Update our benchmarks in README

This seems to benchmark better

Bench.cpp

@@ -258,20 +258,17 @@ void benchConflictSet() {
   }
 }
 
-void benchWorstCaseForRadixRangeRead() {
-  ankerl::nanobench::Bench bench;
+constexpr int kKeyLenForWorstCase = 50;
+
+ConflictSet worstCaseConflictSetForRadixRangeRead(int cardinality) {
   ConflictSet cs{0};
 
-  int64_t version = 0;
-
-  constexpr int kKeyLength = 50;
-
-  for (int i = 0; i < kKeyLength; ++i) {
-    for (int j = 0; j < 256; ++j) {
+  for (int i = 0; i < kKeyLenForWorstCase; ++i) {
+    for (int j = 0; j < cardinality; ++j) {
       auto b = std::vector<uint8_t>(i, 0);
       b.push_back(j);
       auto e = std::vector<uint8_t>(i, 255);
-      e.push_back(j);
+      e.push_back(255 - j);
       weaselab::ConflictSet::WriteRange w[] = {{
                                                    {b.data(), int(b.size())},
                                                    {nullptr, 0},
@@ -280,50 +277,85 @@ void benchWorstCaseForRadixRangeRead() {
                                                    {e.data(), int(e.size())},
                                                    {nullptr, 0},
                                                }};
-      cs.addWrites(w, sizeof(w) / sizeof(w[0]), version);
+      std::sort(std::begin(w), std::end(w),
+                [](const auto &lhs, const auto &rhs) {
+                  int cl = std::min(lhs.begin.len, rhs.begin.len);
+                  if (cl > 0) {
+                    int c = memcmp(lhs.begin.p, rhs.begin.p, cl);
+                    if (c != 0) {
+                      return c < 0;
+                    }
+                  }
+                  return lhs.begin.len < rhs.begin.len;
+                });
+      cs.addWrites(w, sizeof(w) / sizeof(w[0]), 0);
     }
   }
 
   // Defeat short-circuiting on the left
   {
-    auto k = std::vector<uint8_t>(kKeyLength, 0);
+    auto k = std::vector<uint8_t>(kKeyLenForWorstCase, 0);
     weaselab::ConflictSet::WriteRange w[] = {
         {
             {k.data(), int(k.size())},
             {nullptr, 0},
         },
     };
-    cs.addWrites(w, sizeof(w) / sizeof(w[0]), 2);
+    cs.addWrites(w, sizeof(w) / sizeof(w[0]), 1);
   }
 
   // Defeat short-circuiting on the right
   {
-    auto k = std::vector<uint8_t>(kKeyLength, 255);
+    auto k = std::vector<uint8_t>(kKeyLenForWorstCase, 255);
     weaselab::ConflictSet::WriteRange w[] = {
         {
             {k.data(), int(k.size())},
             {nullptr, 0},
         },
     };
-    cs.addWrites(w, sizeof(w) / sizeof(w[0]), 2);
+    cs.addWrites(w, sizeof(w) / sizeof(w[0]), 1);
   }
 
-  auto begin = std::vector<uint8_t>(kKeyLength - 1, 0);
+  return cs;
+}
+
+void benchWorstCaseForRadixRangeRead() {
+  ankerl::nanobench::Bench bench;
+
+  std::unique_ptr<ConflictSet> cs[256];
+  for (int i = 0; i < 256; ++i) {
+    cs[i] =
+        std::make_unique<ConflictSet>(worstCaseConflictSetForRadixRangeRead(i));
+  }
+
+  auto begin = std::vector<uint8_t>(kKeyLenForWorstCase - 1, 0);
   begin.push_back(1);
-  auto end = std::vector<uint8_t>(kKeyLength - 1, 255);
+  auto end = std::vector<uint8_t>(kKeyLenForWorstCase - 1, 255);
   end.push_back(254);
 
   weaselab::ConflictSet::Result result;
   weaselab::ConflictSet::ReadRange r{
-      {begin.data(), int(begin.size())}, {end.data(), int(end.size())}, 1};
+      {begin.data(), int(begin.size())}, {end.data(), int(end.size())}, 0};
 
