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};
18 use mem::{min_align_of, size_of};
20 use num::{CheckedAdd, CheckedMul, is_power_of_two};
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 static 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<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 // For parameterizing over mutability.
169 impl<'t, K, V> Deref<RawTable<K, V>> for &'t RawTable<K, V> {
170 fn deref(&self) -> &RawTable<K, V> {
175 impl<'t, K, V> Deref<RawTable<K, V>> for &'t mut RawTable<K, V> {
176 fn deref(&self) -> &RawTable<K,V> {
181 impl<'t, K, V> DerefMut<RawTable<K, V>> for &'t mut RawTable<K, V> {
182 fn deref_mut(&mut self) -> &mut RawTable<K,V> {
187 // Buckets hold references to the table.
188 impl<K, V, M> FullBucket<K, V, M> {
189 /// Borrow a reference to the table.
190 pub fn table(&self) -> &M {
193 /// Move out the reference to the table.
194 pub fn into_table(self) -> M {
197 /// Get the raw index.
198 pub fn index(&self) -> uint {
203 impl<K, V, M> EmptyBucket<K, V, M> {
204 /// Borrow a reference to the table.
205 pub fn table(&self) -> &M {
208 /// Move out the reference to the table.
209 pub fn into_table(self) -> M {
214 impl<K, V, M> Bucket<K, V, M> {
215 /// Move out the reference to the table.
216 pub fn into_table(self) -> M {
219 /// Get the raw index.
220 pub fn index(&self) -> uint {
225 impl<K, V, M: Deref<RawTable<K, V>>> Bucket<K, V, M> {
226 pub fn new(table: M, hash: &SafeHash) -> Bucket<K, V, M> {
227 Bucket::at_index(table, hash.inspect() as uint)
230 pub fn at_index(table: M, ib_index: uint) -> Bucket<K, V, M> {
231 let ib_index = ib_index & (table.capacity() - 1);
234 table.first_bucket_raw().offset(ib_index as int)
241 pub fn first(table: M) -> Bucket<K, V, M> {
243 raw: table.first_bucket_raw(),
249 /// Reads a bucket at a given index, returning an enum indicating whether
250 /// it's initialized or not. You need to match on this enum to get
251 /// the appropriate types to call most of the other functions in
253 pub fn peek(self) -> BucketState<K, V, M> {
254 match unsafe { *self.raw.hash } {
270 /// Modifies the bucket pointer in place to make it point to the next slot.
271 pub fn next(&mut self) {
272 // Branchless bucket iteration step.
273 // As we reach the end of the table...
274 // We take the current idx: 0111111b
275 // Xor it by its increment: ^ 1000000b
278 // Then AND with the capacity: & 1000000b
280 // to get the backwards offset: 1000000b
281 // ... and it's zero at all other times.
282 let maybe_wraparound_dist = (self.idx ^ (self.idx + 1)) & self.table.capacity();
283 // Finally, we obtain the offset 1 or the offset -cap + 1.
284 let dist = 1i - (maybe_wraparound_dist as int);
289 self.raw = self.raw.offset(dist);
294 impl<K, V, M: Deref<RawTable<K, V>>> EmptyBucket<K, V, M> {
296 pub fn next(self) -> Bucket<K, V, M> {
297 let mut bucket = self.into_bucket();
303 pub fn into_bucket(self) -> Bucket<K, V, M> {
311 pub fn gap_peek(self) -> Option<GapThenFull<K, V, M>> {
312 let gap = EmptyBucket {
318 match self.next().peek() {
330 impl<K, V, M: DerefMut<RawTable<K, V>>> EmptyBucket<K, V, M> {
331 /// Puts given key and value pair, along with the key's hash,
332 /// into this bucket in the hashtable. Note how `self` is 'moved' into
333 /// this function, because this slot will no longer be empty when
334 /// we return! A `FullBucket` is returned for later use, pointing to
335 /// the newly-filled slot in the hashtable.
