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
13 pub use self::BucketState::*;
17 use hash::{Hash, Hasher};
18 use iter::{Iterator, count};
19 use kinds::{Sized, marker};
20 use mem::{min_align_of, size_of};
22 use num::{Int, UnsignedInt};
23 use ops::{Deref, DerefMut, Drop};
24 use option::{Some, None, Option};
25 use ptr::{RawPtr, copy_nonoverlapping_memory, zero_memory};
27 use rt::heap::{allocate, deallocate};
29 const EMPTY_BUCKET: u64 = 0u64;
31 /// The raw hashtable, providing safe-ish access to the unzipped and highly
32 /// optimized arrays of hashes, keys, and values.
34 /// This design uses less memory and is a lot faster than the naive
35 /// `Vec<Option<u64, K, V>>`, because we don't pay for the overhead of an
36 /// option on every element, and we get a generally more cache-aware design.
38 /// Essential invariants of this structure:
40 /// - if t.hashes[i] == EMPTY_BUCKET, then `Bucket::at_index(&t, i).raw`
41 /// points to 'undefined' contents. Don't read from it. This invariant is
42 /// enforced outside this module with the `EmptyBucket`, `FullBucket`,
43 /// and `SafeHash` types.
45 /// - An `EmptyBucket` is only constructed at an index with
46 /// a hash of EMPTY_BUCKET.
48 /// - A `FullBucket` is only constructed at an index with a
49 /// non-EMPTY_BUCKET hash.
51 /// - A `SafeHash` is only constructed for non-`EMPTY_BUCKET` hash. We get
52 /// around hashes of zero by changing them to 0x8000_0000_0000_0000,
53 /// which will likely map to the same bucket, while not being confused
56 /// - All three "arrays represented by pointers" are the same length:
57 /// `capacity`. This is set at creation and never changes. The arrays
58 /// are unzipped to save space (we don't have to pay for the padding
59 /// between odd sized elements, such as in a map from u64 to u8), and
60 /// be more cache aware (scanning through 8 hashes brings in at most
61 /// 2 cache lines, since they're all right beside each other).
63 /// You can kind of think of this module/data structure as a safe wrapper
64 /// around just the "table" part of the hashtable. It enforces some
65 /// invariants at the type level and employs some performance trickery,
66 /// but in general is just a tricked out `Vec<Option<u64, K, V>>`.
67 #[unsafe_no_drop_flag]
68 pub struct RawTable<K, V> {
72 // Because K/V do not appear directly in any of the types in the struct,
73 // inform rustc that in fact instances of K and V are reachable from here.
74 marker: marker::CovariantType<(K,V)>,
77 struct RawBucket<K, V> {
83 pub struct Bucket<K, V, M> {
89 pub struct EmptyBucket<K, V, M> {
95 pub struct FullBucket<K, V, M> {
101 pub type EmptyBucketImm<'table, K, V> = EmptyBucket<K, V, &'table RawTable<K, V>>;
102 pub type FullBucketImm<'table, K, V> = FullBucket<K, V, &'table RawTable<K, V>>;
104 pub type EmptyBucketMut<'table, K, V> = EmptyBucket<K, V, &'table mut RawTable<K, V>>;
105 pub type FullBucketMut<'table, K, V> = FullBucket<K, V, &'table mut RawTable<K, V>>;
107 pub enum BucketState<K, V, M> {
108 Empty(EmptyBucket<K, V, M>),
109 Full(FullBucket<K, V, M>),
112 // A GapThenFull encapsulates the state of two consecutive buckets at once.
113 // The first bucket, called the gap, is known to be empty.
114 // The second bucket is full.
115 struct GapThenFull<K, V, M> {
116 gap: EmptyBucket<K, V, ()>,
117 full: FullBucket<K, V, M>,
120 /// A hash that is not zero, since we use a hash of zero to represent empty
122 #[deriving(PartialEq)]
123 pub struct SafeHash {
128 /// Peek at the hash value, which is guaranteed to be non-zero.
130 pub fn inspect(&self) -> u64 { self.hash }
133 /// We need to remove hashes of 0. That's reserved for empty buckets.
