1 use crate::vec::{Idx, IndexVec};
2 use smallvec::SmallVec;
5 use std::marker::PhantomData;
13 pub const WORD_BYTES: usize = mem::size_of::<Word>();
14 pub const WORD_BITS: usize = WORD_BYTES * 8;
16 /// A fixed-size bitset type with a dense representation.
18 /// NOTE: Use [`GrowableBitSet`] if you need support for resizing after creation.
20 /// `T` is an index type, typically a newtyped `usize` wrapper, but it can also
23 /// All operations that involve an element will panic if the element is equal
24 /// to or greater than the domain size. All operations that involve two bitsets
25 /// will panic if the bitsets have differing domain sizes.
27 /// [`GrowableBitSet`]: struct.GrowableBitSet.html
28 #[derive(Clone, Eq, PartialEq, RustcDecodable, RustcEncodable)]
29 pub struct BitSet<T: Idx> {
32 marker: PhantomData<T>,
35 impl<T: Idx> BitSet<T> {
36 /// Creates a new, empty bitset with a given `domain_size`.
38 pub fn new_empty(domain_size: usize) -> BitSet<T> {
39 let num_words = num_words(domain_size);
42 words: vec![0; num_words],
47 /// Creates a new, filled bitset with a given `domain_size`.
49 pub fn new_filled(domain_size: usize) -> BitSet<T> {
50 let num_words = num_words(domain_size);
51 let mut result = BitSet {
53 words: vec![!0; num_words],
56 result.clear_excess_bits();
60 /// Gets the domain size.
61 pub fn domain_size(&self) -> usize {
65 /// Clear all elements.
67 pub fn clear(&mut self) {
68 for word in &mut self.words {
73 /// Clear excess bits in the final word.
74 fn clear_excess_bits(&mut self) {
75 let num_bits_in_final_word = self.domain_size % WORD_BITS;
76 if num_bits_in_final_word > 0 {
77 let mask = (1 << num_bits_in_final_word) - 1;
78 let final_word_idx = self.words.len() - 1;
79 self.words[final_word_idx] &= mask;
83 /// Efficiently overwrite `self` with `other`.
84 pub fn overwrite(&mut self, other: &BitSet<T>) {
85 assert!(self.domain_size == other.domain_size);
86 self.words.clone_from_slice(&other.words);
89 /// Count the number of set bits in the set.
90 pub fn count(&self) -> usize {
91 self.words.iter().map(|e| e.count_ones() as usize).sum()
94 /// Returns `true` if `self` contains `elem`.
96 pub fn contains(&self, elem: T) -> bool {
97 assert!(elem.index() < self.domain_size);
98 let (word_index, mask) = word_index_and_mask(elem);
99 (self.words[word_index] & mask) != 0
102 /// Is `self` is a (non-strict) superset of `other`?
104 pub fn superset(&self, other: &BitSet<T>) -> bool {
105 assert_eq!(self.domain_size, other.domain_size);
106 self.words.iter().zip(&other.words).all(|(a, b)| (a & b) == *b)
109 /// Is the set empty?
111 pub fn is_empty(&self) -> bool {
112 self.words.iter().all(|a| *a == 0)
115 /// Insert `elem`. Returns whether the set has changed.
117 pub fn insert(&mut self, elem: T) -> bool {
118 assert!(elem.index() < self.domain_size);
119 let (word_index, mask) = word_index_and_mask(elem);
120 let word_ref = &mut self.words[word_index];
121 let word = *word_ref;
122 let new_word = word | mask;
123 *word_ref = new_word;
127 /// Sets all bits to true.
128 pub fn insert_all(&mut self) {
129 for word in &mut self.words {
132 self.clear_excess_bits();
135 /// Returns `true` if the set has changed.
137 pub fn remove(&mut self, elem: T) -> bool {
138 assert!(elem.index() < self.domain_size);
139 let (word_index, mask) = word_index_and_mask(elem);
140 let word_ref = &mut self.words[word_index];
141 let word = *word_ref;
142 let new_word = word & !mask;
143 *word_ref = new_word;
147 /// Sets `self = self | other` and returns `true` if `self` changed
148 /// (i.e., if new bits were added).
