Vectorize all bounds checks for Node256 scan

This seems to benchmark better

Bench.cpp

@@ -258,4 +258,107 @@ void benchConflictSet() {
   }
 }
 
-int main(void) { benchConflictSet(); }
+constexpr int kKeyLenForWorstCase = 50;
+
+ConflictSet worstCaseConflictSetForRadixRangeRead(int cardinality) {
+  ConflictSet cs{0};
+
+  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(255 - j);
+      weaselab::ConflictSet::WriteRange w[] = {{
+                                                   {b.data(), int(b.size())},
+                                                   {nullptr, 0},
+                                               },
+                                               {
+                                                   {e.data(), int(e.size())},
+                                                   {nullptr, 0},
+                                               }};
+      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>(kKeyLenForWorstCase, 0);
+    weaselab::ConflictSet::WriteRange w[] = {
+        {
+            {k.data(), int(k.size())},
+            {nullptr, 0},
+        },
+    };
+    cs.addWrites(w, sizeof(w) / sizeof(w[0]), 1);
+  }
+
+  // Defeat short-circuiting on the right
+  {
+    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]), 1);
+  }
+
+  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>(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())}, 0};
+
+  bench.run("worst case for radix tree", [&]() {
+    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) {
+  benchConflictSet();
+  benchWorstCaseForRadixRangeRead();
+}

CMakeLists.txt

@@ -1,14 +1,16 @@
 cmake_minimum_required(VERSION 3.18)
 project(
   conflict-set
-  VERSION 0.0.5
+  VERSION 0.0.7
   DESCRIPTION
     "A data structure for optimistic concurrency control on ranges of bitwise-lexicographically-ordered keys."
   HOMEPAGE_URL "https://git.weaselab.dev/weaselab/conflict-set"
   LANGUAGES C CXX)
 set(CMAKE_CXX_STANDARD 20)
 
-file(WRITE ${CMAKE_BINARY_DIR}/version.txt ${PROJECT_VERSION})
+file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/version.txt ${PROJECT_VERSION})
+configure_file(${CMAKE_CURRENT_SOURCE_DIR}/version.txt.in
+               ${CMAKE_CURRENT_SOURCE_DIR}/paper/version.txt)
 
 include(CMakePushCheckState)
 include(CheckCXXCompilerFlag)
@@ -47,7 +49,8 @@ if(HAS_FULL_RELRO)
 endif()
 cmake_pop_check_state()
 
-set(version_script_flags LINKER:--version-script=${CMAKE_SOURCE_DIR}/linker.map)
+set(version_script_flags
+    LINKER:--version-script=${CMAKE_CURRENT_SOURCE_DIR}/linker.map)
 cmake_push_check_state()
 list(APPEND CMAKE_REQUIRED_LINK_OPTIONS ${version_script_flags})
 check_cxx_source_compiles("int main(){}" HAS_VERSION_SCRIPT FAIL_REGEX
@@ -59,7 +62,7 @@ option(USE_SIMD_FALLBACK
 
 # This is encouraged according to
 # https://valgrind.org/docs/manual/manual-core-adv.html#manual-core-adv.clientreq
-include_directories(SYSTEM ${CMAKE_SOURCE_DIR}/third_party/valgrind)
+include_directories(SYSTEM ${CMAKE_CURRENT_SOURCE_DIR}/third_party/valgrind)
 
 add_compile_options($<$<COMPILE_LANGUAGE:CXX>:-Wno-invalid-offsetof>)
 
@@ -106,19 +109,20 @@ add_library(${PROJECT_NAME}-object OBJECT ConflictSet.cpp)
 target_compile_options(${PROJECT_NAME}-object PRIVATE -fno-exceptions
                                                       -fvisibility=hidden)
 target_include_directories(${PROJECT_NAME}-object
-                           PRIVATE ${CMAKE_SOURCE_DIR}/include)
+                           PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include)
 
 add_library(${PROJECT_NAME} SHARED $<TARGET_OBJECTS:${PROJECT_NAME}-object>)
 set_target_properties(
   ${PROJECT_NAME} PROPERTIES LIBRARY_OUTPUT_DIRECTORY
-                             "${CMAKE_BINARY_DIR}/radix_tree")
+                             "${CMAKE_CURRENT_BINARY_DIR}/radix_tree")
 if(NOT CMAKE_BUILD_TYPE STREQUAL Debug)
   set_target_properties(${PROJECT_NAME} PROPERTIES LINKER_LANGUAGE C)
 endif()
 
 if(HAS_VERSION_SCRIPT)
-  target_link_options(${PROJECT_NAME} PRIVATE
-                      LINKER:--version-script=${CMAKE_SOURCE_DIR}/linker.map)
+  target_link_options(
+    ${PROJECT_NAME} PRIVATE
+    LINKER:--version-script=${CMAKE_CURRENT_SOURCE_DIR}/linker.map)
 endif()
 
 add_library(${PROJECT_NAME}-static STATIC
@@ -131,7 +135,7 @@ if(APPLE)
   add_custom_command(
     TARGET ${PROJECT_NAME}-static
     PRE_LINK
-    COMMAND ${CMAKE_SOURCE_DIR}/privatize_symbols_macos.sh
+    COMMAND ${CMAKE_CURRENT_SOURCE_DIR}/privatize_symbols_macos.sh
             $<TARGET_OBJECTS:${PROJECT_NAME}-object>)
 else()
   add_custom_command(
@@ -139,21 +143,22 @@ else()
     POST_BUILD
     COMMAND
       ${CMAKE_OBJCOPY}
-      --keep-global-symbols=${CMAKE_SOURCE_DIR}/symbol-exports.txt
+      --keep-global-symbols=${CMAKE_CURRENT_SOURCE_DIR}/symbol-exports.txt
       $<TARGET_FILE:${PROJECT_NAME}-static> || echo
       "Proceeding with all symbols global in static library")
 endif()
 
-set(TEST_FLAGS -Wall -Wextra -Wunreachable-code -Wpedantic -UNDEBUG)
-
 include(CTest)
 
-if(BUILD_TESTING)
+# disable tests if this is being used through e.g. FetchContent
+if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
+
+  set(TEST_FLAGS -Wall -Wextra -Wunreachable-code -Wpedantic -UNDEBUG)
 
   # corpus tests, which are tests curated by libfuzzer. The goal is to get broad
   # coverage with a small number of tests.
 
