1 //! A dynamically-sized view into a contiguous sequence, `[T]`.
3 //! *[See also the slice primitive type](../../std/primitive.slice.html).*
5 //! Slices are a view into a block of memory represented as a pointer and a
10 //! let vec = vec![1, 2, 3];
11 //! let int_slice = &vec[..];
12 //! // coercing an array to a slice
13 //! let str_slice: &[&str] = &["one", "two", "three"];
16 //! Slices are either mutable or shared. The shared slice type is `&[T]`,
17 //! while the mutable slice type is `&mut [T]`, where `T` represents the element
18 //! type. For example, you can mutate the block of memory that a mutable slice
22 //! let x = &mut [1, 2, 3];
24 //! assert_eq!(x, &[1, 7, 3]);
27 //! Here are some of the things this module contains:
31 //! There are several structs that are useful for slices, such as [`Iter`], which
32 //! represents iteration over a slice.
34 //! ## Trait Implementations
36 //! There are several implementations of common traits for slices. Some examples
40 //! * [`Eq`], [`Ord`] - for slices whose element type are [`Eq`] or [`Ord`].
41 //! * [`Hash`] - for slices whose element type is [`Hash`].
45 //! The slices implement `IntoIterator`. The iterator yields references to the
49 //! let numbers = &[0, 1, 2];
50 //! for n in numbers {
51 //! println!("{} is a number!", n);
55 //! The mutable slice yields mutable references to the elements:
58 //! let mut scores = [7, 8, 9];
59 //! for score in &mut scores[..] {
64 //! This iterator yields mutable references to the slice's elements, so while
65 //! the element type of the slice is `i32`, the element type of the iterator is
68 //! * [`.iter`] and [`.iter_mut`] are the explicit methods to return the default
70 //! * Further methods that return iterators are [`.split`], [`.splitn`],
71 //! [`.chunks`], [`.windows`] and more.
73 //! [`Hash`]: core::hash::Hash
74 //! [`.iter`]: ../../std/primitive.slice.html#method.iter
75 //! [`.iter_mut`]: ../../std/primitive.slice.html#method.iter_mut
76 //! [`.split`]: ../../std/primitive.slice.html#method.split
77 //! [`.splitn`]: ../../std/primitive.slice.html#method.splitn
78 //! [`.chunks`]: ../../std/primitive.slice.html#method.chunks
79 //! [`.windows`]: ../../std/primitive.slice.html#method.windows
80 #![stable(feature = "rust1", since = "1.0.0")]
81 // Many of the usings in this module are only used in the test configuration.
82 // It's cleaner to just turn off the unused_imports warning than to fix them.
83 #![cfg_attr(test, allow(unused_imports, dead_code))]
85 use core::borrow::{Borrow, BorrowMut};
86 use core::cmp::Ordering::{self, Less};
87 use core::mem::{self, size_of};
90 use crate::borrow::ToOwned;
91 use crate::boxed::Box;
94 #[unstable(feature = "array_chunks", issue = "74985")]
95 pub use core::slice::ArrayChunks;
96 #[stable(feature = "slice_get_slice", since = "1.28.0")]
97 pub use core::slice::SliceIndex;
98 #[stable(feature = "from_ref", since = "1.28.0")]
99 pub use core::slice::{from_mut, from_ref};
100 #[stable(feature = "rust1", since = "1.0.0")]
101 pub use core::slice::{from_raw_parts, from_raw_parts_mut};
102 #[stable(feature = "rust1", since = "1.0.0")]
103 pub use core::slice::{Chunks, Windows};
104 #[stable(feature = "chunks_exact", since = "1.31.0")]
105 pub use core::slice::{ChunksExact, ChunksExactMut};
106 #[stable(feature = "rust1", since = "1.0.0")]
107 pub use core::slice::{ChunksMut, Split, SplitMut};
108 #[stable(feature = "rust1", since = "1.0.0")]
109 pub use core::slice::{Iter, IterMut};
110 #[stable(feature = "rchunks", since = "1.31.0")]
111 pub use core::slice::{RChunks, RChunksExact, RChunksExactMut, RChunksMut};
112 #[stable(feature = "slice_rsplit", since = "1.27.0")]
113 pub use core::slice::{RSplit, RSplitMut};
114 #[stable(feature = "rust1", since = "1.0.0")]
115 pub use core::slice::{RSplitN, RSplitNMut, SplitN, SplitNMut};
117 ////////////////////////////////////////////////////////////////////////////////
118 // Basic slice extension methods
119 ////////////////////////////////////////////////////////////////////////////////
121 // HACK(japaric) needed for the implementation of `vec!` macro during testing
122 // N.B., see the `hack` module in this file for more details.