   bench.run("worst case for radix tree", [&]() {
-    result = weaselab::ConflictSet::TooOld;
-    cs.check(&r, &result, 1);
-    if (result != weaselab::ConflictSet::Commit) {
-      abort();
+    for (int i = 0; i < 256; ++i) {
+      result = weaselab::ConflictSet::TooOld;
+      cs[i]->check(&r, &result, 1);
+      if (result != weaselab::ConflictSet::Commit) {
+        abort();
+      }
     }
   });
+
+  // for (int i = 0; i < 256; ++i) {
+  //   bench.run("worst case for radix tree, span " + std::to_string(i), [&]() {
+  //     result = weaselab::ConflictSet::TooOld;
+  //     cs[i]->check(&r, &result, 1);
+  //     if (result != weaselab::ConflictSet::Commit) {
+  //       abort();
+  //     }
+  //   });
+  // }
 }
 
 int main(void) {

ConflictSet.cpp

@@ -294,6 +294,12 @@ struct Node48 : Node {
   int8_t nextFree;
   int8_t index[256];
   Child children[kMaxNodes];
+  uint8_t reverseIndex[kMaxNodes];
+  constexpr static int kMaxOfMaxPageSize = 8;
+  constexpr static int kMaxOfMaxShift =
+      std::countr_zero(uint32_t(kMaxOfMaxPageSize));
+  constexpr static int kMaxOfMaxTotalPages = kMaxNodes / kMaxOfMaxPageSize;
+  int64_t maxOfMax[kMaxOfMaxTotalPages];
 
   uint8_t *partialKey() { return (uint8_t *)(this + 1); }
 
@@ -308,6 +314,11 @@ struct Node256 : Node {
   constexpr static auto kType = Type_Node256;
   BitSet bitSet;
   Child children[256];
+  constexpr static int kMaxOfMaxPageSize = 8;
+  constexpr static int kMaxOfMaxShift =
+      std::countr_zero(uint32_t(kMaxOfMaxPageSize));
+  constexpr static int kMaxOfMaxTotalPages = 256 / kMaxOfMaxPageSize;
+  int64_t maxOfMax[kMaxOfMaxTotalPages];
   uint8_t *partialKey() { return (uint8_t *)(this + 1); }
   void copyChildrenAndKeyFrom(const Node48 &other);
   void copyChildrenAndKeyFrom(const Node256 &other);
@@ -405,6 +416,7 @@ inline void Node48::copyChildrenAndKeyFrom(const Node16 &other) {
          kNodeCopySize);
   assert(numChildren == Node16::kMaxNodes);
   memset(index, -1, sizeof(index));
+  memset(children, 0, sizeof(children));
   memcpy(partialKey(), &other + 1, partialKeyLen);
   bitSet.init();
   nextFree = Node16::kMaxNodes;
@@ -415,6 +427,9 @@ inline void Node48::copyChildrenAndKeyFrom(const Node16 &other) {
     children[i] = other.children[i];
     assert(children[i].child->parent == &other);
     children[i].child->parent = this;
+    reverseIndex[i] = x;
+    maxOfMax[i >> Node48::kMaxOfMaxShift] = std::max(
+        maxOfMax[i >> Node48::kMaxOfMaxShift], children[i].childMaxVersion);
     ++i;
   }
 }
@@ -422,13 +437,17 @@ inline void Node48::copyChildrenAndKeyFrom(const Node16 &other) {
 inline void Node48::copyChildrenAndKeyFrom(const Node48 &other) {
   memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
          kNodeCopySize);
-  memcpy(&bitSet, &other.bitSet,
-         sizeof(*this) - sizeof(Node) - sizeof(children));
+  bitSet = other.bitSet;
+  nextFree = other.nextFree;
+  memcpy(index, other.index, sizeof(index));
+  memset(children, 0, sizeof(children));
   for (int i = 0; i < numChildren; ++i) {
     children[i] = other.children[i];
     assert(children[i].child->parent == &other);
     children[i].child->parent = this;
   }
+  memcpy(reverseIndex, other.reverseIndex, sizeof(reverseIndex));
+  memcpy(maxOfMax, other.maxOfMax, sizeof(maxOfMax));
   memcpy(partialKey(), &other + 1, partialKeyLen);
 }
 
