| 98 |
// pointer) and copying memory (`ptr::copy_nonoverlapping`). |
| 99 |
// |
| 100 |
// a. Indexing: |
| 101 |
// 1. We checked the size of the array to >= 2. |
| 102 |
// 2. All the indexing that we will do is always between `0 <= index < len-1` at most. |
| 103 |
// |
| 104 |
// b. Memory copying |
| 105 |
// 1. We are obtaining pointers to references which are guaranteed to be valid. |
| 106 |
// 2. They cannot overlap because we obtain pointers to difference indices of the slice. |
| 107 |
// Namely, `i` and `i+1`. |
| 108 |
// 3. If the slice is properly aligned, the elements are properly aligned. |
| 109 |
// It is the caller's responsibility to make sure the slice is properly aligned. |
| 110 |
// |
| 111 |
// See comments below for further detail. |
| 112 |
unsafe { |
| 113 |
// If the last two elements are out-of-order... |
| 114 |
if len >= 2 && is_less(v.get_unchecked(len - 1), v.get_unchecked(len - 2)) { |
| 115 |
// Read the last element into a stack-allocated variable. If a following comparison |
| 116 |
// operation panics, `hole` will get dropped and automatically write the element back |
| 117 |
// into the slice. |
| 118 |
let tmp = mem::ManuallyDrop::new(ptr::read(v.get_unchecked(len - 1))); |
| 119 |
let v = v.as_mut_ptr(); |
| 120 |
let mut hole = CopyOnDrop::new(&*tmp, v.add(len - 2)); |
| 121 |
ptr::copy_nonoverlapping(v.add(len - 2), v.add(len - 1), 1); |
| 122 |
|
| 123 |
for i in (0..len - 2).rev() { |
| 124 |
if !is_less(&*tmp, &*v.add(i)) { |
| 125 |
break; |
| 126 |
} |
| 127 |
|
| 128 |
// Move `i`-th element one place to the right, thus shifting the hole to the left. |
| 129 |
ptr::copy_nonoverlapping(v.add(i), v.add(i + 1), 1); |
| 130 |
hole.dest = v.add(i); |
| 131 |
} |
| 132 |
// `hole` gets dropped and thus copies `tmp` into the remaining hole in `v`. |
| 133 |
} |
| 134 |
} |
| 135 |
} |
| 136 |
|
| 137 |
/// Partially sorts a slice by shifting several out-of-order elements around. |
| 138 |
/// |
| 139 |
/// Returns `true` if the slice is sorted at the end. This function is *O*(*n*) worst-case. |
| 140 |
#[cold] |
| 141 |
fn partial_insertion_sort<T, F>(v: &mut [T], is_less: &F) -> bool |
| 142 |
where |
| 143 |
F: Fn(&T, &T) -> bool, |
| 144 |
{ |
| 145 |
// Maximum number of adjacent out-of-order pairs that will get shifted. |
| 146 |
const MAX_STEPS: usize = 5; |
| 147 |
// If the slice is shorter than this, don't shift any elements. |
| 148 |
const SHORTEST_SHIFTING: usize = 50; |
| 149 |
|
| 150 |
let len = v.len(); |
| 151 |
let mut i = 1; |
| 152 |
|
| 153 |
for _ in 0..MAX_STEPS { |
| 154 |
// SAFETY: We already explicitly did the bound checking with `i < len`. |
| 155 |
// All our subsequent indexing is only in the range `0 <= index < len` |
| 156 |
unsafe { |
| 157 |
// Find the next pair of adjacent out-of-order elements. |
| 158 |
while i < len && !is_less(v.get_unchecked(i), v.get_unchecked(i - 1)) { |
| 159 |
i += 1; |
| 160 |
} |
| 161 |
} |
| 162 |
|
| 163 |
// Are we done? |
| 164 |
if i == len { |
| 165 |
return true; |
| 166 |
} |
| 167 |
|
| 168 |
// Don't shift elements on short arrays, that has a performance cost. |
| 169 |
if len < SHORTEST_SHIFTING { |
| 170 |
return false; |
| 171 |
} |
| 172 |
|
| 173 |
// Swap the found pair of elements. This puts them in correct order. |
| 174 |
v.swap(i - 1, i); |
| 175 |
|
| 176 |
// Shift the smaller element to the left. |
| 177 |
shift_tail(&mut v[..i], is_less); |
| 178 |
// Shift the greater element to the right. |
| 179 |
shift_head(&mut v[i..], is_less); |
| 180 |
} |
| 181 |
|
| 182 |
// Didn't manage to sort the slice in the limited number of steps. |
| 183 |
false |
| 184 |
} |
| 185 |
|
| 186 |
/// Sorts a slice using insertion sort, which is *O*(*n*^2) worst-case. |
| 187 |
fn insertion_sort<T, F>(v: &mut [T], is_less: &F) |
| 188 |
where |
| 189 |
F: Fn(&T, &T) -> bool, |
| 190 |
{ |
| 191 |
for i in 1..v.len() { |
| 192 |
shift_tail(&mut v[..i + 1], is_less); |
| 193 |
} |
| 194 |
} |
| 195 |
|
| 196 |
/// Sorts `v` using heapsort, which guarantees *O*(*n* \* log(*n*)) worst-case. |
| 197 |
#[cold] |
| 198 |
fn heapsort<T, F>(v: &mut [T], is_less: &F) |
| 199 |
where |
| 200 |
F: Fn(&T, &T) -> bool, |
| 201 |
{ |
| 202 |
// This binary heap respects the invariant `parent >= child`. |
| 203 |
