Add cycles_total

If numChildren + entryPresent is enough, we don't have to copy even if
it would fit in a smaller node.

If we have to copy, we might as well use the smallest acceptable node
type.

Bench.cpp

@@ -361,7 +361,21 @@ void benchWorstCaseForRadixRangeRead() {
 void benchCreateAndDestroy() {
   ankerl::nanobench::Bench bench;
 
-  bench.run("create and destroy", [&]() { ConflictSet cs{0}; });
+  bench.run("create and destroy", [&]() {
+    ConflictSet cs{0};
+    ConflictSet::WriteRange w;
+    uint8_t b[9];
+    b[8] = 0;
+    for (int64_t i = 0; i < 1000; i += 7) {
+      auto x = __builtin_bswap64(i);
+      memcpy(b, &x, 8);
+      w.begin.p = b;
+      w.begin.len = 8;
+      w.end.len = 0;
+      w.end.p = b;
+      cs.addWrites(&w, 1, 1);
+    }
+  });
 }
 
 int main(void) {

ConflictSet.cpp

@@ -195,7 +195,6 @@ struct Node {
   /* end section that's copied to the next node */
 
   uint8_t *partialKey();
-  size_t size() const;
 
   Type getType() const { return type; }
   int32_t getCapacity() const { return partialKeyCapacity; }
@@ -224,6 +223,9 @@ struct Node0 : Node {
 
   void copyChildrenAndKeyFrom(const Node0 &other);
   void copyChildrenAndKeyFrom(const struct Node3 &other);
+  void copyChildrenAndKeyFrom(const struct Node16 &other);
+  void copyChildrenAndKeyFrom(const struct Node48 &other);
+  void copyChildrenAndKeyFrom(const struct Node256 &other);
 
   size_t size() const { return sizeof(Node0) + getCapacity(); }
 };
@@ -240,6 +242,8 @@ struct Node3 : Node {
   void copyChildrenAndKeyFrom(const Node0 &other);
   void copyChildrenAndKeyFrom(const Node3 &other);
   void copyChildrenAndKeyFrom(const struct Node16 &other);
+  void copyChildrenAndKeyFrom(const struct Node48 &other);
+  void copyChildrenAndKeyFrom(const struct Node256 &other);
 
   size_t size() const { return sizeof(Node3) + getCapacity(); }
 };
@@ -256,6 +260,7 @@ struct Node16 : Node {
   void copyChildrenAndKeyFrom(const Node3 &other);
   void copyChildrenAndKeyFrom(const Node16 &other);
   void copyChildrenAndKeyFrom(const struct Node48 &other);
+  void copyChildrenAndKeyFrom(const struct Node256 &other);
 
   size_t size() const { return sizeof(Node16) + getCapacity(); }
 };
@@ -313,6 +318,24 @@ inline void Node0::copyChildrenAndKeyFrom(const Node3 &other) {
   memcpy(partialKey(), &other + 1, partialKeyLen);
 }
 
+inline void Node0::copyChildrenAndKeyFrom(const Node16 &other) {
+  memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
+         kNodeCopySize);
+  memcpy(partialKey(), &other + 1, partialKeyLen);
+}
+
+inline void Node0::copyChildrenAndKeyFrom(const Node48 &other) {
+  memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
+         kNodeCopySize);
+  memcpy(partialKey(), &other + 1, partialKeyLen);
+}
+
+inline void Node0::copyChildrenAndKeyFrom(const Node256 &other) {
+  memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
+         kNodeCopySize);
+  memcpy(partialKey(), &other + 1, partialKeyLen);
+}
+
 inline void Node3::copyChildrenAndKeyFrom(const Node0 &other) {
   memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
          kNodeCopySize);
@@ -344,17 +367,53 @@ inline void Node3::copyChildrenAndKeyFrom(const Node16 &other) {
   }
 }
 
+inline void Node3::copyChildrenAndKeyFrom(const Node48 &other) {
+  memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
+         kNodeCopySize);
+  int i = 0;
+  other.bitSet.forEachSet([&](int c) {
+    // Suppress a false positive -Waggressive-loop-optimizations warning
+    // in gcc
+    assume(i < kMaxNodes);
+    index[i] = c;
+    children[i] = other.children[other.index[c]];
+    childMaxVersion[i] = other.childMaxVersion[other.index[c]];
+    assert(children[i]->parent == &other);
+    children[i]->parent = this;
+    ++i;
+  });
+  memcpy(partialKey(), &other + 1, partialKeyLen);
+}
+
+inline void Node3::copyChildrenAndKeyFrom(const Node256 &other) {
+  memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
+         kNodeCopySize);
+  int i = 0;
+  other.bitSet.forEachSet([&](int c) {
+    // Suppress a false positive -Waggressive-loop-optimizations warning
+    // in gcc
+    assume(i < kMaxNodes);
+    index[i] = c;
+    children[i] = other.children[c];
+    childMaxVersion[i] = other.childMaxVersion[c];
+    assert(children[i]->parent == &other);
+    children[i]->parent = this;
+    ++i;
+  });
+  memcpy(partialKey(), &other + 1, partialKeyLen);
+}
+
 inline void Node16::copyChildrenAndKeyFrom(const Node3 &other) {
   memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
          kNodeCopySize);
-  memcpy(index, other.index, Node3::kMaxNodes);
+  memcpy(index, other.index, other.kMaxNodes);
   memcpy(children, other.children,
-         Node3::kMaxNodes * sizeof(children[0])); // NOLINT
+         other.kMaxNodes * sizeof(children[0])); // NOLINT
   memcpy(childMaxVersion, other.childMaxVersion,
-         Node3::kMaxNodes * sizeof(childMaxVersion[0]));
+         other.kMaxNodes * sizeof(childMaxVersion[0]));
   memcpy(partialKey(), &other + 1, partialKeyLen);
-  assert(numChildren == Node3::kMaxNodes);
-  for (int i = 0; i < Node3::kMaxNodes; ++i) {
+  assert(numChildren == other.kMaxNodes);
+  for (int i = 0; i < other.kMaxNodes; ++i) {
     assert(children[i]->parent == &other);
     children[i]->parent = this;
   }
@@ -380,7 +439,7 @@ inline void Node16::copyChildrenAndKeyFrom(const Node48 &other) {
   other.bitSet.forEachSet([&](int c) {
     // Suppress a false positive -Waggressive-loop-optimizations warning
     // in gcc
-    assume(i < Node16::kMaxNodes);
+    assume(i < kMaxNodes);
     index[i] = c;
     children[i] = other.children[other.index[c]];
     childMaxVersion[i] = other.childMaxVersion[other.index[c]];
@@ -391,10 +450,28 @@ inline void Node16::copyChildrenAndKeyFrom(const Node48 &other) {
   memcpy(partialKey(), &other + 1, partialKeyLen);
 }
 