@@ -436,6 +455,7 @@ inline void Node48::copyChildrenAndKeyFrom(const Node256 &other) {
   memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
          kNodeCopySize);
   memset(index, -1, sizeof(index));
+  memset(children, 0, sizeof(children));
   nextFree = other.numChildren;
   bitSet = other.bitSet;
   int i = 0;
@@ -448,6 +468,9 @@ inline void Node48::copyChildrenAndKeyFrom(const Node256 &other) {
         children[i] = other.children[c];
         assert(children[i].child->parent == &other);
         children[i].child->parent = this;
+        reverseIndex[i] = c;
+        maxOfMax[i >> Node48::kMaxOfMaxShift] = std::max(
+            maxOfMax[i >> Node48::kMaxOfMaxShift], children[i].childMaxVersion);
         ++i;
       },
       0, 256);
@@ -457,13 +480,17 @@ inline void Node48::copyChildrenAndKeyFrom(const Node256 &other) {
 inline void Node256::copyChildrenAndKeyFrom(const Node48 &other) {
   memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
          kNodeCopySize);
-  memset(children, 0, sizeof(children));
   bitSet = other.bitSet;
+  memset(children, 0, sizeof(children));
+  memset(maxOfMax, 0, sizeof(maxOfMax));
   bitSet.forEachInRange(
       [&](int c) {
         children[c] = other.children[other.index[c]];
         assert(children[c].child->parent == &other);
         children[c].child->parent = this;
+        maxOfMax[c >> Node256::kMaxOfMaxShift] =
+            std::max(maxOfMax[c >> Node256::kMaxOfMaxShift],
+                     children[c].childMaxVersion);
       },
       0, 256);
   memcpy(partialKey(), &other + 1, partialKeyLen);
@@ -481,6 +508,7 @@ inline void Node256::copyChildrenAndKeyFrom(const Node256 &other) {
         children[c].child->parent = this;
       },
       0, 256);
+  memcpy(maxOfMax, other.maxOfMax, sizeof(maxOfMax));
   memcpy(partialKey(), &other + 1, partialKeyLen);
 }
 
@@ -532,7 +560,7 @@ template <class T> struct BoundedFreeListAllocator {
   static_assert(std::derived_from<T, Node>);
   static_assert(std::is_trivial_v<T>);
 
-  T *allocate(int partialKeyCapacity) {
+  T *allocate_helper(int partialKeyCapacity) {
     if (freeList != nullptr) {
       T *n = (T *)freeList;
       VALGRIND_MAKE_MEM_DEFINED(freeList, sizeof(freeList));
@@ -560,6 +588,17 @@ template <class T> struct BoundedFreeListAllocator {
     return result;
   }
 
+  T *allocate(int partialKeyCapacity) {
+    T *result = allocate_helper(partialKeyCapacity);
+    if constexpr (!std::is_same_v<T, Node0>) {
+      memset(result->children, 0, sizeof(result->children));
+    }
+    if constexpr (std::is_same_v<T, Node48> || std::is_same_v<T, Node256>) {
+      memset(result->maxOfMax, 0, sizeof(result->maxOfMax));
+    }
+    return result;
+  }
+
   void release(T *p) {
     if (freeListBytes >= kFreeListMaxMemory) {
       removeNode(p);
@@ -732,8 +771,9 @@ Node *&getChildExists(Node *self, uint8_t index) {
   }
 }
 
-// Precondition - an entry for index must exist in the node
-int64_t &maxVersion(Node *n, ConflictSet::Impl *);
+int64_t maxVersion(Node *n, ConflictSet::Impl *);
+
+void setMaxVersion(Node *n, ConflictSet::Impl *, int64_t maxVersion);
 
 Node *&getInTree(Node *n, ConflictSet::Impl *);
 
@@ -1025,6 +1065,7 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
     assert(self48->nextFree < 48);
     int nextFree = self48->nextFree++;
     self48->index[index] = nextFree;
+    self48->reverseIndex[nextFree] = index;
     auto &result = self48->children[nextFree].child;
     result = nullptr;
     return result;
@@ -1210,7 +1251,7 @@ void maybeDownsize(Node *self, NodeAllocators *allocators,
 
