1 use crate::vec::{Idx, IndexVec};
2 use arrayvec::ArrayVec;
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);
40 BitSet { domain_size, words: vec![0; num_words], marker: PhantomData }
43 /// Creates a new, filled bitset with a given `domain_size`.
45 pub fn new_filled(domain_size: usize) -> BitSet<T> {
46 let num_words = num_words(domain_size);
47 let mut result = BitSet { domain_size, words: vec![!0; num_words], marker: PhantomData };
48 result.clear_excess_bits();
52 /// Gets the domain size.
53 pub fn domain_size(&self) -> usize {
57 /// Clear all elements.
59 pub fn clear(&mut self) {
60 for word in &mut self.words {
65 /// Clear excess bits in the final word.
66 fn clear_excess_bits(&mut self) {
67 let num_bits_in_final_word = self.domain_size % WORD_BITS;
68 if num_bits_in_final_word > 0 {
69 let mask = (1 << num_bits_in_final_word) - 1;
70 let final_word_idx = self.words.len() - 1;
71 self.words[final_word_idx] &= mask;
75 /// Efficiently overwrite `self` with `other`.
76 pub fn overwrite(&mut self, other: &BitSet<T>) {
77 assert!(self.domain_size == other.domain_size);
78 self.words.clone_from_slice(&other.words);
81 /// Count the number of set bits in the set.
82 pub fn count(&self) -> usize {
83 self.words.iter().map(|e| e.count_ones() as usize).sum()
86 /// Returns `true` if `self` contains `elem`.
88 pub fn contains(&self, elem: T) -> bool {
89 assert!(elem.index() < self.domain_size);
90 let (word_index, mask) = word_index_and_mask(elem);
91 (self.words[word_index] & mask) != 0
94 /// Is `self` is a (non-strict) superset of `other`?
96 pub fn superset(&self, other: &BitSet<T>) -> bool {
97 assert_eq!(self.domain_size, other.domain_size);
98 self.words.iter().zip(&other.words).all(|(a, b)| (a & b) == *b)
101 /// Is the set empty?
103 pub fn is_empty(&self) -> bool {
104 self.words.iter().all(|a| *a == 0)
107 /// Insert `elem`. Returns whether the set has changed.
109 pub fn insert(&mut self, elem: T) -> bool {
110 assert!(elem.index() < self.domain_size);
111 let (word_index, mask) = word_index_and_mask(elem);
112 let word_ref = &mut self.words[word_index];
113 let word = *word_ref;
114 let new_word = word | mask;
115 *word_ref = new_word;
119 /// Sets all bits to true.
120 pub fn insert_all(&mut self) {
121 for word in &mut self.words {
124 self.clear_excess_bits();
127 /// Returns `true` if the set has changed.
129 pub fn remove(&mut self, elem: T) -> bool {
130 assert!(elem.index() < self.domain_size);
131 let (word_index, mask) = word_index_and_mask(elem);
132 let word_ref = &mut self.words[word_index];
133 let word = *word_ref;
134 let new_word = word & !mask;
135 *word_ref = new_word;
139 /// Sets `self = self | other` and returns `true` if `self` changed
140 /// (i.e., if new bits were added).
141 pub fn union(&mut self, other: &impl UnionIntoBitSet<T>) -> bool {
142 other.union_into(self)
145 /// Sets `self = self - other` and returns `true` if `self` changed.
146 /// (i.e., if any bits were removed).
147 pub fn subtract(&mut self, other: &impl SubtractFromBitSet<T>) -> bool {
148 other.subtract_from(self)
151 /// Sets `self = self & other` and return `true` if `self` changed.
152 /// (i.e., if any bits were removed).
153 pub fn intersect(&mut self, other: &BitSet<T>) -> bool {
154 assert_eq!(self.domain_size, other.domain_size);
155 bitwise(&mut self.words, &other.words, |a, b| a & b)
158 /// Gets a slice of the underlying words.