124 pub use hack::into_vec;
126 // HACK(japaric) needed for the implementation of `Vec::clone` during testing
127 // N.B., see the `hack` module in this file for more details.
129 pub use hack::to_vec;
131 // HACK(japaric): With cfg(test) `impl [T]` is not available, these three
132 // functions are actually methods that are in `impl [T]` but not in
133 // `core::slice::SliceExt` - we need to supply these functions for the
134 // `test_permutations` test
136 use crate::boxed::Box;
139 // We shouldn't add inline attribute to this since this is used in
140 // `vec!` macro mostly and causes perf regression. See #71204 for
141 // discussion and perf results.
142 pub fn into_vec<T>(b: Box<[T]>) -> Vec<T> {
145 let b = Box::into_raw(b);
146 Vec::from_raw_parts(b as *mut T, len, len)
151 pub fn to_vec<T>(s: &[T]) -> Vec<T>
155 let mut vec = Vec::with_capacity(s.len());
156 vec.extend_from_slice(s);
161 #[lang = "slice_alloc"]
166 /// This sort is stable (i.e., does not reorder equal elements) and `O(n * log(n))` worst-case.
168 /// When applicable, unstable sorting is preferred because it is generally faster than stable
169 /// sorting and it doesn't allocate auxiliary memory.
170 /// See [`sort_unstable`](#method.sort_unstable).
172 /// # Current implementation
174 /// The current algorithm is an adaptive, iterative merge sort inspired by
175 /// [timsort](https://en.wikipedia.org/wiki/Timsort).
176 /// It is designed to be very fast in cases where the slice is nearly sorted, or consists of
177 /// two or more sorted sequences concatenated one after another.
179 /// Also, it allocates temporary storage half the size of `self`, but for short slices a
180 /// non-allocating insertion sort is used instead.
185 /// let mut v = [-5, 4, 1, -3, 2];
188 /// assert!(v == [-5, -3, 1, 2, 4]);
190 #[stable(feature = "rust1", since = "1.0.0")]
192 pub fn sort(&mut self)
196 merge_sort(self, |a, b| a.lt(b));
199 /// Sorts the slice with a comparator function.
201 /// This sort is stable (i.e., does not reorder equal elements) and `O(n * log(n))` worst-case.
203 /// The comparator function must define a total ordering for the elements in the slice. If
204 /// the ordering is not total, the order of the elements is unspecified. An order is a
205 /// total order if it is (for all `a`, `b` and `c`):
207 /// * total and antisymmetric: exactly one of `a < b`, `a == b` or `a > b` is true, and
208 /// * transitive, `a < b` and `b < c` implies `a < c`. The same must hold for both `==` and `>`.
210 /// For example, while [`f64`] doesn't implement [`Ord`] because `NaN != NaN`, we can use
211 /// `partial_cmp` as our sort function when we know the slice doesn't contain a `NaN`.
214 /// let mut floats = [5f64, 4.0, 1.0, 3.0, 2.0];
215 /// floats.sort_by(|a, b| a.partial_cmp(b).unwrap());
216 /// assert_eq!(floats, [1.0, 2.0, 3.0, 4.0, 5.0]);
219 /// When applicable, unstable sorting is preferred because it is generally faster than stable
220 /// sorting and it doesn't allocate auxiliary memory.
221 /// See [`sort_unstable_by`](#method.sort_unstable_by).
223 /// # Current implementation
225 /// The current algorithm is an adaptive, iterative merge sort inspired by
226 /// [timsort](https://en.wikipedia.org/wiki/Timsort).
227 /// It is designed to be very fast in cases where the slice is nearly sorted, or consists of
228 /// two or more sorted sequences concatenated one after another.
230 /// Also, it allocates temporary storage half the size of `self`, but for short slices a
231 /// non-allocating insertion sort is used instead.
236 /// let mut v = [5, 4, 1, 3, 2];
237 /// v.sort_by(|a, b| a.cmp(b));
238 /// assert!(v == [1, 2, 3, 4, 5]);
240 /// // reverse sorting
241 /// v.sort_by(|a, b| b.cmp(a));
242 /// assert!(v == [5, 4, 3, 2, 1]);
244 #[stable(feature = "rust1", since = "1.0.0")]
246 pub fn sort_by<F>(&mut self, mut compare: F)
248 F: FnMut(&T, &T) -> Ordering,
250 merge_sort(self, |a, b| compare(a, b) == Less);
253 /// Sorts the slice with a key extraction function.
