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 //! [`Clone`]: ../../std/clone/trait.Clone.html
74 //! [`Eq`]: ../../std/cmp/trait.Eq.html
75 //! [`Ord`]: ../../std/cmp/trait.Ord.html
76 //! [`Iter`]: struct.Iter.html
77 //! [`Hash`]: ../../std/hash/trait.Hash.html
78 //! [`.iter`]: ../../std/primitive.slice.html#method.iter
79 //! [`.iter_mut`]: ../../std/primitive.slice.html#method.iter_mut
80 //! [`.split`]: ../../std/primitive.slice.html#method.split
81 //! [`.splitn`]: ../../std/primitive.slice.html#method.splitn
82 //! [`.chunks`]: ../../std/primitive.slice.html#method.chunks
83 //! [`.windows`]: ../../std/primitive.slice.html#method.windows
84 #![stable(feature = "rust1", since = "1.0.0")]
86 // Many of the usings in this module are only used in the test configuration.
87 // It's cleaner to just turn off the unused_imports warning than to fix them.
88 #![cfg_attr(test, allow(unused_imports, dead_code))]
90 use core::borrow::{Borrow, BorrowMut};
91 use core::cmp::Ordering::{self, Less};
92 use core::mem::{self, size_of};
94 use core::{u8, u16, u32};
96 use crate::borrow::ToOwned;
97 use crate::boxed::Box;
100 #[stable(feature = "rust1", since = "1.0.0")]
101 pub use core::slice::{Chunks, Windows};
102 #[stable(feature = "rust1", since = "1.0.0")]
103 pub use core::slice::{Iter, IterMut};
104 #[stable(feature = "rust1", since = "1.0.0")]
105 pub use core::slice::{SplitMut, ChunksMut, Split};
106 #[stable(feature = "rust1", since = "1.0.0")]
107 pub use core::slice::{SplitN, RSplitN, SplitNMut, RSplitNMut};
108 #[stable(feature = "slice_rsplit", since = "1.27.0")]
109 pub use core::slice::{RSplit, RSplitMut};
110 #[stable(feature = "rust1", since = "1.0.0")]
111 pub use core::slice::{from_raw_parts, from_raw_parts_mut};
112 #[stable(feature = "from_ref", since = "1.28.0")]
113 pub use core::slice::{from_ref, from_mut};
114 #[stable(feature = "slice_get_slice", since = "1.28.0")]
115 pub use core::slice::SliceIndex;
116 #[stable(feature = "chunks_exact", since = "1.31.0")]
117 pub use core::slice::{ChunksExact, ChunksExactMut};
118 #[stable(feature = "rchunks", since = "1.31.0")]
119 pub use core::slice::{RChunks, RChunksMut, RChunksExact, RChunksExactMut};
121 ////////////////////////////////////////////////////////////////////////////////
122 // Basic slice extension methods
123 ////////////////////////////////////////////////////////////////////////////////
125 // HACK(japaric) needed for the implementation of `vec!` macro during testing
126 // NB see the hack module in this file for more details
128 pub use hack::into_vec;
130 // HACK(japaric) needed for the implementation of `Vec::clone` during testing
131 // NB see the hack module in this file for more details
133 pub use hack::to_vec;
135 // HACK(japaric): With cfg(test) `impl [T]` is not available, these three
136 // functions are actually methods that are in `impl [T]` but not in
137 // `core::slice::SliceExt` - we need to supply these functions for the
138 // `test_permutations` test
142 use crate::boxed::Box;
145 use crate::string::ToString;
147 pub fn into_vec<T>(mut b: Box<[T]>) -> Vec<T> {
149 let xs = Vec::from_raw_parts(b.as_mut_ptr(), b.len(), b.len());
156 pub fn to_vec<T>(s: &[T]) -> Vec<T>
159 let mut vector = Vec::with_capacity(s.len());
160 vector.extend_from_slice(s);
165 #[lang = "slice_alloc"]
170 /// This sort is stable (i.e., does not reorder equal elements) and `O(n log n)` worst-case.
172 /// When applicable, unstable sorting is preferred because it is generally faster than stable
173 /// sorting and it doesn't allocate auxiliary memory.
174 /// See [`sort_unstable`](#method.sort_unstable).