134 /// This function wraps up `hash_keyed` to be the only way outside this
135 /// module to generate a SafeHash.
136 pub fn make_hash<Sized? T: Hash<S>, S, H: Hasher<S>>(hasher: &H, t: &T) -> SafeHash {
137 match hasher.hash(t) {
138 // This constant is exceedingly likely to hash to the same
139 // bucket, but it won't be counted as empty! Just so we can maintain
140 // our precious uniform distribution of initial indexes.
141 EMPTY_BUCKET => SafeHash { hash: 0x8000_0000_0000_0000 },
142 h => SafeHash { hash: h },
146 // `replace` casts a `*u64` to a `*SafeHash`. Since we statically
147 // ensure that a `FullBucket` points to an index with a non-zero hash,
148 // and a `SafeHash` is just a `u64` with a different name, this is
151 // This test ensures that a `SafeHash` really IS the same size as a
152 // `u64`. If you need to change the size of `SafeHash` (and
153 // consequently made this test fail), `replace` needs to be
154 // modified to no longer assume this.
156 fn can_alias_safehash_as_u64() {
157 assert_eq!(size_of::<SafeHash>(), size_of::<u64>())
160 impl<K, V> RawBucket<K, V> {
161 unsafe fn offset(self, count: int) -> RawBucket<K, V> {
163 hash: self.hash.offset(count),
164 key: self.key.offset(count),
165 val: self.val.offset(count),
170 // Buckets hold references to the table.
171 impl<K, V, M> FullBucket<K, V, M> {
172 /// Borrow a reference to the table.
173 pub fn table(&self) -> &M {
176 /// Move out the reference to the table.
177 pub fn into_table(self) -> M {
180 /// Get the raw index.
181 pub fn index(&self) -> uint {
186 impl<K, V, M> EmptyBucket<K, V, M> {
187 /// Borrow a reference to the table.
188 pub fn table(&self) -> &M {
191 /// Move out the reference to the table.
192 pub fn into_table(self) -> M {
197 impl<K, V, M> Bucket<K, V, M> {
198 /// Move out the reference to the table.
199 pub fn into_table(self) -> M {
202 /// Get the raw index.
203 pub fn index(&self) -> uint {
208 impl<K, V, M: Deref<RawTable<K, V>>> Bucket<K, V, M> {
209 pub fn new(table: M, hash: &SafeHash) -> Bucket<K, V, M> {
210 Bucket::at_index(table, hash.inspect() as uint)
213 pub fn at_index(table: M, ib_index: uint) -> Bucket<K, V, M> {
214 let ib_index = ib_index & (table.capacity() - 1);
217 table.first_bucket_raw().offset(ib_index as int)
224 pub fn first(table: M) -> Bucket<K, V, M> {
226 raw: table.first_bucket_raw(),
232 /// Reads a bucket at a given index, returning an enum indicating whether
233 /// it's initialized or not. You need to match on this enum to get
234 /// the appropriate types to call most of the other functions in
236 pub fn peek(self) -> BucketState<K, V, M> {
237 match unsafe { *self.raw.hash } {
253 /// Modifies the bucket pointer in place to make it point to the next slot.
254 pub fn next(&mut self) {
255 // Branchless bucket iteration step.
256 // As we reach the end of the table...
257 // We take the current idx: 0111111b
258 // Xor it by its increment: ^ 1000000b
261 // Then AND with the capacity: & 1000000b
263 // to get the backwards offset: 1000000b
264 // ... and it's zero at all other times.
265 let maybe_wraparound_dist = (self.idx ^ (self.idx + 1)) & self.table.capacity();
266 // Finally, we obtain the offset 1 or the offset -cap + 1.
