Merge pull request 'Improve perf/codegen' (#94) from perf into main

Reviewed-on: weaselab/conflict-set#94

.gitea/workflows/ci.yml

@@ -19,15 +19,8 @@ jobs:
 
       - name: Install dependencies
         run: |
-          . /etc/os-release
-          wget -qO- https://apt.llvm.org/llvm-snapshot.gpg.key | sudo tee /etc/apt/trusted.gpg.d/apt.llvm.org.asc
-          echo "deb http://apt.llvm.org/${VERSION_CODENAME}/ llvm-toolchain-${VERSION_CODENAME}-21 main" | sudo tee /etc/apt/sources.list.d/llvm.list
           sudo apt-get update -qq
-          sudo apt-get install -y \
-            clang-21 git nodejs pre-commit
-          for tool in clang clang++; do
-            sudo update-alternatives --install /usr/bin/${tool} ${tool} /usr/bin/${tool}-21 100
-          done
+          sudo apt-get install -y git nodejs pre-commit
 
       - uses: actions/cache@v4
         with:
@@ -45,43 +38,72 @@ jobs:
       matrix:
         include:
           - name: 64-bit-versions
+            runner: ubuntu-latest-amd64
+            arch: amd64
+            llvm_version: "21"
+            msan_url: ""
             cmake_args: -DCMAKE_CXX_FLAGS=-DUSE_64_BIT=1
           - name: debug
-            cmake_args: -DCMAKE_BUILD_TYPE=Debug
+            runner: ubuntu-latest-amd64
+            arch: amd64
+            llvm_version: "21"
+            msan_url: https://minio.weaselab.dev/public/x86_64/msan-toolchain-21.1.8.tar.zst
+            cmake_args: -DCMAKE_BUILD_TYPE=Debug -DMSAN_TOOLCHAIN_PATH=/opt/msan
           - name: simd-fallback
+            runner: ubuntu-latest-amd64
+            arch: amd64
+            llvm_version: "21"
+            msan_url: ""
             cmake_args: -DUSE_SIMD_FALLBACK=ON
           - name: gcc
+            runner: ubuntu-latest-amd64
+            arch: amd64
+            llvm_version: "21"
+            msan_url: ""
             cmake_args: -DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++
-    runs-on: ubuntu-latest-amd64
+          - name: debug-arm64
+            runner: ubuntu-latest-arm64
+            arch: arm64
+            llvm_version: "22"
+            msan_url: https://minio.weaselab.dev/public/aarch64/msan-toolchain-22.1.8.tar.zst
+            cmake_args: -DCMAKE_BUILD_TYPE=Debug -DMSAN_TOOLCHAIN_PATH=/opt/msan
+    runs-on: ${{ matrix.runner }}
     steps:
       - uses: actions/checkout@v4
 
       - uses: actions/cache@v4
         with:
           path: /var/cache/apt/archives
-          key: apt-amd64-${{ hashFiles('.gitea/workflows/ci.yml') }}
+          key: apt-${{ matrix.arch }}-${{ hashFiles('.gitea/workflows/ci.yml') }}
 
-      - name: Install dependencies
+      - name: Install common dependencies
+        run: |
+          sudo apt-get update -qq
+          sudo apt-get install -y \
+            build-essential ccache cmake gcc g++ libc6-dbg \
+            ninja-build python3 valgrind zstd
+          sudo curl -Ls "https://minio.weaselab.dev/public/$(uname -m)/mc.RELEASE.2025-08-13T08-35-41Z" \
+            -o /usr/local/bin/mc && sudo chmod +x /usr/local/bin/mc
+
+      - name: Install LLVM toolchain
+        if: matrix.name != 'gcc'
         run: |
           . /etc/os-release
           wget -qO- https://apt.llvm.org/llvm-snapshot.gpg.key | sudo tee /etc/apt/trusted.gpg.d/apt.llvm.org.asc
-          echo "deb http://apt.llvm.org/${VERSION_CODENAME}/ llvm-toolchain-${VERSION_CODENAME}-21 main" | sudo tee /etc/apt/sources.list.d/llvm.list
+          echo "deb http://apt.llvm.org/${VERSION_CODENAME}/ llvm-toolchain-${VERSION_CODENAME}-${{ matrix.llvm_version }} main" | sudo tee /etc/apt/sources.list.d/llvm.list
           sudo apt-get update -qq
           sudo apt-get install -y \
-            build-essential ccache clang-21 cmake gcc g++ \
-            libc6-dbg llvm-21 lld-21 mold ninja-build python3 valgrind zstd
-          sudo curl -Ls "https://dl.min.io/client/mc/release/linux-amd64/mc" \
-            -o /usr/local/bin/mc && sudo chmod +x /usr/local/bin/mc
+            clang-${{ matrix.llvm_version }} llvm-${{ matrix.llvm_version }} lld-${{ matrix.llvm_version }} mold
           for tool in clang clang++ llvm-ar llvm-nm llvm-ranlib llvm-objcopy llvm-cov llvm-symbolizer lld ld.lld; do
-            sudo update-alternatives --install /usr/bin/${tool} ${tool} /usr/bin/${tool}-21 100
+            sudo update-alternatives --install /usr/bin/${tool} ${tool} /usr/bin/${tool}-${{ matrix.llvm_version }} 100
           done
 
-      - uses: actions/cache@v4
-        with:
-          path: .ccache
-          key: ccache-${{ matrix.name }}-${{ gitea.sha }}
-          restore-keys: |
-            ccache-${{ matrix.name }}-
+      - name: Download MSan toolchain
+        if: matrix.msan_url != ''
+        run: |
+          curl -Ls "${{ matrix.msan_url }}" -o /tmp/msan-toolchain.tar.zst
+          sudo mkdir -p /opt/msan
+          sudo tar --zstd -xf /tmp/msan-toolchain.tar.zst -C /opt/msan
 
       - name: Build
         run: |
@@ -123,8 +145,10 @@ jobs:
         include:
           - runner: ubuntu-latest-amd64
             arch: amd64
+            cmake_args: -DMSAN_TOOLCHAIN_PATH=/opt/msan
           - runner: ubuntu-latest-arm64
             arch: arm64
+            cmake_args: ""
     runs-on: ${{ matrix.runner }}
     steps:
       - uses: actions/checkout@v4
@@ -141,16 +165,23 @@ jobs:
           echo "deb http://apt.llvm.org/${VERSION_CODENAME}/ llvm-toolchain-${VERSION_CODENAME}-21 main" | sudo tee /etc/apt/sources.list.d/llvm.list
           sudo apt-get update -qq
           sudo apt-get install -y \
-            biber build-essential ccache clang-21 cmake devscripts \
-            latexmk libc6-dbg llvm-21 lld-21 mold ninja-build rpm \
-            texlive-bibtex-extra texlive-fonts-recommended \
-            texlive-latex-extra texlive-pictures valgrind zstd
-          sudo curl -Ls "https://dl.min.io/client/mc/release/linux-$(dpkg --print-architecture)/mc" \
+            build-essential ccache clang-21 cmake devscripts \
+            libc6-dbg llvm-21 lld-21 mold ninja-build rpm \
+            valgrind zstd
+          sudo curl -Ls "https://minio.weaselab.dev/public/$(uname -m)/mc.RELEASE.2025-08-13T08-35-41Z" \
             -o /usr/local/bin/mc && sudo chmod +x /usr/local/bin/mc
           for tool in clang clang++ llvm-ar llvm-nm llvm-ranlib llvm-objcopy llvm-cov llvm-symbolizer lld ld.lld; do
             sudo update-alternatives --install /usr/bin/${tool} ${tool} /usr/bin/${tool}-21 100
           done
 
+      - name: Install paper build dependencies
+        if: matrix.arch == 'amd64'
+        run: |
+          sudo apt-get install -y \
+            biber latexmk \
+            texlive-bibtex-extra texlive-fonts-recommended \
+            texlive-latex-extra texlive-pictures
+
       - uses: actions/cache@v4
         with:
           path: .ccache
@@ -158,23 +189,25 @@ jobs:
           restore-keys: |
             ccache-release-${{ matrix.arch }}-
 
+      - name: Download MSan toolchain
+        if: matrix.arch == 'amd64'
+        run: |
+          curl -Ls "https://minio.weaselab.dev/public/x86_64/msan-toolchain-21.1.8.tar.zst" -o /tmp/msan-toolchain.tar.zst
+          sudo mkdir -p /opt/msan
+          sudo tar --zstd -xf /tmp/msan-toolchain.tar.zst -C /opt/msan
+
       - name: Build
         run: |
           export CCACHE_DIR="$GITHUB_WORKSPACE/.ccache"
           rm -rf build
-          cmake -S . -B build -G Ninja -DCMAKE_CXX_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_FLAGS=-DNVALGRIND
+          cmake -S . -B build -G Ninja -DCMAKE_CXX_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_FLAGS=-DNVALGRIND ${{ matrix.cmake_args }}
           ninja -C build
           ccache -s
 