176 /// # Current implementation
178 /// The current algorithm is an adaptive, iterative merge sort inspired by
179 /// [timsort](https://en.wikipedia.org/wiki/Timsort).
180 /// It is designed to be very fast in cases where the slice is nearly sorted, or consists of
181 /// two or more sorted sequences concatenated one after another.
183 /// Also, it allocates temporary storage half the size of `self`, but for short slices a
184 /// non-allocating insertion sort is used instead.
189 /// let mut v = [-5, 4, 1, -3, 2];
192 /// assert!(v == [-5, -3, 1, 2, 4]);
194 #[stable(feature = "rust1", since = "1.0.0")]
196 pub fn sort(&mut self)
199 merge_sort(self, |a, b| a.lt(b));
202 /// Sorts the slice with a comparator function.
204 /// This sort is stable (i.e., does not reorder equal elements) and `O(n log n)` worst-case.
206 /// The comparator function must define a total ordering for the elements in the slice. If
207 /// the ordering is not total, the order of the elements is unspecified. An order is a
208 /// total order if it is (for all `a`, `b` and `c`):
210 /// * total and antisymmetric: exactly one of `a < b`, `a == b` or `a > b` is true, and
211 /// * transitive, `a < b` and `b < c` implies `a < c`. The same must hold for both `==` and `>`.
213 /// For example, while [`f64`] doesn't implement [`Ord`] because `NaN != NaN`, we can use
214 /// `partial_cmp` as our sort function when we know the slice doesn't contain a `NaN`.
217 /// let mut floats = [5f64, 4.0, 1.0, 3.0, 2.0];
218 /// floats.sort_by(|a, b| a.partial_cmp(b).unwrap());
219 /// assert_eq!(floats, [1.0, 2.0, 3.0, 4.0, 5.0]);
222 /// When applicable, unstable sorting is preferred because it is generally faster than stable
223 /// sorting and it doesn't allocate auxiliary memory.
224 /// See [`sort_unstable_by`](#method.sort_unstable_by).
226 /// # Current implementation
228 /// The current algorithm is an adaptive, iterative merge sort inspired by
229 /// [timsort](https://en.wikipedia.org/wiki/Timsort).
230 /// It is designed to be very fast in cases where the slice is nearly sorted, or consists of
231 /// two or more sorted sequences concatenated one after another.
233 /// Also, it allocates temporary storage half the size of `self`, but for short slices a
234 /// non-allocating insertion sort is used instead.
239 /// let mut v = [5, 4, 1, 3, 2];
240 /// v.sort_by(|a, b| a.cmp(b));
241 /// assert!(v == [1, 2, 3, 4, 5]);
243 /// // reverse sorting
244 /// v.sort_by(|a, b| b.cmp(a));
245 /// assert!(v == [5, 4, 3, 2, 1]);
247 #[stable(feature = "rust1", since = "1.0.0")]
249 pub fn sort_by<F>(&mut self, mut compare: F)
250 where F: FnMut(&T, &T) -> Ordering
252 merge_sort(self, |a, b| compare(a, b) == Less);
255 /// Sorts the slice with a key extraction function.
257 /// This sort is stable (i.e., does not reorder equal elements) and `O(m n log(m n))`
258 /// worst-case, where the key function is `O(m)`.
260 /// For expensive key functions (e.g. functions that are not simple property accesses or
261 /// basic operations), [`sort_by_cached_key`](#method.sort_by_cached_key) is likely to be
262 /// significantly faster, as it does not recompute element keys.
264 /// When applicable, unstable sorting is preferred because it is generally faster than stable
265 /// sorting and it doesn't allocate auxiliary memory.
266 /// See [`sort_unstable_by_key`](#method.sort_unstable_by_key).
268 /// # Current implementation
270 /// The current algorithm is an adaptive, iterative merge sort inspired by
271 /// [timsort](https://en.wikipedia.org/wiki/Timsort).
272 /// It is designed to be very fast in cases where the slice is nearly sorted, or consists of
273 /// two or more sorted sequences concatenated one after another.
275 /// Also, it allocates temporary storage half the size of `self`, but for short slices a
276 /// non-allocating insertion sort is used instead.