149 pub fn union(&mut self, other: &impl UnionIntoBitSet<T>) -> bool {
150 other.union_into(self)
153 /// Sets `self = self - other` and returns `true` if `self` changed.
154 /// (i.e., if any bits were removed).
155 pub fn subtract(&mut self, other: &impl SubtractFromBitSet<T>) -> bool {
156 other.subtract_from(self)
159 /// Sets `self = self & other` and return `true` if `self` changed.
160 /// (i.e., if any bits were removed).
161 pub fn intersect(&mut self, other: &BitSet<T>) -> bool {
162 assert_eq!(self.domain_size, other.domain_size);
163 bitwise(&mut self.words, &other.words, |a, b| { a & b })
166 /// Gets a slice of the underlying words.
167 pub fn words(&self) -> &[Word] {
171 /// Iterates over the indices of set bits in a sorted order.
173 pub fn iter(&self) -> BitIter<'_, T> {
174 BitIter::new(&self.words)
177 /// Duplicates the set as a hybrid set.
178 pub fn to_hybrid(&self) -> HybridBitSet<T> {
179 // Note: we currently don't bother trying to make a Sparse set.
180 HybridBitSet::Dense(self.to_owned())
183 /// Set `self = self | other`. In contrast to `union` returns `true` if the set contains at
184 /// least one bit that is not in `other` (i.e. `other` is not a superset of `self`).
186 /// This is an optimization for union of a hybrid bitset.
187 fn reverse_union_sparse(&mut self, sparse: &SparseBitSet<T>) -> bool {
188 assert!(sparse.domain_size == self.domain_size);
189 self.clear_excess_bits();
191 let mut not_already = false;
192 // Index of the current word not yet merged.
193 let mut current_index = 0;
194 // Mask of bits that came from the sparse set in the current word.
195 let mut new_bit_mask = 0;
196 for (word_index, mask) in sparse.iter().map(|x| word_index_and_mask(*x)) {
197 // Next bit is in a word not inspected yet.
198 if word_index > current_index {
199 self.words[current_index] |= new_bit_mask;
200 // Were there any bits in the old word that did not occur in the sparse set?
201 not_already |= (self.words[current_index] ^ new_bit_mask) != 0;
202 // Check all words we skipped for any set bit.
203 not_already |= self.words[current_index+1..word_index].iter().any(|&x| x != 0);
205 current_index = word_index;
206 // Reset bit mask, no bits have been merged yet.
209 // Add bit and mark it as coming from the sparse set.
210 // self.words[word_index] |= mask;
211 new_bit_mask |= mask;
213 self.words[current_index] |= new_bit_mask;
214 // Any bits in the last inspected word that were not in the sparse set?
215 not_already |= (self.words[current_index] ^ new_bit_mask) != 0;
216 // Any bits in the tail? Note `clear_excess_bits` before.
217 not_already |= self.words[current_index+1..].iter().any(|&x| x != 0);
223 /// This is implemented by all the bitsets so that BitSet::union() can be
224 /// passed any type of bitset.
225 pub trait UnionIntoBitSet<T: Idx> {
226 // Performs `other = other | self`.
227 fn union_into(&self, other: &mut BitSet<T>) -> bool;
230 /// This is implemented by all the bitsets so that BitSet::subtract() can be
231 /// passed any type of bitset.
232 pub trait SubtractFromBitSet<T: Idx> {
233 // Performs `other = other - self`.
