Lower kBytesPerKey to 144 by changing Node4 to Node3

ConflictSet.cpp

@@ -176,7 +176,7 @@ int BitSet::firstSetGeq(int i) const {
 
 enum class Type : int8_t {
   Node0,
-  Node4,
+  Node3,
   Node16,
   Node48,
   Node256,
@@ -215,19 +215,19 @@ struct Node0 : Node {
   uint8_t *partialKey() { return (uint8_t *)(this + 1); }
 };
 
-struct Node4 : Node {
-  // Sorted
-  uint8_t index[16]; // 16 so that we can use the same simd index search
-                     // implementation as Node16
-  Child children[4];
-  Node4() { this->type = Type::Node4; }
+struct Node3 : Node {
+  constexpr static auto kMaxNodes = 3;
+  uint8_t index[kMaxNodes];
+  Child children[kMaxNodes];
+  Node3() { this->type = Type::Node3; }
   uint8_t *partialKey() { return (uint8_t *)(this + 1); }
 };
 
 struct Node16 : Node {
   // Sorted
-  uint8_t index[16];
-  Child children[16];
+  constexpr static auto kMaxNodes = 16;
+  uint8_t index[kMaxNodes];
+  Child children[kMaxNodes];
   Node16() { this->type = Type::Node16; }
   uint8_t *partialKey() { return (uint8_t *)(this + 1); }
 };
@@ -262,22 +262,33 @@ std::string getSearchPathPrintable(Node *n);
 
 // Bound memory usage following the analysis in the ART paper
 
-constexpr int kBytesPerKey = 176;
-constexpr int kMinSavingsPerNode = 120;
-constexpr int kMinChildrenNode4 = 2;
-constexpr int kMinChildrenNode16 = 5;
+// Each node with an entry present gets a budget of kBytesPerKey. Node0 always
+// has an entry present.
+constexpr int kBytesPerKey = 144;
+// Induction hypothesis is that each node's surplus is >= kMinNodeSurplus
+constexpr int kMinNodeSurplus = 104;
+constexpr int kMinChildrenNode3 = 2;
+constexpr int kMinChildrenNode16 = 4;
 constexpr int kMinChildrenNode48 = 17;
 constexpr int kMinChildrenNode256 = 49;
 
-static_assert(kMinChildrenNode256 * kMinSavingsPerNode - sizeof(Node256) >=
-              kMinSavingsPerNode);
-static_assert(kMinChildrenNode48 * kMinSavingsPerNode - sizeof(Node48) >=
-              kMinSavingsPerNode);
-static_assert(kMinChildrenNode16 * kMinSavingsPerNode - sizeof(Node16) >=
-              kMinSavingsPerNode);
-static_assert(kMinChildrenNode4 * kMinSavingsPerNode - sizeof(Node4) >=
-              kMinSavingsPerNode);
-static_assert(kBytesPerKey - sizeof(Node0) >= kMinSavingsPerNode);
+constexpr int kNode256Surplus =
+    kMinChildrenNode256 * kMinNodeSurplus - sizeof(Node256);
+static_assert(kNode256Surplus >= kMinNodeSurplus);
+
+constexpr int kNode48Surplus =
+    kMinChildrenNode48 * kMinNodeSurplus - sizeof(Node48);
+static_assert(kNode48Surplus >= kMinNodeSurplus);
+
+constexpr int kNode16Surplus =
+    kMinChildrenNode16 * kMinNodeSurplus - sizeof(Node16);
+static_assert(kNode16Surplus >= kMinNodeSurplus);
+
+constexpr int kNode3Surplus =
+    kMinChildrenNode3 * kMinNodeSurplus - sizeof(Node3);
+static_assert(kNode3Surplus >= kMinNodeSurplus);
+
+static_assert(kBytesPerKey - sizeof(Node0) >= kMinNodeSurplus);
 