159 pub fn words(&self) -> &[Word] {
163 /// Iterates over the indices of set bits in a sorted order.
165 pub fn iter(&self) -> BitIter<'_, T> {
166 BitIter::new(&self.words)
169 /// Duplicates the set as a hybrid set.
170 pub fn to_hybrid(&self) -> HybridBitSet<T> {
171 // Note: we currently don't bother trying to make a Sparse set.
172 HybridBitSet::Dense(self.to_owned())
175 /// Set `self = self | other`. In contrast to `union` returns `true` if the set contains at
176 /// least one bit that is not in `other` (i.e. `other` is not a superset of `self`).
178 /// This is an optimization for union of a hybrid bitset.
179 fn reverse_union_sparse(&mut self, sparse: &SparseBitSet<T>) -> bool {
180 assert!(sparse.domain_size == self.domain_size);
181 self.clear_excess_bits();
183 let mut not_already = false;
184 // Index of the current word not yet merged.
185 let mut current_index = 0;
186 // Mask of bits that came from the sparse set in the current word.
187 let mut new_bit_mask = 0;
188 for (word_index, mask) in sparse.iter().map(|x| word_index_and_mask(*x)) {
189 // Next bit is in a word not inspected yet.
190 if word_index > current_index {
191 self.words[current_index] |= new_bit_mask;
192 // Were there any bits in the old word that did not occur in the sparse set?
193 not_already |= (self.words[current_index] ^ new_bit_mask) != 0;
194 // Check all words we skipped for any set bit.
195 not_already |= self.words[current_index + 1..word_index].iter().any(|&x| x != 0);
197 current_index = word_index;
198 // Reset bit mask, no bits have been merged yet.
201 // Add bit and mark it as coming from the sparse set.
202 // self.words[word_index] |= mask;
203 new_bit_mask |= mask;
205 self.words[current_index] |= new_bit_mask;
206 // Any bits in the last inspected word that were not in the sparse set?
207 not_already |= (self.words[current_index] ^ new_bit_mask) != 0;
208 // Any bits in the tail? Note `clear_excess_bits` before.
209 not_already |= self.words[current_index + 1..].iter().any(|&x| x != 0);
215 /// This is implemented by all the bitsets so that BitSet::union() can be
216 /// passed any type of bitset.
217 pub trait UnionIntoBitSet<T: Idx> {
218 // Performs `other = other | self`.
219 fn union_into(&self, other: &mut BitSet<T>) -> bool;
222 /// This is implemented by all the bitsets so that BitSet::subtract() can be
223 /// passed any type of bitset.
224 pub trait SubtractFromBitSet<T: Idx> {
225 // Performs `other = other - self`.
226 fn subtract_from(&self, other: &mut BitSet<T>) -> bool;
229 impl<T: Idx> UnionIntoBitSet<T> for BitSet<T> {
230 fn union_into(&self, other: &mut BitSet<T>) -> bool {
231 assert_eq!(self.domain_size, other.domain_size);
232 bitwise(&mut other.words, &self.words, |a, b| a | b)
236 impl<T: Idx> SubtractFromBitSet<T> for BitSet<T> {
237 fn subtract_from(&self, other: &mut BitSet<T>) -> bool {
238 assert_eq!(self.domain_size, other.domain_size);
239 bitwise(&mut other.words, &self.words, |a, b| a & !b)
243 impl<T: Idx> fmt::Debug for BitSet<T> {
244 fn fmt(&self, w: &mut fmt::Formatter<'_>) -> fmt::Result {
245 w.debug_list().entries(self.iter()).finish()
249 impl<T: Idx> ToString for BitSet<T> {
250 fn to_string(&self) -> String {
251 let mut result = String::new();
254 // Note: this is a little endian printout of bytes.
256 // i tracks how many bits we have printed so far.
258 for word in &self.words {
259 let mut word = *word;
260 for _ in 0..WORD_BYTES {
261 // for each byte in `word`:
262 let remain = self.domain_size - i;
263 // If less than a byte remains, then mask just that many bits.