       - name: Test
         run: |
           cd build
-          # On arm64, valgrind needs the MAKE_MEM_DEFINED client requests for
-          # https://git.weaselab.dev/weaselab/conflict-set/issues/39, but this
-          # build has -DNVALGRIND, which compiles them out. Skip valgrind
-          # tests here; they run annotated in the test job.
-          ctest --no-compress-output --test-output-size-passed 100000 --test-output-size-failed 100000 ${{ matrix.arch == 'arm64' && '-E valgrind' || '' }} -T Test -j "$(nproc)" --timeout 90 > /dev/null
-
+          ctest --no-compress-output --test-output-size-passed 100000 --test-output-size-failed 100000 -T Test -j "$(nproc)" --timeout 90 > /dev/null
       - name: Package
         run: |
           cd build
@@ -237,7 +270,7 @@ jobs:
           sudo apt-get install -y \
             build-essential ccache clang-21 cmake gcovr \
             libc6-dbg llvm-21 lld-21 mold ninja-build python3 valgrind zstd
-          sudo curl -Ls "https://dl.min.io/client/mc/release/linux-amd64/mc" \
+          sudo curl -Ls "https://minio.weaselab.dev/public/$(uname -m)/mc.RELEASE.2025-08-13T08-35-41Z" \
             -o /usr/local/bin/mc && sudo chmod +x /usr/local/bin/mc
           for tool in clang clang++ llvm-ar llvm-nm llvm-ranlib llvm-objcopy llvm-cov llvm-symbolizer lld ld.lld; do
             sudo update-alternatives --install /usr/bin/${tool} ${tool} /usr/bin/${tool}-21 100

Bench.cpp

@@ -37,15 +37,16 @@ ConflictSet::ReadRange singleton(Arena &arena, TrivialSpan key) {
 }
 
 ConflictSet::ReadRange prefixRange(Arena &arena, TrivialSpan key) {
-  int index;
-  for (index = key.size() - 1; index >= 0; index--)
-    if ((key[index]) != 255)
+  int index = key.size() - 1;
+  for (; index >= 0; index--)
+    if (key[index] != 255)
       break;
 
   // Must not be called with a string that consists only of zero or more '\xff'
-  // bytes.
+  // bytes, or with an empty string (which has no finite upper bound).
   if (index < 0) {
     assert(false);
+    std::abort();
   }
 
   uint8_t *buf = new (arena) uint8_t[index + 1];

CMakeLists.txt

@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.18)
 project(
   conflict-set
-  VERSION 0.0.14
+  VERSION 0.1.0
   DESCRIPTION
     "A data structure for optimistic concurrency control on ranges of bitwise-lexicographically-ordered keys."
   HOMEPAGE_URL "https://git.weaselab.dev/weaselab/conflict-set"
@@ -103,6 +103,13 @@ option(USE_SIMD_FALLBACK
 
 option(DISABLE_TSAN "Disable TSAN" OFF)
 
+set(MSAN_TOOLCHAIN_PATH
+    ""
+    CACHE
+      PATH
+      "Path to an MSan-instrumented libc++ toolchain. When set, an MSan fuzz_driver target is added. Must contain include/c++/v1 and lib. See build_msan_toolchain.sh to produce one."
+)
+
 # This is encouraged according to
 # https://valgrind.org/docs/manual/manual-core-adv.html#manual-core-adv.clientreq
 include_directories(SYSTEM ${CMAKE_CURRENT_SOURCE_DIR}/third_party/valgrind)
@@ -175,6 +182,8 @@ endif()
 
 include(CTest)
 
+option(TEST_SKIP_LIST "Run the slow skip list cross-check tests" OFF)
+
 # disable tests if this is being used through e.g. FetchContent
 if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
 
@@ -217,8 +226,9 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
   target_compile_options(driver_skip_list PRIVATE ${TEST_FLAGS})
   target_link_libraries(driver_skip_list PRIVATE skip_list)
 
-  # enable to test skip list
-  if(0)
+  # Optional cross-check of SkipList.cpp against ReferenceImpl. Opt-in via
+  # -DTEST_SKIP_LIST=ON, since these are slow.
+  if(TEST_SKIP_LIST)
     foreach(TEST ${CORPUS_TESTS})
       get_filename_component(hash ${TEST} NAME)
       add_test(NAME skip_list_${hash} COMMAND driver_skip_list ${TEST})
@@ -253,7 +263,7 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
     endif()
   endif()
 
-  # whitebox tests
+  # whitebox tests asan+ubsan
   add_executable(fuzz_driver ConflictSet.cpp FuzzTestDriver.cpp)
   target_compile_options(fuzz_driver PRIVATE ${TEST_FLAGS})
   if(NOT CMAKE_CROSSCOMPILING)
@@ -268,6 +278,34 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
     add_test(NAME conflict_set_fuzz_${hash} COMMAND fuzz_driver ${TEST})
   endforeach()
 
+  # whitebox tests msan
+  if(MSAN_TOOLCHAIN_PATH)
+    add_executable(fuzz_driver_msan ConflictSet.cpp FuzzTestDriver.cpp)
+    target_compile_options(fuzz_driver_msan PRIVATE ${TEST_FLAGS})
+    if(NOT CMAKE_CROSSCOMPILING)
+      target_compile_options(
+        fuzz_driver_msan
+        PRIVATE -fsanitize=memory -fsanitize-memory-track-origins=2
+                $<$<COMPILE_LANGUAGE:CXX>:-stdlib=libc++>
+                -I${MSAN_TOOLCHAIN_PATH}/include/c++/v1)
+      target_link_options(
+        fuzz_driver_msan
+        PRIVATE
+        -fsanitize=memory
+        -fsanitize-memory-track-origins=2
+        -stdlib=libc++
+        -L${MSAN_TOOLCHAIN_PATH}/lib
+        LINKER:-rpath,${MSAN_TOOLCHAIN_PATH}/lib)
+    endif()
+    target_compile_definitions(fuzz_driver_msan PRIVATE ENABLE_FUZZ)
+    target_include_directories(fuzz_driver_msan
+                               PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include)
+    foreach(TEST ${CORPUS_TESTS})
+      get_filename_component(hash ${TEST} NAME)
+      add_test(NAME conflict_set_msan_${hash} COMMAND fuzz_driver_msan ${TEST})
+    endforeach()
+  endif()
+
   # tsan tests
   if(NOT CMAKE_CROSSCOMPILING AND NOT DISABLE_TSAN)
     add_executable(tsan_driver ConflictSet.cpp FuzzTestDriver.cpp)
@@ -355,6 +393,24 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
                         PROPERTIES CXX_STANDARD_REQUIRED ON)
   add_test(NAME conflict_set_cxx_api_test COMMAND conflict_set_cxx_api_test)
 
+  # SHOW_MEMORY instrumentation, which the rest of the test suite doesn't
+  # compile with
+  add_executable(conflict_set_show_memory_test ConflictSet.cpp
+                                               ShowMemoryTest.cpp)
+  target_compile_options(conflict_set_show_memory_test PRIVATE ${TEST_FLAGS})
+  target_compile_definitions(conflict_set_show_memory_test
+                             PRIVATE SHOW_MEMORY=1)
+  target_include_directories(conflict_set_show_memory_test
+                             PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include)
+  if(NOT CMAKE_CROSSCOMPILING)
+    target_compile_options(conflict_set_show_memory_test
+                           PRIVATE -fsanitize=address,undefined)
+    target_link_options(conflict_set_show_memory_test PRIVATE
+                        -fsanitize=address,undefined)
+  endif()
+  add_test(NAME conflict_set_show_memory_test
+           COMMAND conflict_set_show_memory_test)
+
   # symbol visibility tests
   if(NOT CMAKE_BUILD_TYPE STREQUAL Debug)
     if(APPLE)

ConflictSet.cpp

@@ -80,17 +80,19 @@ using namespace weaselab;
 #define assume assert
 #endif
 
-#if SHOW_MEMORY
+// The SHOW_MEMORY instrumentation. Declared unconditionally so that each of
+// these has a single definition in every configuration (see the definitions at
+// the end of this file), rather than a no-op stub standing in for the real
+// thing.
 void addNode(struct Node *);
 void removeNode(struct Node *);
 void addKey(struct Node *);
 void removeKey(struct Node *);
-#else
-constexpr void addNode(struct Node *) {}
-constexpr void removeNode(struct Node *) {}
-constexpr void addKey(struct Node *) {}
-constexpr void removeKey(struct Node *) {}
-#endif
+// Like `removeKey`, but with a caller-provided search path length. Used when
+// the node's ancestors may already have been released, which makes walking
+// `parent` pointers invalid.
+void removeKey(struct Node *, int64_t searchPathLength);
+int64_t getSearchPathLength(struct Node *);
 