 // setOldestVersion will additionally try to maintain this property:
 //     `max(children, 1) * length >= capacity`
@@ -356,8 +367,8 @@ uint8_t *Node::partialKey() {
   switch (type) {
   case Type::Node0:
     return ((Node0 *)this)->partialKey();
-  case Type::Node4:
-    return ((Node4 *)this)->partialKey();
+  case Type::Node3:
+    return ((Node3 *)this)->partialKey();
   case Type::Node16:
     return ((Node16 *)this)->partialKey();
   case Type::Node48:
@@ -370,13 +381,14 @@ uint8_t *Node::partialKey() {
 
 struct NodeAllocators {
   BoundedFreeListAllocator<Node0> node0;
-  BoundedFreeListAllocator<Node4> node4;
+  BoundedFreeListAllocator<Node3> node3;
   BoundedFreeListAllocator<Node16> node16;
   BoundedFreeListAllocator<Node48> node48;
   BoundedFreeListAllocator<Node256> node256;
 };
 
-int getNodeIndex(Node16 *self, uint8_t index) {
+template <class NodeT> int getNodeIndex(NodeT *self, uint8_t index) {
+  static_assert(std::is_same_v<NodeT, Node3> || std::is_same_v<NodeT, Node16>);
 #ifdef HAS_AVX
   // Based on https://www.the-paper-trail.org/post/art-paper-notes/
 
@@ -388,7 +400,7 @@ int getNodeIndex(Node16 *self, uint8_t index) {
   // keys aren't valid, we'll mask the results to only consider the valid ones
   // below.
   __m128i indices;
-  memcpy(&indices, self->index, sizeof(self->index));
+  memcpy(&indices, self->index, NodeT::kMaxNodes);
   __m128i results = _mm_cmpeq_epi8(key_vec, indices);
 
   // Build a mask to select only the first node->num_children values from the
@@ -411,10 +423,12 @@ int getNodeIndex(Node16 *self, uint8_t index) {
   // https://community.arm.com/arm-community-blogs/b/infrastructure-solutions-blog/posts/porting-x86-vector-bitmask-optimizations-to-arm-neon
 
   uint8x16_t indices;
-  memcpy(&indices, self->index, sizeof(self->index));
+  memcpy(&indices, self->index, NodeT::kMaxNodes);
   // 0xff for each match
   uint16x8_t results =
       vreinterpretq_u16_u8(vceqq_u8(vdupq_n_u8(index), indices));
+  static_assert(NodeT::kMaxNodes <= 16);
+  assume(self->numChildren <= NodeT::kMaxNodes);
   uint64_t mask = self->numChildren == 16
                       ? uint64_t(-1)
                       : (uint64_t(1) << (self->numChildren * 4)) - 1;
@@ -439,8 +453,10 @@ Node *&getChildExists(Node *self, uint8_t index) {
   switch (self->type) {
   case Type::Node0:
     __builtin_unreachable(); // GCOVR_EXCL_LINE
-  case Type::Node4:
-    [[fallthrough]];
+  case Type::Node3: {
+    auto *self3 = static_cast<Node3 *>(self);
+    return self3->children[getNodeIndex(self3, index)].child;
+  }
   case Type::Node16: {
     auto *self16 = static_cast<Node16 *>(self);
     return self16->children[getNodeIndex(self16, index)].child;
@@ -467,8 +483,11 @@ Node *getChild(Node *self, uint8_t index) {
   switch (self->type) {
   case Type::Node0:
     return nullptr;
-  case Type::Node4:
-    [[fallthrough]];
+  case Type::Node3: {
+    auto *self3 = static_cast<Node3 *>(self);
+    int i = getNodeIndex(self3, index);
+    return i < 0 ? nullptr : self3->children[i].child;
+  }
   case Type::Node16: {
     auto *self16 = static_cast<Node16 *>(self);
     int i = getNodeIndex(self16, index);
@@ -486,6 +505,64 @@ Node *getChild(Node *self, uint8_t index) {
   }
 }
 