-  file(GLOB CORPUS_TESTS ${CMAKE_SOURCE_DIR}/corpus/*)
+  file(GLOB CORPUS_TESTS ${CMAKE_CURRENT_SOURCE_DIR}/corpus/*)
 
   # extra testing that relies on shared libraries, which aren't available with
   # wasm
@@ -162,9 +167,11 @@ if(BUILD_TESTING)
     add_library(skip_list SHARED SkipList.cpp)
     target_compile_options(skip_list PRIVATE -fno-exceptions
                                              -fvisibility=hidden)
-    target_include_directories(skip_list PUBLIC ${CMAKE_SOURCE_DIR}/include)
-    set_target_properties(skip_list PROPERTIES LIBRARY_OUTPUT_DIRECTORY
-                                               "${CMAKE_BINARY_DIR}/skip_list")
+    target_include_directories(skip_list
+                               PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
+    set_target_properties(
+      skip_list PROPERTIES LIBRARY_OUTPUT_DIRECTORY
+                           "${CMAKE_CURRENT_BINARY_DIR}/skip_list")
     set_target_properties(skip_list PROPERTIES OUTPUT_NAME ${PROJECT_NAME})
     set_target_properties(
       skip_list PROPERTIES VERSION ${PROJECT_VERSION} SOVERSION
@@ -175,10 +182,11 @@ if(BUILD_TESTING)
     add_library(hash_table SHARED HashTable.cpp)
     target_compile_options(hash_table PRIVATE -fno-exceptions
                                               -fvisibility=hidden)
-    target_include_directories(hash_table PUBLIC ${CMAKE_SOURCE_DIR}/include)
+    target_include_directories(hash_table
+                               PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
     set_target_properties(
       hash_table PROPERTIES LIBRARY_OUTPUT_DIRECTORY
-                            "${CMAKE_BINARY_DIR}/hash_table")
+                            "${CMAKE_CURRENT_BINARY_DIR}/hash_table")
     set_target_properties(hash_table PROPERTIES OUTPUT_NAME ${PROJECT_NAME})
     set_target_properties(
       hash_table PROPERTIES VERSION ${PROJECT_VERSION} SOVERSION
@@ -266,15 +274,19 @@ if(BUILD_TESTING)
     set_property(
       DIRECTORY
       APPEND
-      PROPERTY CMAKE_CONFIGURE_DEPENDS ${CMAKE_SOURCE_DIR}/test_conflict_set.py)
+      PROPERTY CMAKE_CONFIGURE_DEPENDS
+               ${CMAKE_CURRENT_SOURCE_DIR}/test_conflict_set.py)
     execute_process(
-      COMMAND ${Python3_EXECUTABLE} ${CMAKE_SOURCE_DIR}/test_conflict_set.py
-              list OUTPUT_VARIABLE SCRIPT_TESTS)
+      COMMAND ${Python3_EXECUTABLE}
+              ${CMAKE_CURRENT_SOURCE_DIR}/test_conflict_set.py list
+      OUTPUT_VARIABLE SCRIPT_TESTS)
     foreach(TEST ${SCRIPT_TESTS})
       add_test(
         NAME script_test_${TEST}
-        COMMAND ${Python3_EXECUTABLE} ${CMAKE_SOURCE_DIR}/test_conflict_set.py
-                test ${TEST} --build-dir ${CMAKE_BINARY_DIR})
+        COMMAND
+          ${Python3_EXECUTABLE}
+          ${CMAKE_CURRENT_SOURCE_DIR}/test_conflict_set.py test ${TEST}
+          --build-dir ${CMAKE_CURRENT_BINARY_DIR})
     endforeach()
   endif()
 