+inline void Node16::copyChildrenAndKeyFrom(const Node256 &other) {
+  memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
+         kNodeCopySize);
+  int i = 0;
+  other.bitSet.forEachSet([&](int c) {
+    // Suppress a false positive -Waggressive-loop-optimizations warning
+    // in gcc
+    assume(i < kMaxNodes);
+    index[i] = c;
+    children[i] = other.children[c];
+    childMaxVersion[i] = other.childMaxVersion[c];
+    assert(children[i]->parent == &other);
+    children[i]->parent = this;
+    ++i;
+  });
+  memcpy(partialKey(), &other + 1, partialKeyLen);
+}
+
 inline void Node48::copyChildrenAndKeyFrom(const Node16 &other) {
   memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
          kNodeCopySize);
-  assert(numChildren == Node16::kMaxNodes);
+  assert(numChildren == other.kMaxNodes);
   memset(index, -1, sizeof(index));
   memset(children, 0, sizeof(children));
   const auto z = InternalVersionT::zero;
@@ -403,7 +480,7 @@ inline void Node48::copyChildrenAndKeyFrom(const Node16 &other) {
   }
   memcpy(partialKey(), &other + 1, partialKeyLen);
   bitSet.init();
-  nextFree = Node16::kMaxNodes;
+  nextFree = other.kMaxNodes;
   int i = 0;
   for (auto x : other.index) {
     bitSet.set(x);
@@ -413,8 +490,8 @@ inline void Node48::copyChildrenAndKeyFrom(const Node16 &other) {
     assert(children[i]->parent == &other);
     children[i]->parent = this;
     reverseIndex[i] = x;
-    maxOfMax[i >> Node48::kMaxOfMaxShift] =
-        std::max(maxOfMax[i >> Node48::kMaxOfMaxShift], childMaxVersion[i]);
+    maxOfMax[i >> kMaxOfMaxShift] =
+        std::max(maxOfMax[i >> kMaxOfMaxShift], childMaxVersion[i]);
     ++i;
   }
 }
@@ -456,15 +533,15 @@ inline void Node48::copyChildrenAndKeyFrom(const Node256 &other) {
   bitSet.forEachSet([&](int c) {
     // Suppress a false positive -Waggressive-loop-optimizations warning
     // in gcc.
-    assume(i < Node48::kMaxNodes);
+    assume(i < kMaxNodes);
     index[c] = i;
     children[i] = other.children[c];
     childMaxVersion[i] = other.childMaxVersion[c];
     assert(children[i]->parent == &other);
     children[i]->parent = this;
     reverseIndex[i] = c;
-    maxOfMax[i >> Node48::kMaxOfMaxShift] =
-        std::max(maxOfMax[i >> Node48::kMaxOfMaxShift], childMaxVersion[i]);
+    maxOfMax[i >> kMaxOfMaxShift] =
+        std::max(maxOfMax[i >> kMaxOfMaxShift], childMaxVersion[i]);
     ++i;
   });
   memcpy(partialKey(), &other + 1, partialKeyLen);
@@ -487,8 +564,8 @@ inline void Node256::copyChildrenAndKeyFrom(const Node48 &other) {
     childMaxVersion[c] = other.childMaxVersion[other.index[c]];
     assert(children[c]->parent == &other);
     children[c]->parent = this;
-    maxOfMax[c >> Node256::kMaxOfMaxShift] =
-        std::max(maxOfMax[c >> Node256::kMaxOfMaxShift], childMaxVersion[c]);
+    maxOfMax[c >> kMaxOfMaxShift] =
+        std::max(maxOfMax[c >> kMaxOfMaxShift], childMaxVersion[c]);
   });
   memcpy(partialKey(), &other + 1, partialKeyLen);
 }
@@ -693,23 +770,6 @@ uint8_t *Node::partialKey() {
   }
 }
 
-size_t Node::size() const {
-  switch (type) {
-  case Type_Node0:
-    return ((Node0 *)this)->size();
-  case Type_Node3:
-    return ((Node3 *)this)->size();
-  case Type_Node16:
-    return ((Node16 *)this)->size();
-  case Type_Node48:
-    return ((Node48 *)this)->size();
-  case Type_Node256:
-    return ((Node256 *)this)->size();
-  default:                   // GCOVR_EXCL_LINE
-    __builtin_unreachable(); // GCOVR_EXCL_LINE
-  }
-}
-
 // A type that's plumbed along the check call tree. Lifetime ends after each
 // check call.
 struct ReadContext {
@@ -1544,6 +1604,46 @@ void rezero(Node *n, InternalVersionT z) {
   }
 }
 