+template <class NodeT> int getChildGeqSimd(NodeT *self, int child) {
+  static_assert(std::is_same_v<NodeT, Node3> || std::is_same_v<NodeT, Node16>);
+#ifdef HAS_AVX
+  __m128i key_vec = _mm_set1_epi8(child);
+  __m128i indices;
+  memcpy(&indices, self->index, NodeT::kMaxNodes);
+  __m128i results = _mm_cmpeq_epi8(key_vec, _mm_min_epu8(key_vec, indices));
+  int mask = (1 << self->numChildren) - 1;
+  uint32_t bitfield = _mm_movemask_epi8(results) & mask;
+  int result = bitfield == 0 ? -1 : self->index[std::countr_zero(bitfield)];
+  assert(result == [&]() -> int {
+    for (int i = 0; i < self->numChildren; ++i) {
+      if (self->index[i] >= child) {
+        return self->index[i];
+      }
+    }
+    return -1;
+  }());
+  return result;
+#elif defined(HAS_ARM_NEON)
+  uint8x16_t indices;
+  memcpy(&indices, self->index, sizeof(self->index));
+  // 0xff for each leq
+  auto results = vcleq_u8(vdupq_n_u8(child), indices);
+  static_assert(NodeT::kMaxNodes <= 16);
+  assume(self->numChildren <= NodeT::kMaxNodes);
+  uint64_t mask = self->numChildren == 16
+                      ? uint64_t(-1)
+                      : (uint64_t(1) << (self->numChildren * 4)) - 1;
+  // 0xf for each 0xff (within mask)
+  uint64_t bitfield =
+      vget_lane_u64(
+          vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(results), 4)),
+          0) &
+      mask;
+  int simd = bitfield == 0 ? -1 : self->index[std::countr_zero(bitfield) / 4];
+  assert(simd == [&]() -> int {
+    for (int i = 0; i < self->numChildren; ++i) {
+      if (self->index[i] >= child) {
+        return self->index[i];
+      }
+    }
+    return -1;
+  }());
+  return simd;
+#else
+  for (int i = 0; i < self->numChildren; ++i) {
+    if (i > 0) {
+      assert(self->index[i - 1] < self->index[i]);
+    }
+    if (self->index[i] >= child) {
+      return self->index[i];
+    }
+  }
+  return -1;
+#endif
+}
+
 int getChildGeq(Node *self, int child) {
   if (child > 255) {
     return -1;
@@ -493,64 +570,10 @@ int getChildGeq(Node *self, int child) {
   switch (self->type) {
   case Type::Node0:
     return -1;
-  case Type::Node4:
-    [[fallthrough]];
-  case Type::Node16: {
-    auto *self16 = static_cast<Node16 *>(self);
-#ifdef HAS_AVX
-    __m128i key_vec = _mm_set1_epi8(child);
-    __m128i indices;
-    memcpy(&indices, self16->index, sizeof(self16->index));
-    __m128i results = _mm_cmpeq_epi8(key_vec, _mm_min_epu8(key_vec, indices));
-    int mask = (1 << self16->numChildren) - 1;
-    uint32_t bitfield = _mm_movemask_epi8(results) & mask;
-    int result = bitfield == 0 ? -1 : self16->index[std::countr_zero(bitfield)];
-    assert(result == [&]() -> int {
-      for (int i = 0; i < self16->numChildren; ++i) {
-        if (self16->index[i] >= child) {
-          return self16->index[i];
-        }
-      }
-      return -1;
-    }());
-    return result;
-#elif defined(HAS_ARM_NEON)
-    uint8x16_t indices;
-    memcpy(&indices, self16->index, sizeof(self16->index));
-    // 0xff for each leq
-    auto results = vcleq_u8(vdupq_n_u8(child), indices);
-    uint64_t mask = self->numChildren == 16
-                        ? uint64_t(-1)
-                        : (uint64_t(1) << (self->numChildren * 4)) - 1;
-    // 0xf for each 0xff (within mask)
-    uint64_t bitfield =
-        vget_lane_u64(
-            vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(results), 4)),
-            0) &
-        mask;
-    int simd =
-        bitfield == 0 ? -1 : self16->index[std::countr_zero(bitfield) / 4];
-    assert(simd == [&]() -> int {
-      for (int i = 0; i < self->numChildren; ++i) {
-        if (self16->index[i] >= child) {
-          return self16->index[i];
-        }
-      }
-      return -1;
-    }());
-    return simd;
-#else
-    for (int i = 0; i < self->numChildren; ++i) {
-      if (i > 0) {
-        assert(self16->index[i - 1] < self16->index[i]);
-      }
-      if (self16->index[i] >= child) {
-        return self16->index[i];
-      }
-    }
-    return -1;
-#endif
-  }
+  case Type::Node3:
+    return getChildGeqSimd(static_cast<Node3 *>(self), child);
+  case Type::Node16:
+    return getChildGeqSimd(static_cast<Node16 *>(self), child);
   case Type::Node48:
     [[fallthrough]];
   case Type::Node256: {
@@ -561,7 +584,7 @@ int getChildGeq(Node *self, int child) {
   }
 }
 