@@ -308,25 +320,26 @@ if(BUILD_TESTING)
   # symbol visibility tests
   if(NOT WASM AND NOT CMAKE_BUILD_TYPE STREQUAL Debug)
     if(APPLE)
-      set(symbol_exports ${CMAKE_SOURCE_DIR}/apple-symbol-exports.txt)
-      set(symbol_imports ${CMAKE_SOURCE_DIR}/apple-symbol-imports.txt)
+      set(symbol_exports ${CMAKE_CURRENT_SOURCE_DIR}/apple-symbol-exports.txt)
+      set(symbol_imports ${CMAKE_CURRENT_SOURCE_DIR}/apple-symbol-imports.txt)
     else()
-      set(symbol_exports ${CMAKE_SOURCE_DIR}/symbol-exports.txt)
+      set(symbol_exports ${CMAKE_CURRENT_SOURCE_DIR}/symbol-exports.txt)
       if(CMAKE_SYSTEM_PROCESSOR STREQUAL aarch64)
-        set(symbol_imports ${CMAKE_SOURCE_DIR}/aarch64-symbol-imports.txt)
+        set(symbol_imports
+            ${CMAKE_CURRENT_SOURCE_DIR}/aarch64-symbol-imports.txt)
       else()
-        set(symbol_imports ${CMAKE_SOURCE_DIR}/symbol-imports.txt)
+        set(symbol_imports ${CMAKE_CURRENT_SOURCE_DIR}/symbol-imports.txt)
       endif()
     endif()
     add_test(
       NAME conflict_set_shared_symbols
       COMMAND
-        ${CMAKE_SOURCE_DIR}/test_symbols.sh $<TARGET_FILE:${PROJECT_NAME}>
-        ${symbol_exports} ${symbol_imports})
+        ${CMAKE_CURRENT_SOURCE_DIR}/test_symbols.sh
+        $<TARGET_FILE:${PROJECT_NAME}> ${symbol_exports} ${symbol_imports})
     add_test(
       NAME conflict_set_static_symbols
       COMMAND
-        ${CMAKE_SOURCE_DIR}/test_symbols.sh
+        ${CMAKE_CURRENT_SOURCE_DIR}/test_symbols.sh
         $<TARGET_FILE:${PROJECT_NAME}-static> ${symbol_exports}
         ${symbol_imports})
   endif()
@@ -369,13 +382,13 @@ set(CMAKE_OSX_DEPLOYMENT_TARGET 11.0)
 if(APPLE)
   find_program(PANDOC_EXE pandoc)
   if(PANDOC_EXE)
-    execute_process(COMMAND ${PANDOC_EXE} ${CMAKE_SOURCE_DIR}/README.md -o
-                            ${CMAKE_BINARY_DIR}/README.txt)
-    set(CPACK_RESOURCE_FILE_README ${CMAKE_BINARY_DIR}/README.txt)
+    execute_process(COMMAND ${PANDOC_EXE} ${CMAKE_CURRENT_SOURCE_DIR}/README.md
+                            -o ${CMAKE_CURRENT_BINARY_DIR}/README.txt)
+    set(CPACK_RESOURCE_FILE_README ${CMAKE_CURRENT_BINARY_DIR}/README.txt)
   endif()
-  configure_file(${CMAKE_SOURCE_DIR}/LICENSE ${CMAKE_BINARY_DIR}/LICENSE.txt
-                 COPYONLY)
-  set(CPACK_RESOURCE_FILE_LICENSE ${CMAKE_BINARY_DIR}/LICENSE.txt)
+  configure_file(${CMAKE_CURRENT_SOURCE_DIR}/LICENSE
+                 ${CMAKE_CURRENT_BINARY_DIR}/LICENSE.txt COPYONLY)
+  set(CPACK_RESOURCE_FILE_LICENSE ${CMAKE_CURRENT_BINARY_DIR}/LICENSE.txt)
 endif()
 
 include(CPack)

ConflictSet.cpp

@@ -236,11 +236,6 @@ constexpr int kNodeCopyBegin = offsetof(Node, entry);
 constexpr int kNodeCopySize =
     offsetof(Node, parentsIndex) + sizeof(Node::parentsIndex) - kNodeCopyBegin;
 
-struct Child {
-  int64_t childMaxVersion;
-  Node *child;
-};
-
 // copyChildrenAndKeyFrom is responsible for copying all
 // public members of Node, copying the partial key, logically copying the
 // children (converting representation if necessary), and updating all the
@@ -262,7 +257,8 @@ struct Node3 : Node {
   constexpr static auto kType = Type_Node3;
   // Sorted
   uint8_t index[kMaxNodes];
-  Child children[kMaxNodes];
+  Node *children[kMaxNodes];
+  int64_t childMaxVersion[kMaxNodes];
   uint8_t *partialKey() { return (uint8_t *)(this + 1); }
 
   void copyChildrenAndKeyFrom(const Node0 &other);
@@ -277,7 +273,8 @@ struct Node16 : Node {
   constexpr static auto kMaxNodes = 16;
   // Sorted
   uint8_t index[kMaxNodes];
-  Child children[kMaxNodes];
+  Node *children[kMaxNodes];
+  int64_t childMaxVersion[kMaxNodes];
   uint8_t *partialKey() { return (uint8_t *)(this + 1); }
 
   void copyChildrenAndKeyFrom(const Node3 &other);
@@ -293,7 +290,14 @@ struct Node48 : Node {
   BitSet bitSet;
   int8_t nextFree;
   int8_t index[256];
-  Child children[kMaxNodes];
+  Node *children[kMaxNodes];
+  int64_t childMaxVersion[kMaxNodes];
+  uint8_t reverseIndex[kMaxNodes];
+  constexpr static int kMaxOfMaxPageSize = 16;
+  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); }
 
@@ -307,7 +311,13 @@ struct Node48 : Node {
 struct Node256 : Node {
   constexpr static auto kType = Type_Node256;
   BitSet bitSet;
-  Child children[256];
+  Node *children[256];
+  int64_t childMaxVersion[256];
+  constexpr static int kMaxOfMaxPageSize = 16;
+  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);
@@ -339,8 +349,8 @@ inline void Node3::copyChildrenAndKeyFrom(const Node3 &other) {
   memcpy(index, other.index, sizeof(*this) - sizeof(Node));
   memcpy(partialKey(), &other + 1, partialKeyLen);
   for (int i = 0; i < numChildren; ++i) {
-    assert(children[i].child->parent == &other);
-    children[i].child->parent = this;
+    assert(children[i]->parent == &other);
+    children[i]->parent = this;
   }
 }
 
@@ -348,11 +358,13 @@ inline void Node3::copyChildrenAndKeyFrom(const Node16 &other) {
   memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
          kNodeCopySize);
   memcpy(index, other.index, kMaxNodes);
-  memcpy(children, other.children, kMaxNodes * sizeof(Child));
+  memcpy(children, other.children, kMaxNodes * sizeof(children[0])); // NOLINT
+  memcpy(childMaxVersion, other.childMaxVersion,
+         kMaxNodes * sizeof(childMaxVersion[0]));
   memcpy(partialKey(), &other + 1, partialKeyLen);
   for (int i = 0; i < numChildren; ++i) {
-    assert(children[i].child->parent == &other);
-    children[i].child->parent = this;
+    assert(children[i]->parent == &other);
+    children[i]->parent = this;
   }
 }
 