264 let mask = if remain <= 8 { (1 << remain) - 1 } else { 0xFF };
265 assert!(mask <= 0xFF);
266 let byte = word & mask;
268 result.push_str(&format!("{}{:02x}", sep, byte));
285 pub struct BitIter<'a, T: Idx> {
286 /// A copy of the current word, but with any already-visited bits cleared.
287 /// (This lets us use `trailing_zeros()` to find the next set bit.) When it
288 /// is reduced to 0, we move onto the next word.
291 /// The offset (measured in bits) of the current word.
294 /// Underlying iterator over the words.
295 iter: slice::Iter<'a, Word>,
297 marker: PhantomData<T>,
300 impl<'a, T: Idx> BitIter<'a, T> {
302 fn new(words: &'a [Word]) -> BitIter<'a, T> {
303 // We initialize `word` and `offset` to degenerate values. On the first
304 // call to `next()` we will fall through to getting the first word from
305 // `iter`, which sets `word` to the first word (if there is one) and
306 // `offset` to 0. Doing it this way saves us from having to maintain
307 // additional state about whether we have started.
310 offset: usize::MAX - (WORD_BITS - 1),
317 impl<'a, T: Idx> Iterator for BitIter<'a, T> {
319 fn next(&mut self) -> Option<T> {
322 // Get the position of the next set bit in the current word,
323 // then clear the bit.
324 let bit_pos = self.word.trailing_zeros() as usize;
325 let bit = 1 << bit_pos;
327 return Some(T::new(bit_pos + self.offset));
330 // Move onto the next word. `wrapping_add()` is needed to handle
331 // the degenerate initial value given to `offset` in `new()`.
332 let word = self.iter.next()?;
334 self.offset = self.offset.wrapping_add(WORD_BITS);
340 fn bitwise<Op>(out_vec: &mut [Word], in_vec: &[Word], op: Op) -> bool
342 Op: Fn(Word, Word) -> Word,
344 assert_eq!(out_vec.len(), in_vec.len());
345 let mut changed = false;
346 for (out_elem, in_elem) in out_vec.iter_mut().zip(in_vec.iter()) {
347 let old_val = *out_elem;
348 let new_val = op(old_val, *in_elem);
350 changed |= old_val != new_val;
355 const SPARSE_MAX: usize = 8;
357 /// A fixed-size bitset type with a sparse representation and a maximum of
358 /// `SPARSE_MAX` elements. The elements are stored as a sorted `ArrayVec` with
361 /// This type is used by `HybridBitSet`; do not use directly.
362 #[derive(Clone, Debug)]
363 pub struct SparseBitSet<T: Idx> {
365 elems: ArrayVec<[T; SPARSE_MAX]>,
368 impl<T: Idx> SparseBitSet<T> {
369 fn new_empty(domain_size: usize) -> Self {
370 SparseBitSet { domain_size, elems: ArrayVec::new() }
373 fn len(&self) -> usize {
377 fn is_empty(&self) -> bool {
378 self.elems.len() == 0
381 fn contains(&self, elem: T) -> bool {
382 assert!(elem.index() < self.domain_size);
383 self.elems.contains(&elem)
386 fn insert(&mut self, elem: T) -> bool {
387 assert!(elem.index() < self.domain_size);
388 let changed = if let Some(i) = self.elems.iter().position(|&e| e >= elem) {
389 if self.elems[i] == elem {
390 // `elem` is already in the set.
393 // `elem` is smaller than one or more existing elements.
394 self.elems.insert(i, elem);
398 // `elem` is larger than all existing elements.