337 /// Use `make_hash` to construct a `SafeHash` to pass to this function.
338 pub fn put(mut self, hash: SafeHash, key: K, value: V)
339 -> FullBucket<K, V, M> {
341 *self.raw.hash = hash.inspect();
342 ptr::write(self.raw.key, key);
343 ptr::write(self.raw.val, value);
346 self.table.size += 1;
348 FullBucket { raw: self.raw, idx: self.idx, table: self.table }
352 impl<K, V, M: Deref<RawTable<K, V>>> FullBucket<K, V, M> {
354 pub fn next(self) -> Bucket<K, V, M> {
355 let mut bucket = self.into_bucket();
361 pub fn into_bucket(self) -> Bucket<K, V, M> {
369 /// Get the distance between this bucket and the 'ideal' location
370 /// as determined by the key's hash stored in it.
372 /// In the cited blog posts above, this is called the "distance to
373 /// initial bucket", or DIB. Also known as "probe count".
374 pub fn distance(&self) -> uint {
375 // Calculates the distance one has to travel when going from
376 // `hash mod capacity` onwards to `idx mod capacity`, wrapping around
377 // if the destination is not reached before the end of the table.
378 (self.idx - self.hash().inspect() as uint) & (self.table.capacity() - 1)
382 pub fn hash(&self) -> SafeHash {
390 /// Gets references to the key and value at a given index.
391 pub fn read(&self) -> (&K, &V) {
399 impl<K, V, M: DerefMut<RawTable<K, V>>> FullBucket<K, V, M> {
400 /// Removes this bucket's key and value from the hashtable.
402 /// This works similarly to `put`, building an `EmptyBucket` out of the
404 pub fn take(mut self) -> (EmptyBucket<K, V, M>, K, V) {
405 let key = self.raw.key as *const K;
406 let val = self.raw.val as *const V;
408 self.table.size -= 1;
411 *self.raw.hash = EMPTY_BUCKET;
424 pub fn replace(&mut self, h: SafeHash, k: K, v: V) -> (SafeHash, K, V) {
426 let old_hash = ptr::replace(self.raw.hash as *mut SafeHash, h);
427 let old_key = ptr::replace(self.raw.key, k);
428 let old_val = ptr::replace(self.raw.val, v);
430 (old_hash, old_key, old_val)
434 /// Gets mutable references to the key and value at a given index.
435 pub fn read_mut(&mut self) -> (&mut K, &mut V) {
443 impl<'t, K, V, M: Deref<RawTable<K, V>> + 't> FullBucket<K, V, M> {
444 /// Exchange a bucket state for immutable references into the table.
445 /// Because the underlying reference to the table is also consumed,
446 /// no further changes to the structure of the table are possible;
447 /// in exchange for this, the returned references have a longer lifetime
448 /// than the references returned by `read()`.
449 pub fn into_refs(self) -> (&'t K, &'t V) {
457 impl<'t, K, V, M: DerefMut<RawTable<K, V>> + 't> FullBucket<K, V, M> {
458 /// This works similarly to `into_refs`, exchanging a bucket state
459 /// for mutable references into the table.
460 pub fn into_mut_refs(self) -> (&'t mut K, &'t mut V) {
468 impl<K, V, M> BucketState<K, V, M> {
470 pub fn expect_full(self) -> FullBucket<K, V, M> {
473 Empty(..) => fail!("Expected full bucket")
478 impl<K, V, M: Deref<RawTable<K, V>>> GapThenFull<K, V, M> {
480 pub fn full(&self) -> &FullBucket<K, V, M> {
484 pub fn shift(mut self) -> Option<GapThenFull<K, V, M>> {
486 *self.gap.raw.hash = mem::replace(&mut *self.full.raw.hash, EMPTY_BUCKET);
487 copy_nonoverlapping_memory(self.gap.raw.key, self.full.raw.key as *const K, 1);
488 copy_nonoverlapping_memory(self.gap.raw.val, self.full.raw.val as *const V, 1);
491 let FullBucket { raw: prev_raw, idx: prev_idx, .. } = self.full;
493 match self.full.next().peek() {
495 self.gap.raw = prev_raw;
496 self.gap.idx = prev_idx;
508 /// Rounds up to a multiple of a power of two. Returns the closest multiple
509 /// of `target_alignment` that is higher or equal to `unrounded`.