281 /// let mut v = [-5i32, 4, 1, -3, 2];
283 /// v.sort_by_key(|k| k.abs());
284 /// assert!(v == [1, 2, -3, 4, -5]);
286 #[stable(feature = "slice_sort_by_key", since = "1.7.0")]
288 pub fn sort_by_key<K, F>(&mut self, mut f: F)
289 where F: FnMut(&T) -> K, K: Ord
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)
329 where F: FnMut(&T) -> K, K: Ord
331 // Helper macro for indexing our vector by the smallest possible type, to reduce allocation.
332 macro_rules! sort_by_key {
333 ($t:ty, $slice:ident, $f:ident) => ({
334 let mut indices: Vec<_> =
335 $slice.iter().map($f).enumerate().map(|(i, k)| (k, i as $t)).collect();
336 // The elements of `indices` are unique, as they are indexed, so any sort will be
337 // stable with respect to the original slice. We use `sort_unstable` here because
338 // it requires less memory allocation.
339 indices.sort_unstable();
340 for i in 0..$slice.len() {
341 let mut index = indices[i].1;
342 while (index as usize) < i {
343 index = indices[index as usize].1;
345 indices[i].1 = index;
346 $slice.swap(i, index as usize);
351 let sz_u8 = mem::size_of::<(K, u8)>();
352 let sz_u16 = mem::size_of::<(K, u16)>();
353 let sz_u32 = mem::size_of::<(K, u32)>();
354 let sz_usize = mem::size_of::<(K, usize)>();
356 let len = self.len();
357 if len < 2 { return }
358 if sz_u8 < sz_u16 && len <= ( u8::MAX as usize) { return sort_by_key!( u8, self, f) }
359 if sz_u16 < sz_u32 && len <= (u16::MAX as usize) { return sort_by_key!(u16, self, f) }
360 if sz_u32 < sz_usize && len <= (u32::MAX as usize) { return sort_by_key!(u32, self, f) }
361 sort_by_key!(usize, self, f)
364 /// Copies `self` into a new `Vec`.
369 /// let s = [10, 40, 30];
370 /// let x = s.to_vec();
371 /// // Here, `s` and `x` can be modified independently.
373 #[rustc_conversion_suggestion]
374 #[stable(feature = "rust1", since = "1.0.0")]
376 pub fn to_vec(&self) -> Vec<T>
379 // NB see hack module in this file
383 /// Converts `self` into a vector without clones or allocation.
385 /// The resulting vector can be converted back into a box via
386 /// `Vec<T>`'s `into_boxed_slice` method.
391 /// let s: Box<[i32]> = Box::new([10, 40, 30]);
392 /// let x = s.into_vec();
393 /// // `s` cannot be used anymore because it has been converted into `x`.
395 /// assert_eq!(x, vec![10, 40, 30]);
397 #[stable(feature = "rust1", since = "1.0.0")]
399 pub fn into_vec(self: Box<Self>) -> Vec<T> {
400 // NB see hack module in this file
404 /// Creates a vector by repeating a slice `n` times.
408 /// This function will panic if the capacity would overflow.
415 /// #![feature(repeat_generic_slice)]
418 /// assert_eq!([1, 2].repeat(3), vec![1, 2, 1, 2, 1, 2]);
422 /// A panic upon overflow:
425 /// #![feature(repeat_generic_slice)]
427 /// // this will panic at runtime
428 /// b"0123456789abcdef".repeat(usize::max_value());
431 #[unstable(feature = "repeat_generic_slice",
432 reason = "it's on str, why not on slice?",
434 pub fn repeat(&self, n: usize) -> Vec<T> where T: Copy {
439 // If `n` is larger than zero, it can be split as
440 // `n = 2^expn + rem (2^expn > rem, expn >= 0, rem >= 0)`.
441 // `2^expn` is the number represented by the leftmost '1' bit of `n`,
442 // and `rem` is the remaining part of `n`.
444 // Using `Vec` to access `set_len()`.
445 let mut buf = Vec::with_capacity(self.len().checked_mul(n).expect("capacity overflow"));
447 // `2^expn` repetition is done by doubling `buf` `expn`-times.
451 // If `m > 0`, there are remaining bits up to the leftmost '1'.