399 self.elems.push(elem);
402 assert!(self.len() <= SPARSE_MAX);
406 fn remove(&mut self, elem: T) -> bool {
407 assert!(elem.index() < self.domain_size);
408 if let Some(i) = self.elems.iter().position(|&e| e == elem) {
409 self.elems.remove(i);
416 fn to_dense(&self) -> BitSet<T> {
417 let mut dense = BitSet::new_empty(self.domain_size);
418 for elem in self.elems.iter() {
424 fn iter(&self) -> slice::Iter<'_, T> {
429 impl<T: Idx> UnionIntoBitSet<T> for SparseBitSet<T> {
430 fn union_into(&self, other: &mut BitSet<T>) -> bool {
431 assert_eq!(self.domain_size, other.domain_size);
432 let mut changed = false;
433 for elem in self.iter() {
434 changed |= other.insert(*elem);
440 impl<T: Idx> SubtractFromBitSet<T> for SparseBitSet<T> {
441 fn subtract_from(&self, other: &mut BitSet<T>) -> bool {
442 assert_eq!(self.domain_size, other.domain_size);
443 let mut changed = false;
444 for elem in self.iter() {
445 changed |= other.remove(*elem);
451 /// A fixed-size bitset type with a hybrid representation: sparse when there
452 /// are up to a `SPARSE_MAX` elements in the set, but dense when there are more
453 /// than `SPARSE_MAX`.
455 /// This type is especially efficient for sets that typically have a small
456 /// number of elements, but a large `domain_size`, and are cleared frequently.
458 /// `T` is an index type, typically a newtyped `usize` wrapper, but it can also
461 /// All operations that involve an element will panic if the element is equal
462 /// to or greater than the domain size. All operations that involve two bitsets
463 /// will panic if the bitsets have differing domain sizes.
464 #[derive(Clone, Debug)]
465 pub enum HybridBitSet<T: Idx> {
466 Sparse(SparseBitSet<T>),
470 impl<T: Idx> HybridBitSet<T> {
471 pub fn new_empty(domain_size: usize) -> Self {
472 HybridBitSet::Sparse(SparseBitSet::new_empty(domain_size))
475 fn domain_size(&self) -> usize {
477 HybridBitSet::Sparse(sparse) => sparse.domain_size,
478 HybridBitSet::Dense(dense) => dense.domain_size,
482 pub fn clear(&mut self) {
483 let domain_size = self.domain_size();
484 *self = HybridBitSet::new_empty(domain_size);
487 pub fn contains(&self, elem: T) -> bool {
489 HybridBitSet::Sparse(sparse) => sparse.contains(elem),
490 HybridBitSet::Dense(dense) => dense.contains(elem),
494 pub fn superset(&self, other: &HybridBitSet<T>) -> bool {
495 match (self, other) {
496 (HybridBitSet::Dense(self_dense), HybridBitSet::Dense(other_dense)) => {
497 self_dense.superset(other_dense)
500 assert!(self.domain_size() == other.domain_size());
501 other.iter().all(|elem| self.contains(elem))
506 pub fn is_empty(&self) -> bool {
508 HybridBitSet::Sparse(sparse) => sparse.is_empty(),
509 HybridBitSet::Dense(dense) => dense.is_empty(),
513 pub fn insert(&mut self, elem: T) -> bool {
514 // No need to check `elem` against `self.domain_size` here because all
515 // the match cases check it, one way or another.
517 HybridBitSet::Sparse(sparse) if sparse.len() < SPARSE_MAX => {
518 // The set is sparse and has space for `elem`.
521 HybridBitSet::Sparse(sparse) if sparse.contains(elem) => {
522 // The set is sparse and does not have space for `elem`, but
523 // that doesn't matter because `elem` is already present.
526 HybridBitSet::Sparse(sparse) => {
527 // The set is sparse and full. Convert to a dense set.
528 let mut dense = sparse.to_dense();
529 let changed = dense.insert(elem);
531 *self = HybridBitSet::Dense(dense);
534 HybridBitSet::Dense(dense) => dense.insert(elem),
538 pub fn insert_all(&mut self) {
539 let domain_size = self.domain_size();
541 HybridBitSet::Sparse(_) => {
542 *self = HybridBitSet::Dense(BitSet::new_filled(domain_size));
544 HybridBitSet::Dense(dense) => dense.insert_all(),
548 pub fn remove(&mut self, elem: T) -> bool {
549 // Note: we currently don't bother going from Dense back to Sparse.