255 /// This sort is stable (i.e., does not reorder equal elements) and `O(m * n * log(n))`
256 /// worst-case, where the key function is `O(m)`.
258 /// For expensive key functions (e.g. functions that are not simple property accesses or
259 /// basic operations), [`sort_by_cached_key`](#method.sort_by_cached_key) is likely to be
260 /// significantly faster, as it does not recompute element keys.
262 /// When applicable, unstable sorting is preferred because it is generally faster than stable
263 /// sorting and it doesn't allocate auxiliary memory.
264 /// See [`sort_unstable_by_key`](#method.sort_unstable_by_key).
266 /// # Current implementation
268 /// The current algorithm is an adaptive, iterative merge sort inspired by
269 /// [timsort](https://en.wikipedia.org/wiki/Timsort).
270 /// It is designed to be very fast in cases where the slice is nearly sorted, or consists of
271 /// two or more sorted sequences concatenated one after another.
273 /// Also, it allocates temporary storage half the size of `self`, but for short slices a
274 /// non-allocating insertion sort is used instead.
279 /// let mut v = [-5i32, 4, 1, -3, 2];
281 /// v.sort_by_key(|k| k.abs());
282 /// assert!(v == [1, 2, -3, 4, -5]);
284 #[stable(feature = "slice_sort_by_key", since = "1.7.0")]
286 pub fn sort_by_key<K, F>(&mut self, mut f: F)
291 merge_sort(self, |a, b| f(a).lt(&f(b)));
294 /// Sorts the slice with a key extraction function.
296 /// During sorting, the key function is called only once per element.
298 /// This sort is stable (i.e., does not reorder equal elements) and `O(m * n + n * log(n))`
299 /// worst-case, where the key function is `O(m)`.
301 /// For simple key functions (e.g., functions that are property accesses or
302 /// basic operations), [`sort_by_key`](#method.sort_by_key) is likely to be
305 /// # Current implementation
307 /// The current algorithm is based on [pattern-defeating quicksort][pdqsort] by Orson Peters,
308 /// which combines the fast average case of randomized quicksort with the fast worst case of
309 /// heapsort, while achieving linear time on slices with certain patterns. It uses some
310 /// randomization to avoid degenerate cases, but with a fixed seed to always provide
311 /// deterministic behavior.
313 /// In the worst case, the algorithm allocates temporary storage in a `Vec<(K, usize)>` the
314 /// length of the slice.
319 /// let mut v = [-5i32, 4, 32, -3, 2];
321 /// v.sort_by_cached_key(|k| k.to_string());
322 /// assert!(v == [-3, -5, 2, 32, 4]);
325 /// [pdqsort]: https://github.com/orlp/pdqsort
326 #[stable(feature = "slice_sort_by_cached_key", since = "1.34.0")]
328 pub fn sort_by_cached_key<K, F>(&mut self, f: F)
333 // Helper macro for indexing our vector by the smallest possible type, to reduce allocation.
334 macro_rules! sort_by_key {
335 ($t:ty, $slice:ident, $f:ident) => {{
336 let mut indices: Vec<_> =
337 $slice.iter().map($f).enumerate().map(|(i, k)| (k, i as $t)).collect();
338 // The elements of `indices` are unique, as they are indexed, so any sort will be
339 // stable with respect to the original slice. We use `sort_unstable` here because
340 // it requires less memory allocation.
341 indices.sort_unstable();
342 for i in 0..$slice.len() {
343 let mut index = indices[i].1;
344 while (index as usize) < i {
345 index = indices[index as usize].1;
347 indices[i].1 = index;
348 $slice.swap(i, index as usize);
353 let sz_u8 = mem::size_of::<(K, u8)>();
354 let sz_u16 = mem::size_of::<(K, u16)>();
355 let sz_u32 = mem::size_of::<(K, u32)>();
356 let sz_usize = mem::size_of::<(K, usize)>();
358 let len = self.len();
362 if sz_u8 < sz_u16 && len <= (u8::MAX as usize) {
363 return sort_by_key!(u8, self, f);
365 if sz_u16 < sz_u32 && len <= (u16::MAX as usize) {
366 return sort_by_key!(u16, self, f);
368 if sz_u32 < sz_usize && len <= (u32::MAX as usize) {
369 return sort_by_key!(u32, self, f);
371 sort_by_key!(usize, self, f)
374 /// Copies `self` into a new `Vec`.