       // Max versions are stored in the parent, so we need to update it now
       // that we have a new parent.
-      maxVersion(child, impl) = childMaxVersion;
+      setMaxVersion(child, impl, childMaxVersion);
 
       getInTree(self, impl) = child;
       allocators->node3.release(self3);
@@ -1324,9 +1365,12 @@ Node *erase(Node *self, NodeAllocators *allocators, ConflictSet::Impl *impl,
     if (toRemoveChildrenIndex != lastChildrenIndex) {
       parent48->children[toRemoveChildrenIndex] =
           parent48->children[lastChildrenIndex];
-      parent48->index[parent48->children[toRemoveChildrenIndex]
-                          .child->parentsIndex] = toRemoveChildrenIndex;
+      auto parentIndex =
+          parent48->children[toRemoveChildrenIndex].child->parentsIndex;
+      parent48->index[parentIndex] = toRemoveChildrenIndex;
+      parent48->reverseIndex[toRemoveChildrenIndex] = parentIndex;
     }
+    parent48->children[lastChildrenIndex].childMaxVersion = 0;
 
     --parent->numChildren;
 
@@ -1646,56 +1690,81 @@ bool checkMaxBetweenExclusive(Node *n, int begin, int end,
     }
   }
 
+  // [begin, end) is now the half-open interval of children we're interested in.
+
+  const unsigned shiftUpperBound = end - begin;
+  const unsigned shiftAmount = begin;
+  auto inBounds = [&](unsigned c) { return c - shiftAmount < shiftUpperBound; };
+
   switch (n->getType()) {
   case Type_Node0: // GCOVR_EXCL_LINE
     // We would have returned above, after not finding a child
     __builtin_unreachable(); // GCOVR_EXCL_LINE
   case Type_Node3: {
     auto *self = static_cast<Node3 *>(n);
-    for (int i = 0; i < self->numChildren && self->index[i] < end; ++i) {
-      if (begin <= self->index[i]) {
-        if (self->children[i].childMaxVersion > readVersion) {
-          return false;
-        }
-      }
+    bool result = true;
+    for (int i = 0; i < 3; ++i) {
+      result &= !((self->children[i].childMaxVersion > readVersion) &
+                  inBounds(self->index[i]));
     }
+    return result;
   } break;
   case Type_Node16: {
     auto *self = static_cast<Node16 *>(n);
-    for (int i = 0; i < self->numChildren && self->index[i] < end; ++i) {
-      if (begin <= self->index[i]) {
-        if (self->children[i].childMaxVersion > readVersion) {
-          return false;
-        }
-      }
+    bool result = true;
+    for (int i = 0; i < 16; ++i) {
+      result &= !((self->children[i].childMaxVersion > readVersion) &
+                  inBounds(self->index[i]));
     }
+    return result;
   } break;
   case Type_Node48: {
-    bool conflict[256] = {};
     auto *self = static_cast<Node48 *>(n);
-    self->bitSet.forEachInRange(
-        [&](int i) {
-          conflict[i] =
-              self->children[self->index[i]].childMaxVersion > readVersion;
-        },
-        begin, end);
+    // Check all pages
     bool result = true;
-    for (auto c : conflict) {
-      result &= !c;
+    for (int i = 0; i < Node48::kMaxOfMaxTotalPages; ++i) {
+      if (self->maxOfMax[i] > readVersion) {
+        for (int j = 0; j < Node48::kMaxOfMaxPageSize; ++j) {
+          int k = (i << Node48::kMaxOfMaxShift) + j;
+          result &= !(self->children[k].childMaxVersion > readVersion &&
+                      inBounds(self->reverseIndex[k]));
+        }
+      }
     }
     return result;
   }
   case Type_Node256: {
-    bool conflict[256] = {};
     auto *self = static_cast<Node256 *>(n);
-    self->bitSet.forEachInRange(
-        [&](int i) {
-          conflict[i] = self->children[i].childMaxVersion > readVersion;
-        },
-        begin, end);
+    // Check the first page
+    if (self->maxOfMax[begin >> Node256::kMaxOfMaxShift] > readVersion) {
+      bool result = true;
+      for (int i = 0; i < Node256::kMaxOfMaxPageSize; ++i) {
+        int j = (begin & ~(Node256::kMaxOfMaxPageSize - 1)) + i;
+        result &=
+            !((self->children[j].childMaxVersion > readVersion) & inBounds(j));
+      }
+      if (!result) {
+        return result;
+      }
+    }
+    // Check the last page
+    if (end >= 1 &&
+        self->maxOfMax[(end - 1) >> Node256::kMaxOfMaxShift] > readVersion) {
+      bool result = true;
+      for (int i = 0; i < Node256::kMaxOfMaxPageSize; ++i) {
+        int j = ((end - 1) & ~(Node256::kMaxOfMaxPageSize - 1)) + i;
+        result &=
+            !((self->children[j].childMaxVersion > readVersion) & inBounds(j));
+      }
+      if (!result) {
+        return result;
+      }
+    }
+    // Check inner pages
     bool result = true;
-    for (auto c : conflict) {
-      result &= !c;
+    for (int i = (begin >> Node256::kMaxOfMaxShift) + 1;
+         i < ((end - 1) >> Node256::kMaxOfMaxShift); ++i) {
+      result &= self->maxOfMax[i] <= readVersion;
     }
     return result;
   }
@@ -2187,7 +2256,7 @@ template <bool kBegin>
                          allocators) = old;
         old->parent = *self;
         old->parentsIndex = old->partialKey()[partialKeyIndex];
-        maxVersion(old, impl) = oldMaxVersion;
+        setMaxVersion(old, impl, oldMaxVersion);
 