+void mergeWithChild(Node *&self, WriteContext *tls, ConflictSet::Impl *impl,
+                    Node *&dontInvalidate, Node3 *self3) {
+  assert(!self3->entryPresent);
+  auto *child = self3->children[0];
+  int minCapacity = self3->partialKeyLen + 1 + child->partialKeyLen;
+
+  if (minCapacity > child->getCapacity()) {
+    const bool update = child == dontInvalidate;
+    freeAndMakeCapacityAtLeast(child, minCapacity, tls, impl, true);
+    if (update) {
+      dontInvalidate = child;
+    }
+  }
+
+  // Merge partial key with child
+#if DEBUG_VERBOSE && !defined(NDEBUG)
+  fprintf(stderr, "Merge %s into %s\n", getSearchPathPrintable(self).c_str(),
+          getSearchPathPrintable(child).c_str());
+#endif
+
+  InternalVersionT childMaxVersion = self3->childMaxVersion[0];
+
+  // Construct new partial key for child
+  memmove(child->partialKey() + self3->partialKeyLen + 1, child->partialKey(),
+          child->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;
+
+  // Max versions are stored in the parent, so we need to update it now
+  // that we have a new parent.
+  setMaxVersion(child, impl, std::max(childMaxVersion, tls->zero));
+
+  self = child;
+  tls->release(self3);
+}
+
 void maybeDownsize(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
                    Node *&dontInvalidate) {
 
@@ -1557,50 +1657,13 @@ void maybeDownsize(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
   case Type_Node3: {
     auto *self3 = (Node3 *)self;
     if (self->numChildren == 0) {
+      assert(self->entryPresent);
       auto *newSelf = tls->allocate<Node0>(self->partialKeyLen);
       newSelf->copyChildrenAndKeyFrom(*self3);
       getInTree(self, impl) = newSelf;
       tls->release(self3);
     } else if (self->numChildren == 1 && !self->entryPresent) {
-      auto *child = self3->children[0];
-      int minCapacity = self3->partialKeyLen + 1 + child->partialKeyLen;
-
-      if (minCapacity > child->getCapacity()) {
-        const bool update = child == dontInvalidate;
-        freeAndMakeCapacityAtLeast(child, minCapacity, tls, impl, true);
-        if (update) {
-          dontInvalidate = child;
-        }
-      }
-
-      // Merge partial key with child
-#if DEBUG_VERBOSE && !defined(NDEBUG)
-      fprintf(stderr, "Merge %s into %s\n",
-              getSearchPathPrintable(self).c_str(),
-              getSearchPathPrintable(child).c_str());
-#endif
-
-      InternalVersionT childMaxVersion = maxVersion(child, impl);
-
-      // Construct new partial key for child
-      memmove(child->partialKey() + self3->partialKeyLen + 1,
-              child->partialKey(), child->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;
-
-      // Max versions are stored in the parent, so we need to update it now
-      // that we have a new parent.
-      setMaxVersion(child, impl, childMaxVersion);
-      if (child->parent) {
-        rezero(child->parent, tls->zero);
-      }
-
-      getInTree(self, impl) = child;
-      tls->release(self3);
+      mergeWithChild(getInTree(self, impl), tls, impl, dontInvalidate, self3);
     }
   } break;
   case Type_Node16:
@@ -1635,10 +1698,266 @@ void maybeDownsize(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
   }
 }
 