234 fn subtract_from(&self, other: &mut BitSet<T>) -> bool;
237 impl<T: Idx> UnionIntoBitSet<T> for BitSet<T> {
238 fn union_into(&self, other: &mut BitSet<T>) -> bool {
239 assert_eq!(self.domain_size, other.domain_size);
240 bitwise(&mut other.words, &self.words, |a, b| { a | b })
244 impl<T: Idx> SubtractFromBitSet<T> for BitSet<T> {
245 fn subtract_from(&self, other: &mut BitSet<T>) -> bool {
246 assert_eq!(self.domain_size, other.domain_size);
247 bitwise(&mut other.words, &self.words, |a, b| { a & !b })
251 impl<T: Idx> fmt::Debug for BitSet<T> {
252 fn fmt(&self, w: &mut fmt::Formatter<'_>) -> fmt::Result {
254 .entries(self.iter())
259 impl<T: Idx> ToString for BitSet<T> {
260 fn to_string(&self) -> String {
261 let mut result = String::new();
264 // Note: this is a little endian printout of bytes.
266 // i tracks how many bits we have printed so far.
268 for word in &self.words {
269 let mut word = *word;
270 for _ in 0..WORD_BYTES { // for each byte in `word`:
271 let remain = self.domain_size - i;
272 // If less than a byte remains, then mask just that many bits.
273 let mask = if remain <= 8 { (1 << remain) - 1 } else { 0xFF };
274 assert!(mask <= 0xFF);
275 let byte = word & mask;
277 result.push_str(&format!("{}{:02x}", sep, byte));
279 if remain <= 8 { break; }
292 pub struct BitIter<'a, T: Idx> {
293 /// A copy of the current word, but with any already-visited bits cleared.
294 /// (This lets us use `trailing_zeros()` to find the next set bit.) When it
295 /// is reduced to 0, we move onto the next word.
298 /// The offset (measured in bits) of the current word.
301 /// Underlying iterator over the words.
302 iter: slice::Iter<'a, Word>,
304 marker: PhantomData<T>
307 impl<'a, T: Idx> BitIter<'a, T> {
309 fn new(words: &'a [Word]) -> BitIter<'a, T> {
310 // We initialize `word` and `offset` to degenerate values. On the first
311 // call to `next()` we will fall through to getting the first word from
312 // `iter`, which sets `word` to the first word (if there is one) and
313 // `offset` to 0. Doing it this way saves us from having to maintain
314 // additional state about whether we have started.
317 offset: std::usize::MAX - (WORD_BITS - 1),
324 impl<'a, T: Idx> Iterator for BitIter<'a, T> {
326 fn next(&mut self) -> Option<T> {
329 // Get the position of the next set bit in the current word,
330 // then clear the bit.
331 let bit_pos = self.word.trailing_zeros() as usize;
332 let bit = 1 << bit_pos;
334 return Some(T::new(bit_pos + self.offset))
337 // Move onto the next word. `wrapping_add()` is needed to handle
338 // the degenerate initial value given to `offset` in `new()`.
339 let word = self.iter.next()?;
341 self.offset = self.offset.wrapping_add(WORD_BITS);
347 fn bitwise<Op>(out_vec: &mut [Word], in_vec: &[Word], op: Op) -> bool
348 where Op: Fn(Word, Word) -> Word
350 assert_eq!(out_vec.len(), in_vec.len());
351 let mut changed = false;
352 for (out_elem, in_elem) in out_vec.iter_mut().zip(in_vec.iter()) {
353 let old_val = *out_elem;
354 let new_val = op(old_val, *in_elem);
356 changed |= old_val != new_val;
361 const SPARSE_MAX: usize = 8;
363 /// A fixed-size bitset type with a sparse representation and a maximum of
364 /// `SPARSE_MAX` elements. The elements are stored as a sorted `SmallVec` with
365 /// no duplicates; although `SmallVec` can spill its elements to the heap, that
366 /// never happens within this type because of the `SPARSE_MAX` limit.
368 /// This type is used by `HybridBitSet`; do not use directly.