         memmove(old->partialKey(), old->partialKey() + partialKeyIndex + 1,
                 old->partialKeyLen - (partialKeyIndex + 1));
@@ -2211,9 +2280,8 @@ template <bool kBegin>
     }
 
     if constexpr (kBegin) {
-      auto &m = maxVersion(*self, impl);
-      assert(writeVersion >= m);
-      m = writeVersion;
+      assert(maxVersion(*self, impl) <= writeVersion);
+      setMaxVersion(*self, impl, writeVersion);
     }
 
     if (key.size() == 0) {
@@ -2221,9 +2289,8 @@ template <bool kBegin>
     }
 
     if constexpr (!kBegin) {
-      auto &m = maxVersion(*self, impl);
-      assert(writeVersion >= m);
-      m = writeVersion;
+      assert(maxVersion(*self, impl) <= writeVersion);
+      setMaxVersion(*self, impl, writeVersion);
     }
 
     auto &child = getOrCreateChild(*self, key.front(), allocators);
@@ -2234,8 +2301,7 @@ template <bool kBegin>
       child->partialKeyLen = 0;
       child->parent = *self;
       child->parentsIndex = key.front();
-      maxVersion(child, impl) =
-          kBegin ? writeVersion : std::numeric_limits<int64_t>::lowest();
+      setMaxVersion(child, impl, kBegin ? writeVersion : 0);
     }
 
     self = &child;
@@ -2280,7 +2346,7 @@ void addPointWrite(Node *&root, int64_t oldestVersion,
     n->entryPresent = true;
 
     n->entry.pointVersion = writeVersion;
-    maxVersion(n, impl) = writeVersion;
+    setMaxVersion(n, impl, writeVersion);
     n->entry.rangeVersion =
         p != nullptr ? p->entry.rangeVersion : oldestVersion;
   } else {
@@ -2321,9 +2387,8 @@ void addWriteRange(Node *&root, int64_t oldestVersion,
       break;
     }
 
-    auto &m = maxVersion(n, impl);
-    assert(writeVersion >= m);
-    m = writeVersion;
+    assert(maxVersion(n, impl) <= writeVersion);
+    setMaxVersion(n, impl, writeVersion);
 
     remaining = remaining.subspan(n->partialKeyLen + 1,
                                   remaining.size() - (n->partialKeyLen + 1));
@@ -2350,11 +2415,10 @@ void addWriteRange(Node *&root, int64_t oldestVersion,
     beginNode->entry.rangeVersion =
         p != nullptr ? p->entry.rangeVersion : oldestVersion;
     beginNode->entry.pointVersion = writeVersion;
-    maxVersion(beginNode, impl) = writeVersion;
+    assert(maxVersion(beginNode, impl) <= writeVersion);
+    setMaxVersion(beginNode, impl, writeVersion);
   }
-  auto &m = maxVersion(beginNode, impl);
-  assert(writeVersion >= m);
-  m = writeVersion;
+  setMaxVersion(beginNode, impl, writeVersion);
   assert(writeVersion >= beginNode->entry.pointVersion);
   beginNode->entry.pointVersion = writeVersion;
 