+void eraseTree(Node *root, WriteContext *tls) {
+  Arena arena;
+  auto toFree = vector<Node *>(arena);
+  toFree.push_back(root);
+
+  while (toFree.size() > 0) {
+    auto *n = toFree.back();
+    toFree.pop_back();
+    tls->accum.entries_erased += n->entryPresent;
+    ++tls->accum.nodes_released;
+
+    removeNode(n);
+    removeKey(n);
+
+    switch (n->getType()) {
+    case Type_Node0: {
+      auto *n0 = static_cast<Node0 *>(n);
+      tls->release(n0);
+    } break;
+    case Type_Node3: {
+      auto *n3 = static_cast<Node3 *>(n);
+      toFree.append(std::span<Node *>(n3->children, n3->numChildren));
+      tls->release(n3);
+    } break;
+    case Type_Node16: {
+      auto *n16 = static_cast<Node16 *>(n);
+      toFree.append(std::span<Node *>(n16->children, n16->numChildren));
+      tls->release(n16);
+    } break;
+    case Type_Node48: {
+      auto *n48 = static_cast<Node48 *>(n);
+      toFree.append(std::span<Node *>(n48->children, n48->numChildren));
+      tls->release(n48);
+    } break;
+    case Type_Node256: {
+      auto *n256 = static_cast<Node256 *>(n);
+      auto *out = toFree.unsafePrepareAppend(n256->numChildren).data();
+      n256->bitSet.forEachSet([&](int i) { *out++ = n256->children[i]; });
+      assert(out == toFree.end());
+      tls->release(n256);
+    } break;
+    default:                   // GCOVR_EXCL_LINE
+      __builtin_unreachable(); // GCOVR_EXCL_LINE
+    }
+  }
+}
+
+void eraseBetween(Node **inTree, Node3 *n, int begin, int end,
+                  WriteContext *tls) {
+  const unsigned shiftUpperBound = end - begin;
+  const unsigned shiftAmount = begin;
+  auto inBounds = [&](unsigned c) { return c - shiftAmount < shiftUpperBound; };
+  Node **nodeOut = n->children;
+  uint8_t *indexOut = n->index;
+  InternalVersionT *maxVOut = n->childMaxVersion;
+  for (int i = 0; i < n->numChildren; ++i) {
+    if (inBounds(n->index[i])) {
+      eraseTree(n->children[i], tls);
+    } else {
+      *nodeOut++ = n->children[i];
+      *indexOut++ = n->index[i];
+      *maxVOut++ = n->childMaxVersion[i];
+    }
+  }
+  n->numChildren = nodeOut - n->children;
+
+  if (n->numChildren == 0) {
+    auto *newNode = tls->allocate<Node0>(n->partialKeyLen);
+    newNode->copyChildrenAndKeyFrom(*n);
+    tls->release(n);
+    *inTree = newNode;
+  }
+}
+
+void eraseBetween(Node **inTree, Node16 *n, int begin, int end,
+                  WriteContext *tls) {
+  if (end - begin == 256) {
+    for (int i = 0; i < n->numChildren; ++i) {
+      eraseTree(n->children[i], tls);
+    }
+    n->numChildren = 0;
+    auto *newNode = tls->allocate<Node0>(n->partialKeyLen);
+    newNode->copyChildrenAndKeyFrom(*n);
+    tls->release(n);
+    *inTree = newNode;
+    return;
+  }
+  assert(end - begin < 256);
+
+#ifdef HAS_ARM_NEON
+  uint8x16_t indices;
+  memcpy(&indices, n->index, 16);
+  // 0xff for each in bounds
+  auto results =
+      vcltq_u8(vsubq_u8(indices, vdupq_n_u8(begin)), vdupq_n_u8(end - begin));
+  // 0xf for each 0xff
+  uint64_t mask = vget_lane_u64(
+      vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(results), 4)), 0);
+#elif defined(HAS_AVX)
+  __m128i indices;
+  memcpy(&indices, n->index, 16);
+  indices = _mm_sub_epi8(indices, _mm_set1_epi8(begin));
+  uint32_t mask = ~_mm_movemask_epi8(_mm_cmpeq_epi8(
+      indices, _mm_max_epu8(indices, _mm_set1_epi8(end - begin))));
+#else
+  const unsigned shiftUpperBound = end - begin;
+  const unsigned shiftAmount = begin;
+  auto inBounds = [&](unsigned c) { return c - shiftAmount < shiftUpperBound; };
+  uint32_t mask = 0;
+  for (int i = 0; i < 16; ++i) {
+    mask |= inBounds(is[i]) << i;
+  }
+#endif
+  mask &= (decltype(mask)(1) << n->numChildren) - 1;
+
+  if (!mask) {
+    return;
+  }
+
+  int first = std::countr_zero(mask);
+  int count = std::popcount(mask);
+  n->numChildren -= count;
+  for (int i = first; i < first + count; ++i) {
+    eraseTree(n->children[i], tls);
+  }
+  for (int i = first; i < n->numChildren; ++i) {
+    n->children[i] = n->children[i + count];
+    n->childMaxVersion[i] = n->childMaxVersion[i + count];
+    n->index[i] = n->index[i + count];
+  }
+
+  if (n->numChildren + n->entryPresent >= kMinChildrenNode16) {
+    // nop
+  } else if (n->numChildren > 0) {
+    auto *newNode = tls->allocate<Node3>(n->partialKeyLen);
+    newNode->copyChildrenAndKeyFrom(*n);
+    tls->release(n);
+    *inTree = newNode;
+  } else {
+    auto *newNode = tls->allocate<Node0>(n->partialKeyLen);
+    newNode->copyChildrenAndKeyFrom(*n);
+    tls->release(n);
+    *inTree = newNode;
+  }
+}
+
+void eraseBetween(Node **inTree, Node48 *n, int begin, int end,
+                  WriteContext *tls) {
+  for (int i = n->bitSet.firstSetGeq(begin); i >= 0 && i < end;
+       i = n->bitSet.firstSetGeq(i)) {
+    n->bitSet.reset(i);
+    int8_t toRemoveChildrenIndex = std::exchange(n->index[i], -1);
+    int8_t lastChildrenIndex = --n->nextFree;
+    assert(toRemoveChildrenIndex >= 0);
+    assert(lastChildrenIndex >= 0);
+    eraseTree(n->children[toRemoveChildrenIndex], tls);
+    if (toRemoveChildrenIndex != lastChildrenIndex) {
+      n->children[toRemoveChildrenIndex] = n->children[lastChildrenIndex];
+      n->childMaxVersion[toRemoveChildrenIndex] =
+          n->childMaxVersion[lastChildrenIndex];
+      n->maxOfMax[toRemoveChildrenIndex >> Node48::kMaxOfMaxShift] =
+          std::max(n->maxOfMax[toRemoveChildrenIndex >> Node48::kMaxOfMaxShift],
+                   n->childMaxVersion[toRemoveChildrenIndex]);
+      auto parentIndex = n->children[toRemoveChildrenIndex]->parentsIndex;
+      n->index[parentIndex] = toRemoveChildrenIndex;
+      n->reverseIndex[toRemoveChildrenIndex] = parentIndex;
+    }
+    n->childMaxVersion[lastChildrenIndex] = tls->zero;
+    --n->numChildren;
+  }
+
+  if (n->numChildren + n->entryPresent >= kMinChildrenNode48) {
+    // nop
+  } else if (n->numChildren > Node3::kMaxNodes) {
+    auto *newNode = tls->allocate<Node16>(n->partialKeyLen);
+    newNode->copyChildrenAndKeyFrom(*n);
+    tls->release(n);
+    *inTree = newNode;
+  } else if (n->numChildren > 0) {
+    auto *newNode = tls->allocate<Node3>(n->partialKeyLen);
+    newNode->copyChildrenAndKeyFrom(*n);
+    tls->release(n);
+    *inTree = newNode;
+  } else {
+    auto *newNode = tls->allocate<Node0>(n->partialKeyLen);
+    newNode->copyChildrenAndKeyFrom(*n);
+    tls->release(n);
+    *inTree = newNode;
+  }
+}
+
+void eraseBetween(Node **inTree, Node256 *n, int begin, int end,
+                  WriteContext *tls) {
+  for (int i = n->bitSet.firstSetGeq(begin); i >= 0 && i < end;
+       i = n->bitSet.firstSetGeq(i)) {
+    assert(n->children[i] != nullptr);
+    eraseTree(std::exchange(n->children[i], nullptr), tls);
+    n->bitSet.reset(i);
+    --n->numChildren;
+  }
+  if (n->numChildren + n->entryPresent >= kMinChildrenNode256) {
+    // nop
+  } else if (n->numChildren > Node16::kMaxNodes) {
+    auto *newNode = tls->allocate<Node48>(n->partialKeyLen);
+    newNode->copyChildrenAndKeyFrom(*n);
+    tls->release(n);
+    *inTree = newNode;
+  } else if (n->numChildren > Node3::kMaxNodes) {
+    auto *newNode = tls->allocate<Node16>(n->partialKeyLen);
+    newNode->copyChildrenAndKeyFrom(*n);
+    tls->release(n);
+    *inTree = newNode;
+  } else if (n->numChildren > 0) {
+    auto *newNode = tls->allocate<Node3>(n->partialKeyLen);
+    newNode->copyChildrenAndKeyFrom(*n);
+    tls->release(n);
+    *inTree = newNode;
+  } else {
+    auto *newNode = tls->allocate<Node0>(n->partialKeyLen);
+    newNode->copyChildrenAndKeyFrom(*n);
+    tls->release(n);
+    *inTree = newNode;
+  }
+}
+
+// Erase all nodes with a search path starting with n + [child],
+// where child in [begin, end). To avoid the need to propagate erases up the
+// search path, the caller must ensure that the result has at least one child or
+// has entryPresent.
+void eraseBetween(Node *&n, int begin, int end, WriteContext *tls) {
+#if DEBUG_VERBOSE && !defined(NDEBUG)
+  fprintf(stderr, "eraseBetween: %s + [%d,%d)\n",
+          getSearchPathPrintable(n).c_str(), begin, end);
+#endif
+  switch (n->getType()) {
+  case Type_Node0:
+    break;
+  case Type_Node3:
+    eraseBetween(&n, (Node3 *)n, begin, end, tls);
+    break;
+  case Type_Node16:
+    eraseBetween(&n, (Node16 *)n, begin, end, tls);
+    break;
+  case Type_Node48:
+    eraseBetween(&n, (Node48 *)n, begin, end, tls);
+    break;
+  case Type_Node256:
+    eraseBetween(&n, (Node256 *)n, begin, end, tls);
+    break;
+  default:                   // GCOVR_EXCL_LINE
+    __builtin_unreachable(); // GCOVR_EXCL_LINE
+  }
+  assert(n->numChildren > 0 || n->entryPresent);
+}
+
 // Precondition: self is not the root. May invalidate nodes along the search
 // path to self. May invalidate children of self->parent. Returns a pointer to
-// the node after self. If erase invalidates the pointee of `dontInvalidate`, it
-// will update it to its new pointee as well. Precondition: `self->entryPresent`
+// the node after self. If erase invalidates the pointee of `dontInvalidate`,
+// it will update it to its new pointee as well. Precondition:
+// `self->entryPresent`
 Node *erase(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
             bool logical, Node *&dontInvalidate) {
   ++tls->accum.entries_erased;
@@ -1892,8 +2211,8 @@ bool checkPrefixRead(Node *n, const std::span<const uint8_t> key,
         remaining = remaining.subspan(commonLen, remaining.size() - commonLen);
       } else if (n->partialKeyLen > int(remaining.size())) {
         // n is the first physical node greater than remaining, and there's no
-        // eq node. All physical nodes that start with prefix are reachable from
-        // n.
+        // eq node. All physical nodes that start with prefix are reachable
+        // from n.
         if (maxVersion(n, impl) > readVersion) {
           return false;
         }
@@ -1942,8 +2261,8 @@ compare16_32bit_avx512(const InternalVersionT *vs, InternalVersionT rv) {
 #endif
 // GCOVR_EXCL_STOP
 