@@ -360,12 +372,15 @@ inline void Node16::copyChildrenAndKeyFrom(const Node3 &other) {
   memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
          kNodeCopySize);
   memcpy(index, other.index, Node3::kMaxNodes);
-  memcpy(children, other.children, Node3::kMaxNodes * sizeof(Child));
+  memcpy(children, other.children,
+         Node3::kMaxNodes * sizeof(children[0])); // NOLINT
+  memcpy(childMaxVersion, other.childMaxVersion,
+         Node3::kMaxNodes * sizeof(childMaxVersion[0]));
   memcpy(partialKey(), &other + 1, partialKeyLen);
   assert(numChildren == Node3::kMaxNodes);
   for (int i = 0; i < Node3::kMaxNodes; ++i) {
-    assert(children[i].child->parent == &other);
-    children[i].child->parent = this;
+    assert(children[i]->parent == &other);
+    children[i]->parent = this;
   }
 }
 
@@ -375,8 +390,9 @@ inline void Node16::copyChildrenAndKeyFrom(const Node16 &other) {
   memcpy(index, other.index, sizeof(index));
   for (int i = 0; i < numChildren; ++i) {
     children[i] = other.children[i];
-    assert(children[i].child->parent == &other);
-    children[i].child->parent = this;
+    childMaxVersion[i] = other.childMaxVersion[i];
+    assert(children[i]->parent == &other);
+    children[i]->parent = this;
   }
   memcpy(partialKey(), &other + 1, partialKeyLen);
 }
@@ -392,8 +408,9 @@ inline void Node16::copyChildrenAndKeyFrom(const Node48 &other) {
         assume(i < Node16::kMaxNodes);
         index[i] = c;
         children[i] = other.children[other.index[c]];
-        assert(children[i].child->parent == &other);
-        children[i].child->parent = this;
+        childMaxVersion[i] = other.childMaxVersion[other.index[c]];
+        assert(children[i]->parent == &other);
+        children[i]->parent = this;
         ++i;
       },
       0, 256);
@@ -405,6 +422,8 @@ inline void Node48::copyChildrenAndKeyFrom(const Node16 &other) {
          kNodeCopySize);
   assert(numChildren == Node16::kMaxNodes);
   memset(index, -1, sizeof(index));
+  memset(children, 0, sizeof(children));
+  memset(childMaxVersion, 0, sizeof(childMaxVersion));
   memcpy(partialKey(), &other + 1, partialKeyLen);
   bitSet.init();
   nextFree = Node16::kMaxNodes;
@@ -413,8 +432,12 @@ inline void Node48::copyChildrenAndKeyFrom(const Node16 &other) {
     bitSet.set(x);
     index[x] = i;
     children[i] = other.children[i];
-    assert(children[i].child->parent == &other);
-    children[i].child->parent = this;
+    childMaxVersion[i] = other.childMaxVersion[i];
+    assert(children[i]->parent == &other);
+    children[i]->parent = this;
+    reverseIndex[i] = x;
+    maxOfMax[i >> Node48::kMaxOfMaxShift] =
+        std::max(maxOfMax[i >> Node48::kMaxOfMaxShift], childMaxVersion[i]);
     ++i;
   }
 }
@@ -422,13 +445,19 @@ 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));
+  memset(childMaxVersion, 0, sizeof(childMaxVersion));
   for (int i = 0; i < numChildren; ++i) {
     children[i] = other.children[i];
-    assert(children[i].child->parent == &other);
-    children[i].child->parent = this;
+    childMaxVersion[i] = other.childMaxVersion[i];
+    assert(children[i]->parent == &other);
+    children[i]->parent = this;
   }
+  memcpy(reverseIndex, other.reverseIndex, sizeof(reverseIndex));
+  memcpy(maxOfMax, other.maxOfMax, sizeof(maxOfMax));
   memcpy(partialKey(), &other + 1, partialKeyLen);
 }
 
@@ -436,6 +465,8 @@ 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));
+  memset(childMaxVersion, 0, sizeof(childMaxVersion));
   nextFree = other.numChildren;
   bitSet = other.bitSet;
   int i = 0;
@@ -446,8 +477,12 @@ inline void Node48::copyChildrenAndKeyFrom(const Node256 &other) {
         assume(i < Node48::kMaxNodes);
         index[c] = i;
         children[i] = other.children[c];
-        assert(children[i].child->parent == &other);
-        children[i].child->parent = this;
+        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]);
         ++i;
       },
       0, 256);
@@ -457,13 +492,18 @@ 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(childMaxVersion, 0, sizeof(childMaxVersion));
+  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;
+        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]);
       },
       0, 256);
   memcpy(partialKey(), &other + 1, partialKeyLen);
@@ -473,14 +513,17 @@ inline void Node256::copyChildrenAndKeyFrom(const Node256 &other) {
   memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
          kNodeCopySize);
   memset(children, 0, sizeof(children));
+  memset(childMaxVersion, 0, sizeof(childMaxVersion));
   bitSet = other.bitSet;
   bitSet.forEachInRange(
       [&](int c) {
         children[c] = other.children[c];
-        assert(children[c].child->parent == &other);
-        children[c].child->parent = this;
+        childMaxVersion[c] = other.childMaxVersion[c];
+        assert(children[c]->parent == &other);
+        children[c]->parent = this;
       },
       0, 256);
+  memcpy(maxOfMax, other.maxOfMax, sizeof(maxOfMax));
   memcpy(partialKey(), &other + 1, partialKeyLen);
 }
 
