Use our own bitset

VersionedMap.cpp

@@ -8,8 +8,6 @@
 #include <unistd.h>
 #include <unordered_set>
 
-#include <roaring.h>
-
 void *mmapSafe(void *addr, size_t len, int prot, int flags, int fd,
                off_t offset) {
   void *result = mmap(addr, len, prot, flags, fd, offset);
@@ -111,53 +109,30 @@ constexpr uint32_t kUpsizeBytes = 1 << 20;
 constexpr uint32_t kUpsizeNodes = kUpsizeBytes / sizeof(Node);
 static_assert(kUpsizeNodes * sizeof(Node) == kUpsizeBytes);
 
-struct BitSetUnorderedSet {
-  explicit BitSetUnorderedSet(uint32_t size) : s() {}
+struct BitSet {
+  explicit BitSet(uint32_t size) : words((uint64_t *)malloc(size / 64 + 64)) {}
 
-  bool test(uint32_t i) const { return s.find(i) != s.end(); }
+  bool test(uint32_t i) const {
+    return words[i >> 6] & (uint64_t(1) << (i & 63));
+  }
 
   // Returns former value
   bool set(uint32_t i) {
-    auto [it, inserted] = s.insert(i);
+    const auto prev = words[i >> 6];
+    const auto mask = uint64_t(1) << (i & 63);
+    words[i >> 6] |= mask;
     max_ = std::max(i, max_);
-    return !inserted;
+    return prev & mask;
   }
 
   // Returns 0 if set is empty
   uint32_t max() const { return max_; }
 
   template <class F>
-  void iterateAbsentBackwards(F f, uint32_t begin, uint32_t end) const {
-    for (uint32_t i = end - 1; i >= begin; --i) {
-      if (!test(i)) {
-        f(i);
-      }
-    }
-  }
-
-private:
-  uint32_t max_ = 0;
-  std::unordered_set<uint32_t> s;
-};
-
-struct BitSetR {
-  explicit BitSetR(uint32_t size) : s(bitset_create_with_capacity(size)) {}
-
-  bool test(uint32_t i) const { return bitset_get(s, i); }
-
-  // Returns former value
-  bool set(uint32_t i) {
-    max_ = std::max(i, max_);
-    auto result = test(i);
-    bitset_set(s, i);
-    return result;
-  }
-
-  // Returns 0 if set is empty
-  uint32_t max() const { return max_; }
-
-  template <class F>
-  void iterateAbsentBackwards(F f, uint32_t begin, uint32_t end) const {
+  void iterateAbsentApproxBackwards(F f, uint32_t begin, uint32_t end) const {
+    // TODO can this be improved? We can do something with a word at a time
+    // instead of a bit at a time. The first attempt at doing so benchmarked as
+    // slower.
     assert(begin != 0);
     for (uint32_t i = end - 1; i >= begin; --i) {
       if (!test(i)) {
@@ -166,11 +141,11 @@ struct BitSetR {
     }
   }
 
-  ~BitSetR() { bitset_free(s); }
+  ~BitSet() { free(words); }
 
 private:
   uint32_t max_ = 0;
-  bitset_t *s;
+  uint64_t *const words;
 };
 
 struct MemManager {
@@ -212,10 +187,8 @@ struct MemManager {
   }
 
   void gc(const uint32_t *roots, int numRoots, int64_t oldestVersion) {
-    // TODO better bitset?
-
     // Calculate reachable set
-    BitSetR reachable{next};
+    BitSet reachable{next};
     uint32_t stack[1000]; // Much more than bound imposed by max height of tree
     int stackIndex = 0;
     auto tryPush = [&](uint32_t p) {
@@ -272,7 +245,7 @@ struct MemManager {
 
     // Rebuild free list and delref entries
     freeList = 0;
-    reachable.iterateAbsentBackwards(
+    reachable.iterateAbsentApproxBackwards(
         [&](uint32_t i) {
           if (base[i].entry != nullptr) {
             base[i].entry->delref();
@@ -298,6 +271,7 @@ struct VersionedMap::Impl {};
 
 int main() {
   ankerl::nanobench::Bench bench;
+  bench.minEpochIterations(5000);
   weaselab::MemManager mm;
   bench.run("allocate", [&]() {
     auto x = mm.allocate();