1 //===- llvm/ADT/SparseBitVector.h - Efficient Sparse BitVector -*- C++ -*- ===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Daniel Berlin and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the SparseBitVector class. See the doxygen comment for
11 // SparseBitVector for more details on the algorithm used.
13 //===----------------------------------------------------------------------===//
15 #ifndef LLVM_ADT_SPARSEBITVECTOR_H
16 #define LLVM_ADT_SPARSEBITVECTOR_H
21 #include "llvm/Support/DataTypes.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/Support/MathExtras.h"
24 #include "llvm/ADT/ilist"
27 /// SparseBitVector is an implementation of a bitvector that is sparse by only
28 /// storing the elements that have non-zero bits set. In order to make this
29 /// fast for the most common cases, SparseBitVector is implemented as a linked
30 /// list of SparseBitVectorElements. We maintain a pointer to the last
31 /// SparseBitVectorElement accessed (in the form of a list iterator), in order
32 /// to make multiple in-order test/set constant time after the first one is
33 /// executed. Note that using vectors to store SparseBitVectorElement's does
34 /// not work out very well because it causes insertion in the middle to take
35 /// enormous amounts of time with a large amount of bits. Other structures that
36 /// have better worst cases for insertion in the middle (various balanced trees,
37 /// etc) do not perform as well in practice as a linked list with this iterator
38 /// kept up to date. They are also significantly more memory intensive.
41 template <unsigned ElementSize = 128>
42 struct SparseBitVectorElement {
44 typedef unsigned long BitWord;
46 BITWORD_SIZE = sizeof(BitWord) * 8,
47 BITWORDS_PER_ELEMENT = (ElementSize + BITWORD_SIZE - 1) / BITWORD_SIZE,
48 BITS_PER_ELEMENT = ElementSize
51 SparseBitVectorElement<ElementSize> *getNext() const {
54 SparseBitVectorElement<ElementSize> *getPrev() const {
58 void setNext(SparseBitVectorElement<ElementSize> *RHS) {
61 void setPrev(SparseBitVectorElement<ElementSize> *RHS) {
66 SparseBitVectorElement<ElementSize> *Next;
67 SparseBitVectorElement<ElementSize> *Prev;
68 // Index of Element in terms of where first bit starts.
69 unsigned ElementIndex;
70 BitWord Bits[BITWORDS_PER_ELEMENT];
71 // Needed for sentinels
72 SparseBitVectorElement() {
74 memset(&Bits[0], 0, sizeof (BitWord) * BITWORDS_PER_ELEMENT);
77 friend struct ilist_traits<SparseBitVectorElement<ElementSize> >;
79 explicit SparseBitVectorElement(unsigned Idx) {
81 memset(&Bits[0], 0, sizeof (BitWord) * BITWORDS_PER_ELEMENT);
84 ~SparseBitVectorElement() {
88 SparseBitVectorElement(const SparseBitVectorElement &RHS) {
89 ElementIndex = RHS.ElementIndex;
90 std::copy(&RHS.Bits[0], &RHS.Bits[BITWORDS_PER_ELEMENT], Bits);
94 bool operator==(const SparseBitVectorElement &RHS) const {
95 if (ElementIndex != RHS.ElementIndex)
97 for (unsigned i = 0; i < BITWORDS_PER_ELEMENT; ++i)
98 if (Bits[i] != RHS.Bits[i])
103 bool operator!=(const SparseBitVectorElement &RHS) const {
104 return !(*this == RHS);
107 // Return the bits that make up word Idx in our element.
108 BitWord word(unsigned Idx) const {
109 assert (Idx < BITWORDS_PER_ELEMENT);
113 unsigned index() const {
118 for (unsigned i = 0; i < BITWORDS_PER_ELEMENT; ++i)
124 void set(unsigned Idx) {
125 Bits[Idx / BITWORD_SIZE] |= 1L << (Idx % BITWORD_SIZE);
128 bool test_and_set (unsigned Idx) {
129 bool old = test(Idx);
137 void reset(unsigned Idx) {
138 Bits[Idx / BITWORD_SIZE] &= ~(1L << (Idx % BITWORD_SIZE));
141 bool test(unsigned Idx) const {
142 return Bits[Idx / BITWORD_SIZE] & (1L << (Idx % BITWORD_SIZE));
145 unsigned count() const {
146 unsigned NumBits = 0;
147 for (unsigned i = 0; i < BITWORDS_PER_ELEMENT; ++i)
148 if (sizeof(BitWord) == 4)
149 NumBits += CountPopulation_32(Bits[i]);
150 else if (sizeof(BitWord) == 8)
151 NumBits += CountPopulation_64(Bits[i]);
153 assert(0 && "Unsupported!");
157 /// find_first - Returns the index of the first set bit.
158 int find_first() const {
159 for (unsigned i = 0; i < BITWORDS_PER_ELEMENT; ++i)
161 if (sizeof(BitWord) == 4)
162 return i * BITWORD_SIZE + CountTrailingZeros_32(Bits[i]);
163 else if (sizeof(BitWord) == 8)
164 return i * BITWORD_SIZE + CountTrailingZeros_64(Bits[i]);
166 assert(0 && "Unsupported!");
168 assert(0 && "Illegal empty element");
171 /// find_next - Returns the index of the next set bit starting from the
172 /// "Curr" bit. Returns -1 if the next set bit is not found.
173 int find_next(unsigned Curr) const {
174 if (Curr >= BITS_PER_ELEMENT)
177 unsigned WordPos = Curr / BITWORD_SIZE;
178 unsigned BitPos = Curr % BITWORD_SIZE;
179 BitWord Copy = Bits[WordPos];
180 assert (WordPos <= BITWORDS_PER_ELEMENT
181 && "Word Position outside of element");
183 // Mask off previous bits.
184 Copy &= ~0L << BitPos;
187 if (sizeof(BitWord) == 4)
188 return WordPos * BITWORD_SIZE + CountTrailingZeros_32(Copy);
189 else if (sizeof(BitWord) == 8)
190 return WordPos * BITWORD_SIZE + CountTrailingZeros_64(Copy);
192 assert(0 && "Unsupported!");
195 // Check subsequent words.
196 for (unsigned i = WordPos+1; i < BITWORDS_PER_ELEMENT; ++i)
198 if (sizeof(BitWord) == 4)
199 return i * BITWORD_SIZE + CountTrailingZeros_32(Bits[i]);
200 else if (sizeof(BitWord) == 8)
201 return i * BITWORD_SIZE + CountTrailingZeros_64(Bits[i]);
203 assert(0 && "Unsupported!");
208 // Union this element with RHS and return true if this one changed.
209 bool unionWith(const SparseBitVectorElement &RHS) {
210 bool changed = false;
211 for (unsigned i = 0; i < BITWORDS_PER_ELEMENT; ++i) {
212 BitWord old = changed ? 0 : Bits[i];
214 Bits[i] |= RHS.Bits[i];
215 if (!changed && old != Bits[i])
221 // Return true if we have any bits in common with RHS
222 bool intersects(const SparseBitVectorElement &RHS) const {
223 for (unsigned i = 0; i < BITWORDS_PER_ELEMENT; ++i) {
224 if (RHS.Bits[i] & Bits[i])
230 // Intersect this Element with RHS and return true if this one changed.
231 // BecameZero is set to true if this element became all-zero bits.
232 bool intersectWith(const SparseBitVectorElement &RHS,
234 bool changed = false;
238 for (unsigned i = 0; i < BITWORDS_PER_ELEMENT; ++i) {
239 BitWord old = changed ? 0 : Bits[i];
241 Bits[i] &= RHS.Bits[i];
245 if (!changed && old != Bits[i])
248 BecameZero = allzero;
251 // Intersect this Element with the complement of RHS and return true if this
252 // one changed. BecameZero is set to true if this element became all-zero
254 bool intersectWithComplement(const SparseBitVectorElement &RHS,
256 bool changed = false;
260 for (unsigned i = 0; i < BITWORDS_PER_ELEMENT; ++i) {
261 BitWord old = changed ? 0 : Bits[i];
263 Bits[i] &= ~RHS.Bits[i];
267 if (!changed && old != Bits[i])
270 BecameZero = allzero;
273 // Three argument version of intersectWithComplement that intersects
274 // RHS1 & ~RHS2 into this element
275 void intersectWithComplement(const SparseBitVectorElement &RHS1,
276 const SparseBitVectorElement &RHS2,
281 for (unsigned i = 0; i < BITWORDS_PER_ELEMENT; ++i) {
282 Bits[i] = RHS1.Bits[i] & ~RHS2.Bits[i];
286 BecameZero = allzero;
289 // Get a hash value for this element;
290 uint64_t getHashValue() const {
291 uint64_t HashVal = 0;
292 for (unsigned i = 0; i < BITWORDS_PER_ELEMENT; ++i) {
299 template <unsigned ElementSize = 128>
300 class SparseBitVector {
301 typedef ilist<SparseBitVectorElement<ElementSize> > ElementList;
302 typedef typename ElementList::iterator ElementListIter;
303 typedef typename ElementList::const_iterator ElementListConstIter;
305 BITWORD_SIZE = SparseBitVectorElement<ElementSize>::BITWORD_SIZE
308 // Pointer to our current Element.
309 ElementListIter CurrElementIter;
310 ElementList Elements;
312 // This is like std::lower_bound, except we do linear searching from the
314 ElementListIter FindLowerBound(unsigned ElementIndex) {
316 if (Elements.empty()) {
317 CurrElementIter = Elements.begin();
318 return Elements.begin();
321 // Make sure our current iterator is valid.
322 if (CurrElementIter == Elements.end())
325 // Search from our current iterator, either backwards or forwards,
326 // depending on what element we are looking for.
327 ElementListIter ElementIter = CurrElementIter;
328 if (CurrElementIter->index() == ElementIndex) {
330 } else if (CurrElementIter->index() > ElementIndex) {
331 while (ElementIter != Elements.begin()
332 && ElementIter->index() > ElementIndex)
335 while (ElementIter != Elements.end() &&
336 ElementIter->index() < ElementIndex)
339 CurrElementIter = ElementIter;
343 // Iterator to walk set bits in the bitmap. This iterator is a lot uglier
344 // than it would be, in order to be efficient.
345 class SparseBitVectorIterator {
349 const SparseBitVector<ElementSize> *BitVector;
351 // Current element inside of bitmap.
352 ElementListConstIter Iter;
354 // Current bit number inside of our bitmap.
357 // Current word number inside of our element.
360 // Current bits from the element.
361 typename SparseBitVectorElement<ElementSize>::BitWord Bits;
363 // Move our iterator to the first non-zero bit in the bitmap.
364 void AdvanceToFirstNonZero() {
367 if (BitVector->Elements.empty()) {
371 Iter = BitVector->Elements.begin();
372 BitNumber = Iter->index() * ElementSize;
373 unsigned BitPos = Iter->find_first();
375 WordNumber = (BitNumber % ElementSize) / BITWORD_SIZE;
376 Bits = Iter->word(WordNumber);
377 Bits >>= BitPos % BITWORD_SIZE;
380 // Move our iterator to the next non-zero bit.
381 void AdvanceToNextNonZero() {
385 while (Bits && !(Bits & 1)) {
390 // See if we ran out of Bits in this word.
392 int NextSetBitNumber = Iter->find_next(BitNumber % ElementSize) ;
393 // If we ran out of set bits in this element, move to next element.
394 if (NextSetBitNumber == -1 || (BitNumber % ElementSize == 0)) {
398 // We may run out of elements in the bitmap.
399 if (Iter == BitVector->Elements.end()) {
403 // Set up for next non zero word in bitmap.
404 BitNumber = Iter->index() * ElementSize;
405 NextSetBitNumber = Iter->find_first();
406 BitNumber += NextSetBitNumber;
407 WordNumber = (BitNumber % ElementSize) / BITWORD_SIZE;
408 Bits = Iter->word(WordNumber);
409 Bits >>= NextSetBitNumber % BITWORD_SIZE;
411 WordNumber = (NextSetBitNumber % ElementSize) / BITWORD_SIZE;
412 Bits = Iter->word(WordNumber);
413 Bits >>= NextSetBitNumber % BITWORD_SIZE;
414 BitNumber = Iter->index() * ElementSize;
415 BitNumber += NextSetBitNumber;
421 inline SparseBitVectorIterator& operator++() {
424 AdvanceToNextNonZero();
429 inline SparseBitVectorIterator operator++(int) {
430 SparseBitVectorIterator tmp = *this;
435 // Return the current set bit number.
436 unsigned operator*() const {
440 bool operator==(const SparseBitVectorIterator &RHS) const {
441 // If they are both at the end, ignore the rest of the fields.
442 if (AtEnd && RHS.AtEnd)
444 // Otherwise they are the same if they have the same bit number and
446 return AtEnd == RHS.AtEnd && RHS.BitNumber == BitNumber;
448 bool operator!=(const SparseBitVectorIterator &RHS) const {
449 return !(*this == RHS);
451 SparseBitVectorIterator(): BitVector(NULL) {
455 SparseBitVectorIterator(const SparseBitVector<ElementSize> *RHS,
456 bool end = false):BitVector(RHS) {
457 Iter = BitVector->Elements.begin();
462 AdvanceToFirstNonZero();
466 typedef SparseBitVectorIterator iterator;
469 CurrElementIter = Elements.begin ();
475 // SparseBitVector copy ctor.
476 SparseBitVector(const SparseBitVector &RHS) {
477 ElementListConstIter ElementIter = RHS.Elements.begin();
478 while (ElementIter != RHS.Elements.end()) {
479 Elements.push_back(SparseBitVectorElement<ElementSize>(*ElementIter));
483 CurrElementIter = Elements.begin ();
486 // Test, Reset, and Set a bit in the bitmap.
487 bool test(unsigned Idx) {
488 if (Elements.empty())
491 unsigned ElementIndex = Idx / ElementSize;
492 ElementListIter ElementIter = FindLowerBound(ElementIndex);
494 // If we can't find an element that is supposed to contain this bit, there
495 // is nothing more to do.
496 if (ElementIter == Elements.end() ||
497 ElementIter->index() != ElementIndex)
499 return ElementIter->test(Idx % ElementSize);
502 void reset(unsigned Idx) {
503 if (Elements.empty())
506 unsigned ElementIndex = Idx / ElementSize;
507 ElementListIter ElementIter = FindLowerBound(ElementIndex);
509 // If we can't find an element that is supposed to contain this bit, there
510 // is nothing more to do.
511 if (ElementIter == Elements.end() ||
512 ElementIter->index() != ElementIndex)
514 ElementIter->reset(Idx % ElementSize);
516 // When the element is zeroed out, delete it.
517 if (ElementIter->empty()) {
519 Elements.erase(ElementIter);
523 void set(unsigned Idx) {
524 unsigned ElementIndex = Idx / ElementSize;
525 SparseBitVectorElement<ElementSize> *Element;
526 ElementListIter ElementIter;
527 if (Elements.empty()) {
528 Element = new SparseBitVectorElement<ElementSize>(ElementIndex);
529 ElementIter = Elements.insert(Elements.end(), Element);
532 ElementIter = FindLowerBound(ElementIndex);
534 if (ElementIter == Elements.end() ||
535 ElementIter->index() != ElementIndex) {
536 Element = new SparseBitVectorElement<ElementSize>(ElementIndex);
537 // We may have hit the beginning of our SparseBitVector, in which case,
538 // we may need to insert right after this element, which requires moving
539 // the current iterator forward one, because insert does insert before.
540 if (ElementIter != Elements.end() &&
541 ElementIter->index() < ElementIndex)
542 ElementIter = Elements.insert(++ElementIter, Element);
544 ElementIter = Elements.insert(ElementIter, Element);
547 CurrElementIter = ElementIter;
549 ElementIter->set(Idx % ElementSize);
552 bool test_and_set (unsigned Idx) {
553 bool old = test(Idx);
561 bool operator!=(const SparseBitVector &RHS) const {
562 return !(*this == RHS);
565 bool operator==(const SparseBitVector &RHS) const {
566 ElementListConstIter Iter1 = Elements.begin();
567 ElementListConstIter Iter2 = RHS.Elements.begin();
569 for (; Iter1 != Elements.end() && Iter2 != RHS.Elements.end();
571 if (*Iter1 != *Iter2)
574 return Iter1 == Elements.end() && Iter2 == RHS.Elements.end();
577 // Union our bitmap with the RHS and return true if we changed.
578 bool operator|=(const SparseBitVector &RHS) {
579 bool changed = false;
580 ElementListIter Iter1 = Elements.begin();
581 ElementListConstIter Iter2 = RHS.Elements.begin();
583 // If RHS is empty, we are done
584 if (RHS.Elements.empty())
587 while (Iter2 != RHS.Elements.end()) {
588 if (Iter1 == Elements.end() || Iter1->index() > Iter2->index()) {
589 Elements.insert(Iter1,
590 new SparseBitVectorElement<ElementSize>(*Iter2));
593 } else if (Iter1->index() == Iter2->index()) {
594 changed |= Iter1->unionWith(*Iter2);
601 CurrElementIter = Elements.begin();
605 // Intersect our bitmap with the RHS and return true if ours changed.
606 bool operator&=(const SparseBitVector &RHS) {
607 bool changed = false;
608 ElementListIter Iter1 = Elements.begin();
609 ElementListConstIter Iter2 = RHS.Elements.begin();
611 // Check if both bitmaps are empty.
612 if (Elements.empty() && RHS.Elements.empty())
615 // Loop through, intersecting as we go, erasing elements when necessary.
616 while (Iter2 != RHS.Elements.end()) {
617 if (Iter1 == Elements.end())
620 if (Iter1->index() > Iter2->index()) {
622 } else if (Iter1->index() == Iter2->index()) {
624 changed |= Iter1->intersectWith(*Iter2, BecameZero);
626 ElementListIter IterTmp = Iter1;
628 Elements.erase(IterTmp);
634 ElementListIter IterTmp = Iter1;
636 Elements.erase(IterTmp);
639 Elements.erase(Iter1, Elements.end());
640 CurrElementIter = Elements.begin();
644 // Intersect our bitmap with the complement of the RHS and return true if ours
646 bool intersectWithComplement(const SparseBitVector &RHS) {
647 bool changed = false;
648 ElementListIter Iter1 = Elements.begin();
649 ElementListConstIter Iter2 = RHS.Elements.begin();
651 // Check if they are both empty
652 if (Elements.empty() && RHS.Elements.empty())
655 // Loop through, intersecting as we go, erasing elements when necessary.
656 while (Iter2 != RHS.Elements.end()) {
657 if (Iter1 == Elements.end())
660 if (Iter1->index() > Iter2->index()) {
662 } else if (Iter1->index() == Iter2->index()) {
664 changed |= Iter1->intersectWithComplement(*Iter2, BecameZero);
666 ElementListIter IterTmp = Iter1;
668 Elements.erase(IterTmp);
677 CurrElementIter = Elements.begin();
681 bool intersectWithComplement(const SparseBitVector<ElementSize> *RHS) const {
682 return intersectWithComplement(*RHS);
686 // Three argument version of intersectWithComplement. Result of RHS1 & ~RHS2
687 // is stored into this bitmap.
688 void intersectWithComplement(const SparseBitVector<ElementSize> &RHS1,
689 const SparseBitVector<ElementSize> &RHS2)
692 ElementListConstIter Iter1 = RHS1.Elements.begin();
693 ElementListConstIter Iter2 = RHS2.Elements.begin();
695 // Check if they are both empty.
696 if (RHS1.empty() && RHS2.empty())
699 // Loop through, intersecting as we go, erasing elements when necessary.
700 while (Iter2 != RHS2.Elements.end()) {
701 if (Iter1 == RHS1.Elements.end())
704 if (Iter1->index() > Iter2->index()) {
706 } else if (Iter1->index() == Iter2->index()) {
707 bool BecameZero = false;
708 SparseBitVectorElement<ElementSize> *NewElement =
709 new SparseBitVectorElement<ElementSize>(Iter1->index());
710 NewElement->intersectWithComplement(*Iter1, *Iter2, BecameZero);
712 Elements.push_back(NewElement);
723 // copy the remaining elements
724 while (Iter1 != RHS1.Elements.end()) {
725 SparseBitVectorElement<ElementSize> *NewElement =
726 new SparseBitVectorElement<ElementSize>(*Iter1);
727 Elements.push_back(NewElement);
731 CurrElementIter = Elements.begin();
735 void intersectWithComplement(const SparseBitVector<ElementSize> *RHS1,
736 const SparseBitVector<ElementSize> *RHS2) {
737 intersectWithComplement(*RHS1, *RHS2);
740 bool intersects(const SparseBitVector<ElementSize> *RHS) const {
741 return intersects(*RHS);
744 // Return true if we share any bits in common with RHS
745 bool intersects(const SparseBitVector<ElementSize> &RHS) const {
746 ElementListConstIter Iter1 = Elements.begin();
747 ElementListConstIter Iter2 = RHS.Elements.begin();
749 // Check if both bitmaps are empty.
750 if (Elements.empty() && RHS.Elements.empty())
753 // Loop through, intersecting stopping when we hit bits in common.
754 while (Iter2 != RHS.Elements.end()) {
755 if (Iter1 == Elements.end())
758 if (Iter1->index() > Iter2->index()) {
760 } else if (Iter1->index() == Iter2->index()) {
761 if (Iter1->intersects(*Iter2))
772 // Return the first set bit in the bitmap. Return -1 if no bits are set.
773 int find_first() const {
774 if (Elements.empty())
776 const SparseBitVectorElement<ElementSize> &First = *(Elements.begin());
777 return (First.index() * ElementSize) + First.find_first();
780 // Return true if the SparseBitVector is empty
782 return Elements.empty();
785 unsigned count() const {
786 unsigned BitCount = 0;
787 for (ElementListConstIter Iter = Elements.begin();
788 Iter != Elements.end();
790 BitCount += Iter->count();
794 iterator begin() const {
795 return iterator(this);
798 iterator end() const {
799 return iterator(this, ~0);
802 // Get a hash value for this bitmap.
803 uint64_t getHashValue() const {
804 uint64_t HashVal = 0;
805 for (ElementListConstIter Iter = Elements.begin();
806 Iter != Elements.end();
808 HashVal ^= Iter->index();
809 HashVal ^= Iter->getHashValue();
815 // Convenience functions to allow Or and And without dereferencing in the user
818 template <unsigned ElementSize>
819 inline bool operator |=(SparseBitVector<ElementSize> &LHS,
820 const SparseBitVector<ElementSize> *RHS) {
824 template <unsigned ElementSize>
825 inline bool operator |=(SparseBitVector<ElementSize> *LHS,
826 const SparseBitVector<ElementSize> &RHS) {
827 return LHS->operator|=(RHS);
830 template <unsigned ElementSize>
831 inline bool operator &=(SparseBitVector<ElementSize> *LHS,
832 const SparseBitVector<ElementSize> &RHS) {
833 return LHS->operator&=(RHS);
836 template <unsigned ElementSize>
837 inline bool operator &=(SparseBitVector<ElementSize> &LHS,
838 const SparseBitVector<ElementSize> *RHS) {
839 return LHS &= (*RHS);
843 // Dump a SparseBitVector to a stream
844 template <unsigned ElementSize>
845 void dump(const SparseBitVector<ElementSize> &LHS, llvm::OStream &out) {
848 typename SparseBitVector<ElementSize>::iterator bi;
849 for (bi = LHS.begin(); bi != LHS.end(); ++bi) {