-// Returns true if v[i] <= readVersion for all i such that begin <= is[i] < end
-// Preconditions: begin <= end, end - begin < 256
+// Returns true if v[i] <= readVersion for all i such that begin <= is[i] <
+// end Preconditions: begin <= end, end - begin < 256
 template <bool kAVX512>
 bool scan16(const InternalVersionT *vs, const uint8_t *is, int begin, int end,
             InternalVersionT readVersion) {
@@ -2074,9 +2393,10 @@ bool scan16(const InternalVersionT *vs, int begin, int end,
 #endif
 }
 
-// Return whether or not the max version among all keys starting with the search
-// path of n + [child], where child in (begin, end) is <= readVersion. Does not
-// account for the range version of firstGt(searchpath(n) + [end - 1])
+// Return whether or not the max version among all keys starting with the
+// search path of n + [child], where child in (begin, end) is <= readVersion.
+// Does not account for the range version of firstGt(searchpath(n) + [end -
+// 1])
 template <bool kAVX512>
 bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
                                   InternalVersionT readVersion,
@@ -2245,8 +2565,8 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
       } else {
         return true;
       }
-      // [begin, end) is now the half-open interval of children we're interested
-      // in.
+      // [begin, end) is now the half-open interval of children we're
+      // interested in.
       assert(begin < end);
     }
 
@@ -2281,8 +2601,8 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
       } else {
         return true;
       }
-      // [begin, end) is now the half-open interval of children we're interested
-      // in.
+      // [begin, end) is now the half-open interval of children we're
+      // interested in.
       assert(begin < end);
     }
 
@@ -2339,8 +2659,9 @@ checkMaxBetweenExclusive(Node *n, int begin, int end,
 // GCOVR_EXCL_STOP
 __attribute__((target("default")))
 #endif
-bool checkMaxBetweenExclusive(Node *n, int begin, int end,
-                         InternalVersionT readVersion, ReadContext *tls) {
+  bool
+  checkMaxBetweenExclusive(Node * n, int begin, int end,
+                           InternalVersionT readVersion, ReadContext *tls) {
   return checkMaxBetweenExclusiveImpl<false>(n, begin, end, readVersion, tls);
 }
 