551 HybridBitSet::Sparse(sparse) => sparse.remove(elem),
552 HybridBitSet::Dense(dense) => dense.remove(elem),
556 pub fn union(&mut self, other: &HybridBitSet<T>) -> bool {
558 HybridBitSet::Sparse(self_sparse) => {
560 HybridBitSet::Sparse(other_sparse) => {
561 // Both sets are sparse. Add the elements in
562 // `other_sparse` to `self` one at a time. This
563 // may or may not cause `self` to be densified.
564 assert_eq!(self.domain_size(), other.domain_size());
565 let mut changed = false;
566 for elem in other_sparse.iter() {
567 changed |= self.insert(*elem);
571 HybridBitSet::Dense(other_dense) => {
572 // `self` is sparse and `other` is dense. To
573 // merge them, we have two available strategies:
574 // * Densify `self` then merge other
575 // * Clone other then integrate bits from `self`
576 // The second strategy requires dedicated method
577 // since the usual `union` returns the wrong
578 // result. In the dedicated case the computation
579 // is slightly faster if the bits of the sparse
580 // bitset map to only few words of the dense
581 // representation, i.e. indices are near each
584 // Benchmarking seems to suggest that the second
585 // option is worth it.
586 let mut new_dense = other_dense.clone();
587 let changed = new_dense.reverse_union_sparse(self_sparse);
588 *self = HybridBitSet::Dense(new_dense);
594 HybridBitSet::Dense(self_dense) => self_dense.union(other),
598 /// Converts to a dense set, consuming itself in the process.
599 pub fn to_dense(self) -> BitSet<T> {
601 HybridBitSet::Sparse(sparse) => sparse.to_dense(),
602 HybridBitSet::Dense(dense) => dense,
606 pub fn iter(&self) -> HybridIter<'_, T> {
608 HybridBitSet::Sparse(sparse) => HybridIter::Sparse(sparse.iter()),
609 HybridBitSet::Dense(dense) => HybridIter::Dense(dense.iter()),
614 impl<T: Idx> UnionIntoBitSet<T> for HybridBitSet<T> {
615 fn union_into(&self, other: &mut BitSet<T>) -> bool {
617 HybridBitSet::Sparse(sparse) => sparse.union_into(other),
618 HybridBitSet::Dense(dense) => dense.union_into(other),
623 impl<T: Idx> SubtractFromBitSet<T> for HybridBitSet<T> {
624 fn subtract_from(&self, other: &mut BitSet<T>) -> bool {
626 HybridBitSet::Sparse(sparse) => sparse.subtract_from(other),
627 HybridBitSet::Dense(dense) => dense.subtract_from(other),
632 pub enum HybridIter<'a, T: Idx> {
633 Sparse(slice::Iter<'a, T>),
634 Dense(BitIter<'a, T>),
637 impl<'a, T: Idx> Iterator for HybridIter<'a, T> {
640 fn next(&mut self) -> Option<T> {
642 HybridIter::Sparse(sparse) => sparse.next().copied(),
643 HybridIter::Dense(dense) => dense.next(),
648 /// A resizable bitset type with a dense representation.
650 /// `T` is an index type, typically a newtyped `usize` wrapper, but it can also
653 /// All operations that involve an element will panic if the element is equal
654 /// to or greater than the domain size.
655 #[derive(Clone, Debug, PartialEq)]
656 pub struct GrowableBitSet<T: Idx> {
660 impl<T: Idx> GrowableBitSet<T> {
661 /// Ensure that the set can hold at least `min_domain_size` elements.