 // ==================== BEGIN IMPLEMENTATION ====================
 
@@ -116,13 +118,40 @@ struct InternalVersionT {
   constexpr InternalVersionT() = default;
   constexpr explicit InternalVersionT(int64_t value) : value(value) {}
   constexpr int64_t toInt64() const { return value; } // GCOVR_EXCL_LINE
-  constexpr auto operator<=>(const InternalVersionT &rhs) const {
+  constexpr bool operator<(const InternalVersionT &rhs) const {
 #if USE_64_BIT
-    return value <=> rhs.value;
+    return value < rhs.value;
 #else
     // Maintains ordering after overflow, as long as the full-precision versions
     // are within `kMaxCorrectVersionWindow` of eachother.
-    return int32_t(value - rhs.value) <=> 0;
+    return int32_t(value - rhs.value) < 0;
+#endif
+  }
+  constexpr bool operator<=(const InternalVersionT &rhs) const {
+#if USE_64_BIT
+    return value <= rhs.value;
+#else
+    // Maintains ordering after overflow, as long as the full-precision versions
+    // are within `kMaxCorrectVersionWindow` of eachother.
+    return int32_t(value - rhs.value) <= 0;
+#endif
+  }
+  constexpr bool operator>=(const InternalVersionT &rhs) const {
+#if USE_64_BIT
+    return value >= rhs.value;
+#else
+    // Maintains ordering after overflow, as long as the full-precision versions
+    // are within `kMaxCorrectVersionWindow` of eachother.
+    return int32_t(value - rhs.value) >= 0;
+#endif
+  }
+  constexpr bool operator>(const InternalVersionT &rhs) const {
+#if USE_64_BIT
+    return value > rhs.value;
+#else
+    // Maintains ordering after overflow, as long as the full-precision versions
+    // are within `kMaxCorrectVersionWindow` of eachother.
+    return int32_t(value - rhs.value) > 0;
 #endif
   }
   constexpr bool operator==(const InternalVersionT &) const = default;
@@ -307,7 +336,9 @@ struct Node {
   uint8_t *partialKey();
   Type getType() const { return type; }
   int32_t getCapacity() const {
-    assert(!releaseDeferred);
+    // `partialKeyCapacity` is not repurposed by `deferRelease`, so this is also
+    // valid for nodes with `releaseDeferred` set. It's needed to account for
+    // the capacity of a deferred node when it's finally released.
     return partialKeyCapacity;
   }
 
@@ -488,6 +519,7 @@ inline void Node3::copyChildrenAndKeyFrom(const Node16 &other) {
 
 inline void Node16::copyChildrenAndKeyFrom(const Node3 &other) {
   copyCommon(*this, other);
+  memset(index, 0, sizeof(index));
   memcpy(index, other.index, Node3::kMaxNodes);
   memcpy(children, other.children,
          Node3::kMaxNodes * sizeof(children[0])); // NOLINT
@@ -499,6 +531,12 @@ inline void Node16::copyChildrenAndKeyFrom(const Node3 &other) {
     assert(children[i]->parent == &other);
     children[i]->parent = this;
   }
+  // The index bytes beyond numChildren are initialized so that the SIMD loads
+  // of the full 16 bytes are well-defined (reading indeterminate bytes is UB
+  // in C++). valgrind then treats them as undefined so that it can still flag
+  // any read that relies on them.
+  VALGRIND_MAKE_MEM_UNDEFINED(index + numChildren,
+                              sizeof(index) - numChildren * sizeof(index[0]));
 }
 
 inline void Node16::copyChildrenAndKeyFrom(const Node16 &other) {
@@ -515,6 +553,7 @@ inline void Node16::copyChildrenAndKeyFrom(const Node16 &other) {
 
 inline void Node16::copyChildrenAndKeyFrom(const Node48 &other) {
   copyCommon(*this, other);
+  memset(index, 0, sizeof(index));
   int i = 0;
   other.bitSet.forEachSet([&](int c) {
     // Suppress a false positive -Waggressive-loop-optimizations warning
@@ -528,6 +567,12 @@ inline void Node16::copyChildrenAndKeyFrom(const Node48 &other) {
     ++i;
   });
   memcpy(partialKey(), &other + 1, partialKeyLen);
+  // The index bytes beyond numChildren are initialized so that the SIMD loads
+  // of the full 16 bytes are well-defined (reading indeterminate bytes is UB
+  // in C++). valgrind then treats them as undefined so that it can still flag
+  // any read that relies on them.
+  VALGRIND_MAKE_MEM_UNDEFINED(index + numChildren,
+                              sizeof(index) - numChildren * sizeof(index[0]));
 }
 
 inline void Node48::copyChildrenAndKeyFrom(const Node16 &other) {
@@ -535,6 +580,7 @@ inline void Node48::copyChildrenAndKeyFrom(const Node16 &other) {
   assert(numChildren == Node16::kMaxNodes);
   memset(index, -1, sizeof(index));
   memset(children, 0, sizeof(children));
+  memset(reverseIndex, 0, sizeof(reverseIndex));
   const auto z = InternalVersionT::zero;
   for (auto &v : childMaxVersion) {
     v = z;
@@ -554,6 +600,9 @@ inline void Node48::copyChildrenAndKeyFrom(const Node16 &other) {
         std::max(maxOfMax[i >> Node48::kMaxOfMaxShift], childMaxVersion[i]);
     ++i;
   }
+  // The reverseIndex bytes beyond numChildren are initialized so that the
+  // SIMD loads of a full 16-byte page in scan16 are well-defined (reading
+  // indeterminate bytes is UB in C++).
 }
 
 inline void Node48::copyChildrenAndKeyFrom(const Node48 &other) {
@@ -580,6 +629,7 @@ inline void Node48::copyChildrenAndKeyFrom(const Node256 &other) {
   copyCommon(*this, other);
   memset(index, -1, sizeof(index));
   memset(children, 0, sizeof(children));
+  memset(reverseIndex, 0, sizeof(reverseIndex));
   const auto z = InternalVersionT::zero;
   for (auto &v : childMaxVersion) {
     v = z;
@@ -601,6 +651,9 @@ inline void Node48::copyChildrenAndKeyFrom(const Node256 &other) {
     ++i;
   });
   memcpy(partialKey(), &other + 1, partialKeyLen);
+  // The reverseIndex bytes beyond numChildren are initialized so that the
+  // SIMD loads of a full 16-byte page in scan16 are well-defined (reading
+  // indeterminate bytes is UB in C++).
 }
 
 inline void Node256::copyChildrenAndKeyFrom(const Node48 &other) {
@@ -1332,8 +1385,8 @@ TaggedNodePointer getChildGeq(Node *self, int child) {
 
 TaggedNodePointer getFirstChild(Node0 *) { return nullptr; }
 TaggedNodePointer getFirstChild(Node3 *self) {
-  // Improves scan performance
-  self->children[1].prefetch();
+  // Don't prefetch since self->children[1] might be uninitialized and msan
+  // doesn't like that
   return self->children[0];
 }
 TaggedNodePointer getFirstChild(Node16 *self) {
@@ -1992,7 +2045,6 @@ Node *erase(Node *self, WriteContext *writeContext, bool logical) {
       parent48->index[parentIndex] = toRemoveChildrenIndex;
       parent48->reverseIndex[toRemoveChildrenIndex] = parentIndex;
     }
-    parent48->childMaxVersion[lastChildrenIndex] = writeContext->zero;
     VALGRIND_MAKE_MEM_UNDEFINED(parent48->children + lastChildrenIndex,
                                 sizeof(parent48->children[0]));
 
@@ -2087,6 +2139,53 @@ compare16_avx512(const InternalVersionT *vs, InternalVersionT rv) {
 }
 #endif
 