-void setChildrenParents(Node4 *n) {
+void setChildrenParents(Node3 *n) {
   for (int i = 0; i < n->numChildren; ++i) {
     n->children[i].child->parent = n;
   }
@@ -592,8 +615,13 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
   switch (self->type) {
   case Type::Node0:
     break;
-  case Type::Node4:
-    [[fallthrough]];
+  case Type::Node3: {
+    auto *self3 = static_cast<Node3 *>(self);
+    int i = getNodeIndex(self3, index);
+    if (i >= 0) {
+      return self3->children[i].child;
+    }
+  } break;
   case Type::Node16: {
     auto *self16 = static_cast<Node16 *>(self);
     int i = getNodeIndex(self16, index);
@@ -620,43 +648,59 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
   case Type::Node0: {
     auto *self0 = static_cast<Node0 *>(self);
 
-    auto *newSelf = allocators->node4.allocate(self->partialKeyLen);
+    auto *newSelf = allocators->node3.allocate(self->partialKeyLen);
     memcpy((char *)newSelf + kNodeCopyBegin, (char *)self + kNodeCopyBegin,
            kNodeCopySize);
     memcpy(newSelf->partialKey(), self0->partialKey(), self->partialKeyLen);
     allocators->node0.release(self0);
     self = newSelf;
 
-    goto insert16;
+    goto insert3;
   }
-  case Type::Node4: {
-    auto *self4 = static_cast<Node4 *>(self);
-
-    if (self->numChildren == 4) {
+  case Type::Node3: {
+    if (self->numChildren == Node3::kMaxNodes) {
+      auto *self3 = static_cast<Node3 *>(self);
       auto *newSelf = allocators->node16.allocate(self->partialKeyLen);
       memcpy((char *)newSelf + kNodeCopyBegin, (char *)self + kNodeCopyBegin,
              kNodeCopySize);
-      memcpy(newSelf->partialKey(), self4->partialKey(), self->partialKeyLen);
+      memcpy(newSelf->partialKey(), self3->partialKey(), self->partialKeyLen);
       // TODO replace with memcpy?
-      for (int i = 0; i < 4; ++i) {
-        newSelf->index[i] = self4->index[i];
-        newSelf->children[i] = self4->children[i];
+      for (int i = 0; i < Node3::kMaxNodes; ++i) {
+        newSelf->index[i] = self3->index[i];
+        newSelf->children[i] = self3->children[i];
       }
-      allocators->node4.release(self4);
+      allocators->node3.release(self3);
       setChildrenParents(newSelf);
       self = newSelf;
+      goto insert16;
     }
 