453 // `buf.extend(buf)`:
455 ptr::copy_nonoverlapping(
457 (buf.as_mut_ptr() as *mut T).add(buf.len()),
460 // `buf` has capacity of `self.len() * n`.
461 let buf_len = buf.len();
462 buf.set_len(buf_len * 2);
469 // `rem` (`= n - 2^expn`) repetition is done by copying
470 // first `rem` repetitions from `buf` itself.
471 let rem_len = self.len() * n - buf.len(); // `self.len() * rem`
473 // `buf.extend(buf[0 .. rem_len])`:
475 // This is non-overlapping since `2^expn > rem`.
476 ptr::copy_nonoverlapping(
478 (buf.as_mut_ptr() as *mut T).add(buf.len()),
481 // `buf.len() + rem_len` equals to `buf.capacity()` (`= self.len() * n`).
482 let buf_cap = buf.capacity();
483 buf.set_len(buf_cap);
490 #[lang = "slice_u8_alloc"]
493 /// Returns a vector containing a copy of this slice where each byte
494 /// is mapped to its ASCII upper case equivalent.
496 /// ASCII letters 'a' to 'z' are mapped to 'A' to 'Z',
497 /// but non-ASCII letters are unchanged.
499 /// To uppercase the value in-place, use [`make_ascii_uppercase`].
501 /// [`make_ascii_uppercase`]: #method.make_ascii_uppercase
502 #[stable(feature = "ascii_methods_on_intrinsics", since = "1.23.0")]
504 pub fn to_ascii_uppercase(&self) -> Vec<u8> {
505 let mut me = self.to_vec();
506 me.make_ascii_uppercase();
510 /// Returns a vector containing a copy of this slice where each byte
511 /// is mapped to its ASCII lower case equivalent.
513 /// ASCII letters 'A' to 'Z' are mapped to 'a' to 'z',
514 /// but non-ASCII letters are unchanged.
516 /// To lowercase the value in-place, use [`make_ascii_lowercase`].
518 /// [`make_ascii_lowercase`]: #method.make_ascii_lowercase
519 #[stable(feature = "ascii_methods_on_intrinsics", since = "1.23.0")]
521 pub fn to_ascii_lowercase(&self) -> Vec<u8> {
522 let mut me = self.to_vec();
523 me.make_ascii_lowercase();
528 ////////////////////////////////////////////////////////////////////////////////
529 // Extension traits for slices over specific kinds of data
530 ////////////////////////////////////////////////////////////////////////////////
531 #[unstable(feature = "slice_concat_ext",
532 reason = "trait should not have to exist",
534 /// An extension trait for concatenating slices
536 /// While this trait is unstable, the methods are stable. `SliceConcatExt` is
537 /// included in the [standard library prelude], so you can use [`join()`] and
538 /// [`concat()`] as if they existed on `[T]` itself.
540 /// [standard library prelude]: ../../std/prelude/index.html
541 /// [`join()`]: #tymethod.join
542 /// [`concat()`]: #tymethod.concat
543 pub trait SliceConcatExt<T: ?Sized> {
544 #[unstable(feature = "slice_concat_ext",
545 reason = "trait should not have to exist",
547 /// The resulting type after concatenation
550 /// Flattens a slice of `T` into a single value `Self::Output`.
555 /// assert_eq!(["hello", "world"].concat(), "helloworld");
556 /// assert_eq!([[1, 2], [3, 4]].concat(), [1, 2, 3, 4]);
558 #[stable(feature = "rust1", since = "1.0.0")]
559 fn concat(&self) -> Self::Output;
561 /// Flattens a slice of `T` into a single value `Self::Output`, placing a
562 /// given separator between each.