+#ifdef HAS_ARM_NEON
+// Returns a uint8x16_t whose byte i is 0xff if vs[i] > readVersion and 0
+// otherwise. When USE_64_BIT is set the comparison uses the full 64-bit value
+// (via vcgtq_s64); otherwise it falls back to the 32-bit subtract-and-compare
+// used by the original NEON fast path. The caller narrows this 16-byte vector
+// into the nibble-packed (4 bits per version) mask expected by scan16 and
+// checkMaxBetweenExclusiveImpl.
+inline uint8x16_t conflictMask16(const InternalVersionT *vs,
+                                 InternalVersionT readVersion) {
+#if USE_64_BIT
+  int64_t rv;
+  memcpy(&rv, &readVersion, sizeof(rv));
+  const auto rvVec = vdupq_n_s64(rv);
+  int32x2_t r32[8];
+  const auto *vsp = reinterpret_cast<const int64_t *>(vs);
+  for (int j = 0; j < 8; ++j) {
+    r32[j] = vmovn_s64(
+        vreinterpretq_s64_u64(vcgtq_s64(vld1q_s64(vsp + 2 * j), rvVec)));
+  }
+  uint32x4_t w4[4];
+  for (int k = 0; k < 4; ++k) {
+    w4[k] = vreinterpretq_u32_s32(vcombine_s32(r32[2 * k], r32[2 * k + 1]));
+  }
+  return vcombine_u8(
+      vmovn_u16(vcombine_u16(vmovn_u32(w4[0]), vmovn_u32(w4[1]))),
+      vmovn_u16(vcombine_u16(vmovn_u32(w4[2]), vmovn_u32(w4[3]))));
+#else
+  uint32x4_t w4[4];
+  memcpy(w4, vs, sizeof(w4));
+  uint32_t rv;
+  memcpy(&rv, &readVersion, sizeof(rv));
+  const auto rvVec = vdupq_n_u32(rv);
+
+  int32x4_t z;
+  memset(&z, 0, sizeof(z));
+
+  uint16x4_t conflicting[4];
+  for (int i = 0; i < 4; ++i) {
+    conflicting[i] =
+        vmovn_u32(vcgtq_s32(vreinterpretq_s32_u32(vsubq_u32(w4[i], rvVec)), z));
+  }
+  return vcombine_u8(vmovn_u16(vcombine_u16(conflicting[0], conflicting[1])),
+                     vmovn_u16(vcombine_u16(conflicting[2], conflicting[3])));
+#endif
+}
+#endif
+
 // Returns true if v[i] <= readVersion for all i such that begin <= is[i] < end
 // Preconditions: begin <= end, end - begin < 256
 template <bool kAVX512>
@@ -2107,26 +2206,10 @@ bool scan16(const InternalVersionT *vs, const uint8_t *is, int begin, int end,
   uint64_t mask = vget_lane_u64(
       vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(results), 4)), 0);
 
-  uint32x4_t w4[4];
-  memcpy(w4, vs, sizeof(w4));
-  uint32_t rv;
-  memcpy(&rv, &readVersion, sizeof(rv));
-  const auto rvVec = vdupq_n_u32(rv);
-
-  int32x4_t z;
-  memset(&z, 0, sizeof(z));
-
-  uint16x4_t conflicting[4];
-  for (int i = 0; i < 4; ++i) {
-    conflicting[i] =
-        vmovn_u32(vcgtq_s32(vreinterpretq_s32_u32(vsubq_u32(w4[i], rvVec)), z));
-  }
-  auto combined =
-      vcombine_u8(vmovn_u16(vcombine_u16(conflicting[0], conflicting[1])),
-                  vmovn_u16(vcombine_u16(conflicting[2], conflicting[3])));
-
   uint64_t compared = vget_lane_u64(
-      vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(combined), 4)), 0);
+      vreinterpret_u64_u8(vshrn_n_u16(
+          vreinterpretq_u16_u8(conflictMask16(vs, readVersion)), 4)),
+      0);
 
   return !(compared & mask);
 
@@ -2153,12 +2236,14 @@ bool scan16(const InternalVersionT *vs, const uint8_t *is, int begin, int end,
   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;
+    if (vs[i] > readVersion) {
+      compared |= 1u << i;
+      if (inBounds(is[i])) {
+        mask |= 1u << i;
+      }
+    }
   }
   return !(compared & mask);
 
@@ -2174,26 +2259,10 @@ bool scan16(const InternalVersionT *vs, int begin, int end,
   assert(begin <= end);
 
 #if defined(HAS_ARM_NEON)
-  uint32x4_t w4[4];
-  memcpy(w4, vs, sizeof(w4));
-  uint32_t rv;
-  memcpy(&rv, &readVersion, sizeof(rv));
-  const auto rvVec = vdupq_n_u32(rv);
-
-  int32x4_t z;
-  memset(&z, 0, sizeof(z));
-
-  uint16x4_t conflicting[4];
-  for (int i = 0; i < 4; ++i) {
-    conflicting[i] =
-        vmovn_u32(vcgtq_s32(vreinterpretq_s32_u32(vsubq_u32(w4[i], rvVec)), z));
-  }
-  auto combined =
-      vcombine_u8(vmovn_u16(vcombine_u16(conflicting[0], conflicting[1])),
-                  vmovn_u16(vcombine_u16(conflicting[2], conflicting[3])));
-
   uint64_t conflict = vget_lane_u64(
-      vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(combined), 4)), 0);
+      vreinterpret_u64_u8(vshrn_n_u16(
+          vreinterpretq_u16_u8(conflictMask16(vs, readVersion)), 4)),
+      0);
 
   conflict &= end == 16 ? -1 : (uint64_t(1) << (end << 2)) - 1;
   conflict >>= begin << 2;
@@ -2242,6 +2311,8 @@ bool checkMaxBetweenExclusiveImpl(Node3 *n, int begin, int end,
   assert(!(begin == -1 && end == 256));
 
   auto *self = static_cast<Node3 *>(n);
+  assume(1 <= self->numChildren);
+  assume(self->numChildren <= 3);
 
   ++begin;
 
@@ -2250,17 +2321,9 @@ bool checkMaxBetweenExclusiveImpl(Node3 *n, int begin, int end,
   auto inBounds = [&](unsigned c) { return c - shiftAmount < shiftUpperBound; };
 
   uint32_t mask = 0;
-  for (int i = 0; i < Node3::kMaxNodes; ++i) {
+  for (int i = 0; i < self->numChildren; ++i) {
     mask |= inBounds(self->index[i]) << i;
   }
-  mask &= (1 << self->numChildren) - 1;
-#ifdef __aarch64__
-  // The bits surviving the mask above don't derive from uninitialized slots,
-  // but clang 21+ on aarch64 lowers inBounds through flags+csel, which
-  // memcheck models imprecisely, tainting bits the mask provably clears.
-  // https://git.weaselab.dev/weaselab/conflict-set/issues/39
-  VALGRIND_MAKE_MEM_DEFINED(&mask, sizeof(mask));
-#endif
   if (!mask) {
     return true;
   }
@@ -2268,17 +2331,11 @@ bool checkMaxBetweenExclusiveImpl(Node3 *n, int begin, int end,
   const bool firstRangeOk =
       !child->entryPresent || child->entry.rangeVersion <= readVersion;
   uint32_t compared = 0;
-  for (int i = 0; i < Node3::kMaxNodes; ++i) {
+  for (int i = 0; i < self->numChildren; ++i) {
     compared |= (self->childMaxVersion[i] > readVersion) << i;
   }
 
-  uint32_t compared_masked = compared & mask;
-#ifdef __aarch64__
-  // Same imprecise csel modeling as above.
-  // https://git.weaselab.dev/weaselab/conflict-set/issues/39
-  VALGRIND_MAKE_MEM_DEFINED(&compared_masked, sizeof(compared_masked));
-#endif
-  return !compared_masked && firstRangeOk;
+  return !(compared & mask) && firstRangeOk;
 }
 
 template <bool kAVX512>
@@ -2321,26 +2378,11 @@ bool checkMaxBetweenExclusiveImpl(Node16 *n, int begin, int end,
   const bool firstRangeOk =
       !child->entryPresent || child->entry.rangeVersion <= readVersion;
 
-  uint32x4_t w4[4];
-  memcpy(w4, self->childMaxVersion, sizeof(w4));
-  uint32_t rv;
-  memcpy(&rv, &readVersion, sizeof(rv));
-  const auto rvVec = vdupq_n_u32(rv);
-
-  int32x4_t z;
-  memset(&z, 0, sizeof(z));
-
-  uint16x4_t conflicting[4];
-  for (int i = 0; i < 4; ++i) {
-    conflicting[i] =
-        vmovn_u32(vcgtq_s32(vreinterpretq_s32_u32(vsubq_u32(w4[i], rvVec)), z));
-  }
-  auto combined =
-      vcombine_u8(vmovn_u16(vcombine_u16(conflicting[0], conflicting[1])),
-                  vmovn_u16(vcombine_u16(conflicting[2], conflicting[3])));
-
   uint64_t compared = vget_lane_u64(
-      vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(combined), 4)), 0);
+      vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(conflictMask16(
+                                          self->childMaxVersion, readVersion)),
+                                      4)),
+      0);
 
   return !(compared & mask) && firstRangeOk;
 
@@ -2375,10 +2417,9 @@ bool checkMaxBetweenExclusiveImpl(Node16 *n, int begin, int end,
   auto inBounds = [&](unsigned c) { return c - shiftAmount < shiftUpperBound; };
 
   uint32_t mask = 0;
-  for (int i = 0; i < 16; ++i) {
+  for (int i = 0; i < self->numChildren; ++i) {
     mask |= inBounds(self->index[i]) << i;
   }
-  mask &= (1 << self->numChildren) - 1;
   if (!mask) {
     return true;
   }
@@ -2386,7 +2427,7 @@ bool checkMaxBetweenExclusiveImpl(Node16 *n, int begin, int end,
   const bool firstRangeOk =
       !child->entryPresent || child->entry.rangeVersion <= readVersion;
   uint32_t compared = 0;
-  for (int i = 0; i < 16; ++i) {
+  for (int i = 0; i < self->numChildren; ++i) {
     compared |= (self->childMaxVersion[i] > readVersion) << i;
   }
   return !(compared & mask) && firstRangeOk;
@@ -2745,10 +2786,39 @@ checkMaxBetweenExclusiveImpl<true>(Node256 *n, int begin, int end,
   return self;
 }
 