@@ -2369,8 +2433,8 @@ void addWriteRange(Node *&root, int64_t oldestVersion,
     auto *p = nextLogical(endNode);
     endNode->entry.pointVersion =
         p != nullptr ? p->entry.rangeVersion : oldestVersion;
-    auto &m = maxVersion(endNode, impl);
-    m = std::max(m, endNode->entry.pointVersion);
+    auto m = maxVersion(endNode, impl);
+    setMaxVersion(endNode, impl, std::max(m, endNode->entry.pointVersion));
   }
   endNode->entry.rangeVersion = writeVersion;
 
@@ -2570,8 +2634,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   int64_t totalBytes = 0;
 };
 
-// Precondition - an entry for index must exist in the node
-int64_t &maxVersion(Node *n, ConflictSet::Impl *impl) {
+int64_t maxVersion(Node *n, ConflictSet::Impl *impl) {
   int index = n->parentsIndex;
   n = n->parent;
   if (n == nullptr) {
@@ -2605,6 +2668,50 @@ int64_t &maxVersion(Node *n, ConflictSet::Impl *impl) {
   }
 }
 
+void setMaxVersion(Node *n, ConflictSet::Impl *impl, int64_t newMax) {
+  int index = n->parentsIndex;
+  n = n->parent;
+  if (n == nullptr) {
+    impl->rootMaxVersion = newMax;
+    return;
+  }
+  switch (n->getType()) {
+  case Type_Node0:           // GCOVR_EXCL_LINE
+    __builtin_unreachable(); // GCOVR_EXCL_LINE
+  case Type_Node3: {
+    auto *n3 = static_cast<Node3 *>(n);
+    int i = getNodeIndex(n3, index);
+    n3->children[i].childMaxVersion = newMax;
+    return;
+  }
+  case Type_Node16: {
+    auto *n16 = static_cast<Node16 *>(n);
+    int i = getNodeIndex(n16, index);
+    n16->children[i].childMaxVersion = newMax;
+    return;
+  }
+  case Type_Node48: {
+    auto *n48 = static_cast<Node48 *>(n);
+    assert(n48->bitSet.test(index));
+    int i = n48->index[index];
+    n48->children[i].childMaxVersion = newMax;
+    n48->maxOfMax[i >> Node48::kMaxOfMaxShift] =
+        std::max(n48->maxOfMax[i >> Node48::kMaxOfMaxShift], newMax);
+    return;
+  }
+  case Type_Node256: {
+    auto *n256 = static_cast<Node256 *>(n);
+    assert(n256->bitSet.test(index));
+    n256->children[index].childMaxVersion = newMax;
+    n256->maxOfMax[index >> Node256::kMaxOfMaxShift] =
+        std::max(n256->maxOfMax[index >> Node256::kMaxOfMaxShift], newMax);
+    return;
+  }
+  default:                   // GCOVR_EXCL_LINE
+    __builtin_unreachable(); // GCOVR_EXCL_LINE
+  }
+}
+
 Node *&getInTree(Node *n, ConflictSet::Impl *impl) {
   return n->parent == nullptr ? impl->root
                               : getChildExists(n->parent, n->parentsIndex);
@@ -2860,7 +2967,7 @@ Iterator firstGeq(Node *n, std::string_view key) {
 [[maybe_unused]] int64_t checkMaxVersion(Node *root, Node *node,
                                          int64_t oldestVersion, bool &success,
                                          ConflictSet::Impl *impl) {
-  int64_t expected = std::numeric_limits<int64_t>::lowest();
+  int64_t expected = 0;
   if (node->entryPresent) {
     expected = std::max(expected, node->entry.pointVersion);
   }