513 /// Fails if `target_alignment` is not a power of two.
514 fn round_up_to_next(unrounded: uint, target_alignment: uint) -> uint {
515 assert!(is_power_of_two(target_alignment));
516 (unrounded + target_alignment - 1) & !(target_alignment - 1)
521 assert_eq!(round_up_to_next(0, 4), 0);
522 assert_eq!(round_up_to_next(1, 4), 4);
523 assert_eq!(round_up_to_next(2, 4), 4);
524 assert_eq!(round_up_to_next(3, 4), 4);
525 assert_eq!(round_up_to_next(4, 4), 4);
526 assert_eq!(round_up_to_next(5, 4), 8);
529 // Returns a tuple of (key_offset, val_offset),
530 // from the start of a mallocated array.
531 fn calculate_offsets(hashes_size: uint,
532 keys_size: uint, keys_align: uint,
535 let keys_offset = round_up_to_next(hashes_size, keys_align);
536 let end_of_keys = keys_offset + keys_size;
538 let vals_offset = round_up_to_next(end_of_keys, vals_align);
540 (keys_offset, vals_offset)
543 // Returns a tuple of (minimum required malloc alignment, hash_offset,
544 // array_size), from the start of a mallocated array.
545 fn calculate_allocation(hash_size: uint, hash_align: uint,
546 keys_size: uint, keys_align: uint,
547 vals_size: uint, vals_align: uint)
548 -> (uint, uint, uint) {
550 let (_, vals_offset) = calculate_offsets(hash_size,
551 keys_size, keys_align,
553 let end_of_vals = vals_offset + vals_size;
555 let min_align = cmp::max(hash_align, cmp::max(keys_align, vals_align));
557 (min_align, hash_offset, end_of_vals)
561 fn test_offset_calculation() {
562 assert_eq!(calculate_allocation(128, 8, 15, 1, 4, 4), (8, 0, 148));
563 assert_eq!(calculate_allocation(3, 1, 2, 1, 1, 1), (1, 0, 6));
564 assert_eq!(calculate_allocation(6, 2, 12, 4, 24, 8), (8, 0, 48));
565 assert_eq!(calculate_offsets(128, 15, 1, 4), (128, 144));
566 assert_eq!(calculate_offsets(3, 2, 1, 1), (3, 5));
567 assert_eq!(calculate_offsets(6, 12, 4, 8), (8, 24));
570 impl<K, V> RawTable<K, V> {
571 /// Does not initialize the buckets. The caller should ensure they,
572 /// at the very least, set every hash to EMPTY_BUCKET.
573 unsafe fn new_uninitialized(capacity: uint) -> RawTable<K, V> {
578 hashes: 0 as *mut u64,
579 marker: marker::CovariantType,
582 // No need for `checked_mul` before a more restrictive check performed
583 // later in this method.
584 let hashes_size = capacity * size_of::<u64>();
585 let keys_size = capacity * size_of::< K >();
586 let vals_size = capacity * size_of::< V >();
588 // Allocating hashmaps is a little tricky. We need to allocate three
589 // arrays, but since we know their sizes and alignments up front,
590 // we just allocate a single array, and then have the subarrays
593 // This is great in theory, but in practice getting the alignment
594 // right is a little subtle. Therefore, calculating offsets has been
595 // factored out into a different function.