+#if SHOW_MEMORY
+// Length of the key leading to `child`, given the length of the key leading to
+// its parent. Consistent with `getSearchPathLength`.
+int64_t childSearchPathLength(int64_t parentSearchPathLength, Node *child) {
+  return parentSearchPathLength + 1 + child->partialKeyLen;
+}
+#endif
+
 void eraseTree(Node *root, WriteContext *writeContext) {
   Arena arena;
   auto toFree = vector<Node *>(arena);
   toFree.push_back(root);
+  // Length of the key leading to each node in `toFree`, on a parallel stack.
+  // A node is released before its children are visited, so the `parent`
+  // pointers that `getSearchPathLength` walks are dangling by the time a queued
+  // node is popped. Carry the lengths along instead. Only the SHOW_MEMORY
+  // accounting needs them.
+#if SHOW_MEMORY
+  auto searchPathLengths = vector<int64_t>(arena);
+  searchPathLengths.push_back(getSearchPathLength(root));
+#endif
+
+  // Queue `child` for release, where `parentSearchPathLength` is the length of
+  // the key leading to its parent.
+  auto queueChild = [&](Node *child, int64_t parentSearchPathLength) {
+    toFree.push_back(child);
+#if SHOW_MEMORY
+    searchPathLengths.push_back(
+        childSearchPathLength(parentSearchPathLength, child));
+#else
+    (void)parentSearchPathLength;
+#endif
+  };
 
   while (toFree.size() > 0) {
     auto *n = toFree.back();
@@ -2756,7 +2826,13 @@ void eraseTree(Node *root, WriteContext *writeContext) {
     writeContext->accum.entries_erased += n->entryPresent;
     ++writeContext->accum.nodes_released;
 
-    removeKey(n);
+#if SHOW_MEMORY
+    const int64_t searchPathLength = searchPathLengths.back();
+    searchPathLengths.pop_back();
+#else
+    const int64_t searchPathLength = 0;
+#endif
+    removeKey(n, searchPathLength);
 
     switch (n->getType()) {
     case Type_Node0: {
@@ -2766,28 +2842,39 @@ void eraseTree(Node *root, WriteContext *writeContext) {
     case Type_Node3: {
       auto *n3 = static_cast<Node3 *>(n);
       for (int i = 0; i < n3->numChildren; ++i) {
-        toFree.push_back(n3->children[i]);
+        queueChild(n3->children[i], searchPathLength);
       }
       writeContext->release(n3);
     } break;
     case Type_Node16: {
       auto *n16 = static_cast<Node16 *>(n);
       for (int i = 0; i < n16->numChildren; ++i) {
-        toFree.push_back(n16->children[i]);
+        queueChild(n16->children[i], searchPathLength);
       }
       writeContext->release(n16);
     } break;
     case Type_Node48: {
       auto *n48 = static_cast<Node48 *>(n);
       for (int i = 0; i < n48->numChildren; ++i) {
-        toFree.push_back(n48->children[i]);
+        queueChild(n48->children[i], searchPathLength);
       }
       writeContext->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]; });
+#if SHOW_MEMORY
+      auto *outSearchPathLengths =
+          searchPathLengths.unsafePrepareAppend(n256->numChildren).data();
+#endif
+      n256->bitSet.forEachSet([&](int i) {
+        Node *child = n256->children[i];
+        *out++ = child;
+#if SHOW_MEMORY
+        *outSearchPathLengths++ =
+            childSearchPathLength(searchPathLength, child);
+#endif
+      });
       assert(out == toFree.end());
       writeContext->release(n256);
     } break;
@@ -2884,7 +2971,11 @@ horizontalMax16(InternalVersionT *vs, InternalVersionT) {
   return max;
 }
 
-// Precondition: `node->entryPresent`, and node is not the root
+/** @pre `node` is not the root, `node->entryPresent`, and `writeContext !=
+ * nullptr`.
+ * @post `maxVersion(node)` is the max of `writeContext->zero` and the max
+ * version of all logical keys starting with the search path of `node`.
+ */
 void fixupMaxVersion(Node *node, WriteContext *writeContext) {
   assert(node->parent);
   InternalVersionT max;
@@ -2895,9 +2986,13 @@ void fixupMaxVersion(Node *node, WriteContext *writeContext) {
     break;
   case Type_Node3: {
     auto *self3 = static_cast<Node3 *>(node);
-    max = std::max(max,
-                   horizontalMaxUpTo16(self3->childMaxVersion,
-                                       writeContext->zero, self3->numChildren));
+    assume(1 <= self3->numChildren);
+    assume(self3->numChildren <= 3);
+    InternalVersionT m = self3->childMaxVersion[0];
+    for (int i = 1; i < self3->numChildren; ++i) {
+      m = std::max(m, self3->childMaxVersion[i]);
+    }
+    max = std::max(max, m);
   } break;
   case Type_Node16: {
     auto *self16 = static_cast<Node16 *>(node);
@@ -2913,8 +3008,7 @@ void fixupMaxVersion(Node *node, WriteContext *writeContext) {
   } break;
   case Type_Node256: {
     auto *self256 = static_cast<Node256 *>(node);
-    max = std::max(
-        max, horizontalMax16(self256->childMaxVersion, writeContext->zero));
+    max = std::max(max, horizontalMax16(self256->maxOfMax, writeContext->zero));
   } break;
   default:                   // GCOVR_EXCL_LINE
     __builtin_unreachable(); // GCOVR_EXCL_LINE
@@ -3864,17 +3958,17 @@ PRESERVE_NONE void right_side_iter(Job *job, Context *context) {
 void Job::init(const ConflictSet::ReadRange *read, ConflictSet::Result *result,
                Node *root, int64_t oldestVersionFullPrecision) {
   auto begin = TrivialSpan(read->begin.p, read->begin.len);
-  auto end = TrivialSpan(read->end.p, read->end.len);
   if (read->readVersion < oldestVersionFullPrecision) [[unlikely]] {
     *result = ConflictSet::TooOld;
     continuation = complete;
-  } else if (end.size() == 0) {
+  } else if (read->end.len == 0) {
     this->begin = begin;
     this->n = root;
     this->readVersion = InternalVersionT(read->readVersion);
     this->result = result;
     continuation = check::point_read_state_machine::begin;
   } else {
+    auto end = TrivialSpan(read->end.p, read->end.len);
     this->begin = begin;
     this->end = end;
     this->n = root;
@@ -4861,7 +4955,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
         auto *results2 = new (arena) Result[count];
         check::Context context2;
         useSequential(reads, results2, count, context2);
-        assert(memcmp(result, results2, count) == 0);
+        assert(memcmp(result, results2, count * sizeof(Result)) == 0);
         assert(context.readContext == context2.readContext);
 #endif
       }
@@ -4962,15 +5056,17 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
         }
         if (e != nullptr) {
           while (e->releaseDeferred) {
-            e = b->forwardTo;
+            e = e->forwardTo;
           }
         }
         fprintf(stderr, "search path: %s, begin: %s\n",
                 getSearchPathPrintable(b).c_str(),
                 printable(context.results[i].remaining).c_str());
-        fprintf(stderr, "search path: %s, end: %s\n",
-                getSearchPathPrintable(e).c_str(),
-                printable(context.results[i].endRemaining).c_str());
+        if (e != nullptr) {
+          fprintf(stderr, "search path: %s, end: %s\n",
+                  getSearchPathPrintable(e).c_str(),
+                  printable(context.results[i].endRemaining).c_str());
+        }
       }
 #endif
 
@@ -5046,8 +5142,8 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
       for (int i = 0; i < count; ++i) {
         const auto &w = writes[i];
         auto begin = TrivialSpan(w.begin.p, w.begin.len);
-        auto end = TrivialSpan(w.end.p, w.end.len);
         if (w.end.len > 0) {
+          auto end = TrivialSpan(w.end.p, w.end.len);
           addWriteRange(rootParent->children[0], begin, end,
                         InternalVersionT(writeVersion), &writeContext);
         } else {
@@ -5072,7 +5168,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
 
     if (oldestExtantVersion < writeVersion - kMaxCorrectVersionWindow)
         [[unlikely]] {
-      if (writeVersion > newestVersionFullPrecision + kNominalVersionWindow) {
+      if (writeVersion - kNominalVersionWindow > newestVersionFullPrecision) {
         eraseTree(rootParent->children[0], &writeContext);
         init(writeVersion - kNominalVersionWindow);
       }
@@ -5450,8 +5546,9 @@ void internal_addWrites(ConflictSet::Impl *impl,
   impl->totalBytes += mallocBytesDelta;
   impl->memory_bytes.set(impl->totalBytes);
 #if SHOW_MEMORY
+  // Testing-only cross-check of the accounting.
   if (impl->totalBytes != mallocBytes) {
-    abort();
+    abort(); // GCOVR_EXCL_LINE
   }
 #endif
 }
@@ -5462,8 +5559,9 @@ void internal_setOldestVersion(ConflictSet::Impl *impl, int64_t oldestVersion) {
   impl->totalBytes += mallocBytesDelta;
   impl->memory_bytes.set(impl->totalBytes);
 #if SHOW_MEMORY
+  // Testing-only cross-check of the accounting.
   if (impl->totalBytes != mallocBytes) {
-    abort();
+    abort(); // GCOVR_EXCL_LINE
   }
 #endif
 }
@@ -5679,13 +5777,13 @@ std::string getPartialKeyPrintable(Node *n) {
 }
 
 std::string strinc(std::string_view str, bool &ok) {
-  int index;
-  for (index = str.size() - 1; index >= 0; index--)
-    if ((uint8_t &)(str[index]) != 255)
+  int index = static_cast<int>(str.size()) - 1;
+  for (; index >= 0; index--)
+    if (static_cast<uint8_t>(str[index]) != 255)
       break;
 
