1 // Copyright 2014 The Rust Project Developers. See the COPYRIGHT
2 // file at the top-level directory of this distribution and at
3 // http://rust-lang.org/COPYRIGHT.
5 // Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
6 // http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
7 // <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
8 // option. This file may not be copied, modified, or distributed
9 // except according to those terms.
11 // ignore-lexer-test FIXME #15883
15 use hash::{Hash, Hasher};
16 use iter::{Iterator, count};
17 use kinds::{Sized, marker};
18 use mem::{min_align_of, size_of};
20 use num::{Int, UnsignedInt};
21 use ops::{Deref, DerefMut, Drop};
22 use option::{Some, None, Option};
23 use ptr::{RawPtr, copy_nonoverlapping_memory, zero_memory};
25 use rt::heap::{allocate, deallocate};
27 const EMPTY_BUCKET: u64 = 0u64;
29 /// The raw hashtable, providing safe-ish access to the unzipped and highly
30 /// optimized arrays of hashes, keys, and values.
32 /// This design uses less memory and is a lot faster than the naive
33 /// `Vec<Option<u64, K, V>>`, because we don't pay for the overhead of an
34 /// option on every element, and we get a generally more cache-aware design.
36 /// Essential invariants of this structure:
38 /// - if t.hashes[i] == EMPTY_BUCKET, then `Bucket::at_index(&t, i).raw`
39 /// points to 'undefined' contents. Don't read from it. This invariant is
40 /// enforced outside this module with the `EmptyBucket`, `FullBucket`,
41 /// and `SafeHash` types.
43 /// - An `EmptyBucket` is only constructed at an index with
44 /// a hash of EMPTY_BUCKET.
46 /// - A `FullBucket` is only constructed at an index with a
47 /// non-EMPTY_BUCKET hash.
49 /// - A `SafeHash` is only constructed for non-`EMPTY_BUCKET` hash. We get
50 /// around hashes of zero by changing them to 0x8000_0000_0000_0000,
51 /// which will likely map to the same bucket, while not being confused
54 /// - All three "arrays represented by pointers" are the same length:
55 /// `capacity`. This is set at creation and never changes. The arrays
56 /// are unzipped to save space (we don't have to pay for the padding
57 /// between odd sized elements, such as in a map from u64 to u8), and
58 /// be more cache aware (scanning through 8 hashes brings in at most
59 /// 2 cache lines, since they're all right beside each other).
61 /// You can kind of think of this module/data structure as a safe wrapper
62 /// around just the "table" part of the hashtable. It enforces some
63 /// invariants at the type level and employs some performance trickery,
64 /// but in general is just a tricked out `Vec<Option<u64, K, V>>`.
65 #[unsafe_no_drop_flag]
66 pub struct RawTable<K, V> {
70 // Because K/V do not appear directly in any of the types in the struct,
71 // inform rustc that in fact instances of K and V are reachable from here.
72 marker: marker::CovariantType<(K,V)>,
75 struct RawBucket<K, V> {
81 pub struct Bucket<K, V, M> {
87 pub struct EmptyBucket<K, V, M> {
93 pub struct FullBucket<K, V, M> {
99 pub type EmptyBucketImm<'table, K, V> = EmptyBucket<K, V, &'table RawTable<K, V>>;
100 pub type FullBucketImm<'table, K, V> = FullBucket<K, V, &'table RawTable<K, V>>;
102 pub type EmptyBucketMut<'table, K, V> = EmptyBucket<K, V, &'table mut RawTable<K, V>>;
103 pub type FullBucketMut<'table, K, V> = FullBucket<K, V, &'table mut RawTable<K, V>>;
105 pub enum BucketState<K, V, M> {
106 Empty(EmptyBucket<K, V, M>),
107 Full(FullBucket<K, V, M>),
110 // A GapThenFull encapsulates the state of two consecutive buckets at once.
111 // The first bucket, called the gap, is known to be empty.
112 // The second bucket is full.
113 struct GapThenFull<K, V, M> {
114 gap: EmptyBucket<K, V, ()>,
115 full: FullBucket<K, V, M>,
118 /// A hash that is not zero, since we use a hash of zero to represent empty
120 #[deriving(PartialEq)]
121 pub struct SafeHash {
126 /// Peek at the hash value, which is guaranteed to be non-zero.