let sift_down = |v: &mut [T], mut node| { |
| 204 |
loop { |
| 205 |
// Children of `node`. |
| 206 |
let mut child = 2 * node + 1; |
| 207 |
if child >= v.len() { |
| 208 |
break; |
| 209 |
} |
| 210 |
|
| 211 |
// Choose the greater child. |
| 212 |
if child + 1 < v.len() && is_less(&v[child], &v[child + 1]) { |
| 213 |
child += 1; |
| 214 |
} |
| 215 |
|
| 216 |
// Stop if the invariant holds at `node`. |
| 217 |
if !is_less(&v[node], &v[child]) { |
| 218 |
break; |
| 219 |
} |
| 220 |
|
| 221 |
// Swap `node` with the greater child, move one step down, and continue sifting. |
| 222 |
v.swap(node, child); |
| 223 |
node = child; |
| 224 |
} |
| 225 |
}; |
| 226 |
|
| 227 |
// Build the heap in linear time. |
| 228 |
for i in (0..v.len() / 2).rev() { |
| 229 |
sift_down(v, i); |
| 230 |
} |
| 231 |
|
| 232 |
// Pop maximal elements from the heap. |
| 233 |
for i in (1..v.len()).rev() { |
| 234 |
v.swap(0, i); |
| 235 |
sift_down(&mut v[..i], 0); |
| 236 |
} |
| 237 |
} |
| 238 |
|
| 239 |
/// Partitions `v` into elements smaller than `pivot`, followed by elements greater than or equal |
| 240 |
/// to `pivot`. |
| 241 |
/// |
| 242 |
/// Returns the number of elements smaller than `pivot`. |
| 243 |
/// |
| 244 |
/// Partitioning is performed block-by-block in order to minimize the cost of branching operations. |
| 245 |
/// This idea is presented in the [BlockQuicksort][pdf] paper. |
| 246 |
/// |
| 247 |
/// [pdf]: https://drops.dagstuhl.de/opus/volltexte/2016/6389/pdf/LIPIcs-ESA-2016-38.pdf |
| 248 |
fn partition_in_blocks<T, F>(v: &mut [T], pivot: &T, is_less: &F) -> usize |
| 249 |
where |
| 250 |
F: Fn(&T, &T) -> bool, |
| 251 |
{ |
| 252 |
// Number of elements in a typical block. |
| 253 |
const BLOCK: usize = 128; |
| 254 |
|
| 255 |
// The partitioning algorithm repeats the following steps until completion: |
| 256 |
// |
| 257 |
// 1. Trace a block from the left side to identify elements greater than or equal to the pivot. |
| 258 |
// 2. Trace a block from the right side to identify elements smaller than the pivot. |
| 259 |
// 3. Exchange the identified elements between the left and right side. |
| 260 |
// |
| 261 |
// We keep the following variables for a block of elements: |
| 262 |
// |
| 263 |
// 1. `block` - Number of elements in the block. |
| 264 |
// 2. `start` - Start pointer into the `offsets` array. |
| 265 |
// 3. `end` - End pointer into the `offsets` array. |
| 266 |
// 4. `offsets - Indices of out-of-order elements within the block. |
| 267 |
|
| 268 |
// The current block on the left side (from `l` to `l.add(block_l)`). |
| 269 |
let mut l = v.as_mut_ptr(); |
| 270 |
let mut block_l = BLOCK; |
| 271 |
let mut start_l = ptr::null_mut(); |
| 272 |
let mut end_l = ptr::null_mut(); |
| 273 |
let mut offsets_l = [MaybeUninit::<u8>::uninit(); BLOCK]; |
| 274 |
|
| 275 |
// The current block on the right side (from `r.sub(block_r)` to `r`). |
| 276 |
// SAFETY: The documentation for .add() specifically mention that `vec.as_ptr().add(vec.len())` is always safe` |
| 277 |
let mut r = unsafe { l.add(v.len()) }; |
| 278 |
let mut block_r = BLOCK; |
| 279 |
let mut start_r = ptr::null_mut(); |
| 280 |
let mut end_r = ptr::null_mut(); |
| 281 |
let mut offsets_r = [MaybeUninit::<u8>::uninit(); BLOCK]; |
| 282 |
|
| 283 |
// FIXME: When we get VLAs, try creating one array of length `min(v.len(), 2 * BLOCK)` rather |
| 284 |
// than two fixed-size arrays of length `BLOCK`. VLAs might be more cache-efficient. |
| 285 |
|
| 286 |
// Returns the number of elements between pointers `l` (inclusive) and `r` (exclusive). |
| 287 |
fn width<T>(l: *mut T, r: *mut T) -> usize { |
| 288 |
assert!(mem::size_of::<T>() > 0); |
| 289 |
// FIXME: this should *likely* use `offset_from`, but more |
| 290 |
// investigation is needed (including running tests in miri). |
| 291 |
// TODO unstable: (r.addr() - l.addr()) / mem::size_of::<T>() |
| 292 |
(r as usize - l as usize) / mem::size_of::<T>() |
| 293 |
} |
| 294 |
|
| 295 |
loop { |
| 296 |
// We are done with partitioning block-by-block when `l` and `r` get very close. Then we do |
| 297 |
// some patch-up work in order to partition the remaining elements in between. |
| 298 |
let is_done = width(l, r) <= 2 * BLOCK; |
| 299 |
|
| 300 |
if is_done { |
| 301 |