267 let dist = 1i - (maybe_wraparound_dist as int);
272 self.raw = self.raw.offset(dist);
277 impl<K, V, M: Deref<RawTable<K, V>>> EmptyBucket<K, V, M> {
279 pub fn next(self) -> Bucket<K, V, M> {
280 let mut bucket = self.into_bucket();
286 pub fn into_bucket(self) -> Bucket<K, V, M> {
294 pub fn gap_peek(self) -> Option<GapThenFull<K, V, M>> {
295 let gap = EmptyBucket {
301 match self.next().peek() {
313 impl<K, V, M: DerefMut<RawTable<K, V>>> EmptyBucket<K, V, M> {
314 /// Puts given key and value pair, along with the key's hash,
315 /// into this bucket in the hashtable. Note how `self` is 'moved' into
316 /// this function, because this slot will no longer be empty when
317 /// we return! A `FullBucket` is returned for later use, pointing to
318 /// the newly-filled slot in the hashtable.
320 /// Use `make_hash` to construct a `SafeHash` to pass to this function.
321 pub fn put(mut self, hash: SafeHash, key: K, value: V)
322 -> FullBucket<K, V, M> {
324 *self.raw.hash = hash.inspect();
325 ptr::write(self.raw.key, key);
326 ptr::write(self.raw.val, value);
329 self.table.size += 1;
331 FullBucket { raw: self.raw, idx: self.idx, table: self.table }
335 impl<K, V, M: Deref<RawTable<K, V>>> FullBucket<K, V, M> {
337 pub fn next(self) -> Bucket<K, V, M> {
338 let mut bucket = self.into_bucket();
344 pub fn into_bucket(self) -> Bucket<K, V, M> {
352 /// Get the distance between this bucket and the 'ideal' location
353 /// as determined by the key's hash stored in it.
355 /// In the cited blog posts above, this is called the "distance to
356 /// initial bucket", or DIB. Also known as "probe count".
357 pub fn distance(&self) -> uint {
358 // Calculates the distance one has to travel when going from
359 // `hash mod capacity` onwards to `idx mod capacity`, wrapping around
360 // if the destination is not reached before the end of the table.
361 (self.idx - self.hash().inspect() as uint) & (self.table.capacity() - 1)
365 pub fn hash(&self) -> SafeHash {
373 /// Gets references to the key and value at a given index.
374 pub fn read(&self) -> (&K, &V) {
382 impl<K, V, M: DerefMut<RawTable<K, V>>> FullBucket<K, V, M> {
383 /// Removes this bucket's key and value from the hashtable.
385 /// This works similarly to `put`, building an `EmptyBucket` out of the
387 pub fn take(mut self) -> (EmptyBucket<K, V, M>, K, V) {
388 let key = self.raw.key as *const K;
389 let val = self.raw.val as *const V;
391 self.table.size -= 1;
394 *self.raw.hash = EMPTY_BUCKET;
407 pub fn replace(&mut self, h: SafeHash, k: K, v: V) -> (SafeHash, K, V) {
409 let old_hash = ptr::replace(self.raw.hash as *mut SafeHash, h);
410 let old_key = ptr::replace(self.raw.key, k);
411 let old_val = ptr::replace(self.raw.val, v);
413 (old_hash, old_key, old_val)
417 /// Gets mutable references to the key and value at a given index.
418 pub fn read_mut(&mut self) -> (&mut K, &mut V) {
426 impl<'t, K, V, M: Deref<RawTable<K, V>> + 't> FullBucket<K, V, M> {
427 /// Exchange a bucket state for immutable references into the table.
428 /// Because the underlying reference to the table is also consumed,
429 /// no further changes to the structure of the table are possible;
430 /// in exchange for this, the returned references have a longer lifetime
431 /// than the references returned by `read()`.
432 pub fn into_refs(self) -> (&'t K, &'t V) {
440 impl<'t, K, V, M: DerefMut<RawTable<K, V>> + 't> FullBucket<K, V, M> {
441 /// This works similarly to `into_refs`, exchanging a bucket state
442 /// for mutable references into the table.