128 pub fn inspect(&self) -> u64 { self.hash }
131 /// We need to remove hashes of 0. That's reserved for empty buckets.
132 /// This function wraps up `hash_keyed` to be the only way outside this
133 /// module to generate a SafeHash.
134 pub fn make_hash<Sized? T: Hash<S>, S, H: Hasher<S>>(hasher: &H, t: &T) -> SafeHash {
135 match hasher.hash(t) {
136 // This constant is exceedingly likely to hash to the same
137 // bucket, but it won't be counted as empty! Just so we can maintain
138 // our precious uniform distribution of initial indexes.
139 EMPTY_BUCKET => SafeHash { hash: 0x8000_0000_0000_0000 },
140 h => SafeHash { hash: h },
144 // `replace` casts a `*u64` to a `*SafeHash`. Since we statically
145 // ensure that a `FullBucket` points to an index with a non-zero hash,
146 // and a `SafeHash` is just a `u64` with a different name, this is
149 // This test ensures that a `SafeHash` really IS the same size as a
150 // `u64`. If you need to change the size of `SafeHash` (and
151 // consequently made this test fail), `replace` needs to be
152 // modified to no longer assume this.
154 fn can_alias_safehash_as_u64() {
155 assert_eq!(size_of::<SafeHash>(), size_of::<u64>())
158 impl<K, V> RawBucket<K, V> {
159 unsafe fn offset(self, count: int) -> RawBucket<K, V> {
161 hash: self.hash.offset(count),
162 key: self.key.offset(count),
163 val: self.val.offset(count),
168 // Buckets hold references to the table.
169 impl<K, V, M> FullBucket<K, V, M> {
170 /// Borrow a reference to the table.
171 pub fn table(&self) -> &M {
174 /// Move out the reference to the table.
175 pub fn into_table(self) -> M {
178 /// Get the raw index.
179 pub fn index(&self) -> uint {
184 impl<K, V, M> EmptyBucket<K, V, M> {
185 /// Borrow a reference to the table.
186 pub fn table(&self) -> &M {
189 /// Move out the reference to the table.
190 pub fn into_table(self) -> M {
195 impl<K, V, M> Bucket<K, V, M> {
196 /// Move out the reference to the table.
197 pub fn into_table(self) -> M {
200 /// Get the raw index.
201 pub fn index(&self) -> uint {
206 impl<K, V, M: Deref<RawTable<K, V>>> Bucket<K, V, M> {
207 pub fn new(table: M, hash: &SafeHash) -> Bucket<K, V, M> {
208 Bucket::at_index(table, hash.inspect() as uint)
211 pub fn at_index(table: M, ib_index: uint) -> Bucket<K, V, M> {
212 let ib_index = ib_index & (table.capacity() - 1);
215 table.first_bucket_raw().offset(ib_index as int)
222 pub fn first(table: M) -> Bucket<K, V, M> {
224 raw: table.first_bucket_raw(),
230 /// Reads a bucket at a given index, returning an enum indicating whether
231 /// it's initialized or not. You need to match on this enum to get
232 /// the appropriate types to call most of the other functions in
234 pub fn peek(self) -> BucketState<K, V, M> {
235 match unsafe { *self.raw.hash } {
251 /// Modifies the bucket pointer in place to make it point to the next slot.
252 pub fn next(&mut self) {
253 // Branchless bucket iteration step.
254 // As we reach the end of the table...
255 // We take the current idx: 0111111b
256 // Xor it by its increment: ^ 1000000b
259 // Then AND with the capacity: & 1000000b
261 // to get the backwards offset: 1000000b
262 // ... and it's zero at all other times.
263 let maybe_wraparound_dist = (self.idx ^ (self.idx + 1)) & self.table.capacity();
264 // Finally, we obtain the offset 1 or the offset -cap + 1.