567 /// assert_eq!(["hello", "world"].join(" "), "hello world");
568 /// assert_eq!([[1, 2], [3, 4]].join(&0), [1, 2, 0, 3, 4]);
570 #[stable(feature = "rename_connect_to_join", since = "1.3.0")]
571 fn join(&self, sep: &T) -> Self::Output;
573 #[stable(feature = "rust1", since = "1.0.0")]
574 #[rustc_deprecated(since = "1.3.0", reason = "renamed to join")]
575 fn connect(&self, sep: &T) -> Self::Output;
578 #[unstable(feature = "slice_concat_ext",
579 reason = "trait should not have to exist",
581 impl<T: Clone, V: Borrow<[T]>> SliceConcatExt<T> for [V] {
582 type Output = Vec<T>;
584 fn concat(&self) -> Vec<T> {
585 let size = self.iter().map(|slice| slice.borrow().len()).sum();
586 let mut result = Vec::with_capacity(size);
588 result.extend_from_slice(v.borrow())
593 fn join(&self, sep: &T) -> Vec<T> {
594 let mut iter = self.iter();
595 let first = match iter.next() {
596 Some(first) => first,
597 None => return vec![],
599 let size = self.iter().map(|slice| slice.borrow().len()).sum::<usize>() + self.len() - 1;
600 let mut result = Vec::with_capacity(size);
601 result.extend_from_slice(first.borrow());
604 result.push(sep.clone());
605 result.extend_from_slice(v.borrow())
610 fn connect(&self, sep: &T) -> Vec<T> {
615 ////////////////////////////////////////////////////////////////////////////////
616 // Standard trait implementations for slices
617 ////////////////////////////////////////////////////////////////////////////////
619 #[stable(feature = "rust1", since = "1.0.0")]
620 impl<T> Borrow<[T]> for Vec<T> {
621 fn borrow(&self) -> &[T] {
626 #[stable(feature = "rust1", since = "1.0.0")]
627 impl<T> BorrowMut<[T]> for Vec<T> {
628 fn borrow_mut(&mut self) -> &mut [T] {
633 #[stable(feature = "rust1", since = "1.0.0")]
634 impl<T: Clone> ToOwned for [T] {
637 fn to_owned(&self) -> Vec<T> {
642 fn to_owned(&self) -> Vec<T> {
646 fn clone_into(&self, target: &mut Vec<T>) {
647 // drop anything in target that will not be overwritten
648 target.truncate(self.len());
649 let len = target.len();
651 // reuse the contained values' allocations/resources.
652 target.clone_from_slice(&self[..len]);
654 // target.len <= self.len due to the truncate above, so the
655 // slice here is always in-bounds.
656 target.extend_from_slice(&self[len..]);
660 ////////////////////////////////////////////////////////////////////////////////
662 ////////////////////////////////////////////////////////////////////////////////
664 /// Inserts `v[0]` into pre-sorted sequence `v[1..]` so that whole `v[..]` becomes sorted.
666 /// This is the integral subroutine of insertion sort.
667 fn insert_head<T, F>(v: &mut [T], is_less: &mut F)
668 where F: FnMut(&T, &T) -> bool
670 if v.len() >= 2 && is_less(&v[1], &v[0]) {
672 // There are three ways to implement insertion here:
674 // 1. Swap adjacent elements until the first one gets to its final destination.
675 // However, this way we copy data around more than is necessary. If elements are big
676 // structures (costly to copy), this method will be slow.
678 // 2. Iterate until the right place for the first element is found. Then shift the
679 // elements succeeding it to make room for it and finally place it into the
680 // remaining hole. This is a good method.
682 // 3. Copy the first element into a temporary variable. Iterate until the right place
683 // for it is found. As we go along, copy every traversed element into the slot
684 // preceding it. Finally, copy data from the temporary variable into the remaining
685 // hole. This method is very good. Benchmarks demonstrated slightly better
686 // performance than with the 2nd method.
688 // All methods were benchmarked, and the 3rd showed best results. So we chose that one.
689 let mut tmp = mem::ManuallyDrop::new(ptr::read(&v[0]));
691 // Intermediate state of the insertion process is always tracked by `hole`, which
692 // serves two purposes:
693 // 1. Protects integrity of `v` from panics in `is_less`.
694 // 2. Fills the remaining hole in `v` in the end.
698 // If `is_less` panics at any point during the process, `hole` will get dropped and
699 // fill the hole in `v` with `tmp`, thus ensuring that `v` still holds every object it
700 // initially held exactly once.
701 let mut hole = InsertionHole {
705 ptr::copy_nonoverlapping(&v[1], &mut v[0], 1);
707 for i in 2..v.len() {
708 if !is_less(&v[i], &*tmp) {
711 ptr::copy_nonoverlapping(&v[i], &mut v[i - 1], 1);
712 hole.dest = &mut v[i];
714 // `hole` gets dropped and thus copies `tmp` into the remaining hole in `v`.