379 /// let s = [10, 40, 30];
380 /// let x = s.to_vec();
381 /// // Here, `s` and `x` can be modified independently.
383 #[rustc_conversion_suggestion]
384 #[stable(feature = "rust1", since = "1.0.0")]
386 pub fn to_vec(&self) -> Vec<T>
390 // N.B., see the `hack` module in this file for more details.
394 /// Converts `self` into a vector without clones or allocation.
396 /// The resulting vector can be converted back into a box via
397 /// `Vec<T>`'s `into_boxed_slice` method.
402 /// let s: Box<[i32]> = Box::new([10, 40, 30]);
403 /// let x = s.into_vec();
404 /// // `s` cannot be used anymore because it has been converted into `x`.
406 /// assert_eq!(x, vec![10, 40, 30]);
408 #[stable(feature = "rust1", since = "1.0.0")]
410 pub fn into_vec(self: Box<Self>) -> Vec<T> {
411 // N.B., see the `hack` module in this file for more details.
415 /// Creates a vector by repeating a slice `n` times.
419 /// This function will panic if the capacity would overflow.
426 /// assert_eq!([1, 2].repeat(3), vec![1, 2, 1, 2, 1, 2]);
429 /// A panic upon overflow:
432 /// // this will panic at runtime
433 /// b"0123456789abcdef".repeat(usize::MAX);
435 #[stable(feature = "repeat_generic_slice", since = "1.40.0")]
436 pub fn repeat(&self, n: usize) -> Vec<T>
444 // If `n` is larger than zero, it can be split as
445 // `n = 2^expn + rem (2^expn > rem, expn >= 0, rem >= 0)`.
446 // `2^expn` is the number represented by the leftmost '1' bit of `n`,
447 // and `rem` is the remaining part of `n`.
449 // Using `Vec` to access `set_len()`.
450 let capacity = self.len().checked_mul(n).expect("capacity overflow");
451 let mut buf = Vec::with_capacity(capacity);
453 // `2^expn` repetition is done by doubling `buf` `expn`-times.
457 // If `m > 0`, there are remaining bits up to the leftmost '1'.
459 // `buf.extend(buf)`:
461 ptr::copy_nonoverlapping(
463 (buf.as_mut_ptr() as *mut T).add(buf.len()),
466 // `buf` has capacity of `self.len() * n`.
467 let buf_len = buf.len();
468 buf.set_len(buf_len * 2);
475 // `rem` (`= n - 2^expn`) repetition is done by copying
476 // first `rem` repetitions from `buf` itself.
477 let rem_len = capacity - buf.len(); // `self.len() * rem`
479 // `buf.extend(buf[0 .. rem_len])`:
481 // This is non-overlapping since `2^expn > rem`.
482 ptr::copy_nonoverlapping(
484 (buf.as_mut_ptr() as *mut T).add(buf.len()),
487 // `buf.len() + rem_len` equals to `buf.capacity()` (`= self.len() * n`).
488 buf.set_len(capacity);
494 /// Flattens a slice of `T` into a single value `Self::Output`.
499 /// assert_eq!(["hello", "world"].concat(), "helloworld");
500 /// assert_eq!([[1, 2], [3, 4]].concat(), [1, 2, 3, 4]);
502 #[stable(feature = "rust1", since = "1.0.0")]
503 pub fn concat<Item: ?Sized>(&self) -> <Self as Concat<Item>>::Output
510 /// Flattens a slice of `T` into a single value `Self::Output`, placing a
511 /// given separator between each.
516 /// assert_eq!(["hello", "world"].join(" "), "hello world");
517 /// assert_eq!([[1, 2], [3, 4]].join(&0), [1, 2, 0, 3, 4]);
518 /// assert_eq!([[1, 2], [3, 4]].join(&[0, 0][..]), [1, 2, 0, 0, 3, 4]);
520 #[stable(feature = "rename_connect_to_join", since = "1.3.0")]
521 pub fn join<Separator>(&self, sep: Separator) -> <Self as Join<Separator>>::Output
523 Self: Join<Separator>,
525 Join::join(self, sep)
528 /// Flattens a slice of `T` into a single value `Self::Output`, placing a
529 /// given separator between each.