   // Must not be called with a string that consists only of zero or more
-  // '\xff' bytes.
+  // '\xff' bytes, and the empty string has no successor.
   if (index < 0) {
     ok = false;
     return {};
@@ -5693,7 +5791,8 @@ std::string strinc(std::string_view str, bool &ok) {
   ok = true;
 
   auto r = std::string(str.substr(0, index + 1));
-  ((uint8_t &)r[r.size() - 1])++;
+  auto &last = r[r.size() - 1];
+  last = static_cast<char>(static_cast<uint8_t>(last) + 1);
   return r;
 }
 
@@ -5928,6 +6027,14 @@ checkMaxVersion(Node *root, Node *node, InternalVersionT oldestVersion,
 
 } // namespace
 
+// The SHOW_MEMORY instrumentation. Each of these functions is defined exactly
+// once, with the accounting compiled out when SHOW_MEMORY is disabled. Defining
+// them once (rather than defining no-op stubs when SHOW_MEMORY is disabled)
+// keeps every configuration of this file reporting each function on the same
+// line, which merging coverage from objects compiled with different options
+// requires: gcovr, merging the SHOW_MEMORY test with the rest of the build,
+// otherwise aborts with "Got function ... on multiple lines". The optimizer
+// removes the resulting empty calls.
 #if SHOW_MEMORY
 
 int64_t nodeBytes = 0;
@@ -5956,7 +6063,10 @@ int64_t getNodeSize(struct Node *n) {
   }
 }
 
+#endif
+
 int64_t getSearchPathLength(Node *n) {
+#if SHOW_MEMORY
   assert(n != nullptr);
   int64_t result = 0;
   for (;;) {
@@ -5968,9 +6078,14 @@ int64_t getSearchPathLength(Node *n) {
     n = n->parent;
   }
   return result;
+#else
+  (void)n;
+  return 0;
+#endif
 }
 
 void addNode(Node *n) {
+#if SHOW_MEMORY
   nodeBytes += getNodeSize(n);
   partialCapacityBytes += n->getCapacity();
   if (nodeBytes > peakNodeBytes) {
@@ -5979,14 +6094,22 @@ void addNode(Node *n) {
   if (partialCapacityBytes > peakPartialCapacityBytes) {
     peakPartialCapacityBytes = partialCapacityBytes;
   }
+#else
+  (void)n;
+#endif
 }
 
 void removeNode(Node *n) {
+#if SHOW_MEMORY
   nodeBytes -= getNodeSize(n);
   partialCapacityBytes -= n->getCapacity();
+#else
+  (void)n;
+#endif
 }
 
 void addKey(Node *n) {
+#if SHOW_MEMORY
   if (!n->entryPresent) {
     ++totalKeys;
     keyBytes += getSearchPathLength(n);
@@ -5997,15 +6120,36 @@ void addKey(Node *n) {
       peakKeyBytes = keyBytes;
     }
   }
+#else
+  (void)n;
+#endif
 }
 
 void removeKey(Node *n) {
+#if SHOW_MEMORY
   if (n->entryPresent) {
     --totalKeys;
     keyBytes -= getSearchPathLength(n);
   }
+#else
+  (void)n;
+#endif
 }
 
+void removeKey(Node *n, int64_t searchPathLength) {
+#if SHOW_MEMORY
+  if (n->entryPresent) {
+    --totalKeys;
+    keyBytes -= searchPathLength;
+  }
+#else
+  (void)n;
+  (void)searchPathLength;
+#endif
+}
+
+#if SHOW_MEMORY
+
 struct __attribute__((visibility("default"))) PeakPrinter {
   ~PeakPrinter() {
     printf("--- radix_tree ---\n");

Dockerfile

@@ -1,46 +0,0 @@
-FROM ubuntu:rolling
-
-WORKDIR /tmp
-ENV HOME=/tmp
-RUN chmod -R 777 /tmp
-
-# Install apt dependencies
-RUN apt-get update
-RUN apt-get upgrade -y
-RUN TZ=America/Los_Angeles DEBIAN_FRONTEND=noninteractive apt-get install -y \
-    biber \
-    build-essential \
-    ccache \
-    clang \
-    cmake \
-    curl \
-    devscripts \
-    gcovr \
-    git \
-    latexmk \
-    libc6-dbg \
-    llvm \
-    mold \
-    ninja-build \
-    nodejs \
-    pre-commit \
-    python3-requests \
-    rpm \
-    texlive-bibtex-extra \
-    texlive-fonts-recommended \
-    texlive-latex-extra \
-    texlive-pictures \
-    valgrind \
-    wget \
-    zstd
-
-# MinIO client, for uploading build artifacts
-RUN curl -Ls "https://dl.min.io/client/mc/release/linux-$(dpkg --print-architecture)/mc" \
-    -o /usr/local/bin/mc && chmod +x /usr/local/bin/mc
-
-ENV CC=clang
-ENV CXX=clang++
-
-# Try to have all the pre-commit hooks we'll need already initialized
-COPY .pre-commit-config.yaml /tmp/
-RUN git init && pre-commit install-hooks

HashTable.cpp

@@ -59,6 +59,12 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     while (keyUpdates > 0) {
       if (iter == map.end()) {
         iter = map.begin();
+        // The map is empty, so there are no more entries to garbage-collect.
+        // Break rather than spinning forever (keyUpdates is charged at 2x
+        // the entry count, so it may still be > 0 after emptying the map).
+        if (iter == map.end()) {
+          break;
+        }
       }
       for (; iter != map.end(); --keyUpdates) {
         if (iter->second <= oldestVersion) {
@@ -96,7 +102,9 @@ void ConflictSet::setOldestVersion(int64_t oldestVersion) {
   return impl->setOldestVersion(oldestVersion);
 }
 
-int64_t ConflictSet::getBytes() const { return -1; }
+// The hash_table implementation does not track memory usage, so return 0 to
+// satisfy the API contract that getBytes() returns a non-negative value.
+int64_t ConflictSet::getBytes() const { return 0; }
 
 void ConflictSet::getMetricsV1(MetricsV1 **metrics, int *count) const {
   *metrics = nullptr;
@@ -161,7 +169,10 @@ __attribute__((__visibility__("default"))) void ConflictSet_destroy(void *cs) {
 }
 __attribute__((__visibility__("default"))) int64_t
 ConflictSet_getBytes(void *cs) {
-  using Impl = ConflictSet::Impl;
-  return -1;
+  (void)cs;
+  // The hash_table implementation does not track memory usage, so return 0 to
+  // satisfy the API contract that ConflictSet_getBytes returns a non-negative
+  // value.
+  return 0;
 }
 }

Internal.h

@@ -23,7 +23,9 @@ using namespace weaselab;
 #include <callgrind.h>
 
 #define DEBUG_VERBOSE 0
+#ifndef SHOW_MEMORY
 #define SHOW_MEMORY 0
+#endif
 
 // std::span is not trivially constructible. We want a span that leaves its
 // members uninitialized for performance reasons.
@@ -96,6 +98,11 @@ static_assert(std::is_trivial_v<TrivialSpan>);
 #if SHOW_MEMORY
 inline int64_t mallocBytes = 0;
 inline int64_t peakMallocBytes = 0;
+// Defined in ConflictSet.cpp
+extern int64_t nodeBytes;
+extern int64_t partialCapacityBytes;
+extern int64_t totalKeys;
+extern int64_t keyBytes;
 #endif
 
 inline thread_local int64_t mallocBytesDelta = 0;