718 // When dropped, copies from `src` into `dest`.
719 struct InsertionHole<T> {
724 impl<T> Drop for InsertionHole<T> {
726 unsafe { ptr::copy_nonoverlapping(self.src, self.dest, 1); }
731 /// Merges non-decreasing runs `v[..mid]` and `v[mid..]` using `buf` as temporary storage, and
732 /// stores the result into `v[..]`.
736 /// The two slices must be non-empty and `mid` must be in bounds. Buffer `buf` must be long enough
737 /// to hold a copy of the shorter slice. Also, `T` must not be a zero-sized type.
738 unsafe fn merge<T, F>(v: &mut [T], mid: usize, buf: *mut T, is_less: &mut F)
739 where F: FnMut(&T, &T) -> bool
742 let v = v.as_mut_ptr();
743 let v_mid = v.add(mid);
744 let v_end = v.add(len);
746 // The merge process first copies the shorter run into `buf`. Then it traces the newly copied
747 // run and the longer run forwards (or backwards), comparing their next unconsumed elements and
748 // copying the lesser (or greater) one into `v`.
750 // As soon as the shorter run is fully consumed, the process is done. If the longer run gets
751 // consumed first, then we must copy whatever is left of the shorter run into the remaining
754 // Intermediate state of the process is always tracked by `hole`, which serves two purposes:
755 // 1. Protects integrity of `v` from panics in `is_less`.
756 // 2. Fills the remaining hole in `v` if the longer run gets consumed first.
760 // If `is_less` panics at any point during the process, `hole` will get dropped and fill the
761 // hole in `v` with the unconsumed range in `buf`, thus ensuring that `v` still holds every
762 // object it initially held exactly once.
765 if mid <= len - mid {
766 // The left run is shorter.
767 ptr::copy_nonoverlapping(v, buf, mid);
774 // Initially, these pointers point to the beginnings of their arrays.
775 let left = &mut hole.start;
776 let mut right = v_mid;
777 let out = &mut hole.dest;
779 while *left < hole.end && right < v_end {
780 // Consume the lesser side.
781 // If equal, prefer the left run to maintain stability.
782 let to_copy = if is_less(&*right, &**left) {
783 get_and_increment(&mut right)
785 get_and_increment(left)
787 ptr::copy_nonoverlapping(to_copy, get_and_increment(out), 1);
790 // The right run is shorter.
791 ptr::copy_nonoverlapping(v_mid, buf, len - mid);
794 end: buf.add(len - mid),
798 // Initially, these pointers point past the ends of their arrays.
799 let left = &mut hole.dest;
800 let right = &mut hole.end;
803 while v < *left && buf < *right {
804 // Consume the greater side.
805 // If equal, prefer the right run to maintain stability.
806 let to_copy = if is_less(&*right.offset(-1), &*left.offset(-1)) {
807 decrement_and_get(left)
809 decrement_and_get(right)
811 ptr::copy_nonoverlapping(to_copy, decrement_and_get(&mut out), 1);
814 // Finally, `hole` gets dropped. If the shorter run was not fully consumed, whatever remains of
815 // it will now be copied into the hole in `v`.
817 unsafe fn get_and_increment<T>(ptr: &mut *mut T) -> *mut T {
819 *ptr = ptr.offset(1);
823 unsafe fn decrement_and_get<T>(ptr: &mut *mut T) -> *mut T {
824 *ptr = ptr.offset(-1);
828 // When dropped, copies the range `start..end` into `dest..`.
829 struct MergeHole<T> {
835 impl<T> Drop for MergeHole<T> {
837 // `T` is not a zero-sized type, so it's okay to divide by its size.
838 let len = (self.end as usize - self.start as usize) / mem::size_of::<T>();
839 unsafe { ptr::copy_nonoverlapping(self.start, self.dest, len); }
844 /// This merge sort borrows some (but not all) ideas from TimSort, which is described in detail
845 /// [here](http://svn.python.org/projects/python/trunk/Objects/listsort.txt).