@@ -532,7 +575,7 @@ 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 +603,18 @@ 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));
+      memset(result->childMaxVersion, 0, sizeof(result->childMaxVersion));
+    }
+    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);
@@ -711,29 +766,30 @@ Node *&getChildExists(Node *self, uint8_t index) {
     __builtin_unreachable(); // GCOVR_EXCL_LINE
   case Type_Node3: {
     auto *self3 = static_cast<Node3 *>(self);
-    return self3->children[getNodeIndex(self3, index)].child;
+    return self3->children[getNodeIndex(self3, index)];
   }
   case Type_Node16: {
     auto *self16 = static_cast<Node16 *>(self);
-    return self16->children[getNodeIndex(self16, index)].child;
+    return self16->children[getNodeIndex(self16, index)];
   }
   case Type_Node48: {
     auto *self48 = static_cast<Node48 *>(self);
     assert(self48->bitSet.test(index));
-    return self48->children[self48->index[index]].child;
+    return self48->children[self48->index[index]];
   }
   case Type_Node256: {
     auto *self256 = static_cast<Node256 *>(self);
     assert(self256->bitSet.test(index));
-    return self256->children[index].child;
+    return self256->children[index];
   }
   default:                   // GCOVR_EXCL_LINE
     __builtin_unreachable(); // GCOVR_EXCL_LINE
   }
 }
 
-// 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 *);
 
@@ -744,21 +800,21 @@ Node *getChild(Node *self, uint8_t index) {
   case Type_Node3: {
     auto *self3 = static_cast<Node3 *>(self);
     int i = getNodeIndex(self3, index);
-    return i < 0 ? nullptr : self3->children[i].child;
+    return i < 0 ? nullptr : self3->children[i];
   }
   case Type_Node16: {
     auto *self16 = static_cast<Node16 *>(self);
     int i = getNodeIndex(self16, index);
-    return i < 0 ? nullptr : self16->children[i].child;
+    return i < 0 ? nullptr : self16->children[i];
   }
   case Type_Node48: {
     auto *self48 = static_cast<Node48 *>(self);
     int i = self48->index[index];
-    return i < 0 ? nullptr : self48->children[i].child;
+    return i < 0 ? nullptr : self48->children[i];
   }
   case Type_Node256: {
     auto *self256 = static_cast<Node256 *>(self);
-    return self256->children[index].child;
+    return self256->children[index];
   }
   default:                   // GCOVR_EXCL_LINE
     __builtin_unreachable(); // GCOVR_EXCL_LINE
@@ -874,26 +930,26 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
     auto *self3 = static_cast<Node3 *>(self);
     int i = getNodeIndex(self3, index);
     if (i >= 0) {
-      return self3->children[i].child;
+      return self3->children[i];
     }
   } break;
   case Type_Node16: {
     auto *self16 = static_cast<Node16 *>(self);
     int i = getNodeIndex(self16, index);
     if (i >= 0) {
-      return self16->children[i].child;
+      return self16->children[i];
     }
   } break;
   case Type_Node48: {
     auto *self48 = static_cast<Node48 *>(self);
     int secondIndex = self48->index[index];
     if (secondIndex >= 0) {
-      return self48->children[secondIndex].child;
+      return self48->children[secondIndex];
     }
   } break;
   case Type_Node256: {
     auto *self256 = static_cast<Node256 *>(self);
-    if (auto &result = self256->children[index].child; result != nullptr) {
+    if (auto &result = self256->children[index]; result != nullptr) {
       return result;
     }
   } break;
@@ -931,12 +987,16 @@ Node *&getOrCreateChild(Node *&self, uint8_t 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));
+                (self->numChildren - (i + 1)) *
+                    sizeof(self3->children[0])); // NOLINT
+        memmove(self3->childMaxVersion + i + 1, self3->childMaxVersion + i,
+                (self->numChildren - (i + 1)) *
+                    sizeof(self3->childMaxVersion[0]));
         break;
       }
     }
     self3->index[i] = index;
-    auto &result = self3->children[i].child;
+    auto &result = self3->children[i];
     result = nullptr;
     return result;
   }
@@ -968,7 +1028,10 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
       memmove(self16->index + i + 1, self16->index + i,
               self->numChildren - (i + 1));
       memmove(self16->children + i + 1, self16->children + i,
-              (self->numChildren - (i + 1)) * sizeof(Child));
+              (self->numChildren - (i + 1)) * sizeof(self16->children[0]));
+      memmove(self16->childMaxVersion + i + 1, self16->childMaxVersion + i,
+              (self->numChildren - (i + 1)) *
+                  sizeof(self16->childMaxVersion[0]));
     }
 #elif defined(HAS_ARM_NEON)
     uint8x16_t indices;
@@ -988,7 +1051,11 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
       memmove(self16->index + i + 1, self16->index + i,
               self->numChildren - (i + 1));
       memmove(self16->children + i + 1, self16->children + i,
-              (self->numChildren - (i + 1)) * sizeof(Child));
+              (self->numChildren - (i + 1)) *
+                  sizeof(self16->children[0])); // NOLINT
+      memmove(self16->childMaxVersion + i + 1, self16->childMaxVersion + i,
+              (self->numChildren - (i + 1)) *
+                  sizeof(self16->childMaxVersion[0]));
     }
 #else
     int i = 0;
@@ -997,13 +1064,16 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
         memmove(self16->index + i + 1, self16->index + i,
                 self->numChildren - (i + 1));
         memmove(self16->children + i + 1, self16->children + i,
-                (self->numChildren - (i + 1)) * sizeof(Child));
+                (self->numChildren - (i + 1)) * sizeof(self16->children[0]));
+        memmove(self16->childMaxVersion + i + 1, self16->childMaxVersion + i,
+                (self->numChildren - (i + 1)) *
+                    sizeof(self16->childMaxVersion[0]));
         break;
       }
     }
 #endif
     self16->index[i] = index;
-    auto &result = self16->children[i].child;
+    auto &result = self16->children[i];
     result = nullptr;
     return result;
   }
@@ -1025,7 +1095,8 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
     assert(self48->nextFree < 48);
     int nextFree = self48->nextFree++;
     self48->index[index] = nextFree;
-    auto &result = self48->children[nextFree].child;
+    self48->reverseIndex[nextFree] = index;
+    auto &result = self48->children[nextFree];
     result = nullptr;
     return result;
   }
@@ -1035,7 +1106,7 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
     auto *self256 = static_cast<Node256 *>(self);
     ++self->numChildren;
     self256->bitSet.set(index);
-    return self256->children[index].child;
+    return self256->children[index];
   }
   default:                   // GCOVR_EXCL_LINE
     __builtin_unreachable(); // GCOVR_EXCL_LINE
@@ -1178,7 +1249,7 @@ void maybeDownsize(Node *self, NodeAllocators *allocators,
       getInTree(self, impl) = newSelf;
       allocators->node3.release(self3);
     } else if (self->numChildren == 1 && !self->entryPresent) {
-      auto *child = self3->children[0].child;
+      auto *child = self3->children[0];
       int minCapacity = self3->partialKeyLen + 1 + child->partialKeyLen;
 