@@ -2361,8 +2682,8 @@ Vector<uint8_t> getSearchPath(Arena &arena, Node *n) {
   return result;
 } // GCOVR_EXCL_LINE
 
-// Return true if the max version among all keys that start with key + [child],
-// where begin < child < end, is <= readVersion.
+// Return true if the max version among all keys that start with key +
+// [child], where begin < child < end, is <= readVersion.
 //
 // Precondition: transitively, no child of n has a search path that's a longer
 // prefix of key than n
@@ -2432,8 +2753,8 @@ downLeftSpine:
 }
 
 namespace {
-// Return true if the max version among all keys that start with key[:prefixLen]
-// that are >= key is <= readVersion
+// Return true if the max version among all keys that start with
+// key[:prefixLen] that are >= key is <= readVersion
 struct CheckRangeLeftSide {
   CheckRangeLeftSide(Node *n, std::span<const uint8_t> key, int prefixLen,
                      InternalVersionT readVersion, ReadContext *tls)
@@ -2547,8 +2868,8 @@ struct CheckRangeLeftSide {
   }
 };
 
-// Return true if the max version among all keys that start with key[:prefixLen]
-// that are < key is <= readVersion
+// Return true if the max version among all keys that start with
+// key[:prefixLen] that are < key is <= readVersion
 struct CheckRangeRightSide {
   CheckRangeRightSide(Node *n, std::span<const uint8_t> key, int prefixLen,
                       InternalVersionT readVersion, ReadContext *tls)
@@ -2794,9 +3115,9 @@ checkMaxBetweenExclusiveImpl<true>(Node *n, int begin, int end,
                                    InternalVersionT readVersion, ReadContext *);
 #endif
 
-// Returns a pointer the pointer to the newly inserted node in the tree.  Caller
-// must set `entryPresent`, and `entry` fields.  All nodes along the search path
-// of the result will have `maxVersion` set to `writeVersion` as a
+// Returns a pointer the pointer to the newly inserted node in the tree.
+// Caller must set `entryPresent`, and `entry` fields.  All nodes along the
+// search path of the result will have `maxVersion` set to `writeVersion` as a
 // postcondition. Nodes along the search path may be invalidated.
 [[nodiscard]]
 Node **insert(Node **self, std::span<const uint8_t> key,
@@ -2812,31 +3133,6 @@ Node **insert(Node **self, std::span<const uint8_t> key,
   return self;
 }
 
-void destroyTree(Node *root) {
-  Arena arena;
-  auto toFree = vector<Node *>(arena);
-  toFree.push_back(root);
-
-#if SHOW_MEMORY
-  for (auto *iter = root; iter != nullptr; iter = nextPhysical(iter)) {
-    removeNode(iter);
-    removeKey(iter);
-  }
-#endif
-
-  while (toFree.size() > 0) {
-    auto *n = toFree.back();
-    toFree.pop_back();
-    // Add all children to toFree
-    for (auto c = getChildGeq(n, 0); c != nullptr;
-         c = getChildGeq(n, c->parentsIndex + 1)) {
-      assert(c != nullptr);
-      toFree.push_back(c);
-    }
-    safe_free(n, n->size());
-  }
-}
-
 void addPointWrite(Node *&root, std::span<const uint8_t> key,
                    InternalVersionT writeVersion, WriteContext *tls,
                    ConflictSet::Impl *impl) {
@@ -2911,55 +3207,93 @@ void addWriteRange(Node *&root, std::span<const uint8_t> begin,
 
   Node **useAsRoot =
       insert(&root, begin.subspan(0, lcp), writeVersion, tls, impl);
+  begin = begin.subspan(lcp, begin.size() - lcp);
+  end = end.subspan(lcp, end.size() - lcp);
 
-  int consumed = lcp;
-
-  begin = begin.subspan(consumed, begin.size() - consumed);
-  end = end.subspan(consumed, end.size() - consumed);
-
-  auto *beginNode = *insert(useAsRoot, begin, writeVersion, tls, impl);
-
-  const bool insertedBegin = !beginNode->entryPresent;
-
-  addKey(beginNode);
-  beginNode->entryPresent = true;
-
-  if (insertedBegin) {
-    ++tls->accum.entries_inserted;
-    auto *p = nextLogical(beginNode);
-    beginNode->entry.rangeVersion =
-        p == nullptr ? tls->zero : std::max(p->entry.rangeVersion, tls->zero);
+  if (beginIsPrefix) {
+    auto *beginNode = *useAsRoot;
+    addKey(beginNode);
+    if (!beginNode->entryPresent) {
+      ++tls->accum.entries_inserted;
+      auto *p = nextLogical(beginNode);
+      beginNode->entry.rangeVersion =
+          p == nullptr ? tls->zero : std::max(p->entry.rangeVersion, tls->zero);
+      beginNode->entryPresent = true;
+    }
     beginNode->entry.pointVersion = writeVersion;
-  }
-  assert(writeVersion >= beginNode->entry.pointVersion);
-  beginNode->entry.pointVersion = writeVersion;
 
-  auto *endNode = *insert(useAsRoot, end, writeVersion, tls, impl);
-
-  const bool insertedEnd = !endNode->entryPresent;
-
-  addKey(endNode);
-  endNode->entryPresent = true;
-
-  if (insertedEnd) {
-    ++tls->accum.entries_inserted;
-    auto *p = nextLogical(endNode);
-    endNode->entry.pointVersion =
-        p == nullptr ? tls->zero : std::max(p->entry.rangeVersion, tls->zero);
-    if (beginIsPrefix) {
+    auto *endNode = *insert(useAsRoot, end, writeVersion, tls, impl);
+    addKey(endNode);
+    if (!endNode->entryPresent) {
+      ++tls->accum.entries_inserted;
+      auto *p = nextLogical(endNode);
+      endNode->entry.pointVersion =
+          p == nullptr ? tls->zero : std::max(p->entry.rangeVersion, tls->zero);
       // beginNode may have been invalidated when inserting end
       beginNode = *useAsRoot;
-      assert(beginNode->entryPresent);
+      endNode->entryPresent = true;
     }
-  }
-  endNode->entry.rangeVersion = writeVersion;
+    endNode->entry.rangeVersion = writeVersion;
 