534 /// # #![allow(deprecated)]
535 /// assert_eq!(["hello", "world"].connect(" "), "hello world");
536 /// assert_eq!([[1, 2], [3, 4]].connect(&0), [1, 2, 0, 3, 4]);
538 #[stable(feature = "rust1", since = "1.0.0")]
539 #[rustc_deprecated(since = "1.3.0", reason = "renamed to join")]
540 pub fn connect<Separator>(&self, sep: Separator) -> <Self as Join<Separator>>::Output
542 Self: Join<Separator>,
544 Join::join(self, sep)
548 #[lang = "slice_u8_alloc"]
551 /// Returns a vector containing a copy of this slice where each byte
552 /// is mapped to its ASCII upper case equivalent.
554 /// ASCII letters 'a' to 'z' are mapped to 'A' to 'Z',
555 /// but non-ASCII letters are unchanged.
557 /// To uppercase the value in-place, use [`make_ascii_uppercase`].
559 /// [`make_ascii_uppercase`]: u8::make_ascii_uppercase
560 #[stable(feature = "ascii_methods_on_intrinsics", since = "1.23.0")]
562 pub fn to_ascii_uppercase(&self) -> Vec<u8> {
563 let mut me = self.to_vec();
564 me.make_ascii_uppercase();
568 /// Returns a vector containing a copy of this slice where each byte
569 /// is mapped to its ASCII lower case equivalent.
571 /// ASCII letters 'A' to 'Z' are mapped to 'a' to 'z',
572 /// but non-ASCII letters are unchanged.
574 /// To lowercase the value in-place, use [`make_ascii_lowercase`].
576 /// [`make_ascii_lowercase`]: u8::make_ascii_lowercase
577 #[stable(feature = "ascii_methods_on_intrinsics", since = "1.23.0")]
579 pub fn to_ascii_lowercase(&self) -> Vec<u8> {
580 let mut me = self.to_vec();
581 me.make_ascii_lowercase();
586 ////////////////////////////////////////////////////////////////////////////////
587 // Extension traits for slices over specific kinds of data
588 ////////////////////////////////////////////////////////////////////////////////
590 /// Helper trait for [`[T]::concat`](../../std/primitive.slice.html#method.concat).
592 /// Note: the `Item` type parameter is not used in this trait,
593 /// but it allows impls to be more generic.
594 /// Without it, we get this error:
597 /// error[E0207]: the type parameter `T` is not constrained by the impl trait, self type, or predica
598 /// --> src/liballoc/slice.rs:608:6
600 /// 608 | impl<T: Clone, V: Borrow<[T]>> Concat for [V] {
601 /// | ^ unconstrained type parameter
604 /// This is because there could exist `V` types with multiple `Borrow<[_]>` impls,
605 /// such that multiple `T` types would apply:
608 /// # #[allow(dead_code)]
609 /// pub struct Foo(Vec<u32>, Vec<String>);
611 /// impl std::borrow::Borrow<[u32]> for Foo {
612 /// fn borrow(&self) -> &[u32] { &self.0 }
615 /// impl std::borrow::Borrow<[String]> for Foo {
616 /// fn borrow(&self) -> &[String] { &self.1 }
619 #[unstable(feature = "slice_concat_trait", issue = "27747")]
620 pub trait Concat<Item: ?Sized> {
621 #[unstable(feature = "slice_concat_trait", issue = "27747")]
622 /// The resulting type after concatenation
625 /// Implementation of [`[T]::concat`](../../std/primitive.slice.html#method.concat)
626 #[unstable(feature = "slice_concat_trait", issue = "27747")]
627 fn concat(slice: &Self) -> Self::Output;
630 /// Helper trait for [`[T]::join`](../../std/primitive.slice.html#method.join)
631 #[unstable(feature = "slice_concat_trait", issue = "27747")]
632 pub trait Join<Separator> {
633 #[unstable(feature = "slice_concat_trait", issue = "27747")]
634 /// The resulting type after concatenation
637 /// Implementation of [`[T]::join`](../../std/primitive.slice.html#method.join)
638 #[unstable(feature = "slice_concat_trait", issue = "27747")]
639 fn join(slice: &Self, sep: Separator) -> Self::Output;
642 #[unstable(feature = "slice_concat_ext", issue = "27747")]
643 impl<T: Clone, V: Borrow<[T]>> Concat<T> for [V] {
644 type Output = Vec<T>;
646 fn concat(slice: &Self) -> Vec<T> {
647 let size = slice.iter().map(|slice| slice.borrow().len()).sum();
648 let mut result = Vec::with_capacity(size);
650 result.extend_from_slice(v.borrow())
656 #[unstable(feature = "slice_concat_ext", issue = "27747")]
657 impl<T: Clone, V: Borrow<[T]>> Join<&T> for [V] {