265 let dist = 1i - (maybe_wraparound_dist as int);
270 self.raw = self.raw.offset(dist);
275 impl<K, V, M: Deref<RawTable<K, V>>> EmptyBucket<K, V, M> {
277 pub fn next(self) -> Bucket<K, V, M> {
278 let mut bucket = self.into_bucket();
284 pub fn into_bucket(self) -> Bucket<K, V, M> {
292 pub fn gap_peek(self) -> Option<GapThenFull<K, V, M>> {
293 let gap = EmptyBucket {
299 match self.next().peek() {
311 impl<K, V, M: DerefMut<RawTable<K, V>>> EmptyBucket<K, V, M> {
312 /// Puts given key and value pair, along with the key's hash,
313 /// into this bucket in the hashtable. Note how `self` is 'moved' into
314 /// this function, because this slot will no longer be empty when
315 /// we return! A `FullBucket` is returned for later use, pointing to
316 /// the newly-filled slot in the hashtable.
318 /// Use `make_hash` to construct a `SafeHash` to pass to this function.
319 pub fn put(mut self, hash: SafeHash, key: K, value: V)
320 -> FullBucket<K, V, M> {
322 *self.raw.hash = hash.inspect();
323 ptr::write(self.raw.key, key);
324 ptr::write(self.raw.val, value);
327 self.table.size += 1;
329 FullBucket { raw: self.raw, idx: self.idx, table: self.table }
333 impl<K, V, M: Deref<RawTable<K, V>>> FullBucket<K, V, M> {
335 pub fn next(self) -> Bucket<K, V, M> {
336 let mut bucket = self.into_bucket();
342 pub fn into_bucket(self) -> Bucket<K, V, M> {
350 /// Get the distance between this bucket and the 'ideal' location
351 /// as determined by the key's hash stored in it.
353 /// In the cited blog posts above, this is called the "distance to
354 /// initial bucket", or DIB. Also known as "probe count".
355 pub fn distance(&self) -> uint {
356 // Calculates the distance one has to travel when going from
357 // `hash mod capacity` onwards to `idx mod capacity`, wrapping around
358 // if the destination is not reached before the end of the table.
359 (self.idx - self.hash().inspect() as uint) & (self.table.capacity() - 1)
363 pub fn hash(&self) -> SafeHash {
371 /// Gets references to the key and value at a given index.
372 pub fn read(&self) -> (&K, &V) {
380 impl<K, V, M: DerefMut<RawTable<K, V>>> FullBucket<K, V, M> {
381 /// Removes this bucket's key and value from the hashtable.
383 /// This works similarly to `put`, building an `EmptyBucket` out of the
385 pub fn take(mut self) -> (EmptyBucket<K, V, M>, K, V) {
386 let key = self.raw.key as *const K;
387 let val = self.raw.val as *const V;
389 self.table.size -= 1;
392 *self.raw.hash = EMPTY_BUCKET;
405 pub fn replace(&mut self, h: SafeHash, k: K, v: V) -> (SafeHash, K, V) {
407 let old_hash = ptr::replace(self.raw.hash as *mut SafeHash, h);
408 let old_key = ptr::replace(self.raw.key, k);
409 let old_val = ptr::replace(self.raw.val, v);
411 (old_hash, old_key, old_val)
415 /// Gets mutable references to the key and value at a given index.
416 pub fn read_mut(&mut self) -> (&mut K, &mut V) {
424 impl<'t, K, V, M: Deref<RawTable<K, V>> + 't> FullBucket<K, V, M> {
425 /// Exchange a bucket state for immutable references into the table.
426 /// Because the underlying reference to the table is also consumed,
427 /// no further changes to the structure of the table are possible;
428 /// in exchange for this, the returned references have a longer lifetime
429 /// than the references returned by `read()`.
430 pub fn into_refs(self) -> (&'t K, &'t V) {
438 impl<'t, K, V, M: DerefMut<RawTable<K, V>> + 't> FullBucket<K, V, M> {
439 /// This works similarly to `into_refs`, exchanging a bucket state
440 /// for mutable references into the table.