596 let (malloc_alignment, hash_offset, size) =
597 calculate_allocation(
598 hashes_size, min_align_of::<u64>(),
599 keys_size, min_align_of::< K >(),
600 vals_size, min_align_of::< V >());
602 // One check for overflow that covers calculation and rounding of size.
603 let size_of_bucket = size_of::<u64>().checked_add(&size_of::<K>()).unwrap()
604 .checked_add(&size_of::<V>()).unwrap();
605 assert!(size >= capacity.checked_mul(&size_of_bucket)
606 .expect("capacity overflow"),
607 "capacity overflow");
609 let buffer = allocate(size, malloc_alignment);
611 let hashes = buffer.offset(hash_offset as int) as *mut u64;
617 marker: marker::CovariantType,
621 fn first_bucket_raw(&self) -> RawBucket<K, V> {
622 let hashes_size = self.capacity * size_of::<u64>();
623 let keys_size = self.capacity * size_of::<K>();
625 let buffer = self.hashes as *mut u8;
626 let (keys_offset, vals_offset) = calculate_offsets(hashes_size,
627 keys_size, min_align_of::<K>(),
628 min_align_of::<V>());
633 key: buffer.offset(keys_offset as int) as *mut K,
634 val: buffer.offset(vals_offset as int) as *mut V
639 /// Creates a new raw table from a given capacity. All buckets are
641 #[allow(experimental)]
642 pub fn new(capacity: uint) -> RawTable<K, V> {
644 let ret = RawTable::new_uninitialized(capacity);
645 zero_memory(ret.hashes, capacity);
650 /// The hashtable's capacity, similar to a vector's.
651 pub fn capacity(&self) -> uint {
655 /// The number of elements ever `put` in the hashtable, minus the number
656 /// of elements ever `take`n.
657 pub fn size(&self) -> uint {
661 fn raw_buckets(&self) -> RawBuckets<K, V> {
663 raw: self.first_bucket_raw(),
665 self.hashes.offset(self.capacity as int)
670 pub fn iter(&self) -> Entries<K, V> {
672 iter: self.raw_buckets(),
673 elems_left: self.size(),
677 pub fn iter_mut(&mut self) -> MutEntries<K, V> {
679 iter: self.raw_buckets(),
680 elems_left: self.size(),
684 pub fn into_iter(self) -> MoveEntries<K, V> {
686 iter: self.raw_buckets(),
691 /// Returns an iterator that copies out each entry. Used while the table
692 /// is being dropped.
693 unsafe fn rev_move_buckets(&mut self) -> RevMoveBuckets<K, V> {
694 let raw_bucket = self.first_bucket_raw();
696 raw: raw_bucket.offset(self.capacity as int),
697 hashes_end: raw_bucket.hash,
698 elems_left: self.size
703 /// A raw iterator. The basis for some other iterators in this module. Although
704 /// this interface is safe, it's not used outside this module.
705 struct RawBuckets<'a, K, V> {
706 raw: RawBucket<K, V>,
710 impl<'a, K, V> Iterator<RawBucket<K, V>> for RawBuckets<'a, K, V> {
711 fn next(&mut self) -> Option<RawBucket<K, V>> {
712 while self.raw.hash != self.hashes_end {
714 // We are swapping out the pointer to a bucket and replacing
715 // it with the pointer to the next one.
716 let prev = ptr::replace(&mut self.raw, self.raw.offset(1));
717 if *prev.hash != EMPTY_BUCKET {
727 /// An iterator that moves out buckets in reverse order. It leaves the table
728 /// in an an inconsistent state and should only be used for dropping
729 /// the table's remaining entries. It's used in the implementation of Drop.