       if (minCapacity > child->getCapacity()) {
@@ -1210,7 +1281,7 @@ 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);
@@ -1291,7 +1362,11 @@ Node *erase(Node *self, NodeAllocators *allocators, ConflictSet::Impl *impl,
             sizeof(parent3->index[0]) *
                 (parent->numChildren - (nodeIndex + 1)));
     memmove(parent3->children + nodeIndex, parent3->children + nodeIndex + 1,
-            sizeof(parent3->children[0]) *
+            sizeof(parent3->children[0]) * // NOLINT
+                (parent->numChildren - (nodeIndex + 1)));
+    memmove(parent3->childMaxVersion + nodeIndex,
+            parent3->childMaxVersion + nodeIndex + 1,
+            sizeof(parent3->childMaxVersion[0]) *
                 (parent->numChildren - (nodeIndex + 1)));
 
     --parent->numChildren;
@@ -1305,7 +1380,11 @@ Node *erase(Node *self, NodeAllocators *allocators, ConflictSet::Impl *impl,
             sizeof(parent16->index[0]) *
                 (parent->numChildren - (nodeIndex + 1)));
     memmove(parent16->children + nodeIndex, parent16->children + nodeIndex + 1,
-            sizeof(parent16->children[0]) *
+            sizeof(parent16->children[0]) * // NOLINT
+                (parent->numChildren - (nodeIndex + 1)));
+    memmove(parent16->childMaxVersion + nodeIndex,
+            parent16->childMaxVersion + nodeIndex + 1,
+            sizeof(parent16->childMaxVersion[0]) *
                 (parent->numChildren - (nodeIndex + 1)));
 
     --parent->numChildren;
@@ -1324,9 +1403,18 @@ 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;
+      parent48->childMaxVersion[toRemoveChildrenIndex] =
+          parent48->childMaxVersion[lastChildrenIndex];
+      parent48->maxOfMax[toRemoveChildrenIndex >> Node48::kMaxOfMaxShift] =
+          std::max(parent48->maxOfMax[toRemoveChildrenIndex >>
+                                      Node48::kMaxOfMaxShift],
+                   parent48->childMaxVersion[toRemoveChildrenIndex]);
+      auto parentIndex =
+          parent48->children[toRemoveChildrenIndex]->parentsIndex;
+      parent48->index[parentIndex] = toRemoveChildrenIndex;
+      parent48->reverseIndex[toRemoveChildrenIndex] = parentIndex;
     }
+    parent48->childMaxVersion[lastChildrenIndex] = 0;
 
     --parent->numChildren;
 
@@ -1336,7 +1424,7 @@ Node *erase(Node *self, NodeAllocators *allocators, ConflictSet::Impl *impl,
   case Type_Node256: {
     auto *parent256 = static_cast<Node256 *>(parent);
     parent256->bitSet.reset(parentsIndex);
-    parent256->children[parentsIndex].child = nullptr;
+    parent256->children[parentsIndex] = nullptr;
 
     --parent->numChildren;
 