// Number of remaining elements (still not compared to the pivot). |
| 302 |
let mut rem = width(l, r); |
| 303 |
if start_l < end_l || start_r < end_r { |
| 304 |
rem -= BLOCK; |
| 305 |
} |
| 306 |
|
| 307 |
// Adjust block sizes so that the left and right block don't overlap, but get perfectly |
| 308 |
// aligned to cover the whole remaining gap. |
| 309 |
if start_l < end_l { |
| 310 |
block_r = rem; |
| 311 |
} else if start_r < end_r { |
| 312 |
block_l = rem; |
| 313 |
} else { |
| 314 |
// There were the same number of elements to switch on both blocks during the last |
| 315 |
// iteration, so there are no remaining elements on either block. Cover the remaining |
| 316 |
// items with roughly equally-sized blocks. |
| 317 |
block_l = rem / 2; |
| 318 |
block_r = rem - block_l; |
| 319 |
} |
| 320 |
debug_assert!(block_l <= BLOCK && block_r <= BLOCK); |
| 321 |
debug_assert!(width(l, r) == block_l + block_r); |
| 322 |
} |
| 323 |
|
| 324 |
if start_l == end_l { |
| 325 |
// Trace `block_l` elements from the left side. |
| 326 |
// TODO unstable: start_l = MaybeUninit::slice_as_mut_ptr(&mut offsets_l); |
| 327 |
start_l = offsets_l.as_mut_ptr() as *mut u8; |
| 328 |
end_l = start_l; |
| 329 |
let mut elem = l; |
| 330 |
|
| 331 |
for i in 0..block_l { |
| 332 |
// SAFETY: The unsafety operations below involve the usage of the `offset`. |
| 333 |
// According to the conditions required by the function, we satisfy them because: |
| 334 |
// 1. `offsets_l` is stack-allocated, and thus considered separate allocated object. |
| 335 |
// 2. The function `is_less` returns a `bool`. |
| 336 |
// Casting a `bool` will never overflow `isize`. |
| 337 |
// 3. We have guaranteed that `block_l` will be `<= BLOCK`. |
| 338 |
// Plus, `end_l` was initially set to the begin pointer of `offsets_` which was declared on the stack. |
| 339 |
// 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. |
| 340 |
// Another unsafety operation here is dereferencing `elem`. |
| 341 |
// However, `elem` was initially the begin pointer to the slice which is always valid. |
| 342 |
unsafe { |
| 343 |
// Branchless comparison. |
| 344 |
*end_l = i as u8; |
| 345 |
end_l = end_l.offset(!is_less(&*elem, pivot) as isize); |
| 346 |
elem = elem.offset(1); |
| 347 |
} |
| 348 |
} |
| 349 |
} |
| 350 |
|
| 351 |
if start_r == end_r { |
| 352 |
// Trace `block_r` elements from the right side. |
| 353 |
// TODO unstable: start_r = MaybeUninit::slice_as_mut_ptr(&mut offsets_r); |
| 354 |
start_r = offsets_r.as_mut_ptr() as *mut u8; |
| 355 |
end_r = start_r; |
| 356 |
let mut elem = r; |
| 357 |
|
| 358 |
for i in 0..block_r { |
| 359 |
// SAFETY: The unsafety operations below involve the usage of the `offset`. |
| 360 |
// According to the conditions required by the function, we satisfy them because: |
| 361 |
// 1. `offsets_r` is stack-allocated, and thus considered separate allocated object. |
| 362 |
// 2. The function `is_less` returns a `bool`. |
| 363 |
// Casting a `bool` will never overflow `isize`. |
| 364 |
// 3. We have guaranteed that `block_r` will be `<= BLOCK`. |
| 365 |
// Plus, `end_r` was initially set to the begin pointer of `offsets_` which was declared on the stack. |
| 366 |
// 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. |
| 367 |
// Another unsafety operation here is dereferencing `elem`. |
| 368 |
// However, `elem` was initially `1 * sizeof(T)` past the end and we decrement it by `1 * sizeof(T)` before accessing it. |
| 369 |
// Plus, `block_r` was asserted to be less than `BLOCK` and `elem` will therefore at most be pointing to the beginning of the slice. |
| 370 |
unsafe { |
| 371 |
// Branchless comparison. |
| 372 |
elem = elem.offset(-1); |
| 373 |
*end_r = i as u8; |
| 374 |
end_r = end_r.offset(is_less(&*elem, pivot) as isize); |
| 375 |
} |
| 376 |
} |
| 377 |
} |
| 378 |
|
| 379 |
// Number of out-of-order elements to swap between the left and right side. |
| 380 |
let count = Ord::min(width(start_l, end_l), width(start_r, end_r)); |
| 381 |
|
| 382 |
if count > 0 { |
| 383 |
macro_rules! left { |
| 384 |
() => { |
| 385 |
l.offset(*start_l as isize) |
| 386 |
}; |
| 387 |
} |
| 388 |
macro_rules! right { |
| 389 |
() => { |
| 390 |
r.offset(-(*start_r as isize) - 1) |
| 391 |
}; |
| 392 |
} |
| 393 |
|
