1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929
//! Slice sorting
//!
//! This module contains a sorting algorithm based on Orson Peters' pattern-defeating quicksort,
//! published at: <https://github.com/orlp/pdqsort>
//!
//! Unstable sorting is compatible with libcore because it doesn't allocate memory, unlike our
//! stable sorting implementation.
use crate::cmp;
use crate::mem::{self, MaybeUninit};
use crate::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: &mut F)
where
F: FnMut(&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: &mut F)
where
F: FnMut(&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: &mut F) -> bool
where
F: FnMut(&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: &mut F)
where
F: FnMut(&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]
#[unstable(feature = "sort_internals", reason = "internal to sort module", issue = "none")]
pub fn heapsort<T, F>(v: &mut [T], mut is_less: F)
where
F: FnMut(&T, &T) -> bool,
{
// This binary heap respects the invariant `parent >= child`.
let mut 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: &mut F) -> usize
where
F: FnMut(&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).
(r.addr() - l.addr()) / 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.
start_l = MaybeUninit::slice_as_mut_ptr(&mut offsets_l);
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.
start_r = MaybeUninit::slice_as_mut_ptr(&mut offsets_r);
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.offset(block_l as isize) };
}
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: &mut F) -> (usize, bool)
where
F: FnMut(&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: &mut F) -> usize
where
F: FnMut(&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: &mut F) -> (usize, bool)
where
F: FnMut(&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.
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: &mut F, mut pred: Option<&'a T>, mut limit: u32)
where
F: FnMut(&T, &T) -> bool,
{
// Slices of up to this length get sorted using insertion sort.
const MAX_INSERTION: usize = 20;
// 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(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 = &pivot[0];
// 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;
}
}
}
/// Sorts `v` using pattern-defeating quicksort, which is *O*(*n* \* log(*n*)) worst-case.
pub fn quicksort<T, F>(v: &mut [T], mut is_less: F)
where
F: FnMut(&T, &T) -> bool,
{
// 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, &mut is_less, None, limit);
}
fn partition_at_index_loop<'a, T, F>(
mut v: &'a mut [T],
mut index: usize,
is_less: &mut F,
mut pred: Option<&'a T>,
) where
F: FnMut(&T, &T) -> bool,
{
loop {
// For slices of up to this length it's probably faster to simply sort them.
const MAX_INSERTION: usize = 10;
if v.len() <= MAX_INSERTION {
insertion_sort(v, is_less);
return;
}
// Choose a pivot
let (pivot, _) = choose_pivot(v, is_less);
// 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(p) = pred {
if !is_less(p, &v[pivot]) {
let mid = partition_equal(v, pivot, is_less);
// If we've passed our index, then we're good.
if mid > index {
return;
}
// Otherwise, continue sorting elements greater than the pivot.
v = &mut v[mid..];
index = index - mid;
pred = None;
continue;
}
}
let (mid, _) = partition(v, pivot, is_less);
// 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 = &pivot[0];
if mid < index {
v = right;
index = index - mid - 1;
pred = Some(pivot);
} else if mid > index {
v = left;
} else {
// If mid == index, then we're done, since partition() guaranteed that all elements
// after mid are greater than or equal to mid.
return;
}
}
}
pub fn partition_at_index<T, F>(
v: &mut [T],
index: usize,
mut is_less: F,
) -> (&mut [T], &mut T, &mut [T])
where
F: FnMut(&T, &T) -> bool,
{
use cmp::Ordering::Greater;
use cmp::Ordering::Less;
if index >= v.len() {
panic!("partition_at_index index {} greater than length of slice {}", index, v.len());
}
if mem::size_of::<T>() == 0 {
// Sorting has no meaningful behavior on zero-sized types. Do nothing.
} else if index == v.len() - 1 {
// Find max element and place it in the last position of the array. We're free to use
// `unwrap()` here because we know v must not be empty.
let (max_index, _) = v
.iter()
.enumerate()
.max_by(|&(_, x), &(_, y)| if is_less(x, y) { Less } else { Greater })
.unwrap();
v.swap(max_index, index);
} else if index == 0 {
// Find min element and place it in the first position of the array. We're free to use
// `unwrap()` here because we know v must not be empty.
let (min_index, _) = v
.iter()
.enumerate()
.min_by(|&(_, x), &(_, y)| if is_less(x, y) { Less } else { Greater })
.unwrap();
v.swap(min_index, index);
} else {
partition_at_index_loop(v, index, &mut is_less, None);
}
let (left, right) = v.split_at_mut(index);
let (pivot, right) = right.split_at_mut(1);
let pivot = &mut pivot[0];
(left, pivot, right)
}