@@ -1621,75 +1709,183 @@ downLeftSpine:
   }
 }
 
-// Return the max version among all keys starting with the search path of n +
-// [child], where child in (begin, end). Does not account for the range version
-// of firstGt(searchpath(n) + [end - 1])
-int64_t maxBetweenExclusive(Node *n, int begin, int end) {
+// Returns true if all in-bounds vs are <= readVersion
+bool scan16(const int64_t *vs, const uint8_t *is, int begin, int end,
+            int64_t readVersion) {
+
+#ifdef HAS_ARM_NEON
+
+  assert(end - begin < 256);
+  uint8x16_t indices;
+  memcpy(&indices, is, 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);
+
+  uint64_t compared = 0;
+  for (int i = 0; i < 16; ++i) {
+    compared |= (uint64_t(vs[i] > readVersion) << (i << 2));
+  }
+  return !(compared & mask);
+
+#else
+
+  const unsigned shiftUpperBound = end - begin;
+  const unsigned shiftAmount = begin;
+  auto inBounds = [&](unsigned c) { return c - shiftAmount < shiftUpperBound; };
+
+  uint32_t compared = 0;
+  for (int i = 0; i < 16; ++i) {
+    compared |= ((vs[i] > readVersion) << i);
+  }
+  uint32_t mask = 0;
+  for (int i = 0; i < 16; ++i) {
+    mask |= inBounds(is[i]) << i;
+  }
+  return !(compared & mask);
+
+#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])
+bool checkMaxBetweenExclusive(Node *n, int begin, int end,
+                              int64_t readVersion) {
   assume(-1 <= begin);
   assume(begin <= 256);
   assume(-1 <= end);
   assume(end <= 256);
   assume(begin < end);
-  int64_t result = std::numeric_limits<int64_t>::lowest();
+
+  assert(!(begin == -1 && end == 256));
+
   {
     int c = getChildGeq(n, begin + 1);
     if (c >= 0 && c < end) {
       auto *child = getChildExists(n, c);
       if (child->entryPresent) {
-        result = std::max(result, child->entry.rangeVersion);
+        if (!(child->entry.rangeVersion <= readVersion)) {
+          return false;
+        };
       }
       begin = c;
     } else {
-      return result;
+      return true;
     }
   }
+
+  // [begin, end) is now the half-open interval of children we're interested in.
+  assert(begin < end);
+
+  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]) {
-        result = std::max(result, self->children[i].childMaxVersion);
-      }
+    bool result = true;
+    for (int i = 0; i < 3; ++i) {
+      result &= !((self->childMaxVersion[i] > 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]) {
-        result = std::max(result, self->children[i].childMaxVersion);
-      }
-    }
+
+    return scan16(self->childMaxVersion, self->index, begin, end, readVersion);
+
   } break;
   case Type_Node48: {
     auto *self = static_cast<Node48 *>(n);
-    self->bitSet.forEachInRange(
-        [&](int i) {
-          result =
-              std::max(result, self->children[self->index[i]].childMaxVersion);
-        },
-        begin, end);
-    break;
+    // Check all pages
+    static_assert(Node48::kMaxOfMaxPageSize == 16);
+    for (int i = 0; i < Node48::kMaxOfMaxTotalPages; ++i) {
+      if (self->maxOfMax[i] > readVersion) {
+        if (!scan16(self->childMaxVersion + (i << Node48::kMaxOfMaxShift),
+                    self->reverseIndex + (i << Node48::kMaxOfMaxShift), begin,
+                    end, readVersion)) {
+          return false;
+        }
+      }
+    }
+    return true;
   }
   case Type_Node256: {
     auto *self = static_cast<Node256 *>(n);
-    self->bitSet.forEachInRange(
-        [&](int i) {
-          result = std::max(result, self->children[i].childMaxVersion);
-        },
-        begin, end);
-    break;
+    if (end <= 0) {
+      return true;
+    }
+    const int firstPage = begin >> Node256::kMaxOfMaxShift;
+    const int lastPage = (end - 1) >> Node256::kMaxOfMaxShift;
+    if (firstPage == lastPage) {
+      if (self->maxOfMax[firstPage] <= readVersion) {
+        return true;
+      }
+      uint64_t conflict = 0;
+      // Check all in page
+      for (int i = 0; i < Node256::kMaxOfMaxPageSize; ++i) {
+        conflict |=
+            (self->childMaxVersion[(firstPage << Node256::kMaxOfMaxShift) + i] >
+             readVersion)
+            << i;
+      }
+      // Mask away out of bounds
+      const int intraPageBegin = begin & (Node256::kMaxOfMaxPageSize - 1);
+      const int intraPageEnd = end - (lastPage << Node256::kMaxOfMaxShift);
+      conflict &= (1 << intraPageEnd) - 1;
+      conflict >>= intraPageBegin;
+      return !conflict;
+    }
+    // Check the first page
+    if (self->maxOfMax[firstPage] > readVersion) {
+      uint64_t conflict = 0;
+      for (int i = 0; i < Node256::kMaxOfMaxPageSize; ++i) {
+        int j = (firstPage << Node256::kMaxOfMaxShift) + i;
+        conflict |= (self->childMaxVersion[j] > readVersion) << i;
+      }
+      const int intraPageBegin = begin & (Node256::kMaxOfMaxPageSize - 1);
+      conflict >>= intraPageBegin;
+      if (conflict) {
+        return false;
+      }
+    }
+    // Check the last page
+    if (self->maxOfMax[lastPage] > readVersion) {
+      uint64_t conflict = 0;
+      for (int i = 0; i < Node256::kMaxOfMaxPageSize; ++i) {
+        int j = (lastPage << Node256::kMaxOfMaxShift) + i;
+        conflict |= (self->childMaxVersion[j] > readVersion) << i;
+      }
+      const int intraPageEnd = end - (lastPage << Node256::kMaxOfMaxShift);
+      conflict &= (1 << intraPageEnd) - 1;
+      if (conflict) {
+        return false;
+      }
+    }
+    uint64_t conflict = 0;
+    // Check all possible inner pages
+    for (int i = 1; i < Node256::kMaxOfMaxTotalPages - 1; ++i) {
+      conflict |= (self->maxOfMax[i] > readVersion) << i;
+    }
+    // Only keep inner pages
+    const int innerPageBegin = (begin >> Node256::kMaxOfMaxShift) + 1;
+    const int innerPageEnd = (end - 1) >> Node256::kMaxOfMaxShift;
+    conflict &= (1 << innerPageEnd) - 1;
+    conflict >>= innerPageBegin;
+    return !conflict;
   }
   default:                   // GCOVR_EXCL_LINE
     __builtin_unreachable(); // GCOVR_EXCL_LINE
   }
-#if DEBUG_VERBOSE && !defined(NDEBUG)
-  fprintf(stderr, "At `%s', max version in (%02x, %02x) is %" PRId64 "\n",
-          getSearchPathPrintable(n).c_str(), begin, end, result);
-#endif
-  return result;
 }
 
 Vector<uint8_t> getSearchPath(Arena &arena, Node *n) {
@@ -1722,7 +1918,7 @@ bool checkRangeStartsWith(Node *n, std::span<const uint8_t> key, int begin,
 #endif
   auto remaining = key;
   if (remaining.size() == 0) {
-    return maxBetweenExclusive(n, begin, end) <= readVersion;
+    return checkMaxBetweenExclusive(n, begin, end, readVersion);
   }
 
   auto *child = getChild(n, remaining[0]);
@@ -1821,7 +2017,7 @@ struct CheckRangeLeftSide {
       }
 
       if (searchPathLen >= prefixLen) {
-        if (maxBetweenExclusive(n, remaining[0], 256) > readVersion) {
+        if (!checkMaxBetweenExclusive(n, remaining[0], 256, readVersion)) {
           ok = false;
           return true;
         }
@@ -1963,7 +2159,7 @@ struct CheckRangeRightSide {
           return true;
         }
 