441 pub fn into_mut_refs(self) -> (&'t mut K, &'t mut V) {
449 impl<K, V, M> BucketState<K, V, M> {
451 pub fn expect_full(self) -> FullBucket<K, V, M> {
454 Empty(..) => panic!("Expected full bucket")
459 impl<K, V, M: Deref<RawTable<K, V>>> GapThenFull<K, V, M> {
461 pub fn full(&self) -> &FullBucket<K, V, M> {
465 pub fn shift(mut self) -> Option<GapThenFull<K, V, M>> {
467 *self.gap.raw.hash = mem::replace(&mut *self.full.raw.hash, EMPTY_BUCKET);
468 copy_nonoverlapping_memory(self.gap.raw.key, self.full.raw.key as *const K, 1);
469 copy_nonoverlapping_memory(self.gap.raw.val, self.full.raw.val as *const V, 1);
472 let FullBucket { raw: prev_raw, idx: prev_idx, .. } = self.full;
474 match self.full.next().peek() {
476 self.gap.raw = prev_raw;
477 self.gap.idx = prev_idx;
489 /// Rounds up to a multiple of a power of two. Returns the closest multiple
490 /// of `target_alignment` that is higher or equal to `unrounded`.
494 /// Panics if `target_alignment` is not a power of two.
495 fn round_up_to_next(unrounded: uint, target_alignment: uint) -> uint {
496 assert!(target_alignment.is_power_of_two());
497 (unrounded + target_alignment - 1) & !(target_alignment - 1)
502 assert_eq!(round_up_to_next(0, 4), 0);
503 assert_eq!(round_up_to_next(1, 4), 4);
504 assert_eq!(round_up_to_next(2, 4), 4);
505 assert_eq!(round_up_to_next(3, 4), 4);
506 assert_eq!(round_up_to_next(4, 4), 4);
507 assert_eq!(round_up_to_next(5, 4), 8);
510 // Returns a tuple of (key_offset, val_offset),
511 // from the start of a mallocated array.
512 fn calculate_offsets(hashes_size: uint,
513 keys_size: uint, keys_align: uint,
516 let keys_offset = round_up_to_next(hashes_size, keys_align);
517 let end_of_keys = keys_offset + keys_size;
519 let vals_offset = round_up_to_next(end_of_keys, vals_align);
521 (keys_offset, vals_offset)
524 // Returns a tuple of (minimum required malloc alignment, hash_offset,
525 // array_size), from the start of a mallocated array.
526 fn calculate_allocation(hash_size: uint, hash_align: uint,
527 keys_size: uint, keys_align: uint,
528 vals_size: uint, vals_align: uint)
529 -> (uint, uint, uint) {
531 let (_, vals_offset) = calculate_offsets(hash_size,
532 keys_size, keys_align,
534 let end_of_vals = vals_offset + vals_size;
536 let min_align = cmp::max(hash_align, cmp::max(keys_align, vals_align));
538 (min_align, hash_offset, end_of_vals)
542 fn test_offset_calculation() {
543 assert_eq!(calculate_allocation(128, 8, 15, 1, 4, 4), (8, 0, 148));
544 assert_eq!(calculate_allocation(3, 1, 2, 1, 1, 1), (1, 0, 6));
545 assert_eq!(calculate_allocation(6, 2, 12, 4, 24, 8), (8, 0, 48));
546 assert_eq!(calculate_offsets(128, 15, 1, 4), (128, 144));
547 assert_eq!(calculate_offsets(3, 2, 1, 1), (3, 5));
548 assert_eq!(calculate_offsets(6, 12, 4, 8), (8, 24));
551 impl<K, V> RawTable<K, V> {
552 /// Does not initialize the buckets. The caller should ensure they,
553 /// at the very least, set every hash to EMPTY_BUCKET.
554 unsafe fn new_uninitialized(capacity: uint) -> RawTable<K, V> {
559 hashes: 0 as *mut u64,
560 marker: marker::CovariantType,
563 // No need for `checked_mul` before a more restrictive check performed
564 // later in this method.
565 let hashes_size = capacity * size_of::<u64>();
566 let keys_size = capacity * size_of::< K >();
567 let vals_size = capacity * size_of::< V >();
569 // Allocating hashmaps is a little tricky. We need to allocate three
570 // arrays, but since we know their sizes and alignments up front,
571 // we just allocate a single array, and then have the subarrays
574 // This is great in theory, but in practice getting the alignment
575 // right is a little subtle. Therefore, calculating offsets has been
576 // factored out into a different function.