658 type Output = Vec<T>;
660 fn join(slice: &Self, sep: &T) -> Vec<T> {
661 let mut iter = slice.iter();
662 let first = match iter.next() {
663 Some(first) => first,
664 None => return vec![],
666 let size = slice.iter().map(|v| v.borrow().len()).sum::<usize>() + slice.len() - 1;
667 let mut result = Vec::with_capacity(size);
668 result.extend_from_slice(first.borrow());
671 result.push(sep.clone());
672 result.extend_from_slice(v.borrow())
678 #[unstable(feature = "slice_concat_ext", issue = "27747")]
679 impl<T: Clone, V: Borrow<[T]>> Join<&[T]> for [V] {
680 type Output = Vec<T>;
682 fn join(slice: &Self, sep: &[T]) -> Vec<T> {
683 let mut iter = slice.iter();
684 let first = match iter.next() {
685 Some(first) => first,
686 None => return vec![],
689 slice.iter().map(|v| v.borrow().len()).sum::<usize>() + sep.len() * (slice.len() - 1);
690 let mut result = Vec::with_capacity(size);
691 result.extend_from_slice(first.borrow());
694 result.extend_from_slice(sep);
695 result.extend_from_slice(v.borrow())
701 ////////////////////////////////////////////////////////////////////////////////
702 // Standard trait implementations for slices
703 ////////////////////////////////////////////////////////////////////////////////
705 #[stable(feature = "rust1", since = "1.0.0")]
706 impl<T> Borrow<[T]> for Vec<T> {
707 fn borrow(&self) -> &[T] {
712 #[stable(feature = "rust1", since = "1.0.0")]
713 impl<T> BorrowMut<[T]> for Vec<T> {
714 fn borrow_mut(&mut self) -> &mut [T] {
719 #[stable(feature = "rust1", since = "1.0.0")]
720 impl<T: Clone> ToOwned for [T] {
723 fn to_owned(&self) -> Vec<T> {
728 fn to_owned(&self) -> Vec<T> {
732 fn clone_into(&self, target: &mut Vec<T>) {
733 // drop anything in target that will not be overwritten
734 target.truncate(self.len());
736 // target.len <= self.len due to the truncate above, so the
737 // slices here are always in-bounds.
738 let (init, tail) = self.split_at(target.len());
740 // reuse the contained values' allocations/resources.
741 target.clone_from_slice(init);
742 target.extend_from_slice(tail);
746 ////////////////////////////////////////////////////////////////////////////////
748 ////////////////////////////////////////////////////////////////////////////////
750 /// Inserts `v[0]` into pre-sorted sequence `v[1..]` so that whole `v[..]` becomes sorted.
752 /// This is the integral subroutine of insertion sort.
753 fn insert_head<T, F>(v: &mut [T], is_less: &mut F)
755 F: FnMut(&T, &T) -> bool,
757 if v.len() >= 2 && is_less(&v[1], &v[0]) {
759 // There are three ways to implement insertion here:
761 // 1. Swap adjacent elements until the first one gets to its final destination.
762 // However, this way we copy data around more than is necessary. If elements are big
763 // structures (costly to copy), this method will be slow.
765 // 2. Iterate until the right place for the first element is found. Then shift the
766 // elements succeeding it to make room for it and finally place it into the
767 // remaining hole. This is a good method.
769 // 3. Copy the first element into a temporary variable. Iterate until the right place
770 // for it is found. As we go along, copy every traversed element into the slot
771 // preceding it. Finally, copy data from the temporary variable into the remaining
772 // hole. This method is very good. Benchmarks demonstrated slightly better
773 // performance than with the 2nd method.
775 // All methods were benchmarked, and the 3rd showed best results. So we chose that one.
776 let mut tmp = mem::ManuallyDrop::new(ptr::read(&v[0]));
778 // Intermediate state of the insertion process is always tracked by `hole`, which
779 // serves two purposes:
780 // 1. Protects integrity of `v` from panics in `is_less`.
781 // 2. Fills the remaining hole in `v` in the end.
785 // If `is_less` panics at any point during the process, `hole` will get dropped and
786 // fill the hole in `v` with `tmp`, thus ensuring that `v` still holds every object it
787 // initially held exactly once.