-  for (beginNode = nextLogical(beginNode); beginNode != endNode;
-       beginNode = erase(beginNode, tls, impl, /*logical*/ true, endNode)) {
-  }
+    for (beginNode = nextLogical(beginNode); beginNode != endNode;
+         beginNode = erase(beginNode, tls, impl, /*logical*/ true, endNode)) {
+    }
 
-  // Inserting end trashed endNode's maxVersion. Fix that
-  fixupMaxVersion(endNode, impl, tls);
+    // Inserting end trashed endNode's maxVersion. Fix that
+    fixupMaxVersion(endNode, impl, tls);
+  } else /*!beginIsPrefix*/ {
+    auto *beginNode = *insert(useAsRoot, begin, writeVersion, tls, impl);
+    addKey(beginNode);
+    if (!beginNode->entryPresent) {
+      ++tls->accum.entries_inserted;
+      auto *p = nextLogical(beginNode);
+      beginNode->entry.rangeVersion =
+          p == nullptr ? tls->zero : std::max(p->entry.rangeVersion, tls->zero);
+      beginNode->entryPresent = true;
+    }
+    beginNode->entry.pointVersion = writeVersion;
+
+    auto *endNode = *insert(useAsRoot, end, writeVersion, tls, impl);
+    addKey(endNode);
+    if (!endNode->entryPresent) {
+      ++tls->accum.entries_inserted;
+      auto *p = nextLogical(endNode);
+      endNode->entry.pointVersion =
+          p == nullptr ? tls->zero : std::max(p->entry.rangeVersion, tls->zero);
+      endNode->entryPresent = true;
+    }
+    endNode->entry.rangeVersion = writeVersion;
+
+    eraseBetween(*useAsRoot, begin[0] + 1, end[0], tls);
+
+    // // Erase along left spine
+    // for (auto [n, key] = std::make_tuple(useAsRoot, begin);;) {
+    //   auto before = key;
+    //   auto **child = &getOrCreateChild(*n, key, writeVersion, tls);
+    //   if (key.size() > 0) {
+    //     eraseBetween(*child, int(key[0]) + 1, 256, tls);
+    //     if ((*child)->numChildren == 1 && !(*child)->entryPresent) {
+    //       Node *dummy = nullptr;
+    //       mergeWithChild(*child, tls, impl, dummy,
+    //                      static_cast<Node3 *>(*child));
+    //       key = before;
+    //       continue;
+    //     }
+    //   } else {
+    //     eraseBetween(*child, 0, 256, tls);
+    //     beginNode = *child;
+    //     break;
+    //   }
+    //   n = child;
+    // }
+
+    for (beginNode = nextLogical(beginNode); beginNode != endNode;
+         beginNode = erase(beginNode, tls, impl, /*logical*/ true, endNode)) {
+    }
+
+    // Inserting end trashed endNode's maxVersion. Fix that
+    fixupMaxVersion(endNode, impl, tls);
+  }
 }
 
 Node *firstGeqPhysical(Node *n, const std::span<const uint8_t> key) {
@@ -3070,7 +3404,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     if (oldestExtantVersion < writeVersion - kMaxCorrectVersionWindow)
         [[unlikely]] {
       if (writeVersion > newestVersionFullPrecision + kNominalVersionWindow) {
-        destroyTree(root);
+        eraseTree(root, &tls);
         init(writeVersion - kNominalVersionWindow);
       }
 
@@ -3133,17 +3467,17 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     int64_t set_oldest_iterations_accum = 0;
     for (; fuel > 0 && n != nullptr; ++set_oldest_iterations_accum) {
       rezero(n, oldestVersion);
-      // The "make sure gc keeps up with writes" calculations assume that we're
-      // scanning key by key, not node by node. Make sure we only spend fuel
-      // when there's a logical entry.
+      // The "make sure gc keeps up with writes" calculations assume that
+      // we're scanning key by key, not node by node. Make sure we only spend
+      // fuel when there's a logical entry.
       fuel -= n->entryPresent;
       if (n->entryPresent && std::max(n->entry.pointVersion,
                                       n->entry.rangeVersion) <= oldestVersion) {
         // Any transaction n would have prevented from committing is
         // going to fail with TooOld anyway.
 
-        // There's no way to insert a range such that range version of the right
-        // node is greater than the point version of the left node
+        // There's no way to insert a range such that range version of the
+        // right node is greater than the point version of the left node
         assert(n->entry.rangeVersion <= oldestVersion);
         Node *dummy = nullptr;
         n = erase(n, &tls, this, /*logical*/ false, dummy);
@@ -3180,9 +3514,9 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     this->oldestVersion = oldestVersion;
     InternalVersionT::zero = tls.zero = oldestVersion;
 #ifdef NDEBUG
-    // This is here for performance reasons, since we want to amortize the cost
-    // of storing the search path as a string. In tests, we want to exercise the
-    // rest of the code often.
+    // This is here for performance reasons, since we want to amortize the
+    // cost of storing the search path as a string. In tests, we want to
+    // exercise the rest of the code often.
     if (keyUpdates < 100) {
       return;
     }
@@ -3238,7 +3572,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     initMetrics();
   }
   ~Impl() {
-    destroyTree(root);
+    eraseTree(root, &tls);
     safe_free(metrics, metricsCount * sizeof(metrics[0]));
   }
 