730 struct RevMoveBuckets<'a, K, V> {
731 raw: RawBucket<K, V>,
732 hashes_end: *mut u64,
736 impl<'a, K, V> Iterator<(K, V)> for RevMoveBuckets<'a, K, V> {
737 fn next(&mut self) -> Option<(K, V)> {
738 if self.elems_left == 0 {
743 debug_assert!(self.raw.hash != self.hashes_end);
746 self.raw = self.raw.offset(-1);
748 if *self.raw.hash != EMPTY_BUCKET {
749 self.elems_left -= 1;
751 ptr::read(self.raw.key as *const K),
752 ptr::read(self.raw.val as *const V)
760 /// Iterator over shared references to entries in a table.
761 pub struct Entries<'a, K: 'a, V: 'a> {
762 iter: RawBuckets<'a, K, V>,
766 /// Iterator over mutable references to entries in a table.
767 pub struct MutEntries<'a, K: 'a, V: 'a> {
768 iter: RawBuckets<'a, K, V>,
772 /// Iterator over the entries in a table, consuming the table.
773 pub struct MoveEntries<K, V> {
774 table: RawTable<K, V>,
775 iter: RawBuckets<'static, K, V>
778 impl<'a, K, V> Iterator<(&'a K, &'a V)> for Entries<'a, K, V> {
779 fn next(&mut self) -> Option<(&'a K, &'a V)> {
780 self.iter.next().map(|bucket| {
781 self.elems_left -= 1;
789 fn size_hint(&self) -> (uint, Option<uint>) {
790 (self.elems_left, Some(self.elems_left))
794 impl<'a, K, V> Iterator<(&'a K, &'a mut V)> for MutEntries<'a, K, V> {
795 fn next(&mut self) -> Option<(&'a K, &'a mut V)> {
796 self.iter.next().map(|bucket| {
797 self.elems_left -= 1;
805 fn size_hint(&self) -> (uint, Option<uint>) {
806 (self.elems_left, Some(self.elems_left))
810 impl<K, V> Iterator<(SafeHash, K, V)> for MoveEntries<K, V> {
811 fn next(&mut self) -> Option<(SafeHash, K, V)> {
812 self.iter.next().map(|bucket| {
813 self.table.size -= 1;
819 ptr::read(bucket.key as *const K),
820 ptr::read(bucket.val as *const V)
826 fn size_hint(&self) -> (uint, Option<uint>) {
827 let size = self.table.size();
832 impl<K: Clone, V: Clone> Clone for RawTable<K, V> {
833 fn clone(&self) -> RawTable<K, V> {
835 let mut new_ht = RawTable::new_uninitialized(self.capacity());
838 let cap = self.capacity();
839 let mut new_buckets = Bucket::first(&mut new_ht);
840 let mut buckets = Bucket::first(self);
841 while buckets.index() != cap {
842 match buckets.peek() {
845 let (k, v) = full.read();
846 (full.hash(), k.clone(), v.clone())
848 *new_buckets.raw.hash = h.inspect();
849 mem::overwrite(new_buckets.raw.key, k);
850 mem::overwrite(new_buckets.raw.val, v);
853 *new_buckets.raw.hash = EMPTY_BUCKET;
861 new_ht.size = self.size();
869 impl<K, V> Drop for RawTable<K, V> {
871 if self.hashes.is_null() {
874 // This is done in reverse because we've likely partially taken
875 // some elements out with `.into_iter()` from the front.
876 // Check if the size is 0, so we don't do a useless scan when
877 // dropping empty tables such as on resize.
878 // Also avoid double drop of elements that have been already moved out.
880 for _ in self.rev_move_buckets() {}
883 let hashes_size = self.capacity * size_of::<u64>();
884 let keys_size = self.capacity * size_of::<K>();
885 let vals_size = self.capacity * size_of::<V>();
886 let (align, _, size) = calculate_allocation(hashes_size, min_align_of::<u64>(),
887 keys_size, min_align_of::<K>(),
888 vals_size, min_align_of::<V>());
891 deallocate(self.hashes as *mut u8, size, align);
892 // Remember how everything was allocated out of one buffer
893 // during initialization? We only need one call to free here.