| 394 |
// Instead of swapping one pair at the time, it is more efficient to perform a cyclic |
| 395 |
// permutation. This is not strictly equivalent to swapping, but produces a similar |
| 396 |
// result using fewer memory operations. |
| 397 |
|
| 398 |
// SAFETY: The use of `ptr::read` is valid because there is at least one element in |
| 399 |
// both `offsets_l` and `offsets_r`, so `left!` is a valid pointer to read from. |
| 400 |
// |
| 401 |
// The uses of `left!` involve calls to `offset` on `l`, which points to the |
| 402 |
// beginning of `v`. All the offsets pointed-to by `start_l` are at most `block_l`, so |
| 403 |
// these `offset` calls are safe as all reads are within the block. The same argument |
| 404 |
// applies for the uses of `right!`. |
| 405 |
// |
| 406 |
// The calls to `start_l.offset` are valid because there are at most `count-1` of them, |
| 407 |
// plus the final one at the end of the unsafe block, where `count` is the minimum number |
| 408 |
// of collected offsets in `offsets_l` and `offsets_r`, so there is no risk of there not |
| 409 |
// being enough elements. The same reasoning applies to the calls to `start_r.offset`. |
| 410 |
// |
| 411 |
// The calls to `copy_nonoverlapping` are safe because `left!` and `right!` are guaranteed |
| 412 |
// not to overlap, and are valid because of the reasoning above. |
| 413 |
unsafe { |
| 414 |
let tmp = ptr::read(left!()); |
| 415 |
ptr::copy_nonoverlapping(right!(), left!(), 1); |
| 416 |
|
| 417 |
for _ in 1..count { |
| 418 |
start_l = start_l.offset(1); |
| 419 |
ptr::copy_nonoverlapping(left!(), right!(), 1); |
| 420 |
start_r = start_r.offset(1); |
| 421 |
ptr::copy_nonoverlapping(right!(), left!(), 1); |
| 422 |
} |
| 423 |
|
| 424 |
ptr::copy_nonoverlapping(&tmp, right!(), 1); |
| 425 |
mem::forget(tmp); |
| 426 |
start_l = start_l.offset(1); |
| 427 |
start_r = start_r.offset(1); |
| 428 |
} |
| 429 |
} |
| 430 |
|
| 431 |
if start_l == end_l { |
| 432 |
// All out-of-order elements in the left block were moved. Move to the next block. |
| 433 |
|
| 434 |
// block-width-guarantee |
| 435 |
// SAFETY: if `!is_done` then the slice width is guaranteed to be at least `2*BLOCK` wide. There |
| 436 |
// are at most `BLOCK` elements in `offsets_l` because of its size, so the `offset` operation is |
| 437 |
// safe. Otherwise, the debug assertions in the `is_done` case guarantee that |
| 438 |
// `width(l, r) == block_l + block_r`, namely, that the block sizes have been adjusted to account |
| 439 |
// for the smaller number of remaining elements. |
| 440 |
l = unsafe { l.add(block_l) }; |
| 441 |
} |
| 442 |
|
| 443 |
if start_r == end_r { |
| 444 |
// All out-of-order elements in the right block were moved. Move to the previous block. |
| 445 |
|
| 446 |
// SAFETY: Same argument as [block-width-guarantee]. Either this is a full block `2*BLOCK`-wide, |
| 447 |
// or `block_r` has been adjusted for the last handful of elements. |
| 448 |
r = unsafe { r.offset(-(block_r as isize)) }; |
| 449 |
} |
| 450 |
|
| 451 |
if is_done { |
| 452 |
break; |
| 453 |
} |
| 454 |
} |
| 455 |
|
| 456 |
// All that remains now is at most one block (either the left or the right) with out-of-order |
| 457 |
// elements that need to be moved. Such remaining elements can be simply shifted to the end |
| 458 |
// within their block. |
| 459 |
|
| 460 |
if start_l < end_l { |
| 461 |
// The left block remains. |
| 462 |
// Move its remaining out-of-order elements to the far right. |
| 463 |
debug_assert_eq!(width(l, r), block_l); |
| 464 |
while start_l < end_l { |
| 465 |
// remaining-elements-safety |
| 466 |
// SAFETY: while the loop condition holds there are still elements in `offsets_l`, so it |
| 467 |
// is safe to point `end_l` to the previous element. |
| 468 |
// |
| 469 |
// The `ptr::swap` is safe if both its arguments are valid for reads and writes: |
| 470 |
// - Per the debug assert above, the distance between `l` and `r` is `block_l` |
| 471 |
// elements, so there can be at most `block_l` remaining offsets between `start_l` |
| 472 |
// and `end_l`. This means `r` will be moved at most `block_l` steps back, which |
| 473 |
// makes the `r.offset` calls valid (at that point `l == r`). |
| 474 |
// - `offsets_l` contains valid offsets into `v` collected during the partitioning of |