443 pub fn into_mut_refs(self) -> (&'t mut K, &'t mut V) {
451 impl<K, V, M> BucketState<K, V, M> {
453 pub fn expect_full(self) -> FullBucket<K, V, M> {
456 Empty(..) => panic!("Expected full bucket")
461 impl<K, V, M: Deref<RawTable<K, V>>> GapThenFull<K, V, M> {
463 pub fn full(&self) -> &FullBucket<K, V, M> {
467 pub fn shift(mut self) -> Option<GapThenFull<K, V, M>> {
469 *self.gap.raw.hash = mem::replace(&mut *self.full.raw.hash, EMPTY_BUCKET);
470 copy_nonoverlapping_memory(self.gap.raw.key, self.full.raw.key as *const K, 1);
471 copy_nonoverlapping_memory(self.gap.raw.val, self.full.raw.val as *const V, 1);
474 let FullBucket { raw: prev_raw, idx: prev_idx, .. } = self.full;
476 match self.full.next().peek() {
478 self.gap.raw = prev_raw;
479 self.gap.idx = prev_idx;
491 /// Rounds up to a multiple of a power of two. Returns the closest multiple
492 /// of `target_alignment` that is higher or equal to `unrounded`.
496 /// Panics if `target_alignment` is not a power of two.
497 fn round_up_to_next(unrounded: uint, target_alignment: uint) -> uint {
498 assert!(target_alignment.is_power_of_two());
499 (unrounded + target_alignment - 1) & !(target_alignment - 1)
504 assert_eq!(round_up_to_next(0, 4), 0);
505 assert_eq!(round_up_to_next(1, 4), 4);
506 assert_eq!(round_up_to_next(2, 4), 4);
507 assert_eq!(round_up_to_next(3, 4), 4);
508 assert_eq!(round_up_to_next(4, 4), 4);
509 assert_eq!(round_up_to_next(5, 4), 8);
512 // Returns a tuple of (key_offset, val_offset),
513 // from the start of a mallocated array.
514 fn calculate_offsets(hashes_size: uint,
515 keys_size: uint, keys_align: uint,
518 let keys_offset = round_up_to_next(hashes_size, keys_align);
519 let end_of_keys = keys_offset + keys_size;
521 let vals_offset = round_up_to_next(end_of_keys, vals_align);
523 (keys_offset, vals_offset)
526 // Returns a tuple of (minimum required malloc alignment, hash_offset,
527 // array_size), from the start of a mallocated array.
528 fn calculate_allocation(hash_size: uint, hash_align: uint,
529 keys_size: uint, keys_align: uint,
530 vals_size: uint, vals_align: uint)
531 -> (uint, uint, uint) {
533 let (_, vals_offset) = calculate_offsets(hash_size,
534 keys_size, keys_align,
536 let end_of_vals = vals_offset + vals_size;
538 let min_align = cmp::max(hash_align, cmp::max(keys_align, vals_align));
540 (min_align, hash_offset, end_of_vals)
544 fn test_offset_calculation() {
545 assert_eq!(calculate_allocation(128, 8, 15, 1, 4, 4), (8, 0, 148));
546 assert_eq!(calculate_allocation(3, 1, 2, 1, 1, 1), (1, 0, 6));
547 assert_eq!(calculate_allocation(6, 2, 12, 4, 24, 8), (8, 0, 48));
548 assert_eq!(calculate_offsets(128, 15, 1, 4), (128, 144));
549 assert_eq!(calculate_offsets(3, 2, 1, 1), (3, 5));
550 assert_eq!(calculate_offsets(6, 12, 4, 8), (8, 24));
553 impl<K, V> RawTable<K, V> {
554 /// Does not initialize the buckets. The caller should ensure they,
555 /// at the very least, set every hash to EMPTY_BUCKET.
556 unsafe fn new_uninitialized(capacity: uint) -> RawTable<K, V> {
561 hashes: 0 as *mut u64,
562 marker: marker::CovariantType,
565 // No need for `checked_mul` before a more restrictive check performed
566 // later in this method.
567 let hashes_size = capacity * size_of::<u64>();
568 let keys_size = capacity * size_of::< K >();
569 let vals_size = capacity * size_of::< V >();
571 // Allocating hashmaps is a little tricky. We need to allocate three
572 // arrays, but since we know their sizes and alignments up front,
573 // we just allocate a single array, and then have the subarrays
576 // This is great in theory, but in practice getting the alignment
577 // right is a little subtle. Therefore, calculating offsets has been
578 // factored out into a different function.