-    goto insert16;
+  insert3:
+    auto *self3 = static_cast<Node3 *>(self);
+    ++self->numChildren;
+    int i = 0;
+    for (; i < self->numChildren - 1; ++i) {
+      if (int(self3->index[i]) > int(index)) {
+        memmove(self3->index + i + 1, self3->index + i,
+                self->numChildren - (i + 1));
+        memmove(self3->children + i + 1, self3->children + i,
+                (self->numChildren - (i + 1)) * sizeof(Child));
+        break;
+      }
+    }
+    self3->index[i] = index;
+    auto &result = self3->children[i].child;
+    result = nullptr;
+    return result;
   }
   case Type::Node16: {
-    if (self->numChildren == 16) {
+    if (self->numChildren == Node16::kMaxNodes) {
       auto *self16 = static_cast<Node16 *>(self);
       auto *newSelf = allocators->node48.allocate(self->partialKeyLen);
       memcpy((char *)newSelf + kNodeCopyBegin, (char *)self + kNodeCopyBegin,
              kNodeCopySize);
       memcpy(newSelf->partialKey(), self16->partialKey(), self->partialKeyLen);
-      newSelf->nextFree = 16;
+      newSelf->nextFree = Node16::kMaxNodes;
       int i = 0;
       for (auto x : self16->index) {
         newSelf->bitSet.set(x);
@@ -664,7 +708,7 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
         newSelf->index[x] = i;
         ++i;
       }
-      assert(i == 16);
+      assert(i == Node16::kMaxNodes);
       allocators->node16.release(self16);
       setChildrenParents(newSelf);
       self = newSelf;
@@ -673,6 +717,7 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
 
   insert16:
     auto *self16 = static_cast<Node16 *>(self);
+    assert(self->type == Type::Node16);
 
     ++self->numChildren;
     int i = 0;
@@ -773,23 +818,23 @@ void makeCapacityAtLeast(Node *&self, int capacity, NodeAllocators *allocators,
     }
     self = newSelf;
   } break;
-  case Type::Node4: {
-    auto *self4 = (Node4 *)self;
-    auto *newSelf = allocators->node4.allocate(capacity);
+  case Type::Node3: {
+    auto *self3 = (Node3 *)self;
+    auto *newSelf = allocators->node3.allocate(capacity);
     memcpy((char *)newSelf + kNodeCopyBegin, (char *)self + kNodeCopyBegin,
            kNodeCopySize);
-    memcpy(newSelf->partialKey(), self4->partialKey(), self->partialKeyLen);
+    memcpy(newSelf->partialKey(), self3->partialKey(), self->partialKeyLen);
     // TODO replace with memcpy?
-    for (int i = 0; i < 4; ++i) {
-      newSelf->index[i] = self4->index[i];
-      newSelf->children[i] = self4->children[i];
+    for (int i = 0; i < Node3::kMaxNodes; ++i) {
+      newSelf->index[i] = self3->index[i];
+      newSelf->children[i] = self3->children[i];
     }
     getInTree(self, impl) = newSelf;
     setChildrenParents(newSelf);
     if constexpr (kUseFreeList) {
-      allocators->node4.release(self4);
+      allocators->node3.release(self3);
     } else {
-      free(self4);
+      free(self3);
     }
     self = newSelf;
   } break;
@@ -800,7 +845,7 @@ void makeCapacityAtLeast(Node *&self, int capacity, NodeAllocators *allocators,
            kNodeCopySize);
     memcpy(newSelf->partialKey(), self16->partialKey(), self->partialKeyLen);
     // TODO replace with memcpy?
-    for (int i = 0; i < 16; ++i) {
+    for (int i = 0; i < Node16::kMaxNodes; ++i) {
       newSelf->index[i] = self16->index[i];
       newSelf->children[i] = self16->children[i];
     }
@@ -883,20 +928,20 @@ void maybeDownsize(Node *self, NodeAllocators *allocators,
   switch (self->type) {
   case Type::Node0:
     __builtin_unreachable(); // GCOVR_EXCL_LINE
-  case Type::Node4: {
-    auto *self4 = (Node4 *)self;
+  case Type::Node3: {
+    auto *self3 = (Node3 *)self;
     if (self->numChildren == 0) {
       auto *newSelf = allocators->node0.allocate(self->partialKeyLen);
       memcpy((char *)newSelf + kNodeCopyBegin, (char *)self + kNodeCopyBegin,
              kNodeCopySize);
-      memcpy(newSelf->partialKey(), self4->partialKey(), self->partialKeyLen);
+      memcpy(newSelf->partialKey(), self3->partialKey(), self->partialKeyLen);
 
       getInTree(self, impl) = newSelf;
-      allocators->node4.release(self4);
+      allocators->node3.release(self3);
     } else if (self->numChildren == 1) {
       if (!self->entryPresent) {
-        auto *child = self4->children[0].child;
-        int minCapacity = self4->partialKeyLen + 1 + child->partialKeyLen;
+        auto *child = self3->children[0].child;
+        int minCapacity = self3->partialKeyLen + 1 + child->partialKeyLen;
 
         if (minCapacity > child->partialKeyCapacity) {
           const bool update = child == dontInvalidate;
@@ -917,11 +962,11 @@ void maybeDownsize(Node *self, NodeAllocators *allocators,
         int64_t childMaxVersion = maxVersion(child, impl);
 