| 475 |
// the last block, so the `l.offset` calls are valid. |
| 476 |
unsafe { |
| 477 |
end_l = end_l.offset(-1); |
| 478 |
ptr::swap(l.offset(*end_l as isize), r.offset(-1)); |
| 479 |
r = r.offset(-1); |
| 480 |
} |
| 481 |
} |
| 482 |
width(v.as_mut_ptr(), r) |
| 483 |
} else if start_r < end_r { |
| 484 |
// The right block remains. |
| 485 |
// Move its remaining out-of-order elements to the far left. |
| 486 |
debug_assert_eq!(width(l, r), block_r); |
| 487 |
while start_r < end_r { |
| 488 |
// SAFETY: See the reasoning in [remaining-elements-safety]. |
| 489 |
unsafe { |
| 490 |
end_r = end_r.offset(-1); |
| 491 |
ptr::swap(l, r.offset(-(*end_r as isize) - 1)); |
| 492 |
l = l.offset(1); |
| 493 |
} |
| 494 |
} |
| 495 |
width(v.as_mut_ptr(), l) |
| 496 |
} else { |
| 497 |
// Nothing else to do, we're done. |
| 498 |
width(v.as_mut_ptr(), l) |
| 499 |
} |
| 500 |
} |
| 501 |
|
| 502 |
/// Partitions `v` into elements smaller than `v[pivot]`, followed by elements greater than or |
| 503 |
/// equal to `v[pivot]`. |
| 504 |
/// |
| 505 |
/// Returns a tuple of: |
| 506 |
/// |
| 507 |
/// 1. Number of elements smaller than `v[pivot]`. |
| 508 |
/// 2. True if `v` was already partitioned. |
| 509 |
fn partition<T, F>(v: &mut [T], pivot: usize, is_less: &F) -> (usize, bool) |
| 510 |
where |
| 511 |
F: Fn(&T, &T) -> bool, |
| 512 |
{ |
| 513 |
let (mid, was_partitioned) = { |
| 514 |
// Place the pivot at the beginning of slice. |
| 515 |
v.swap(0, pivot); |
| 516 |
let (pivot, v) = v.split_at_mut(1); |
| 517 |
let pivot = &mut pivot[0]; |
| 518 |
|
| 519 |
// Read the pivot into a stack-allocated variable for efficiency. If a following comparison |
| 520 |
// operation panics, the pivot will be automatically written back into the slice. |
| 521 |
|
| 522 |
// SAFETY: `pivot` is a reference to the first element of `v`, so `ptr::read` is safe. |
| 523 |
let tmp = mem::ManuallyDrop::new(unsafe { ptr::read(pivot) }); |
| 524 |
let _pivot_guard = unsafe { CopyOnDrop::new(&*tmp, pivot) }; |
| 525 |
let pivot = &*tmp; |
| 526 |
|
| 527 |
// Find the first pair of out-of-order elements. |
| 528 |
let mut l = 0; |
| 529 |
let mut r = v.len(); |
| 530 |
|
| 531 |
// SAFETY: The unsafety below involves indexing an array. |
| 532 |
// For the first one: We already do the bounds checking here with `l < r`. |
| 533 |
// For the second one: We initially have `l == 0` and `r == v.len()` and we checked that `l < r` at every indexing operation. |
| 534 |
// From here we know that `r` must be at least `r == l` which was shown to be valid from the first one. |
| 535 |
unsafe { |
| 536 |
// Find the first element greater than or equal to the pivot. |
| 537 |
while l < r && is_less(v.get_unchecked(l), pivot) { |
| 538 |
l += 1; |
| 539 |
} |
| 540 |
|
| 541 |
// Find the last element smaller that the pivot. |
| 542 |
while l < r && !is_less(v.get_unchecked(r - 1), pivot) { |
| 543 |
r -= 1; |
| 544 |
} |
| 545 |
} |
| 546 |
|
| 547 |
( |
| 548 |
l + partition_in_blocks(&mut v[l..r], pivot, is_less), |
| 549 |
l >= r, |
| 550 |
) |
| 551 |
|
| 552 |
// `_pivot_guard` goes out of scope and writes the pivot (which is a stack-allocated |
| 553 |
// variable) back into the slice where it originally was. This step is critical in ensuring |
| 554 |
// safety! |
| 555 |
}; |
| 556 |
|
| 557 |
// Place the pivot between the two partitions. |
| 558 |
v.swap(0, mid); |
| 559 |
|
| 560 |
(mid, was_partitioned) |
| 561 |
} |
| 562 |
|
| 563 |
/// Partitions `v` into elements equal to `v[pivot]` followed by elements greater than `v[pivot]`. |
| 564 |
/// |
| 565 |
/// Returns the number of elements equal to the pivot. It is assumed that `v` does not contain |
| 566 |
/// elements smaller than the pivot. |
| 567 |
fn partition_equal<T, F>(v: &mut [T], pivot: usize, is_less: &F) -> usize |
| 568 |
where |
| 569 |
F: Fn(&T, &T) -> bool, |
| 570 |
{ |
| 571 |
// Place the pivot at the beginning of slice. |
| 572 |
v.swap(0, pivot); |
| 573 |
let (pivot, v) = v.split_at_mut(1); |
| 574 |
let pivot = &mut pivot[0]; |
| 575 |
|
| 576 |
// Read the pivot into a stack-allocated variable for efficiency. If a following comparison |
| 577 |
// operation panics, the pivot will be automatically written back into the slice. |