579 let (malloc_alignment, hash_offset, size) =
580 calculate_allocation(
581 hashes_size, min_align_of::<u64>(),
582 keys_size, min_align_of::< K >(),
583 vals_size, min_align_of::< V >());
585 // One check for overflow that covers calculation and rounding of size.
586 let size_of_bucket = size_of::<u64>().checked_add(size_of::<K>()).unwrap()
587 .checked_add(size_of::<V>()).unwrap();
588 assert!(size >= capacity.checked_mul(size_of_bucket)
589 .expect("capacity overflow"),
590 "capacity overflow");
592 let buffer = allocate(size, malloc_alignment);
593 if buffer.is_null() { ::alloc::oom() }
595 let hashes = buffer.offset(hash_offset as int) as *mut u64;
601 marker: marker::CovariantType,
605 fn first_bucket_raw(&self) -> RawBucket<K, V> {
606 let hashes_size = self.capacity * size_of::<u64>();
607 let keys_size = self.capacity * size_of::<K>();
609 let buffer = self.hashes as *mut u8;
610 let (keys_offset, vals_offset) = calculate_offsets(hashes_size,
611 keys_size, min_align_of::<K>(),
612 min_align_of::<V>());
617 key: buffer.offset(keys_offset as int) as *mut K,
618 val: buffer.offset(vals_offset as int) as *mut V
623 /// Creates a new raw table from a given capacity. All buckets are
625 #[allow(experimental)]
626 pub fn new(capacity: uint) -> RawTable<K, V> {
628 let ret = RawTable::new_uninitialized(capacity);
629 zero_memory(ret.hashes, capacity);
634 /// The hashtable's capacity, similar to a vector's.
635 pub fn capacity(&self) -> uint {
639 /// The number of elements ever `put` in the hashtable, minus the number
640 /// of elements ever `take`n.
641 pub fn size(&self) -> uint {
645 fn raw_buckets(&self) -> RawBuckets<K, V> {
647 raw: self.first_bucket_raw(),
649 self.hashes.offset(self.capacity as int)
651 marker: marker::ContravariantLifetime,
655 pub fn iter(&self) -> Entries<K, V> {
657 iter: self.raw_buckets(),
658 elems_left: self.size(),
662 pub fn iter_mut(&mut self) -> MutEntries<K, V> {
664 iter: self.raw_buckets(),
665 elems_left: self.size(),
669 pub fn into_iter(self) -> MoveEntries<K, V> {
670 let RawBuckets { raw, hashes_end, .. } = self.raw_buckets();
671 // Replace the marker regardless of lifetime bounds on parameters.
675 hashes_end: hashes_end,
676 marker: marker::ContravariantLifetime,
682 /// Returns an iterator that copies out each entry. Used while the table
683 /// is being dropped.
684 unsafe fn rev_move_buckets(&mut self) -> RevMoveBuckets<K, V> {
685 let raw_bucket = self.first_bucket_raw();
687 raw: raw_bucket.offset(self.capacity as int),
688 hashes_end: raw_bucket.hash,
689 elems_left: self.size,
690 marker: marker::ContravariantLifetime,
695 /// A raw iterator. The basis for some other iterators in this module. Although
696 /// this interface is safe, it's not used outside this module.
697 struct RawBuckets<'a, K, V> {
698 raw: RawBucket<K, V>,
699 hashes_end: *mut u64,
700 marker: marker::ContravariantLifetime<'a>,
703 impl<'a, K, V> Iterator<RawBucket<K, V>> for RawBuckets<'a, K, V> {
704 fn next(&mut self) -> Option<RawBucket<K, V>> {
705 while self.raw.hash != self.hashes_end {
707 // We are swapping out the pointer to a bucket and replacing
708 // it with the pointer to the next one.
709 let prev = ptr::replace(&mut self.raw, self.raw.offset(1));
710 if *prev.hash != EMPTY_BUCKET {
720 /// An iterator that moves out buckets in reverse order. It leaves the table
721 /// in an an inconsistent state and should only be used for dropping
722 /// the table's remaining entries. It's used in the implementation of Drop.