369 #[derive(Clone, Debug)]
370 pub struct SparseBitSet<T: Idx> {
372 elems: SmallVec<[T; SPARSE_MAX]>,
375 impl<T: Idx> SparseBitSet<T> {
376 fn new_empty(domain_size: usize) -> Self {
379 elems: SmallVec::new()
383 fn len(&self) -> usize {
387 fn is_empty(&self) -> bool {
388 self.elems.len() == 0
391 fn contains(&self, elem: T) -> bool {
392 assert!(elem.index() < self.domain_size);
393 self.elems.contains(&elem)
396 fn insert(&mut self, elem: T) -> bool {
397 assert!(elem.index() < self.domain_size);
398 let changed = if let Some(i) = self.elems.iter().position(|&e| e >= elem) {
399 if self.elems[i] == elem {
400 // `elem` is already in the set.
403 // `elem` is smaller than one or more existing elements.
404 self.elems.insert(i, elem);
408 // `elem` is larger than all existing elements.
409 self.elems.push(elem);
412 assert!(self.len() <= SPARSE_MAX);
416 fn remove(&mut self, elem: T) -> bool {
417 assert!(elem.index() < self.domain_size);
418 if let Some(i) = self.elems.iter().position(|&e| e == elem) {
419 self.elems.remove(i);
426 fn to_dense(&self) -> BitSet<T> {
427 let mut dense = BitSet::new_empty(self.domain_size);
428 for elem in self.elems.iter() {
434 fn iter(&self) -> slice::Iter<'_, T> {
439 impl<T: Idx> UnionIntoBitSet<T> for SparseBitSet<T> {
440 fn union_into(&self, other: &mut BitSet<T>) -> bool {
441 assert_eq!(self.domain_size, other.domain_size);
442 let mut changed = false;
443 for elem in self.iter() {
444 changed |= other.insert(*elem);
450 impl<T: Idx> SubtractFromBitSet<T> for SparseBitSet<T> {
451 fn subtract_from(&self, other: &mut BitSet<T>) -> bool {
452 assert_eq!(self.domain_size, other.domain_size);
453 let mut changed = false;
454 for elem in self.iter() {
455 changed |= other.remove(*elem);
461 /// A fixed-size bitset type with a hybrid representation: sparse when there
462 /// are up to a `SPARSE_MAX` elements in the set, but dense when there are more
463 /// than `SPARSE_MAX`.
465 /// This type is especially efficient for sets that typically have a small
466 /// number of elements, but a large `domain_size`, and are cleared frequently.
468 /// `T` is an index type, typically a newtyped `usize` wrapper, but it can also
471 /// All operations that involve an element will panic if the element is equal
472 /// to or greater than the domain size. All operations that involve two bitsets
473 /// will panic if the bitsets have differing domain sizes.
474 #[derive(Clone, Debug)]
475 pub enum HybridBitSet<T: Idx> {
476 Sparse(SparseBitSet<T>),
480 impl<T: Idx> HybridBitSet<T> {
481 pub fn new_empty(domain_size: usize) -> Self {
482 HybridBitSet::Sparse(SparseBitSet::new_empty(domain_size))
485 fn domain_size(&self) -> usize {
487 HybridBitSet::Sparse(sparse) => sparse.domain_size,
488 HybridBitSet::Dense(dense) => dense.domain_size,
492 pub fn clear(&mut self) {
493 let domain_size = self.domain_size();
494 *self = HybridBitSet::new_empty(domain_size);
497 pub fn contains(&self, elem: T) -> bool {
499 HybridBitSet::Sparse(sparse) => sparse.contains(elem),
500 HybridBitSet::Dense(dense) => dense.contains(elem),
504 pub fn superset(&self, other: &HybridBitSet<T>) -> bool {
505 match (self, other) {
506 (HybridBitSet::Dense(self_dense), HybridBitSet::Dense(other_dense)) => {
507 self_dense.superset(other_dense)
510 assert!(self.domain_size() == other.domain_size());
511 other.iter().all(|elem| self.contains(elem))
516 pub fn is_empty(&self) -> bool {
518 HybridBitSet::Sparse(sparse) => sparse.is_empty(),
519 HybridBitSet::Dense(dense) => dense.is_empty(),
523 pub fn insert(&mut self, elem: T) -> bool {
524 // No need to check `elem` against `self.domain_size` here because all
525 // the match cases check it, one way or another.