| 578 |
// SAFETY: The pointer here is valid because it is obtained from a reference to a slice. |
| 579 |
let tmp = mem::ManuallyDrop::new(unsafe { ptr::read(pivot) }); |
| 580 |
let _pivot_guard = unsafe { CopyOnDrop::new(&*tmp, pivot) }; |
| 581 |
let pivot = &*tmp; |
| 582 |
|
| 583 |
// Now partition the slice. |
| 584 |
let mut l = 0; |
| 585 |
let mut r = v.len(); |
| 586 |
loop { |
| 587 |
// SAFETY: The unsafety below involves indexing an array. |
| 588 |
// For the first one: We already do the bounds checking here with `l < r`. |
| 589 |
// For the second one: We initially have `l == 0` and `r == v.len()` and we checked that `l < r` at every indexing operation. |
| 590 |
// From here we know that `r` must be at least `r == l` which was shown to be valid from the first one. |
| 591 |
unsafe { |
| 592 |
// Find the first element greater than the pivot. |
| 593 |
while l < r && !is_less(pivot, v.get_unchecked(l)) { |
| 594 |
l += 1; |
| 595 |
} |
| 596 |
|
| 597 |
// Find the last element equal to the pivot. |
| 598 |
while l < r && is_less(pivot, v.get_unchecked(r - 1)) { |
| 599 |
r -= 1; |
| 600 |
} |
| 601 |
|
| 602 |
// Are we done? |
| 603 |
if l >= r { |
| 604 |
break; |
| 605 |
} |
| 606 |
|
| 607 |
// Swap the found pair of out-of-order elements. |
| 608 |
r -= 1; |
| 609 |
let ptr = v.as_mut_ptr(); |
| 610 |
ptr::swap(ptr.add(l), ptr.add(r)); |
| 611 |
l += 1; |
| 612 |
} |
| 613 |
} |
| 614 |
|
| 615 |
// We found `l` elements equal to the pivot. Add 1 to account for the pivot itself. |
| 616 |
l + 1 |
| 617 |
|
| 618 |
// `_pivot_guard` goes out of scope and writes the pivot (which is a stack-allocated variable) |
| 619 |
// back into the slice where it originally was. This step is critical in ensuring safety! |
| 620 |
} |
| 621 |
|
| 622 |
/// Scatters some elements around in an attempt to break patterns that might cause imbalanced |
| 623 |
/// partitions in quicksort. |
| 624 |
#[cold] |
| 625 |
fn break_patterns<T>(v: &mut [T]) { |
| 626 |
let len = v.len(); |
| 627 |
if len >= 8 { |
| 628 |
// Pseudorandom number generator from the "Xorshift RNGs" paper by George Marsaglia. |
| 629 |
let mut random = len as u32; |
| 630 |
let mut gen_u32 = || { |
| 631 |
random ^= random << 13; |
| 632 |
random ^= random >> 17; |
| 633 |
random ^= random << 5; |
| 634 |
random |
| 635 |
}; |
| 636 |
let mut gen_usize = || { |
| 637 |
if usize::BITS <= 32 { |
| 638 |
gen_u32() as usize |
| 639 |
} else { |
| 640 |
(((gen_u32() as u64) << 32) | (gen_u32() as u64)) as usize |
| 641 |
} |
| 642 |
}; |
| 643 |
|
| 644 |
// Take random numbers modulo this number. |
| 645 |
// The number fits into `usize` because `len` is not greater than `isize::MAX`. |
| 646 |
let modulus = len.next_power_of_two(); |
| 647 |
|
| 648 |
// Some pivot candidates will be in the nearby of this index. Let's randomize them. |
| 649 |
let pos = len / 4 * 2; |
| 650 |
|
| 651 |
for i in 0..3 { |
| 652 |
// Generate a random number modulo `len`. However, in order to avoid costly operations |
| 653 |
// we first take it modulo a power of two, and then decrease by `len` until it fits |
| 654 |
// into the range `[0, len - 1]`. |
| 655 |
let mut other = gen_usize() & (modulus - 1); |
| 656 |
|
| 657 |
// `other` is guaranteed to be less than `2 * len`. |
| 658 |
if other >= len { |
| 659 |
other -= len; |
| 660 |
} |
| 661 |
|
| 662 |
v.swap(pos - 1 + i, other); |
| 663 |
} |
| 664 |
} |
| 665 |
} |
| 666 |
|
| 667 |
/// Chooses a pivot in `v` and returns the index and `true` if the slice is likely already sorted. |
| 668 |
/// |
| 669 |
/// Elements in `v` might be reordered in the process. |
| 670 |
fn choose_pivot<T, F>(v: &mut [T], is_less: &F) -> (usize, bool) |
| 671 |
where |
| 672 |
F: Fn(&T, &T) -> bool, |
| 673 |
{ |
| 674 |
// Minimum length to choose the median-of-medians method. |
| 675 |
// Shorter slices use the simple median-of-three method. |
| 676 |
const SHORTEST_MEDIAN_OF_MEDIANS: usize = 50; |
| 677 |
// Maximum number of swaps that can be performed in this function. |
| 678 |
const MAX_SWAPS: usize = 4 * 3; |
| 679 |
|
| 680 |
let len = v.len(); |
| 681 |
|
| 682 |
// Three indices near which we are going to choose a pivot. |
| 683 |
#[allow(clippy::identity_op)] |
| 684 |
let mut a = len / 4 * 1; |
| 685 |