662 pub fn ensure(&mut self, min_domain_size: usize) {
663 if self.bit_set.domain_size < min_domain_size {
664 self.bit_set.domain_size = min_domain_size;
667 let min_num_words = num_words(min_domain_size);
668 if self.bit_set.words.len() < min_num_words {
669 self.bit_set.words.resize(min_num_words, 0)
673 pub fn new_empty() -> GrowableBitSet<T> {
674 GrowableBitSet { bit_set: BitSet::new_empty(0) }
677 pub fn with_capacity(capacity: usize) -> GrowableBitSet<T> {
678 GrowableBitSet { bit_set: BitSet::new_empty(capacity) }
681 /// Returns `true` if the set has changed.
683 pub fn insert(&mut self, elem: T) -> bool {
684 self.ensure(elem.index() + 1);
685 self.bit_set.insert(elem)
689 pub fn contains(&self, elem: T) -> bool {
690 let (word_index, mask) = word_index_and_mask(elem);
691 if let Some(word) = self.bit_set.words.get(word_index) { (word & mask) != 0 } else { false }
695 /// A fixed-size 2D bit matrix type with a dense representation.
697 /// `R` and `C` are index types used to identify rows and columns respectively;
698 /// typically newtyped `usize` wrappers, but they can also just be `usize`.
700 /// All operations that involve a row and/or column index will panic if the
701 /// index exceeds the relevant bound.
702 #[derive(Clone, Eq, PartialEq, RustcDecodable, RustcEncodable)]
703 pub struct BitMatrix<R: Idx, C: Idx> {
707 marker: PhantomData<(R, C)>,
710 impl<R: Idx, C: Idx> BitMatrix<R, C> {
711 /// Creates a new `rows x columns` matrix, initially empty.
712 pub fn new(num_rows: usize, num_columns: usize) -> BitMatrix<R, C> {
713 // For every element, we need one bit for every other
714 // element. Round up to an even number of words.
715 let words_per_row = num_words(num_columns);
719 words: vec![0; num_rows * words_per_row],
724 /// Creates a new matrix, with `row` used as the value for every row.
725 pub fn from_row_n(row: &BitSet<C>, num_rows: usize) -> BitMatrix<R, C> {
726 let num_columns = row.domain_size();
727 let words_per_row = num_words(num_columns);
728 assert_eq!(words_per_row, row.words().len());
732 words: iter::repeat(row.words()).take(num_rows).flatten().cloned().collect(),
737 pub fn rows(&self) -> impl Iterator<Item = R> {
738 (0..self.num_rows).map(R::new)
741 /// The range of bits for a given row.
742 fn range(&self, row: R) -> (usize, usize) {
743 let words_per_row = num_words(self.num_columns);
744 let start = row.index() * words_per_row;
745 (start, start + words_per_row)
748 /// Sets the cell at `(row, column)` to true. Put another way, insert
749 /// `column` to the bitset for `row`.
751 /// Returns `true` if this changed the matrix.
752 pub fn insert(&mut self, row: R, column: C) -> bool {
753 assert!(row.index() < self.num_rows && column.index() < self.num_columns);
754 let (start, _) = self.range(row);
755 let (word_index, mask) = word_index_and_mask(column);
756 let words = &mut self.words[..];
757 let word = words[start + word_index];
758 let new_word = word | mask;
759 words[start + word_index] = new_word;
763 /// Do the bits from `row` contain `column`? Put another way, is
764 /// the matrix cell at `(row, column)` true? Put yet another way,
765 /// if the matrix represents (transitive) reachability, can
766 /// `row` reach `column`?
767 pub fn contains(&self, row: R, column: C) -> bool {
768 assert!(row.index() < self.num_rows && column.index() < self.num_columns);
769 let (start, _) = self.range(row);
770 let (word_index, mask) = word_index_and_mask(column);
771 (self.words[start + word_index] & mask) != 0
774 /// Returns those indices that are true in rows `a` and `b`. This
775 /// is an O(n) operation where `n` is the number of elements
776 /// (somewhat independent from the actual size of the
777 /// intersection, in particular).