LongestCommonPrefix.h

@@ -96,7 +96,7 @@ inline int firstNeqStride(const uint8_t *ap, const uint8_t *bp) {
 // GCOVR_EXCL_START
 #if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
 __attribute__((target("avx512f,avx512bw"))) inline int
-longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
+longestCommonPrefixLong(const uint8_t *ap, const uint8_t *bp, int cl) {
   int i = 0;
   int end = cl & ~63;
   while (i < end) {
@@ -123,7 +123,7 @@ __attribute__((target("default")))
 // GCOVR_EXCL_STOP
 
 inline int
-longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
+longestCommonPrefixLong(const uint8_t *ap, const uint8_t *bp, int cl) {
   if (!(cl >= 0)) {
     __builtin_unreachable(); // GCOVR_EXCL_LINE
   }
@@ -182,4 +182,72 @@ longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
   }
 
   return i;
-}
+}
+// Use in-bounds native-word loads for short common lengths. The second load
+// overlaps the first; because the first load matched, any mismatch found in
+// the overlap is impossible and the returned absolute index is still correct.
+// Precondition: cl <= 2 * sizeof(uint64_t)
+__attribute__((always_inline)) static inline int
+longestCommonPrefixShort(const uint8_t *ap, const uint8_t *bp, int cl) {
+  assert(cl <= 2 * int(sizeof(uint64_t)));
+  if (cl >= 8) {
+    uint64_t a0;
+    uint64_t b0;
+    memcpy(&a0, ap, sizeof(a0));
+    memcpy(&b0, bp, sizeof(b0));
+    if (uint64_t mismatched = a0 ^ b0) {
+      return std::countr_zero(mismatched) / 8;
+    }
+    if (cl == 8) {
+      return 8;
+    }
+
+    uint64_t a1;
+    uint64_t b1;
+    memcpy(&a1, ap + cl - 8, sizeof(a1));
+    memcpy(&b1, bp + cl - 8, sizeof(b1));
+    if (uint64_t mismatched = a1 ^ b1) {
+      return cl - 8 + std::countr_zero(mismatched) / 8;
+    }
+    return cl;
+  }
+
+  if (cl >= 4) {
+    uint32_t a0;
+    uint32_t b0;
+    memcpy(&a0, ap, sizeof(a0));
+    memcpy(&b0, bp, sizeof(b0));
+    if (uint32_t mismatched = a0 ^ b0) {
+      return std::countr_zero(mismatched) / 8;
+    }
+    if (cl == 4) {
+      return 4;
+    }
+
+    uint32_t a1;
+    uint32_t b1;
+    memcpy(&a1, ap + cl - 4, sizeof(a1));
+    memcpy(&b1, bp + cl - 4, sizeof(b1));
+    if (uint32_t mismatched = a1 ^ b1) {
+      return cl - 4 + std::countr_zero(mismatched) / 8;
+    }
+    return cl;
+  }
+
+  int i = 0;
+  while (i < cl && ap[i] == bp[i]) {
+    ++i;
+  }
+  return i;
+}
+
+__attribute__((always_inline)) static inline int
+longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
+  if (!(cl >= 0)) {
+    __builtin_unreachable(); // GCOVR_EXCL_LINE
+  }
+  if (cl <= 2 * int(sizeof(uint64_t))) {
+    return longestCommonPrefixShort(ap, bp, cl);
+  }
+  return longestCommonPrefixLong(ap, bp, cl);
+}

README.md

@@ -7,10 +7,10 @@ Hardware for all benchmarks is an AMD Ryzen 9 7900 with (2x32GB) 5600MT/s CL28-3
 ```
 $ clang++ --version
 
-Ubuntu clang version 20.0.0 (++20241120082228+86734c857724-1~exp1~20241120202359.554)
+Ubuntu clang version 21.1.8 (6ubuntu1)
 Target: x86_64-pc-linux-gnu
 Thread model: posix
-InstalledDir: /usr/lib/llvm-20/bin
+InstalledDir: /usr/lib/llvm-21/bin
 ```
 
 # Microbenchmark
@@ -19,30 +19,30 @@ InstalledDir: /usr/lib/llvm-20/bin
 
 |               ns/op |                op/s |    err% |          ins/op |          cyc/op |    IPC |         bra/op |   miss% |     total | benchmark
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|              161.29 |        6,200,056.17 |    0.1% |        3,014.03 |          831.04 |  3.627 |         504.59 |    0.0% |      1.93 | `point reads`
-|              158.32 |        6,316,160.64 |    0.1% |        2,954.16 |          815.80 |  3.621 |         490.17 |    0.0% |      1.89 | `prefix reads`
-|              237.39 |        4,212,409.50 |    0.2% |        3,592.41 |        1,233.96 |  2.911 |         629.31 |    0.0% |      2.84 | `range reads`
-|              442.11 |        2,261,878.94 |    0.0% |        4,450.57 |        2,314.25 |  1.923 |         707.92 |    2.1% |      5.28 | `point writes`
-|              439.89 |        2,273,308.53 |    0.1% |        4,410.22 |        2,302.29 |  1.916 |         694.74 |    2.1% |      5.25 | `prefix writes`
-|              290.96 |        3,436,936.78 |    0.0% |        2,315.38 |        1,528.68 |  1.515 |         396.69 |    3.3% |      3.49 | `range writes`
-|              476.93 |        2,096,762.02 |    0.6% |        6,999.33 |        2,484.94 |  2.817 |       1,251.73 |    1.3% |      0.06 | `monotonic increasing point writes`
-|          131,736.57 |            7,590.91 |    1.1% |      807,444.50 |      704,941.71 |  1.145 |     144,584.60 |    0.9% |      0.01 | `worst case for radix tree`
-|               45.50 |       21,978,369.95 |    1.1% |          902.00 |          232.36 |  3.882 |         132.00 |    0.0% |      0.01 | `create and destroy`
+|              165.93 |        6,026,520.04 |    0.1% |        3,107.03 |          610.42 |  5.090 |         558.59 |    0.0% |      1.98 | `point reads`
+|              163.08 |        6,132,046.57 |    0.1% |        3,036.76 |          599.68 |  5.064 |         539.35 |    0.0% |      1.95 | `prefix reads`
+|              250.26 |        3,995,834.21 |    0.1% |        3,722.71 |          920.25 |  4.045 |         692.00 |    0.0% |      2.99 | `range reads`
+|              359.64 |        2,780,581.63 |    0.1% |        4,523.64 |        1,322.63 |  3.420 |         720.22 |    2.1% |      4.30 | `point writes`
+|              358.20 |        2,791,726.93 |    0.1% |        4,484.57 |        1,317.23 |  3.405 |         705.00 |    2.0% |      4.28 | `prefix writes`
+|              193.91 |        5,157,081.29 |    0.1% |        2,224.10 |          713.33 |  3.118 |         377.17 |    3.3% |      2.32 | `range writes`
+|              422.03 |        2,369,513.38 |    1.1% |        6,855.96 |        1,551.55 |  4.419 |       1,227.82 |    1.4% |      0.05 | `monotonic increasing point writes`
+|          142,467.00 |            7,019.17 |    3.8% |      812,047.33 |      508,535.40 |  1.597 |     151,246.50 |    1.0% |      0.01 | `worst case for radix tree`
+|               37.66 |       26,554,217.74 |    0.8% |          701.00 |          138.80 |  5.050 |         102.00 |    0.0% |      0.01 | `create and destroy`
 
 ## Radix tree (this implementation)
 
 
 |               ns/op |                op/s |    err% |          ins/op |          cyc/op |    IPC |         bra/op |   miss% |     total | benchmark
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|               12.36 |       80,885,626.43 |    0.2% |          243.56 |           63.62 |  3.828 |          31.07 |    0.6% |      0.15 | `point reads`
-|               14.18 |       70,502,196.81 |    0.1% |          297.72 |           73.13 |  4.071 |          40.31 |    0.5% |      0.17 | `prefix reads`
-|               33.44 |       29,901,623.04 |    0.1% |          767.90 |          172.42 |  4.454 |         101.32 |    0.2% |      0.40 | `range reads`
-|               19.48 |       51,342,564.70 |    0.3% |          374.45 |          100.43 |  3.728 |          48.92 |    0.5% |      0.23 | `point writes`
-|               37.46 |       26,694,471.44 |    0.1% |          672.00 |          193.14 |  3.479 |         101.28 |    0.3% |      0.45 | `prefix writes`
-|               38.78 |       25,784,784.34 |    0.0% |          738.26 |          199.93 |  3.693 |         111.59 |    0.1% |      0.47 | `range writes`
-|               76.05 |       13,148,995.74 |    0.7% |        1,450.77 |          397.16 |  3.653 |         275.72 |    0.0% |      0.01 | `monotonic increasing point writes`
-|          286,920.33 |            3,485.29 |    0.4% |    4,117,948.00 |    1,521,352.00 |  2.707 |     714,833.00 |    0.1% |      0.01 | `worst case for radix tree`
-|               95.66 |       10,453,798.72 |    0.5% |        1,986.00 |          495.04 |  4.012 |         315.00 |    0.0% |      0.01 | `create and destroy`
+|               12.25 |       81,664,669.29 |    0.4% |          233.49 |           44.94 |  5.196 |          31.23 |    0.7% |      0.15 | `point reads`
+|               13.35 |       74,903,950.49 |    0.1% |          271.75 |           49.10 |  5.534 |          40.21 |    0.4% |      0.16 | `prefix reads`
+|               31.37 |       31,874,583.62 |    0.1% |          700.39 |          115.43 |  6.067 |         100.83 |    0.2% |      0.37 | `range reads`
+|               19.55 |       51,153,057.80 |    0.1% |          366.51 |           71.90 |  5.098 |          47.92 |    0.5% |      0.23 | `point writes`
+|               36.92 |       27,088,344.17 |    0.1% |          645.11 |          135.72 |  4.753 |         100.31 |    0.4% |      0.44 | `prefix writes`
+|               36.84 |       27,143,553.08 |    0.1% |          691.04 |          135.42 |  5.103 |         114.01 |    0.1% |      0.45 | `range writes`
+|               74.62 |       13,400,467.72 |    0.9% |        1,418.37 |          274.20 |  5.173 |         273.39 |    0.1% |      0.01 | `monotonic increasing point writes`
+|          352,671.00 |            2,835.50 |   15.3% |    3,687,535.00 |    1,287,119.00 |  2.865 |     712,030.00 |    0.1% |      0.01 | :wavy_dash: `worst case for radix tree` (Unstable with ~1.0 iters. Increase `minEpochIterations` to e.g. 10)
+|               96.63 |       10,348,729.42 |    1.2% |        1,766.00 |          355.53 |  4.967 |         283.00 |    0.0% |      0.01 | `create and destroy`
 