let mut b = len / 4 * 2; |
| 686 |
let mut c = len / 4 * 3; |
| 687 |
|
| 688 |
// Counts the total number of swaps we are about to perform while sorting indices. |
| 689 |
let mut swaps = 0; |
| 690 |
|
| 691 |
if len >= 8 { |
| 692 |
// Swaps indices so that `v[a] <= v[b]`. |
| 693 |
// SAFETY: `len >= 8` so there are at least two elements in the neighborhoods of |
| 694 |
// `a`, `b` and `c`. This means the three calls to `sort_adjacent` result in |
| 695 |
// corresponding calls to `sort3` with valid 3-item neighborhoods around each |
| 696 |
// pointer, which in turn means the calls to `sort2` are done with valid |
| 697 |
// references. Thus the `v.get_unchecked` calls are safe, as is the `ptr::swap` |
| 698 |
// call. |
| 699 |
let mut sort2 = |a: &mut usize, b: &mut usize| unsafe { |
| 700 |
if is_less(v.get_unchecked(*b), v.get_unchecked(*a)) { |
| 701 |
ptr::swap(a, b); |
| 702 |
swaps += 1; |
| 703 |
} |
| 704 |
}; |
| 705 |
|
| 706 |
// Swaps indices so that `v[a] <= v[b] <= v[c]`. |
| 707 |
let mut sort3 = |a: &mut usize, b: &mut usize, c: &mut usize| { |
| 708 |
sort2(a, b); |
| 709 |
sort2(b, c); |
| 710 |
sort2(a, b); |
| 711 |
}; |
| 712 |
|
| 713 |
if len >= SHORTEST_MEDIAN_OF_MEDIANS { |
| 714 |
// Finds the median of `v[a - 1], v[a], v[a + 1]` and stores the index into `a`. |
| 715 |
let mut sort_adjacent = |a: &mut usize| { |
| 716 |
let tmp = *a; |
| 717 |
sort3(&mut (tmp - 1), a, &mut (tmp + 1)); |
| 718 |
}; |
| 719 |
|
| 720 |
// Find medians in the neighborhoods of `a`, `b`, and `c`. |
| 721 |
sort_adjacent(&mut a); |
| 722 |
sort_adjacent(&mut b); |
| 723 |
sort_adjacent(&mut c); |
| 724 |
} |
| 725 |
|
| 726 |
// Find the median among `a`, `b`, and `c`. |
| 727 |
sort3(&mut a, &mut b, &mut c); |
| 728 |
} |
| 729 |
|
| 730 |
if swaps < MAX_SWAPS { |
| 731 |
(b, swaps == 0) |
| 732 |
} else { |
| 733 |
// The maximum number of swaps was performed. Chances are the slice is descending or mostly |
| 734 |
// descending, so reversing will probably help sort it faster. |
| 735 |
v.reverse(); |
| 736 |
(len - 1 - b, true) |
| 737 |
} |
| 738 |
} |
| 739 |
|
| 740 |
/// Sorts `v` recursively. |
| 741 |
/// |
| 742 |
/// If the slice had a predecessor in the original array, it is specified as `pred`. |
| 743 |
/// |
| 744 |
/// `limit` is the number of allowed imbalanced partitions before switching to `heapsort`. If zero, |
| 745 |
/// this function will immediately switch to heapsort. |
| 746 |
fn recurse<'a, T, F>(mut v: &'a mut [T], is_less: &F, mut pred: Option<&'a mut T>, mut limit: u32) |
| 747 |
where |
| 748 |
T: Send, |
| 749 |
F: Fn(&T, &T) -> bool + Sync, |
| 750 |
{ |
| 751 |
// Slices of up to this length get sorted using insertion sort. |
| 752 |
const MAX_INSERTION: usize = 20; |
| 753 |
// If both partitions are up to this length, we continue sequentially. This number is as small |
| 754 |
// as possible but so that the overhead of Rayon's task scheduling is still negligible. |
| 755 |
const MAX_SEQUENTIAL: usize = 2000; |
| 756 |
|
| 757 |
// True if the last partitioning was reasonably balanced. |
| 758 |
let mut was_balanced = true; |
| 759 |
// True if the last partitioning didn't shuffle elements (the slice was already partitioned). |
| 760 |
let mut was_partitioned = true; |
| 761 |
|
| 762 |
loop { |
| 763 |
let len = v.len(); |
| 764 |
|
| 765 |
// Very short slices get sorted using insertion sort. |
| 766 |
if len <= MAX_INSERTION { |
| 767 |
insertion_sort(v, is_less); |
| 768 |
return; |
| 769 |
} |
| 770 |
|
| 771 |
// If too many bad pivot choices were made, simply fall back to heapsort in order to |
| 772 |
// guarantee `O(n * log(n))` worst-case. |
| 773 |
if limit == 0 { |
| 774 |
heapsort(v, is_less); |
| 775 |
return; |
| 776 |
} |
| 777 |
|
| 778 |
// If the last partitioning was imbalanced, try breaking patterns in the slice by shuffling |
| 779 |
// some elements around. Hopefully we'll choose a better pivot this time. |
| 780 |
if !was_balanced { |
| 781 |
break_patterns(v); |
| 782 |
limit -= 1; |
| 783 |
} |
| 784 |
|
| 785 |
// Choose a pivot and try guessing whether the slice is already sorted. |
| 786 |
let (pivot, likely_sorted) = choose_pivot(v, is_less); |
| 787 |
|
| 788 |
// If the last partitioning was decently balanced and didn't shuffle elements, and if pivot |