778 pub fn intersect_rows(&self, row1: R, row2: R) -> Vec<C> {
779 assert!(row1.index() < self.num_rows && row2.index() < self.num_rows);
780 let (row1_start, row1_end) = self.range(row1);
781 let (row2_start, row2_end) = self.range(row2);
782 let mut result = Vec::with_capacity(self.num_columns);
783 for (base, (i, j)) in (row1_start..row1_end).zip(row2_start..row2_end).enumerate() {
784 let mut v = self.words[i] & self.words[j];
785 for bit in 0..WORD_BITS {
790 result.push(C::new(base * WORD_BITS + bit));
798 /// Adds the bits from row `read` to the bits from row `write`, and
799 /// returns `true` if anything changed.
801 /// This is used when computing transitive reachability because if
802 /// you have an edge `write -> read`, because in that case
803 /// `write` can reach everything that `read` can (and
804 /// potentially more).
805 pub fn union_rows(&mut self, read: R, write: R) -> bool {
806 assert!(read.index() < self.num_rows && write.index() < self.num_rows);
807 let (read_start, read_end) = self.range(read);
808 let (write_start, write_end) = self.range(write);
809 let words = &mut self.words[..];
810 let mut changed = false;
811 for (read_index, write_index) in (read_start..read_end).zip(write_start..write_end) {
812 let word = words[write_index];
813 let new_word = word | words[read_index];
814 words[write_index] = new_word;
815 changed |= word != new_word;
820 /// Adds the bits from `with` to the bits from row `write`, and
821 /// returns `true` if anything changed.
822 pub fn union_row_with(&mut self, with: &BitSet<C>, write: R) -> bool {
823 assert!(write.index() < self.num_rows);
824 assert_eq!(with.domain_size(), self.num_columns);
825 let (write_start, write_end) = self.range(write);
826 let mut changed = false;
827 for (read_index, write_index) in (0..with.words().len()).zip(write_start..write_end) {
828 let word = self.words[write_index];
829 let new_word = word | with.words()[read_index];
830 self.words[write_index] = new_word;
831 changed |= word != new_word;
836 /// Sets every cell in `row` to true.
837 pub fn insert_all_into_row(&mut self, row: R) {
838 assert!(row.index() < self.num_rows);
839 let (start, end) = self.range(row);
840 let words = &mut self.words[..];
841 for index in start..end {
844 self.clear_excess_bits(row);
847 /// Clear excess bits in the final word of the row.
848 fn clear_excess_bits(&mut self, row: R) {
849 let num_bits_in_final_word = self.num_columns % WORD_BITS;
850 if num_bits_in_final_word > 0 {
851 let mask = (1 << num_bits_in_final_word) - 1;
852 let (_, end) = self.range(row);
853 let final_word_idx = end - 1;
854 self.words[final_word_idx] &= mask;
858 /// Gets a slice of the underlying words.
859 pub fn words(&self) -> &[Word] {
863 /// Iterates through all the columns set to true in a given row of
865 pub fn iter(&self, row: R) -> BitIter<'_, C> {
866 assert!(row.index() < self.num_rows);
867 let (start, end) = self.range(row);
868 BitIter::new(&self.words[start..end])
871 /// Returns the number of elements in `row`.
872 pub fn count(&self, row: R) -> usize {
873 let (start, end) = self.range(row);
874 self.words[start..end].iter().map(|e| e.count_ones() as usize).sum()
878 impl<R: Idx, C: Idx> fmt::Debug for BitMatrix<R, C> {
879 fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
880 /// Forces its contents to print in regular mode instead of alternate mode.
881 struct OneLinePrinter<T>(T);
882 impl<T: fmt::Debug> fmt::Debug for OneLinePrinter<T> {
883 fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
884 write!(fmt, "{:?}", self.0)
888 write!(fmt, "BitMatrix({}x{}) ", self.num_rows, self.num_columns)?;
889 let items = self.rows().flat_map(|r| self.iter(r).map(move |c| (r, c)));
890 fmt.debug_set().entries(items.map(OneLinePrinter)).finish()
894 /// A fixed-column-size, variable-row-size 2D bit matrix with a moderately
895 /// sparse representation.