788 let mut hole = InsertionHole { src: &mut *tmp, dest: &mut v[1] };
789 ptr::copy_nonoverlapping(&v[1], &mut v[0], 1);
791 for i in 2..v.len() {
792 if !is_less(&v[i], &*tmp) {
795 ptr::copy_nonoverlapping(&v[i], &mut v[i - 1], 1);
796 hole.dest = &mut v[i];
798 // `hole` gets dropped and thus copies `tmp` into the remaining hole in `v`.
802 // When dropped, copies from `src` into `dest`.
803 struct InsertionHole<T> {
808 impl<T> Drop for InsertionHole<T> {
811 ptr::copy_nonoverlapping(self.src, self.dest, 1);
817 /// Merges non-decreasing runs `v[..mid]` and `v[mid..]` using `buf` as temporary storage, and
818 /// stores the result into `v[..]`.
822 /// The two slices must be non-empty and `mid` must be in bounds. Buffer `buf` must be long enough
823 /// to hold a copy of the shorter slice. Also, `T` must not be a zero-sized type.
824 unsafe fn merge<T, F>(v: &mut [T], mid: usize, buf: *mut T, is_less: &mut F)
826 F: FnMut(&T, &T) -> bool,
829 let v = v.as_mut_ptr();
830 let (v_mid, v_end) = unsafe { (v.add(mid), v.add(len)) };
832 // The merge process first copies the shorter run into `buf`. Then it traces the newly copied
833 // run and the longer run forwards (or backwards), comparing their next unconsumed elements and
834 // copying the lesser (or greater) one into `v`.
836 // As soon as the shorter run is fully consumed, the process is done. If the longer run gets
837 // consumed first, then we must copy whatever is left of the shorter run into the remaining
840 // Intermediate state of the process is always tracked by `hole`, which serves two purposes:
841 // 1. Protects integrity of `v` from panics in `is_less`.
842 // 2. Fills the remaining hole in `v` if the longer run gets consumed first.
846 // If `is_less` panics at any point during the process, `hole` will get dropped and fill the
847 // hole in `v` with the unconsumed range in `buf`, thus ensuring that `v` still holds every
848 // object it initially held exactly once.
851 if mid <= len - mid {
852 // The left run is shorter.
854 ptr::copy_nonoverlapping(v, buf, mid);
855 hole = MergeHole { start: buf, end: buf.add(mid), dest: v };
858 // Initially, these pointers point to the beginnings of their arrays.
859 let left = &mut hole.start;
860 let mut right = v_mid;
861 let out = &mut hole.dest;
863 while *left < hole.end && right < v_end {
864 // Consume the lesser side.
865 // If equal, prefer the left run to maintain stability.
867 let to_copy = if is_less(&*right, &**left) {
868 get_and_increment(&mut right)
870 get_and_increment(left)
872 ptr::copy_nonoverlapping(to_copy, get_and_increment(out), 1);
876 // The right run is shorter.
878 ptr::copy_nonoverlapping(v_mid, buf, len - mid);
879 hole = MergeHole { start: buf, end: buf.add(len - mid), dest: v_mid };
882 // Initially, these pointers point past the ends of their arrays.
883 let left = &mut hole.dest;
884 let right = &mut hole.end;
887 while v < *left && buf < *right {
888 // Consume the greater side.
889 // If equal, prefer the right run to maintain stability.
891 let to_copy = if is_less(&*right.offset(-1), &*left.offset(-1)) {
892 decrement_and_get(left)
894 decrement_and_get(right)
896 ptr::copy_nonoverlapping(to_copy, decrement_and_get(&mut out), 1);
900 // Finally, `hole` gets dropped. If the shorter run was not fully consumed, whatever remains of
901 // it will now be copied into the hole in `v`.
903 unsafe fn get_and_increment<T>(ptr: &mut *mut T) -> *mut T {
905 *ptr = unsafe { ptr.offset(1) };
909 unsafe fn decrement_and_get<T>(ptr: &mut *mut T) -> *mut T {
910 *ptr = unsafe { ptr.offset(-1) };
914 // When dropped, copies the range `start..end` into `dest..`.
915 struct MergeHole<T> {
921 impl<T> Drop for MergeHole<T> {
923 // `T` is not a zero-sized type, so it's okay to divide by its size.
924 let len = (self.end as usize - self.start as usize) / mem::size_of::<T>();
926 ptr::copy_nonoverlapping(self.start, self.dest, len);
932 /// This merge sort borrows some (but not all) ideas from TimSort, which is described in detail
933 /// [here](http://svn.python.org/projects/python/trunk/Objects/listsort.txt).