527 HybridBitSet::Sparse(sparse) if sparse.len() < SPARSE_MAX => {
528 // The set is sparse and has space for `elem`.
531 HybridBitSet::Sparse(sparse) if sparse.contains(elem) => {
532 // The set is sparse and does not have space for `elem`, but
533 // that doesn't matter because `elem` is already present.
536 HybridBitSet::Sparse(sparse) => {
537 // The set is sparse and full. Convert to a dense set.
538 let mut dense = sparse.to_dense();
539 let changed = dense.insert(elem);
541 *self = HybridBitSet::Dense(dense);
544 HybridBitSet::Dense(dense) => dense.insert(elem),
548 pub fn insert_all(&mut self) {
549 let domain_size = self.domain_size();
551 HybridBitSet::Sparse(_) => {
552 *self = HybridBitSet::Dense(BitSet::new_filled(domain_size));
554 HybridBitSet::Dense(dense) => dense.insert_all(),
558 pub fn remove(&mut self, elem: T) -> bool {
559 // Note: we currently don't bother going from Dense back to Sparse.
561 HybridBitSet::Sparse(sparse) => sparse.remove(elem),
562 HybridBitSet::Dense(dense) => dense.remove(elem),
566 pub fn union(&mut self, other: &HybridBitSet<T>) -> bool {
568 HybridBitSet::Sparse(self_sparse) => {
570 HybridBitSet::Sparse(other_sparse) => {
571 // Both sets are sparse. Add the elements in
572 // `other_sparse` to `self` one at a time. This
573 // may or may not cause `self` to be densified.
574 assert_eq!(self.domain_size(), other.domain_size());
575 let mut changed = false;
576 for elem in other_sparse.iter() {
577 changed |= self.insert(*elem);
581 HybridBitSet::Dense(other_dense) => {
582 // `self` is sparse and `other` is dense. To
583 // merge them, we have two available strategies:
584 // * Densify `self` then merge other
585 // * Clone other then integrate bits from `self`
586 // The second strategy requires dedicated method
587 // since the usual `union` returns the wrong
588 // result. In the dedicated case the computation
589 // is slightly faster if the bits of the sparse
590 // bitset map to only few words of the dense
591 // representation, i.e. indices are near each
594 // Benchmarking seems to suggest that the second
595 // option is worth it.
596 let mut new_dense = other_dense.clone();
597 let changed = new_dense.reverse_union_sparse(self_sparse);
598 *self = HybridBitSet::Dense(new_dense);
604 HybridBitSet::Dense(self_dense) => self_dense.union(other),
608 /// Converts to a dense set, consuming itself in the process.
609 pub fn to_dense(self) -> BitSet<T> {
611 HybridBitSet::Sparse(sparse) => sparse.to_dense(),
612 HybridBitSet::Dense(dense) => dense,
616 pub fn iter(&self) -> HybridIter<'_, T> {
618 HybridBitSet::Sparse(sparse) => HybridIter::Sparse(sparse.iter()),
619 HybridBitSet::Dense(dense) => HybridIter::Dense(dense.iter()),
624 impl<T: Idx> UnionIntoBitSet<T> for HybridBitSet<T> {
625 fn union_into(&self, other: &mut BitSet<T>) -> bool {
627 HybridBitSet::Sparse(sparse) => sparse.union_into(other),
628 HybridBitSet::Dense(dense) => dense.union_into(other),
633 impl<T: Idx> SubtractFromBitSet<T> for HybridBitSet<T> {
634 fn subtract_from(&self, other: &mut BitSet<T>) -> bool {
636 HybridBitSet::Sparse(sparse) => sparse.subtract_from(other),
637 HybridBitSet::Dense(dense) => dense.subtract_from(other),
642 pub enum HybridIter<'a, T: Idx> {
643 Sparse(slice::Iter<'a, T>),
644 Dense(BitIter<'a, T>),
647 impl<'a, T: Idx> Iterator for HybridIter<'a, T> {
650 fn next(&mut self) -> Option<T> {
652 HybridIter::Sparse(sparse) => sparse.next().copied(),
653 HybridIter::Dense(dense) => dense.next(),
658 /// A resizable bitset type with a dense representation.