897 /// Initially, every row has no explicit representation. If any bit within a
898 /// row is set, the entire row is instantiated as `Some(<HybridBitSet>)`.
899 /// Furthermore, any previously uninstantiated rows prior to it will be
900 /// instantiated as `None`. Those prior rows may themselves become fully
901 /// instantiated later on if any of their bits are set.
903 /// `R` and `C` are index types used to identify rows and columns respectively;
904 /// typically newtyped `usize` wrappers, but they can also just be `usize`.
905 #[derive(Clone, Debug)]
906 pub struct SparseBitMatrix<R, C>
912 rows: IndexVec<R, Option<HybridBitSet<C>>>,
915 impl<R: Idx, C: Idx> SparseBitMatrix<R, C> {
916 /// Creates a new empty sparse bit matrix with no rows or columns.
917 pub fn new(num_columns: usize) -> Self {
918 Self { num_columns, rows: IndexVec::new() }
921 fn ensure_row(&mut self, row: R) -> &mut HybridBitSet<C> {
922 // Instantiate any missing rows up to and including row `row` with an
923 // empty HybridBitSet.
924 self.rows.ensure_contains_elem(row, || None);
926 // Then replace row `row` with a full HybridBitSet if necessary.
927 let num_columns = self.num_columns;
928 self.rows[row].get_or_insert_with(|| HybridBitSet::new_empty(num_columns))
931 /// Sets the cell at `(row, column)` to true. Put another way, insert
932 /// `column` to the bitset for `row`.
934 /// Returns `true` if this changed the matrix.
935 pub fn insert(&mut self, row: R, column: C) -> bool {
936 self.ensure_row(row).insert(column)
939 /// Do the bits from `row` contain `column`? Put another way, is
940 /// the matrix cell at `(row, column)` true? Put yet another way,
941 /// if the matrix represents (transitive) reachability, can
942 /// `row` reach `column`?
943 pub fn contains(&self, row: R, column: C) -> bool {
944 self.row(row).map_or(false, |r| r.contains(column))
947 /// Adds the bits from row `read` to the bits from row `write`, and
948 /// returns `true` if anything changed.
950 /// This is used when computing transitive reachability because if
951 /// you have an edge `write -> read`, because in that case
952 /// `write` can reach everything that `read` can (and
953 /// potentially more).
954 pub fn union_rows(&mut self, read: R, write: R) -> bool {
955 if read == write || self.row(read).is_none() {
959 self.ensure_row(write);
960 if let (Some(read_row), Some(write_row)) = self.rows.pick2_mut(read, write) {
961 write_row.union(read_row)
967 /// Union a row, `from`, into the `into` row.
968 pub fn union_into_row(&mut self, into: R, from: &HybridBitSet<C>) -> bool {
969 self.ensure_row(into).union(from)
972 /// Insert all bits in the given row.
973 pub fn insert_all_into_row(&mut self, row: R) {
974 self.ensure_row(row).insert_all();
977 pub fn rows(&self) -> impl Iterator<Item = R> {
981 /// Iterates through all the columns set to true in a given row of
983 pub fn iter<'a>(&'a self, row: R) -> impl Iterator<Item = C> + 'a {
984 self.row(row).into_iter().flat_map(|r| r.iter())
987 pub fn row(&self, row: R) -> Option<&HybridBitSet<C>> {
988 if let Some(Some(row)) = self.rows.get(row) { Some(row) } else { None }
993 fn num_words<T: Idx>(domain_size: T) -> usize {
994 (domain_size.index() + WORD_BITS - 1) / WORD_BITS
998 fn word_index_and_mask<T: Idx>(elem: T) -> (usize, Word) {
999 let elem = elem.index();
1000 let word_index = elem / WORD_BITS;
1001 let mask = 1 << (elem % WORD_BITS);