847 /// The algorithm identifies strictly descending and non-descending subsequences, which are called
848 /// natural runs. There is a stack of pending runs yet to be merged. Each newly found run is pushed
849 /// onto the stack, and then some pairs of adjacent runs are merged until these two invariants are
852 /// 1. for every `i` in `1..runs.len()`: `runs[i - 1].len > runs[i].len`
853 /// 2. for every `i` in `2..runs.len()`: `runs[i - 2].len > runs[i - 1].len + runs[i].len`
855 /// The invariants ensure that the total running time is `O(n log n)` worst-case.
856 fn merge_sort<T, F>(v: &mut [T], mut is_less: F)
857 where F: FnMut(&T, &T) -> bool
859 // Slices of up to this length get sorted using insertion sort.
860 const MAX_INSERTION: usize = 20;
861 // Very short runs are extended using insertion sort to span at least this many elements.
862 const MIN_RUN: usize = 10;
864 // Sorting has no meaningful behavior on zero-sized types.
865 if size_of::<T>() == 0 {
871 // Short arrays get sorted in-place via insertion sort to avoid allocations.
872 if len <= MAX_INSERTION {
874 for i in (0..len-1).rev() {
875 insert_head(&mut v[i..], &mut is_less);
881 // Allocate a buffer to use as scratch memory. We keep the length 0 so we can keep in it
882 // shallow copies of the contents of `v` without risking the dtors running on copies if
883 // `is_less` panics. When merging two sorted runs, this buffer holds a copy of the shorter run,
884 // which will always have length at most `len / 2`.
885 let mut buf = Vec::with_capacity(len / 2);
887 // In order to identify natural runs in `v`, we traverse it backwards. That might seem like a
888 // strange decision, but consider the fact that merges more often go in the opposite direction
889 // (forwards). According to benchmarks, merging forwards is slightly faster than merging
890 // backwards. To conclude, identifying runs by traversing backwards improves performance.
891 let mut runs = vec![];
894 // Find the next natural run, and reverse it if it's strictly descending.
895 let mut start = end - 1;
899 if is_less(v.get_unchecked(start + 1), v.get_unchecked(start)) {
900 while start > 0 && is_less(v.get_unchecked(start),
901 v.get_unchecked(start - 1)) {
904 v[start..end].reverse();
906 while start > 0 && !is_less(v.get_unchecked(start),
907 v.get_unchecked(start - 1)) {
914 // Insert some more elements into the run if it's too short. Insertion sort is faster than
915 // merge sort on short sequences, so this significantly improves performance.
916 while start > 0 && end - start < MIN_RUN {
918 insert_head(&mut v[start..end], &mut is_less);
921 // Push this run onto the stack.
928 // Merge some pairs of adjacent runs to satisfy the invariants.
929 while let Some(r) = collapse(&runs) {
930 let left = runs[r + 1];
933 merge(&mut v[left.start .. right.start + right.len], left.len, buf.as_mut_ptr(),
938 len: left.len + right.len,
944 // Finally, exactly one run must remain in the stack.
945 debug_assert!(runs.len() == 1 && runs[0].start == 0 && runs[0].len == len);
947 // Examines the stack of runs and identifies the next pair of runs to merge. More specifically,
948 // if `Some(r)` is returned, that means `runs[r]` and `runs[r + 1]` must be merged next. If the
949 // algorithm should continue building a new run instead, `None` is returned.
951 // TimSort is infamous for its buggy implementations, as described here:
952 // http://envisage-project.eu/timsort-specification-and-verification/
954 // The gist of the story is: we must enforce the invariants on the top four runs on the stack.
955 // Enforcing them on just top three is not sufficient to ensure that the invariants will still
956 // hold for *all* runs in the stack.
958 // This function correctly checks invariants for the top four runs. Additionally, if the top
959 // run starts at index 0, it will always demand a merge operation until the stack is fully
960 // collapsed, in order to complete the sort.
962 fn collapse(runs: &[Run]) -> Option<usize> {
964 if n >= 2 && (runs[n - 1].start == 0 ||
965 runs[n - 2].len <= runs[n - 1].len ||
966 (n >= 3 && runs[n - 3].len <= runs[n - 2].len + runs[n - 1].len) ||
967 (n >= 4 && runs[n - 4].len <= runs[n - 3].len + runs[n - 2].len)) {
968 if n >= 3 && runs[n - 3].len < runs[n - 1].len {
978 #[derive(Clone, Copy)]