577 let (malloc_alignment, hash_offset, size) =
578 calculate_allocation(
579 hashes_size, min_align_of::<u64>(),
580 keys_size, min_align_of::< K >(),
581 vals_size, min_align_of::< V >());
583 // One check for overflow that covers calculation and rounding of size.
584 let size_of_bucket = size_of::<u64>().checked_add(size_of::<K>()).unwrap()
585 .checked_add(size_of::<V>()).unwrap();
586 assert!(size >= capacity.checked_mul(size_of_bucket)
587 .expect("capacity overflow"),
588 "capacity overflow");
590 let buffer = allocate(size, malloc_alignment);
591 if buffer.is_null() { ::alloc::oom() }
593 let hashes = buffer.offset(hash_offset as int) as *mut u64;
599 marker: marker::CovariantType,
603 fn first_bucket_raw(&self) -> RawBucket<K, V> {
604 let hashes_size = self.capacity * size_of::<u64>();
605 let keys_size = self.capacity * size_of::<K>();
607 let buffer = self.hashes as *mut u8;
608 let (keys_offset, vals_offset) = calculate_offsets(hashes_size,
609 keys_size, min_align_of::<K>(),
610 min_align_of::<V>());
615 key: buffer.offset(keys_offset as int) as *mut K,
616 val: buffer.offset(vals_offset as int) as *mut V
621 /// Creates a new raw table from a given capacity. All buckets are
623 #[allow(experimental)]
624 pub fn new(capacity: uint) -> RawTable<K, V> {
626 let ret = RawTable::new_uninitialized(capacity);
627 zero_memory(ret.hashes, capacity);
632 /// The hashtable's capacity, similar to a vector's.
633 pub fn capacity(&self) -> uint {
637 /// The number of elements ever `put` in the hashtable, minus the number
638 /// of elements ever `take`n.
639 pub fn size(&self) -> uint {
643 fn raw_buckets(&self) -> RawBuckets<K, V> {
645 raw: self.first_bucket_raw(),
647 self.hashes.offset(self.capacity as int)
649 marker: marker::ContravariantLifetime,
653 pub fn iter(&self) -> Entries<K, V> {
655 iter: self.raw_buckets(),
656 elems_left: self.size(),
660 pub fn iter_mut(&mut self) -> MutEntries<K, V> {
662 iter: self.raw_buckets(),
663 elems_left: self.size(),
667 pub fn into_iter(self) -> MoveEntries<K, V> {
668 let RawBuckets { raw, hashes_end, .. } = self.raw_buckets();
669 // Replace the marker regardless of lifetime bounds on parameters.
673 hashes_end: hashes_end,
674 marker: marker::ContravariantLifetime,
680 /// Returns an iterator that copies out each entry. Used while the table
681 /// is being dropped.
682 unsafe fn rev_move_buckets(&mut self) -> RevMoveBuckets<K, V> {
683 let raw_bucket = self.first_bucket_raw();
685 raw: raw_bucket.offset(self.capacity as int),
686 hashes_end: raw_bucket.hash,
687 elems_left: self.size,
688 marker: marker::ContravariantLifetime,
693 /// A raw iterator. The basis for some other iterators in this module. Although
694 /// this interface is safe, it's not used outside this module.
695 struct RawBuckets<'a, K, V> {
696 raw: RawBucket<K, V>,
697 hashes_end: *mut u64,
698 marker: marker::ContravariantLifetime<'a>,
701 impl<'a, K, V> Iterator<RawBucket<K, V>> for RawBuckets<'a, K, V> {
702 fn next(&mut self) -> Option<RawBucket<K, V>> {
703 while self.raw.hash != self.hashes_end {
705 // We are swapping out the pointer to a bucket and replacing
706 // it with the pointer to the next one.
707 let prev = ptr::replace(&mut self.raw, self.raw.offset(1));
708 if *prev.hash != EMPTY_BUCKET {
718 /// An iterator that moves out buckets in reverse order. It leaves the table
719 /// in an inconsistent state and should only be used for dropping
720 /// the table's remaining entries. It's used in the implementation of Drop.