         // Construct new partial key for child
-        memmove(child->partialKey() + self4->partialKeyLen + 1,
+        memmove(child->partialKey() + self3->partialKeyLen + 1,
                 child->partialKey(), child->partialKeyLen);
-        memcpy(child->partialKey(), self4->partialKey(), self->partialKeyLen);
-        child->partialKey()[self4->partialKeyLen] = self4->index[0];
-        child->partialKeyLen += 1 + self4->partialKeyLen;
+        memcpy(child->partialKey(), self3->partialKey(), self->partialKeyLen);
+        child->partialKey()[self3->partialKeyLen] = self3->index[0];
+        child->partialKeyLen += 1 + self3->partialKeyLen;
 
         child->parent = self->parent;
         child->parentsIndex = self->parentsIndex;
@@ -931,19 +976,19 @@ void maybeDownsize(Node *self, NodeAllocators *allocators,
         maxVersion(child, impl) = childMaxVersion;
 
         getInTree(self, impl) = child;
-        allocators->node4.release(self4);
+        allocators->node3.release(self3);
       }
     }
   } break;
   case Type::Node16:
     if (self->numChildren + int(self->entryPresent) < kMinChildrenNode16) {
       auto *self16 = (Node16 *)self;
-      auto *newSelf = allocators->node4.allocate(self->partialKeyLen);
+      auto *newSelf = allocators->node3.allocate(self->partialKeyLen);
       memcpy((char *)newSelf + kNodeCopyBegin, (char *)self + kNodeCopyBegin,
              kNodeCopySize);
       memcpy(newSelf->partialKey(), self16->partialKey(), self->partialKeyLen);
       // TODO replace with memcpy?
-      for (int i = 0; i < 4; ++i) {
+      for (int i = 0; i < Node3::kMaxNodes; ++i) {
         newSelf->index[i] = self16->index[i];
         newSelf->children[i] = self16->children[i];
       }
@@ -965,7 +1010,7 @@ void maybeDownsize(Node *self, NodeAllocators *allocators,
           [&](int c) {
             // Suppress a false positive -Waggressive-loop-optimizations warning
             // in gcc. `assume` doesn't work for some reason.
-            if (!(i < 16)) {
+            if (!(i < Node16::kMaxNodes)) {
               __builtin_unreachable(); // GCOVR_EXCL_LINE
             }
             newSelf->index[i] = c;
@@ -1034,8 +1079,8 @@ Node *erase(Node *self, NodeAllocators *allocators, ConflictSet::Impl *impl,
   case Type::Node0:
     allocators->node0.release((Node0 *)self);
     break;
-  case Type::Node4:
-    allocators->node4.release((Node4 *)self);
+  case Type::Node3:
+    allocators->node3.release((Node3 *)self);
     break;
   case Type::Node16:
     allocators->node16.release((Node16 *)self);
@@ -1051,8 +1096,17 @@ Node *erase(Node *self, NodeAllocators *allocators, ConflictSet::Impl *impl,
   switch (parent->type) {
   case Type::Node0:
     __builtin_unreachable(); // GCOVR_EXCL_LINE
-  case Type::Node4:
-    [[fallthrough]];
+  case Type::Node3: {
+    auto *parent3 = static_cast<Node3 *>(parent);
+    int nodeIndex = getNodeIndex(parent3, parentsIndex);
+    assert(nodeIndex >= 0);
+    memmove(parent3->index + nodeIndex, parent3->index + nodeIndex + 1,
+            sizeof(parent3->index[0]) *
+                (parent->numChildren - (nodeIndex + 1)));
+    memmove(parent3->children + nodeIndex, parent3->children + nodeIndex + 1,
+            sizeof(parent3->children[0]) *
+                (parent->numChildren - (nodeIndex + 1)));
+  } break;
   case Type::Node16: {
     auto *parent16 = static_cast<Node16 *>(parent);
     int nodeIndex = getNodeIndex(parent16, parentsIndex);
@@ -1399,8 +1453,14 @@ int64_t maxBetweenExclusive(Node *n, int begin, int end) {
   case Type::Node0:
     // We would have returned above, after not finding a child
     __builtin_unreachable(); // GCOVR_EXCL_LINE
-  case Type::Node4:
-    [[fallthrough]];
+  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]) {
+        result = std::max(result, self->children[i].childMaxVersion);
+      }
+    }
+  } break;
   case Type::Node16: {
     auto *self = static_cast<Node16 *>(n);
     for (int i = 0; i < self->numChildren && self->index[i] < end; ++i) {
@@ -1408,8 +1468,7 @@ int64_t maxBetweenExclusive(Node *n, int begin, int end) {
         result = std::max(result, self->children[i].childMaxVersion);
       }
     }
-    break;
-  }
+  } break;
   case Type::Node48: {
     auto *self = static_cast<Node48 *>(n);
     self->bitSet.forEachInRange(
@@ -1898,7 +1957,7 @@ template <bool kBegin>
         int64_t oldMaxVersion = maxVersion(old, impl);
 