660 /// `T` is an index type, typically a newtyped `usize` wrapper, but it can also
663 /// All operations that involve an element will panic if the element is equal
664 /// to or greater than the domain size.
665 #[derive(Clone, Debug, PartialEq)]
666 pub struct GrowableBitSet<T: Idx> {
670 impl<T: Idx> GrowableBitSet<T> {
671 /// Ensure that the set can hold at least `min_domain_size` elements.
672 pub fn ensure(&mut self, min_domain_size: usize) {
673 if self.bit_set.domain_size < min_domain_size {
674 self.bit_set.domain_size = min_domain_size;
677 let min_num_words = num_words(min_domain_size);
678 if self.bit_set.words.len() < min_num_words {
679 self.bit_set.words.resize(min_num_words, 0)
683 pub fn new_empty() -> GrowableBitSet<T> {
684 GrowableBitSet { bit_set: BitSet::new_empty(0) }
687 pub fn with_capacity(capacity: usize) -> GrowableBitSet<T> {
688 GrowableBitSet { bit_set: BitSet::new_empty(capacity) }
691 /// Returns `true` if the set has changed.
693 pub fn insert(&mut self, elem: T) -> bool {
694 self.ensure(elem.index() + 1);
695 self.bit_set.insert(elem)
699 pub fn contains(&self, elem: T) -> bool {
700 let (word_index, mask) = word_index_and_mask(elem);
701 if let Some(word) = self.bit_set.words.get(word_index) {
709 /// A fixed-size 2D bit matrix type with a dense representation.
711 /// `R` and `C` are index types used to identify rows and columns respectively;
712 /// typically newtyped `usize` wrappers, but they can also just be `usize`.
714 /// All operations that involve a row and/or column index will panic if the
715 /// index exceeds the relevant bound.
716 #[derive(Clone, Debug, Eq, PartialEq, RustcDecodable, RustcEncodable)]
717 pub struct BitMatrix<R: Idx, C: Idx> {
721 marker: PhantomData<(R, C)>,
724 impl<R: Idx, C: Idx> BitMatrix<R, C> {
725 /// Creates a new `rows x columns` matrix, initially empty.
726 pub fn new(num_rows: usize, num_columns: usize) -> BitMatrix<R, C> {
727 // For every element, we need one bit for every other
728 // element. Round up to an even number of words.
729 let words_per_row = num_words(num_columns);
733 words: vec![0; num_rows * words_per_row],
738 /// Creates a new matrix, with `row` used as the value for every row.
739 pub fn from_row_n(row: &BitSet<C>, num_rows: usize) -> BitMatrix<R, C> {
740 let num_columns = row.domain_size();
741 let words_per_row = num_words(num_columns);
742 assert_eq!(words_per_row, row.words().len());
746 words: iter::repeat(row.words()).take(num_rows).flatten().cloned().collect(),
751 pub fn rows(&self) -> impl Iterator<Item = R> {
752 (0..self.num_rows).map(R::new)
755 /// The range of bits for a given row.
756 fn range(&self, row: R) -> (usize, usize) {
757 let words_per_row = num_words(self.num_columns);
758 let start = row.index() * words_per_row;
759 (start, start + words_per_row)
762 /// Sets the cell at `(row, column)` to true. Put another way, insert
763 /// `column` to the bitset for `row`.
765 /// Returns `true` if this changed the matrix.
766 pub fn insert(&mut self, row: R, column: C) -> bool {
767 assert!(row.index() < self.num_rows && column.index() < self.num_columns);
768 let (start, _) = self.range(row);
769 let (word_index, mask) = word_index_and_mask(column);
770 let words = &mut self.words[..];
771 let word = words[start + word_index];
772 let new_word = word | mask;
773 words[start + word_index] = new_word;
777 /// Do the bits from `row` contain `column`? Put another way, is
778 /// the matrix cell at `(row, column)` true? Put yet another way,
779 /// if the matrix represents (transitive) reachability, can
780 /// `row` reach `column`?