935 /// The algorithm identifies strictly descending and non-descending subsequences, which are called
936 /// natural runs. There is a stack of pending runs yet to be merged. Each newly found run is pushed
937 /// onto the stack, and then some pairs of adjacent runs are merged until these two invariants are
940 /// 1. for every `i` in `1..runs.len()`: `runs[i - 1].len > runs[i].len`
941 /// 2. for every `i` in `2..runs.len()`: `runs[i - 2].len > runs[i - 1].len + runs[i].len`
943 /// The invariants ensure that the total running time is `O(n * log(n))` worst-case.
944 fn merge_sort<T, F>(v: &mut [T], mut is_less: F)
946 F: FnMut(&T, &T) -> bool,
948 // Slices of up to this length get sorted using insertion sort.
949 const MAX_INSERTION: usize = 20;
950 // Very short runs are extended using insertion sort to span at least this many elements.
951 const MIN_RUN: usize = 10;
953 // Sorting has no meaningful behavior on zero-sized types.
954 if size_of::<T>() == 0 {
960 // Short arrays get sorted in-place via insertion sort to avoid allocations.
961 if len <= MAX_INSERTION {
963 for i in (0..len - 1).rev() {
964 insert_head(&mut v[i..], &mut is_less);
970 // Allocate a buffer to use as scratch memory. We keep the length 0 so we can keep in it
971 // shallow copies of the contents of `v` without risking the dtors running on copies if
972 // `is_less` panics. When merging two sorted runs, this buffer holds a copy of the shorter run,
973 // which will always have length at most `len / 2`.
974 let mut buf = Vec::with_capacity(len / 2);
976 // In order to identify natural runs in `v`, we traverse it backwards. That might seem like a
977 // strange decision, but consider the fact that merges more often go in the opposite direction
978 // (forwards). According to benchmarks, merging forwards is slightly faster than merging
979 // backwards. To conclude, identifying runs by traversing backwards improves performance.
980 let mut runs = vec![];
983 // Find the next natural run, and reverse it if it's strictly descending.
984 let mut start = end - 1;
988 if is_less(v.get_unchecked(start + 1), v.get_unchecked(start)) {
989 while start > 0 && is_less(v.get_unchecked(start), v.get_unchecked(start - 1)) {
992 v[start..end].reverse();
994 while start > 0 && !is_less(v.get_unchecked(start), v.get_unchecked(start - 1))
1002 // Insert some more elements into the run if it's too short. Insertion sort is faster than
1003 // merge sort on short sequences, so this significantly improves performance.
1004 while start > 0 && end - start < MIN_RUN {
1006 insert_head(&mut v[start..end], &mut is_less);
1009 // Push this run onto the stack.
1010 runs.push(Run { start, len: end - start });
1013 // Merge some pairs of adjacent runs to satisfy the invariants.
1014 while let Some(r) = collapse(&runs) {
1015 let left = runs[r + 1];
1016 let right = runs[r];
1019 &mut v[left.start..right.start + right.len],
1025 runs[r] = Run { start: left.start, len: left.len + right.len };
1030 // Finally, exactly one run must remain in the stack.
1031 debug_assert!(runs.len() == 1 && runs[0].start == 0 && runs[0].len == len);
1033 // Examines the stack of runs and identifies the next pair of runs to merge. More specifically,
1034 // if `Some(r)` is returned, that means `runs[r]` and `runs[r + 1]` must be merged next. If the
1035 // algorithm should continue building a new run instead, `None` is returned.
1037 // TimSort is infamous for its buggy implementations, as described here:
1038 // http://envisage-project.eu/timsort-specification-and-verification/
1040 // The gist of the story is: we must enforce the invariants on the top four runs on the stack.
1041 // Enforcing them on just top three is not sufficient to ensure that the invariants will still
1042 // hold for *all* runs in the stack.
1044 // This function correctly checks invariants for the top four runs. Additionally, if the top
1045 // run starts at index 0, it will always demand a merge operation until the stack is fully
1046 // collapsed, in order to complete the sort.
1048 fn collapse(runs: &[Run]) -> Option<usize> {
1051 && (runs[n - 1].start == 0
1052 || runs[n - 2].len <= runs[n - 1].len
1053 || (n >= 3 && runs[n - 3].len <= runs[n - 2].len + runs[n - 1].len)
1054 || (n >= 4 && runs[n - 4].len <= runs[n - 3].len + runs[n - 2].len))
1056 if n >= 3 && runs[n - 3].len < runs[n - 1].len { Some(n - 3) } else { Some(n - 2) }
1062 #[derive(Clone, Copy)]