         // *self will have one child
-        *self = allocators->node4.allocate(partialKeyIndex);
+        *self = allocators->node3.allocate(partialKeyIndex);
 
         memcpy((char *)*self + kNodeCopyBegin, (char *)old + kNodeCopyBegin,
                kNodeCopySize);
@@ -2261,8 +2320,11 @@ int64_t &maxVersion(Node *n, ConflictSet::Impl *impl) {
   switch (n->type) {
   case Type::Node0:
     __builtin_unreachable(); // GCOVR_EXCL_LINE
-  case Type::Node4:
-    [[fallthrough]];
+  case Type::Node3: {
+    auto *n3 = static_cast<Node3 *>(n);
+    int i = getNodeIndex(n3, index);
+    return n3->children[i].childMaxVersion;
+  }
   case Type::Node16: {
     auto *n16 = static_cast<Node16 *>(n);
     int i = getNodeIndex(n16, index);
@@ -2314,23 +2376,6 @@ ConflictSet::~ConflictSet() {
   }
 }
 
-#if SHOW_MEMORY
-__attribute__((visibility("default"))) void showMemory(const ConflictSet &cs) {
-  ConflictSet::Impl *impl;
-  memcpy(&impl, &cs, sizeof(impl)); // NOLINT
-  fprintf(stderr, "Max Node0 memory usage: %" PRId64 "\n",
-          impl->allocators.node0.highWaterMarkBytes());
-  fprintf(stderr, "Max Node4 memory usage: %" PRId64 "\n",
-          impl->allocators.node4.highWaterMarkBytes());
-  fprintf(stderr, "Max Node16 memory usage: %" PRId64 "\n",
-          impl->allocators.node16.highWaterMarkBytes());
-  fprintf(stderr, "Max Node48 memory usage: %" PRId64 "\n",
-          impl->allocators.node48.highWaterMarkBytes());
-  fprintf(stderr, "Max Node256 memory usage: %" PRId64 "\n",
-          impl->allocators.node256.highWaterMarkBytes());
-}
-#endif
-
 ConflictSet::ConflictSet(ConflictSet &&other) noexcept
     : impl(std::exchange(other.impl, nullptr)) {}
 
@@ -2553,8 +2598,8 @@ Iterator firstGeq(Node *n, std::string_view key) {
   case Type::Node0:
     minNumChildren = 0;
     break;
-  case Type::Node4:
-    minNumChildren = kMinChildrenNode4;
+  case Type::Node3:
+    minNumChildren = kMinChildrenNode3;
     break;
   case Type::Node16:
     minNumChildren = kMinChildrenNode16;