Small performance boost.

Cargo.toml

@@ -17,5 +17,8 @@ serde_test = "1.0"
 name = "bench"
 harness = false
 
+[lib]
+bench = false
+
 [profile.release]
 lto = true

benches/bench.rs

@@ -17,16 +17,27 @@ fn default_set<C: FromIterator<u64>>(n: usize) -> C {
     (0..n).map(|_| rng.next_u64()).collect()
 }
 
+fn contains(data: &[u64]) {
+    let set = BitSet::new();
+
+    for v in data {
+        set.contains(v);
+    }
+}
+
+fn insert(data: &[u64]) {
+    let mut set = BitSet::new();
+
+    for v in data {
+        set.insert(*v);
+    }
+}
+
 fn criterion_benchmark(c: &mut Criterion) {
     let data: Vec<_> = default_set(1000);
 
-    c.bench_function("insert", |b| b.iter(|| {
+    c.bench_function("contains", |b| b.iter(|| contains(&data)));
-        let mut set = BitSet::new();
+    c.bench_function("insert", |b| b.iter(|| insert(&data)));
-
-        for v in &data {
-            set.insert(*v);
-        }
-    }));
 }
 
 criterion_group!(benches, criterion_benchmark);

src/lib.rs

@@ -76,10 +76,12 @@ impl BitSet {
         }
     }
 
+    #[inline]
     pub fn len(&self) -> usize {
         self.nbits
     }
 
+    #[inline]
     pub fn is_empty(&self) -> bool {
         self.nbits == 0
     }
@@ -89,6 +91,7 @@ impl BitSet {
         self.nbits = 0;
     }
 
+    #[inline]
     pub fn contains(&self, value: &u64) -> bool {
         let (block, bit) = block_bit(value, &BITS);
 
@@ -98,6 +101,7 @@ impl BitSet {
         }
     }
 
+    #[inline]
     pub fn is_subset(&self, other: &BitSet) -> bool {
         if self.nbits > other.nbits {
             false
@@ -116,6 +120,7 @@ impl BitSet {
         other.is_subset(self)
     }
 
+    #[inline]
     pub fn insert(&mut self, value: u64) -> bool {
         let (block, bit) = block_bit(&value, &BITS);
         let block = self.blocks.entry(block).or_insert(0);
@@ -132,6 +137,7 @@ impl BitSet {
         }
     }
 
+    #[inline]
     pub fn remove(&mut self, value: &u64) -> bool {
         let (block, bit) = block_bit(value, &BITS);
         let block = self.blocks.entry(block).or_insert(0);
@@ -148,15 +154,17 @@ impl BitSet {
         }
     }
 
+    #[inline]
     pub fn union_with(&mut self, other: &Self) {
         for (key, value) in &other.blocks {
-            *self.blocks.entry(*key).or_insert(0) |= value;
+            let block = self.blocks.entry(*key).or_insert(0);
-        }
 
-        self.nbits = self.blocks
+            let n = block.count_ones();
-            .values()
+
-            .map(|block| block.count_ones() as usize)
+            *block |= value;
-            .sum();
+
+            self.nbits += (block.count_ones() - n) as usize;
+        }
     }
 }
 
@@ -206,13 +214,14 @@ impl ops::BitOr for BitSet {
     #[inline]
     fn bitor(mut self, rhs: Self) -> Self {
         for (key, value) in &rhs.blocks {
-            *self.blocks.entry(*key).or_insert(0) |= value;
+            let block = self.blocks.entry(*key).or_insert(0);
-        }
 
-        self.nbits = self.blocks
+            let n = block.count_ones();
-            .values()
+
-            .map(|block| block.count_ones() as usize)
+            *block |= value;
-            .sum();
+
+            self.nbits += (block.count_ones() - n) as usize;
+        }
 
         self
     }
@@ -224,13 +233,14 @@ impl<'a> ops::BitOr<&'a Self> for BitSet {
     #[inline]
     fn bitor(mut self, rhs: &'a Self) -> Self {
         for (key, value) in &rhs.blocks {
-            *self.blocks.entry(*key).or_insert(0) |= value;
+            let block = self.blocks.entry(*key).or_insert(0);
-        }
 
-        self.nbits = self.blocks
+            let n = block.count_ones();
-            .values()
+
-            .map(|block| block.count_ones() as usize)
+            *block |= value;
-            .sum();
+
+            self.nbits += (block.count_ones() - n) as usize;
+        }
 
         self
     }