781 pub fn contains(&self, row: R, column: C) -> bool {
782 assert!(row.index() < self.num_rows && column.index() < self.num_columns);
783 let (start, _) = self.range(row);
784 let (word_index, mask) = word_index_and_mask(column);
785 (self.words[start + word_index] & mask) != 0
788 /// Returns those indices that are true in rows `a` and `b`. This
789 /// is an O(n) operation where `n` is the number of elements
790 /// (somewhat independent from the actual size of the
791 /// intersection, in particular).
792 pub fn intersect_rows(&self, row1: R, row2: R) -> Vec<C> {
793 assert!(row1.index() < self.num_rows && row2.index() < self.num_rows);
794 let (row1_start, row1_end) = self.range(row1);
795 let (row2_start, row2_end) = self.range(row2);
796 let mut result = Vec::with_capacity(self.num_columns);
797 for (base, (i, j)) in (row1_start..row1_end).zip(row2_start..row2_end).enumerate() {
798 let mut v = self.words[i] & self.words[j];
799 for bit in 0..WORD_BITS {
804 result.push(C::new(base * WORD_BITS + bit));
812 /// Adds the bits from row `read` to the bits from row `write`, and
813 /// returns `true` if anything changed.
815 /// This is used when computing transitive reachability because if
816 /// you have an edge `write -> read`, because in that case
817 /// `write` can reach everything that `read` can (and
818 /// potentially more).
819 pub fn union_rows(&mut self, read: R, write: R) -> bool {
820 assert!(read.index() < self.num_rows && write.index() < self.num_rows);
821 let (read_start, read_end) = self.range(read);
822 let (write_start, write_end) = self.range(write);
823 let words = &mut self.words[..];
824 let mut changed = false;
825 for (read_index, write_index) in (read_start..read_end).zip(write_start..write_end) {
826 let word = words[write_index];
827 let new_word = word | words[read_index];
828 words[write_index] = new_word;
829 changed |= word != new_word;
834 /// Adds the bits from `with` to the bits from row `write`, and
835 /// returns `true` if anything changed.
836 pub fn union_row_with(&mut self, with: &BitSet<C>, write: R) -> bool {
837 assert!(write.index() < self.num_rows);
838 assert_eq!(with.domain_size(), self.num_columns);
839 let (write_start, write_end) = self.range(write);
840 let mut changed = false;
841 for (read_index, write_index) in (0..with.words().len()).zip(write_start..write_end) {
842 let word = self.words[write_index];
843 let new_word = word | with.words()[read_index];
844 self.words[write_index] = new_word;
845 changed |= word != new_word;
850 /// Sets every cell in `row` to true.
851 pub fn insert_all_into_row(&mut self, row: R) {
852 assert!(row.index() < self.num_rows);
853 let (start, end) = self.range(row);
854 let words = &mut self.words[..];
855 for index in start..end {
858 self.clear_excess_bits(row);
861 /// Clear excess bits in the final word of the row.
862 fn clear_excess_bits(&mut self, row: R) {
863 let num_bits_in_final_word = self.num_columns % WORD_BITS;
864 if num_bits_in_final_word > 0 {
865 let mask = (1 << num_bits_in_final_word) - 1;
866 let (_, end) = self.range(row);
867 let final_word_idx = end - 1;
868 self.words[final_word_idx] &= mask;
872 /// Gets a slice of the underlying words.
873 pub fn words(&self) -> &[Word] {
877 /// Iterates through all the columns set to true in a given row of
879 pub fn iter(&self, row: R) -> BitIter<'_, C> {
880 assert!(row.index() < self.num_rows);
881 let (start, end) = self.range(row);
882 BitIter::new(&self.words[start..end])
885 /// Returns the number of elements in `row`.