README.md

@@ -60,27 +60,27 @@ Performance counters:
 
 |               ns/op |                op/s |    err% |     total | benchmark
 |--------------------:|--------------------:|--------:|----------:|:----------
-|              255.20 |        3,918,570.44 |    0.1% |      0.01 | `point reads`
-|              269.35 |        3,712,633.44 |    0.8% |      0.01 | `prefix reads`
-|              502.72 |        1,989,186.40 |    0.4% |      0.01 | `range reads`
-|              456.85 |        2,188,902.27 |    0.6% |      0.01 | `point writes`
-|              444.81 |        2,248,148.60 |    0.7% |      0.01 | `prefix writes`
-|              250.00 |        4,000,000.00 |    1.7% |      0.02 | `range writes`
-|              566.51 |        1,765,186.07 |    0.5% |      0.01 | `monotonic increasing point writes`
-|              226.41 |        4,416,703.74 |    0.5% |      0.01 | `worst case for radix tree`
+|              256.89 |        3,892,784.92 |    0.3% |      0.01 | `point reads`
+|              272.90 |        3,664,395.04 |    0.2% |      0.01 | `prefix reads`
+|              507.22 |        1,971,549.50 |    0.7% |      0.01 | `range reads`
+|              452.66 |        2,209,181.91 |    0.5% |      0.01 | `point writes`
+|              438.09 |        2,282,619.96 |    0.4% |      0.01 | `prefix writes`
+|              253.33 |        3,947,420.36 |    2.5% |      0.02 | `range writes`
+|              574.07 |        1,741,936.71 |    0.3% |      0.01 | `monotonic increasing point writes`
+|          151,562.50 |            6,597.94 |    1.5% |      0.01 | `worst case for radix tree`
 
 ## Radix tree (this implementation)
 
 |               ns/op |                op/s |    err% |     total | benchmark
 |--------------------:|--------------------:|--------:|----------:|:----------
-|               19.60 |       51,025,020.51 |    0.1% |      0.01 | `point reads`
-|               55.62 |       17,980,734.93 |    0.7% |      0.01 | `prefix reads`
-|              174.86 |        5,718,896.02 |    0.4% |      0.01 | `range reads`
-|               28.27 |       35,372,166.39 |    0.1% |      0.01 | `point writes`
-|               43.85 |       22,804,171.49 |    0.5% |      0.01 | `prefix writes`
-|               49.59 |       20,165,355.92 |    0.9% |      0.01 | `range writes`
-|               92.04 |       10,864,732.33 |    3.6% |      0.01 | `monotonic increasing point writes`
-|            6,937.00 |          144,154.53 |    0.2% |      0.01 | `worst case for radix tree`
+|               19.83 |       50,420,955.28 |    0.1% |      0.01 | `point reads`
+|               55.95 |       17,872,542.40 |    0.5% |      0.01 | `prefix reads`
+|               88.28 |       11,327,709.50 |    0.4% |      0.01 | `range reads`
+|               29.15 |       34,309,531.64 |    0.5% |      0.01 | `point writes`
+|               42.36 |       23,607,424.27 |    1.1% |      0.01 | `prefix writes`
+|               50.00 |       20,000,000.00 |    0.0% |      0.01 | `range writes`
+|               93.52 |       10,692,413.79 |    3.3% |      0.01 | `monotonic increasing point writes`
+|        2,388,417.00 |              418.69 |    0.4% |      0.03 | `worst case for radix tree`
 
 # "Real data" test
 

include/ConflictSet.h

@@ -21,7 +21,7 @@ limitations under the License.
 #ifdef __cplusplus
 namespace weaselab {
 /** A data structure for optimistic concurrency control on ranges of
- * bitwise-lexicographically-ordered keys.
+ * bitwise-lexicographically-ordered keys. All versions must be >= 0.
  *
  * Thread safety:
  *   - It's safe to operate on two different ConflictSets in two different
@@ -106,7 +106,7 @@ private:
 #else
 
 /** A data structure for optimistic concurrency control on ranges of
- * bitwise-lexicographically-ordered keys.
+ * bitwise-lexicographically-ordered keys. All versions must be >= 0.
  *
  * Thread safety:
  *   - It's safe to operate on two different ConflictSets in two different