@@ -3306,8 +3640,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
           "Total number of checks where the result is \"too old\"");
   COUNTER(check_bytes_total, "Total number of key bytes checked");
   COUNTER(point_writes_total, "Total number of point writes");
-  COUNTER(range_writes_total,
-          "Total number of range writes (includes prefix writes)");
+  COUNTER(range_writes_total, "Total number of range writes");
   GAUGE(memory_bytes, "Total number of bytes in use");
   COUNTER(nodes_allocated_total,
           "The total number of physical tree nodes allocated");
@@ -3321,16 +3654,14 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
           "The total number of entries inserted in the tree");
   COUNTER(entries_erased_total,
           "The total number of entries erased from the tree");
-  COUNTER(
-      gc_iterations_total,
-      "The total number of iterations of the main loop for garbage collection");
+  COUNTER(gc_iterations_total, "The total number of iterations of the main "
+                               "loop for garbage collection");
   COUNTER(write_bytes_total, "Total number of key bytes in calls to addWrites");
   GAUGE(oldest_version,
         "The lowest version that doesn't result in \"TooOld\" for checks");
   GAUGE(newest_version, "The version of the most recent call to addWrites");
-  GAUGE(
-      oldest_extant_version,
-      "A lower bound on the lowest version associated with an existing entry");
+  GAUGE(oldest_extant_version, "A lower bound on the lowest version "
+                               "associated with an existing entry");
   // ==================== END METRICS DEFINITIONS ====================
 #undef GAUGE
 #undef COUNTER
@@ -3416,6 +3747,7 @@ InternalVersionT exchangeMaxVersion(Node *n, InternalVersionT newMax) {
 }
 
 void setMaxVersion(Node *n, ConflictSet::Impl *impl, InternalVersionT newMax) {
+  assert(newMax >= InternalVersionT::zero);
   int index = n->parentsIndex;
   n = n->parent;
   if (n == nullptr) {
@@ -3705,8 +4037,8 @@ std::string strinc(std::string_view str, bool &ok) {
     if ((uint8_t &)(str[index]) != 255)
       break;
 
-  // Must not be called with a string that consists only of zero or more '\xff'
-  // bytes.
+  // Must not be called with a string that consists only of zero or more
+  // '\xff' bytes.
   if (index < 0) {
     ok = false;
     return {};
@@ -3926,7 +4258,7 @@ checkMaxVersion(Node *root, Node *node, InternalVersionT oldestVersion,
   bool success = true;
 
   if (node->partialKeyLen > 0) {
-    fprintf(stderr, "Root cannot have a partial key");
+    fprintf(stderr, "Root cannot have a partial key\n");
     success = false;
   }
   checkParentPointers(node, success);

Internal.h

@@ -273,6 +273,16 @@ template <class T> struct Vector {
     size_ += slice.size();
   }
 
+  // Caller must write to the returned slice
+  std::span<T> unsafePrepareAppend(int appendSize) {
+    if (size_ + appendSize > capacity) {
+      grow(std::max<int>(size_ + appendSize, capacity * 2));
+    }
+    auto result = std::span<T>(t + size_, appendSize);
+    size_ += appendSize;
+    return result;
+  }
+
   void push_back(const T &t) { append(std::span<const T>(&t, 1)); }
 
   T *begin() { return t; }

ServerBench.cpp

@@ -164,6 +164,63 @@ double toSeconds(timeval t) {
   return double(t.tv_sec) + double(t.tv_usec) * 1e-6;
 }
 
+#include <linux/perf_event.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/ioctl.h>
+#include <sys/syscall.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#ifdef __linux__
+struct PerfCounter {
+  explicit PerfCounter(int event) {
+    struct perf_event_attr pe;
+
+    memset(&pe, 0, sizeof(pe));
+    pe.type = PERF_TYPE_HARDWARE;
+    pe.size = sizeof(pe);
+    pe.config = event;
+    pe.inherit = 1;
+    pe.exclude_kernel = 1;
+    pe.exclude_hv = 1;
+
+    fd = perf_event_open(&pe, 0, -1, -1, 0);
+    if (fd == -1) {
+      fprintf(stderr, "Error opening leader %llx\n", pe.config);
+      exit(EXIT_FAILURE);
+    }
+  }
+
+  int64_t total() {
+    int64_t count;
+    if (read(fd, &count, sizeof(count)) != sizeof(count)) {
+      perror("read instructions from perf");
+      abort();
+    }
+    return count;
+  }
+
+  ~PerfCounter() { close(fd); }
+
+private:
+  int fd;
+  static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
+                              int cpu, int group_fd, unsigned long flags) {
+    int ret;
+
+    ret = syscall(SYS_perf_event_open, hw_event, pid, cpu, group_fd, flags);
+    return ret;
+  }
+};
+#else
+struct PerfCounter {
+  explicit PerPerfCounter(int) {}
+  int64_t total() { return 0; }
+};
+#endif
+
 int main(int argc, char **argv) {
   if (argc != 3) {
     goto fail;
@@ -176,6 +233,8 @@ int main(int argc, char **argv) {
     int metricsCount;
     cs.getMetricsV1(&metrics, &metricsCount);
 
+    PerfCounter instructions{PERF_COUNT_HW_INSTRUCTIONS};
+    PerfCounter cycles{PERF_COUNT_HW_CPU_CYCLES};
     auto w = std::thread{workload, &cs};
 
     for (;;) {
@@ -203,6 +262,16 @@ int main(int argc, char **argv) {
               "transactions_total ";
       body += std::to_string(transactions.load(std::memory_order_relaxed));
       body += "\n";
+      body += "# HELP instructions_total Total number of instructions\n"
+              "# TYPE instructions_total counter\n"
+              "instructions_total ";
+      body += std::to_string(instructions.total());
+      body += "\n";
+      body += "# HELP cycles_total Total number of cycles\n"
+              "# TYPE cycles_total counter\n"
+              "cycles_total ";
+      body += std::to_string(cycles.total());
+      body += "\n";
 
       for (int i = 0; i < metricsCount; ++i) {
         body += "# HELP ";