723 struct RevMoveBuckets<'a, K, V> {
724 raw: RawBucket<K, V>,
725 hashes_end: *mut u64,
727 marker: marker::ContravariantLifetime<'a>,
730 impl<'a, K, V> Iterator<(K, V)> for RevMoveBuckets<'a, K, V> {
731 fn next(&mut self) -> Option<(K, V)> {
732 if self.elems_left == 0 {
737 debug_assert!(self.raw.hash != self.hashes_end);
740 self.raw = self.raw.offset(-1);
742 if *self.raw.hash != EMPTY_BUCKET {
743 self.elems_left -= 1;
745 ptr::read(self.raw.key as *const K),
746 ptr::read(self.raw.val as *const V)
754 /// Iterator over shared references to entries in a table.
755 pub struct Entries<'a, K: 'a, V: 'a> {
756 iter: RawBuckets<'a, K, V>,
760 /// Iterator over mutable references to entries in a table.
761 pub struct MutEntries<'a, K: 'a, V: 'a> {
762 iter: RawBuckets<'a, K, V>,
766 /// Iterator over the entries in a table, consuming the table.
767 pub struct MoveEntries<K, V> {
768 table: RawTable<K, V>,
769 iter: RawBuckets<'static, K, V>
772 impl<'a, K, V> Iterator<(&'a K, &'a V)> for Entries<'a, K, V> {
773 fn next(&mut self) -> Option<(&'a K, &'a V)> {
774 self.iter.next().map(|bucket| {
775 self.elems_left -= 1;
783 fn size_hint(&self) -> (uint, Option<uint>) {
784 (self.elems_left, Some(self.elems_left))
788 impl<'a, K, V> Iterator<(&'a K, &'a mut V)> for MutEntries<'a, K, V> {
789 fn next(&mut self) -> Option<(&'a K, &'a mut V)> {
790 self.iter.next().map(|bucket| {
791 self.elems_left -= 1;
799 fn size_hint(&self) -> (uint, Option<uint>) {
800 (self.elems_left, Some(self.elems_left))
804 impl<K, V> Iterator<(SafeHash, K, V)> for MoveEntries<K, V> {
805 fn next(&mut self) -> Option<(SafeHash, K, V)> {
806 self.iter.next().map(|bucket| {
807 self.table.size -= 1;
813 ptr::read(bucket.key as *const K),
814 ptr::read(bucket.val as *const V)
820 fn size_hint(&self) -> (uint, Option<uint>) {
821 let size = self.table.size();
826 impl<K: Clone, V: Clone> Clone for RawTable<K, V> {
827 fn clone(&self) -> RawTable<K, V> {
829 let mut new_ht = RawTable::new_uninitialized(self.capacity());
832 let cap = self.capacity();
833 let mut new_buckets = Bucket::first(&mut new_ht);
834 let mut buckets = Bucket::first(self);
835 while buckets.index() != cap {
836 match buckets.peek() {
839 let (k, v) = full.read();
840 (full.hash(), k.clone(), v.clone())
842 *new_buckets.raw.hash = h.inspect();
843 ptr::write(new_buckets.raw.key, k);
844 ptr::write(new_buckets.raw.val, v);
847 *new_buckets.raw.hash = EMPTY_BUCKET;
855 new_ht.size = self.size();
863 impl<K, V> Drop for RawTable<K, V> {
865 if self.hashes.is_null() {
868 // This is done in reverse because we've likely partially taken
869 // some elements out with `.into_iter()` from the front.
870 // Check if the size is 0, so we don't do a useless scan when
871 // dropping empty tables such as on resize.
872 // Also avoid double drop of elements that have been already moved out.
874 for _ in self.rev_move_buckets() {}
877 let hashes_size = self.capacity * size_of::<u64>();
878 let keys_size = self.capacity * size_of::<K>();
879 let vals_size = self.capacity * size_of::<V>();
880 let (align, _, size) = calculate_allocation(hashes_size, min_align_of::<u64>(),
881 keys_size, min_align_of::<K>(),
882 vals_size, min_align_of::<V>());
885 deallocate(self.hashes as *mut u8, size, align);
886 // Remember how everything was allocated out of one buffer
887 // during initialization? We only need one call to free here.