| 789 |
// selection predicts the slice is likely already sorted... |
| 790 |
if was_balanced && was_partitioned && likely_sorted { |
| 791 |
// Try identifying several out-of-order elements and shifting them to correct |
| 792 |
// positions. If the slice ends up being completely sorted, we're done. |
| 793 |
if partial_insertion_sort(v, is_less) { |
| 794 |
return; |
| 795 |
} |
| 796 |
} |
| 797 |
|
| 798 |
// If the chosen pivot is equal to the predecessor, then it's the smallest element in the |
| 799 |
// slice. Partition the slice into elements equal to and elements greater than the pivot. |
| 800 |
// This case is usually hit when the slice contains many duplicate elements. |
| 801 |
if let Some(ref p) = pred { |
| 802 |
if !is_less(p, &v[pivot]) { |
| 803 |
let mid = partition_equal(v, pivot, is_less); |
| 804 |
|
| 805 |
// Continue sorting elements greater than the pivot. |
| 806 |
v = &mut v[mid..]; |
| 807 |
continue; |
| 808 |
} |
| 809 |
} |
| 810 |
|
| 811 |
// Partition the slice. |
| 812 |
let (mid, was_p) = partition(v, pivot, is_less); |
| 813 |
was_balanced = Ord::min(mid, len - mid) >= len / 8; |
| 814 |
was_partitioned = was_p; |
| 815 |
|
| 816 |
// Split the slice into `left`, `pivot`, and `right`. |
| 817 |
let (left, right) = v.split_at_mut(mid); |
| 818 |
let (pivot, right) = right.split_at_mut(1); |
| 819 |
let pivot = &mut pivot[0]; |
| 820 |
|
| 821 |
if Ord::max(left.len(), right.len()) <= MAX_SEQUENTIAL { |
| 822 |
// Recurse into the shorter side only in order to minimize the total number of recursive |
| 823 |
// calls and consume less stack space. Then just continue with the longer side (this is |
| 824 |
// akin to tail recursion). |
| 825 |
if left.len() < right.len() { |
| 826 |
recurse(left, is_less, pred, limit); |
| 827 |
v = right; |
| 828 |
pred = Some(pivot); |
| 829 |
} else { |
| 830 |
recurse(right, is_less, Some(pivot), limit); |
| 831 |
v = left; |
| 832 |
} |
| 833 |
} else { |
| 834 |
// Sort the left and right half in parallel. |
| 835 |
rayon_core::join( |
| 836 |
|| recurse(left, is_less, pred, limit), |
| 837 |
|| recurse(right, is_less, Some(pivot), limit), |
| 838 |
); |
| 839 |
break; |
| 840 |
} |
| 841 |
} |
| 842 |
} |
| 843 |
|
| 844 |
/// Sorts `v` using pattern-defeating quicksort in parallel. |
| 845 |
/// |
| 846 |
/// The algorithm is unstable, in-place, and *O*(*n* \* log(*n*)) worst-case. |
| 847 |
pub(super) fn par_quicksort<T, F>(v: &mut [T], is_less: F) |
| 848 |
where |
| 849 |
T: Send, |
| 850 |
F: Fn(&T, &T) -> bool + Sync, |
| 851 |
{ |
| 852 |
// Sorting has no meaningful behavior on zero-sized types. |
| 853 |
if mem::size_of::<T>() == 0 { |
| 854 |
return; |
| 855 |
} |
| 856 |
|
| 857 |
// Limit the number of imbalanced partitions to `floor(log2(len)) + 1`. |
| 858 |
let limit = usize::BITS - v.len().leading_zeros(); |
| 859 |
|
| 860 |
recurse(v, &is_less, None, limit); |
| 861 |
} |
| 862 |
|
| 863 |
#[cfg(test)] |
| 864 |
mod tests { |
| 865 |
use super::heapsort; |
| 866 |
use rand::distributions::Uniform; |
| 867 |
use rand::{thread_rng, Rng}; |
| 868 |
|
| 869 |
#[test] |
| 870 |
fn test_heapsort() { |
| 871 |
let rng = &mut thread_rng(); |
| 872 |
|
| 873 |
for len in (0..25).chain(500..501) { |
| 874 |
for &modulus in &[5, 10, 100] { |
| 875 |
let dist = Uniform::new(0, modulus); |
| 876 |
for _ in 0..100 { |
| 877 |
let v: Vec<i32> = rng.sample_iter(&dist).take(len).collect(); |
| 878 |
|
| 879 |
// Test heapsort using `<` operator. |
| 880 |
let mut tmp = v.clone(); |
| 881 |
heapsort(&mut tmp, &|a, b| a < b); |
| 882 |
assert!(tmp.windows(2).all(|w| w[0] <= w[1])); |
| 883 |
|
| 884 |
// Test heapsort using `>` operator. |
| 885 |
let mut tmp = v.clone(); |
| 886 |
heapsort(&mut tmp, &|a, b| a > b); |
| 887 |
assert!(tmp.windows(2).all(|w| w[0] >= w[1])); |
| 888 |
} |
| 889 |
} |
| 890 |
} |
| 891 |
|
| 892 |
// Sort using a completely random comparison function. |
| 893 |
// This will reorder the elements *somehow*, but won't panic. |
| 894 |
let mut v: Vec<_> = (0..100).collect(); |
| 895 |
heapsort(&mut v, &|_, _| thread_rng().gen()); |
| 896 |
heapsort(&mut v, &|a, b| a < b); |
| 897 |
|
| 898 |
for (i, &entry) in v.iter().enumerate() { |
| 899 |
assert_eq!(entry, i); |
| 900 |
} |
| 901 |
} |
| 902 |
} |