721 struct RevMoveBuckets<'a, K, V> {
722 raw: RawBucket<K, V>,
723 hashes_end: *mut u64,
725 marker: marker::ContravariantLifetime<'a>,
728 impl<'a, K, V> Iterator<(K, V)> for RevMoveBuckets<'a, K, V> {
729 fn next(&mut self) -> Option<(K, V)> {
730 if self.elems_left == 0 {
735 debug_assert!(self.raw.hash != self.hashes_end);
738 self.raw = self.raw.offset(-1);
740 if *self.raw.hash != EMPTY_BUCKET {
741 self.elems_left -= 1;
743 ptr::read(self.raw.key as *const K),
744 ptr::read(self.raw.val as *const V)
752 /// Iterator over shared references to entries in a table.
753 pub struct Entries<'a, K: 'a, V: 'a> {
754 iter: RawBuckets<'a, K, V>,
758 /// Iterator over mutable references to entries in a table.
759 pub struct MutEntries<'a, K: 'a, V: 'a> {
760 iter: RawBuckets<'a, K, V>,
764 /// Iterator over the entries in a table, consuming the table.
765 pub struct MoveEntries<K, V> {
766 table: RawTable<K, V>,
767 iter: RawBuckets<'static, K, V>
770 impl<'a, K, V> Iterator<(&'a K, &'a V)> for Entries<'a, K, V> {
771 fn next(&mut self) -> Option<(&'a K, &'a V)> {
772 self.iter.next().map(|bucket| {
773 self.elems_left -= 1;
781 fn size_hint(&self) -> (uint, Option<uint>) {
782 (self.elems_left, Some(self.elems_left))
786 impl<'a, K, V> Iterator<(&'a K, &'a mut V)> for MutEntries<'a, K, V> {
787 fn next(&mut self) -> Option<(&'a K, &'a mut V)> {
788 self.iter.next().map(|bucket| {
789 self.elems_left -= 1;
797 fn size_hint(&self) -> (uint, Option<uint>) {
798 (self.elems_left, Some(self.elems_left))
802 impl<K, V> Iterator<(SafeHash, K, V)> for MoveEntries<K, V> {
803 fn next(&mut self) -> Option<(SafeHash, K, V)> {
804 self.iter.next().map(|bucket| {
805 self.table.size -= 1;
811 ptr::read(bucket.key as *const K),
812 ptr::read(bucket.val as *const V)
818 fn size_hint(&self) -> (uint, Option<uint>) {
819 let size = self.table.size();
824 impl<K: Clone, V: Clone> Clone for RawTable<K, V> {
825 fn clone(&self) -> RawTable<K, V> {
827 let mut new_ht = RawTable::new_uninitialized(self.capacity());
830 let cap = self.capacity();
831 let mut new_buckets = Bucket::first(&mut new_ht);
832 let mut buckets = Bucket::first(self);
833 while buckets.index() != cap {
834 match buckets.peek() {
837 let (k, v) = full.read();
838 (full.hash(), k.clone(), v.clone())
840 *new_buckets.raw.hash = h.inspect();
841 ptr::write(new_buckets.raw.key, k);
842 ptr::write(new_buckets.raw.val, v);
845 *new_buckets.raw.hash = EMPTY_BUCKET;
853 new_ht.size = self.size();
861 impl<K, V> Drop for RawTable<K, V> {
863 if self.hashes.is_null() {
866 // This is done in reverse because we've likely partially taken
867 // some elements out with `.into_iter()` from the front.
868 // Check if the size is 0, so we don't do a useless scan when
869 // dropping empty tables such as on resize.
870 // Also avoid double drop of elements that have been already moved out.
872 for _ in self.rev_move_buckets() {}
875 let hashes_size = self.capacity * size_of::<u64>();
876 let keys_size = self.capacity * size_of::<K>();
877 let vals_size = self.capacity * size_of::<V>();
878 let (align, _, size) = calculate_allocation(hashes_size, min_align_of::<u64>(),
879 keys_size, min_align_of::<K>(),
880 vals_size, min_align_of::<V>());
883 deallocate(self.hashes as *mut u8, size, align);
884 // Remember how everything was allocated out of one buffer
885 // during initialization? We only need one call to free here.