-        if (maxBetweenExclusive(n, -1, remaining[0]) > readVersion) {
+        if (!checkMaxBetweenExclusive(n, -1, remaining[0], readVersion)) {
           ok = false;
           return true;
         }
@@ -2174,7 +2370,7 @@ 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));
@@ -2198,9 +2394,8 @@ 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) {
@@ -2208,9 +2403,8 @@ 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);
@@ -2221,8 +2415,7 @@ 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;
@@ -2267,7 +2460,7 @@ 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 {
@@ -2308,9 +2501,8 @@ 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));
@@ -2337,11 +2529,10 @@ 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;
 
@@ -2356,8 +2547,8 @@ 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;
 
@@ -2557,8 +2748,7 @@ 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) {
@@ -2570,22 +2760,66 @@ int64_t &maxVersion(Node *n, ConflictSet::Impl *impl) {
   case Type_Node3: {
     auto *n3 = static_cast<Node3 *>(n);
     int i = getNodeIndex(n3, index);
-    return n3->children[i].childMaxVersion;
+    return n3->childMaxVersion[i];
   }
   case Type_Node16: {
     auto *n16 = static_cast<Node16 *>(n);
     int i = getNodeIndex(n16, index);
-    return n16->children[i].childMaxVersion;
+    return n16->childMaxVersion[i];
   }
   case Type_Node48: {
     auto *n48 = static_cast<Node48 *>(n);
     assert(n48->bitSet.test(index));
-    return n48->children[n48->index[index]].childMaxVersion;
+    return n48->childMaxVersion[n48->index[index]];
   }
   case Type_Node256: {
     auto *n256 = static_cast<Node256 *>(n);
     assert(n256->bitSet.test(index));
-    return n256->children[index].childMaxVersion;
+    return n256->childMaxVersion[index];
+  }
+  default:                   // GCOVR_EXCL_LINE
+    __builtin_unreachable(); // GCOVR_EXCL_LINE
+  }
+}
+
+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->childMaxVersion[i] = newMax;
+    return;
+  }
+  case Type_Node16: {
+    auto *n16 = static_cast<Node16 *>(n);
+    int i = getNodeIndex(n16, index);
+    n16->childMaxVersion[i] = newMax;
+    return;
+  }
+  case Type_Node48: {
+    auto *n48 = static_cast<Node48 *>(n);
+    assert(n48->bitSet.test(index));
+    int i = n48->index[index];
+    n48->childMaxVersion[i] = 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->childMaxVersion[index] = newMax;
+    n256->maxOfMax[index >> Node256::kMaxOfMaxShift] =
+        std::max(n256->maxOfMax[index >> Node256::kMaxOfMaxShift], newMax);
+    return;
   }
   default:                   // GCOVR_EXCL_LINE
     __builtin_unreachable(); // GCOVR_EXCL_LINE
@@ -2847,7 +3081,7 @@ 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

@@ -58,27 +58,29 @@ Performance counters:
 
 ## Skip list
 
-|  ns/op |         op/s | err% | total | benchmark                           |
-| -----: | -----------: | ---: | ----: | :---------------------------------- |
-| 246.99 | 4,048,700.59 | 0.2% |  0.01 | `point reads`                       |
-| 260.16 | 3,843,784.65 | 0.1% |  0.01 | `prefix reads`                      |
-| 493.35 | 2,026,953.19 | 0.1% |  0.01 | `range reads`                       |
-| 462.05 | 2,164,289.23 | 0.6% |  0.01 | `point writes`                      |
-| 448.19 | 2,231,205.25 | 0.9% |  0.01 | `prefix writes`                     |
-| 255.83 | 3,908,845.72 | 1.5% |  0.02 | `range writes`                      |
-| 582.63 | 1,716,349.02 | 1.3% |  0.01 | `monotonic increasing point writes` |
+|               ns/op |                op/s |    err% |     total | benchmark
+|--------------------:|--------------------:|--------:|----------:|:----------
+|              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                           |
-| -----: | ------------: | ---: | ----: | :---------------------------------- |
-|  19.42 | 51,483,206.67 | 0.3% |  0.01 | `point reads`                       |
-|  58.43 | 17,115,612.57 | 0.1% |  0.01 | `prefix reads`                      |
-| 216.09 |  4,627,766.60 | 0.2% |  0.01 | `range reads`                       |
-|  28.35 | 35,267,567.72 | 0.2% |  0.01 | `point writes`                      |
-|  43.43 | 23,026,226.17 | 0.2% |  0.01 | `prefix writes`                     |
-|  50.00 | 20,000,000.00 | 0.0% |  0.01 | `range writes`                      |
-|  92.38 | 10,824,863.69 | 4.1% |  0.01 | `monotonic increasing point writes` |
+|               ns/op |                op/s |    err% |     total | benchmark
+|--------------------:|--------------------:|--------:|----------:|:----------
+|               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
 

apple-symbol-imports.txt

@@ -1,3 +1,5 @@
+___stack_chk_fail
+___stack_chk_guard
 __tlv_bootstrap
 _abort
 _bzero

conflict_set.py

@@ -115,7 +115,9 @@ class ConflictSet:
 
     def check(self, *reads: ReadRange) -> list[Result]:
         r = (ctypes.c_int * len(reads))()
-        self._lib.ConflictSet_check(self.p, *reads, r, 1)
+        self._lib.ConflictSet_check(
+            self.p, (ReadRange * len(reads))(*reads), r, len(reads)
+        )
         return [Result(x) for x in r]
 
     def setOldestVersion(self, version: int) -> None: