diff --git a/src/librustc_data_structures/bit_set.rs b/src/librustc_data_structures/bit_set.rs index 7a11ca070071b..3430d83f23f5e 100644 --- a/src/librustc_data_structures/bit_set.rs +++ b/src/librustc_data_structures/bit_set.rs @@ -5,6 +5,10 @@ use std::iter; use std::marker::PhantomData; use std::mem; use std::slice; +#[cfg(test)] +extern crate test; +#[cfg(test)] +use test::Bencher; pub type Word = u64; pub const WORD_BYTES: usize = mem::size_of::(); @@ -177,6 +181,45 @@ impl BitSet { // Note: we currently don't bother trying to make a Sparse set. HybridBitSet::Dense(self.to_owned()) } + + /// Set `self = self | other`. In contrast to `union` returns `true` if the set contains at + /// least one bit that is not in `other` (i.e. `other` is not a superset of `self`). + /// + /// This is an optimization for union of a hybrid bitset. + fn reverse_union_sparse(&mut self, sparse: &SparseBitSet) -> bool { + assert!(sparse.domain_size == self.domain_size); + self.clear_excess_bits(); + + let mut not_already = false; + // Index of the current word not yet merged. + let mut current_index = 0; + // Mask of bits that came from the sparse set in the current word. + let mut new_bit_mask = 0; + for (word_index, mask) in sparse.iter().map(|x| word_index_and_mask(*x)) { + // Next bit is in a word not inspected yet. + if word_index > current_index { + self.words[current_index] |= new_bit_mask; + // Were there any bits in the old word that did not occur in the sparse set? + not_already |= (self.words[current_index] ^ new_bit_mask) != 0; + // Check all words we skipped for any set bit. + not_already |= self.words[current_index+1..word_index].iter().any(|&x| x != 0); + // Update next word. + current_index = word_index; + // Reset bit mask, no bits have been merged yet. + new_bit_mask = 0; + } + // Add bit and mark it as coming from the sparse set. + // self.words[word_index] |= mask; + new_bit_mask |= mask; + } + self.words[current_index] |= new_bit_mask; + // Any bits in the last inspected word that were not in the sparse set? + not_already |= (self.words[current_index] ^ new_bit_mask) != 0; + // Any bits in the tail? Note `clear_excess_bits` before. + not_already |= self.words[current_index+1..].iter().any(|&x| x != 0); + + not_already + } } /// This is implemented by all the bitsets so that BitSet::union() can be @@ -514,10 +557,22 @@ impl HybridBitSet { changed } HybridBitSet::Dense(other_dense) => { - // `self` is sparse and `other` is dense. Densify - // `self` and then do the bitwise union. - let mut new_dense = self_sparse.to_dense(); - let changed = new_dense.union(other_dense); + // `self` is sparse and `other` is dense. To + // merge them, we have two available strategies: + // * Densify `self` then merge other + // * Clone other then integrate bits from `self` + // The second strategy requires dedicated method + // since the usual `union` returns the wrong + // result. In the dedicated case the computation + // is slightly faster if the bits of the sparse + // bitset map to only few words of the dense + // representation, i.e. indices are near each + // other. + // + // Benchmarking seems to suggest that the second + // option is worth it. + let mut new_dense = other_dense.clone(); + let changed = new_dense.reverse_union_sparse(self_sparse); *self = HybridBitSet::Dense(new_dense); changed } @@ -1214,3 +1269,87 @@ fn sparse_matrix_iter() { } assert!(iter.next().is_none()); } + +/// Merge dense hybrid set into empty sparse hybrid set. +#[bench] +fn union_hybrid_sparse_empty_to_dense(b: &mut Bencher) { + let mut pre_dense: HybridBitSet = HybridBitSet::new_empty(256); + for i in 0..10 { + assert!(pre_dense.insert(i)); + } + let pre_sparse: HybridBitSet = HybridBitSet::new_empty(256); + b.iter(|| { + let dense = pre_dense.clone(); + let mut sparse = pre_sparse.clone(); + sparse.union(&dense); + }) +} + +/// Merge dense hybrid set into full hybrid set with same indices. +#[bench] +fn union_hybrid_sparse_full_to_dense(b: &mut Bencher) { + let mut pre_dense: HybridBitSet = HybridBitSet::new_empty(256); + for i in 0..10 { + assert!(pre_dense.insert(i)); + } + let mut pre_sparse: HybridBitSet = HybridBitSet::new_empty(256); + for i in 0..SPARSE_MAX { + assert!(pre_sparse.insert(i)); + } + b.iter(|| { + let dense = pre_dense.clone(); + let mut sparse = pre_sparse.clone(); + sparse.union(&dense); + }) +} + +/// Merge dense hybrid set into full hybrid set with indices over the whole domain. +#[bench] +fn union_hybrid_sparse_domain_to_dense(b: &mut Bencher) { + let mut pre_dense: HybridBitSet = HybridBitSet::new_empty(SPARSE_MAX*64); + for i in 0..10 { + assert!(pre_dense.insert(i)); + } + let mut pre_sparse: HybridBitSet = HybridBitSet::new_empty(SPARSE_MAX*64); + for i in 0..SPARSE_MAX { + assert!(pre_sparse.insert(i*64)); + } + b.iter(|| { + let dense = pre_dense.clone(); + let mut sparse = pre_sparse.clone(); + sparse.union(&dense); + }) +} + +/// Merge dense hybrid set into empty hybrid set where the domain is very small. +#[bench] +fn union_hybrid_sparse_empty_small_domain(b: &mut Bencher) { + let mut pre_dense: HybridBitSet = HybridBitSet::new_empty(SPARSE_MAX); + for i in 0..SPARSE_MAX { + assert!(pre_dense.insert(i)); + } + let pre_sparse: HybridBitSet = HybridBitSet::new_empty(SPARSE_MAX); + b.iter(|| { + let dense = pre_dense.clone(); + let mut sparse = pre_sparse.clone(); + sparse.union(&dense); + }) +} + +/// Merge dense hybrid set into full hybrid set where the domain is very small. +#[bench] +fn union_hybrid_sparse_full_small_domain(b: &mut Bencher) { + let mut pre_dense: HybridBitSet = HybridBitSet::new_empty(SPARSE_MAX); + for i in 0..SPARSE_MAX { + assert!(pre_dense.insert(i)); + } + let mut pre_sparse: HybridBitSet = HybridBitSet::new_empty(SPARSE_MAX); + for i in 0..SPARSE_MAX { + assert!(pre_sparse.insert(i)); + } + b.iter(|| { + let dense = pre_dense.clone(); + let mut sparse = pre_sparse.clone(); + sparse.union(&dense); + }) +}