886 pub fn count(&self, row: R) -> usize {
887 let (start, end) = self.range(row);
888 self.words[start..end].iter().map(|e| e.count_ones() as usize).sum()
892 /// A fixed-column-size, variable-row-size 2D bit matrix with a moderately
893 /// sparse representation.
895 /// Initially, every row has no explicit representation. If any bit within a
896 /// row is set, the entire row is instantiated as `Some(<HybridBitSet>)`.
897 /// Furthermore, any previously uninstantiated rows prior to it will be
898 /// instantiated as `None`. Those prior rows may themselves become fully
899 /// instantiated later on if any of their bits are set.
901 /// `R` and `C` are index types used to identify rows and columns respectively;
902 /// typically newtyped `usize` wrappers, but they can also just be `usize`.
903 #[derive(Clone, Debug)]
904 pub struct SparseBitMatrix<R, C>
910 rows: IndexVec<R, Option<HybridBitSet<C>>>,
913 impl<R: Idx, C: Idx> SparseBitMatrix<R, C> {
914 /// Creates a new empty sparse bit matrix with no rows or columns.
915 pub fn new(num_columns: usize) -> Self {
918 rows: IndexVec::new(),
922 fn ensure_row(&mut self, row: R) -> &mut HybridBitSet<C> {
923 // Instantiate any missing rows up to and including row `row` with an
924 // empty HybridBitSet.
925 self.rows.ensure_contains_elem(row, || None);
927 // Then replace row `row` with a full HybridBitSet if necessary.
928 let num_columns = self.num_columns;
929 self.rows[row].get_or_insert_with(|| HybridBitSet::new_empty(num_columns))
932 /// Sets the cell at `(row, column)` to true. Put another way, insert
933 /// `column` to the bitset for `row`.
935 /// Returns `true` if this changed the matrix.
936 pub fn insert(&mut self, row: R, column: C) -> bool {
937 self.ensure_row(row).insert(column)
940 /// Do the bits from `row` contain `column`? Put another way, is
941 /// the matrix cell at `(row, column)` true? Put yet another way,
942 /// if the matrix represents (transitive) reachability, can
943 /// `row` reach `column`?
944 pub fn contains(&self, row: R, column: C) -> bool {
945 self.row(row).map_or(false, |r| r.contains(column))
948 /// Adds the bits from row `read` to the bits from row `write`, and
949 /// returns `true` if anything changed.
951 /// This is used when computing transitive reachability because if
952 /// you have an edge `write -> read`, because in that case
953 /// `write` can reach everything that `read` can (and
954 /// potentially more).
955 pub fn union_rows(&mut self, read: R, write: R) -> bool {
956 if read == write || self.row(read).is_none() {
960 self.ensure_row(write);
961 if let (Some(read_row), Some(write_row)) = self.rows.pick2_mut(read, write) {
962 write_row.union(read_row)
968 /// Union a row, `from`, into the `into` row.
969 pub fn union_into_row(&mut self, into: R, from: &HybridBitSet<C>) -> bool {
970 self.ensure_row(into).union(from)
973 /// Insert all bits in the given row.
974 pub fn insert_all_into_row(&mut self, row: R) {
975 self.ensure_row(row).insert_all();
978 pub fn rows(&self) -> impl Iterator<Item = R> {
982 /// Iterates through all the columns set to true in a given row of
984 pub fn iter<'a>(&'a self, row: R) -> impl Iterator<Item = C> + 'a {
985 self.row(row).into_iter().flat_map(|r| r.iter())
988 pub fn row(&self, row: R) -> Option<&HybridBitSet<C>> {
989 if let Some(Some(row)) = self.rows.get(row) {
998 fn num_words<T: Idx>(domain_size: T) -> usize {
999 (domain_size.index() + WORD_BITS - 1) / WORD_BITS
1003 fn word_index_and_mask<T: Idx>(elem: T) -> (usize, Word) {
1004 let elem = elem.index();
1005 let word_index = elem / WORD_BITS;
1006 let mask = 1 << (elem % WORD_BITS);