 # "Real data" test
 
@@ -51,13 +51,13 @@ Point queries only. Gc ratio is the ratio of time spent doing garbage collection
 ## skip list
 
 ```
-Check: 4.53508 seconds, 371.81 MB/s, Add: 3.81222 seconds, 150.919 MB/s, Gc ratio: 33.66%, Peak idle memory: 5.61007e+06
+Check: 4.69875 seconds, 347.017 MB/s, Add: 3.39475 seconds, 165.153 MB/s, Gc ratio: 38.7414%, Peak idle memory: 5.51852e+06
 ```
 
 ## radix tree
 
 ```
-Check: 0.957735 seconds, 1760.6 MB/s, Add: 1.19942 seconds, 479.678 MB/s, Gc ratio: 38.6069%, Peak idle memory: 2.05667e+06
+Check: 0.976745 seconds, 1669.37 MB/s, Add: 1.20464 seconds, 465.411 MB/s, Gc ratio: 39.4158%, Peak idle memory: 2.0226e+06
 ```
 
 ## hash table
@@ -65,6 +65,6 @@ Check: 0.957735 seconds, 1760.6 MB/s, Add: 1.19942 seconds, 479.678 MB/s, Gc rat
 (The hash table implementation doesn't work on range queries, and its purpose is to provide an idea of how fast point queries can be)
 
 ```
-Check: 0.804598 seconds, 2095.69 MB/s, Add: 0.671221 seconds, 857.147 MB/s, Gc ratio: 35.0034%, Peak idle memory: 0
+Check: 0.864589 seconds, 1885.92 MB/s, Add: 0.634604 seconds, 883.47 MB/s, Gc ratio: 41.184%, Peak idle memory: 0
 ```
 

RealDataBench.cpp

@@ -1,5 +1,6 @@
 #include <ConflictSet.h>
 
+#include <algorithm>
 #include <cerrno>
 #include <chrono>
 #include <cstdio>
@@ -77,10 +78,10 @@ int main(int argc, const char **argv) {
       begin = end + 1;
       end = (uint8_t *)memchr(begin, '\n', size);
 
-      if (line.size() > 0 && line[0] == 'P') {
-        write = line.subspan(2, line.size());
-      } else if (line.size() > 0 && line[0] == 'L') {
-        reads.push_back(line.subspan(2, line.size()));
+      if (line.size() >= 2 && line[0] == 'P') {
+        write = line.subspan(2, line.size() - 2);
+      } else if (line.size() >= 2 && line[0] == 'L') {
+        reads.push_back(line.subspan(2, line.size() - 2));
       } else if (line.empty()) {
         {
           readRanges.resize(reads.size());
@@ -90,7 +91,7 @@ int main(int argc, const char **argv) {
             iter->begin.len = read.size();
             checkBytes += read.size();
             iter->end.len = 0;
-            iter->readVersion = version - 100;
+            iter->readVersion = std::max<int64_t>(0, version - 100);
             ++iter;
           }
         }
@@ -121,7 +122,7 @@ int main(int argc, const char **argv) {
         }
 
         timer = now();
-        cs.setOldestVersion(version - 10000);
+        cs.setOldestVersion(std::max<int64_t>(0, version - 10000));
         gcTime += now() - timer;
       }
     }

ServerBench.cpp

@@ -12,6 +12,7 @@
 #include <sys/ioctl.h>
 #include <sys/resource.h>
 #include <sys/socket.h>
+#include <sys/syscall.h>
 #include <sys/types.h>
 #include <sys/uio.h>
 #include <thread>

ShowMemoryTest.cpp

@@ -0,0 +1,115 @@
+// Exercises the SHOW_MEMORY instrumentation, which the other tests don't
+// compile with (this target passes -DSHOW_MEMORY=1). The instrumentation has
+// its own consistency checks, which abort on failure, and asserts are enabled
+// by TEST_FLAGS. Only one ConflictSet may exist at a time, since the
+// accounting is process-wide.
+#include "ConflictSet.h"
+#include "Internal.h"
+
+#include <cassert>
+#include <cstdint>
+#include <string>
+
+using namespace weaselab;
+
+namespace {
+
+void writePoint(ConflictSet &cs, const std::string &key, int64_t writeVersion) {
+  ConflictSet::WriteRange w{{(const uint8_t *)key.data(), int(key.size())},
+                            {nullptr, 0}};
+  cs.addWrites(&w, 1, writeVersion);
+}
+
+ConflictSet::Result checkPoint(const ConflictSet &cs, const std::string &key,
+                               int64_t readVersion) {
+  ConflictSet::ReadRange r{{(const uint8_t *)key.data(), int(key.size())},
+                           {nullptr, 0},
+                           readVersion};
+  ConflictSet::Result result;
+  cs.check(&r, &result, 1);
+  return result;
+}
+
+// Everything a ConflictSet allocated must have been accounted for and released
+// by the time it's destroyed.
+void checkFullyReleased() {
+  assert(nodeBytes == 0);
+  assert(partialCapacityBytes == 0);
+  assert(totalKeys == 0);
+  assert(keyBytes == 0);
+  assert(mallocBytes == 0);
+}
+
+} // namespace
+
+int main() {
+  // Splitting a node replaces it, which defers the release of the old node.
+  // Releasing a deferred node must account for its capacity, and the key it
+  // held must be accounted for without walking `parent` pointers of nodes that
+  // have already been released.
+  {
+    ConflictSet cs{0};
+    writePoint(cs, "a", 1);
+    writePoint(cs, "ab", 2);
+    assert(checkPoint(cs, "a", 0) == ConflictSet::Conflict);
+    assert(checkPoint(cs, "ab", 0) == ConflictSet::Conflict);
+    assert(checkPoint(cs, "b", 0) == ConflictSet::Commit);
+    assert(nodeBytes > 0);
+    assert(keyBytes > 0);
+  }
+  checkFullyReleased();
+
+  // Exceeding the version window erases the tree mid-workload.
+  {
+    ConflictSet cs{0};
+    writePoint(cs, "abc", 1);
+    const int64_t version = int64_t(1) << 40;
+    writePoint(cs, "abc", version);
+    assert(checkPoint(cs, "abc", 0) == ConflictSet::TooOld);
+    assert(checkPoint(cs, "abc", version) == ConflictSet::Commit);
+  }
+  checkFullyReleased();
+
+  // Erase a tree with many nodes, both mid-workload and at destruction.
+  {
+    ConflictSet cs{0};
+    for (int i = 0; i < 1000; ++i) {
+      writePoint(cs, "prefix/" + std::to_string(i), i + 1);
+    }
+    const int64_t version = int64_t(1) << 40;
+    writePoint(cs, "prefix/1", version);
+    assert(checkPoint(cs, "prefix/1", version) == ConflictSet::Commit);
+
+    for (int i = 0; i < 1000; ++i) {
+      writePoint(cs, "other/" + std::to_string(i), version + i + 1);
+    }
+    assert(checkPoint(cs, "other/999", version) == ConflictSet::Conflict);
+  }
+  checkFullyReleased();
+
+  // Grow a node through every node type, so that the accounting for each node
+  // type's size is exercised, then erase the entries one at a time so that
+  // nodes are released individually rather than as part of a whole tree.
+  {
+    ConflictSet cs{0};
+    std::string key(1, '\0');
+    for (int i = 0; i < 256; ++i) {
+      key[0] = char(i);
+      writePoint(cs, key, i + 1);
+    }
+    for (int i = 0; i < 256; ++i) {
+      key[0] = char(i);
+      assert(checkPoint(cs, key, i) == ConflictSet::Conflict);
+    }
+
+    // Garbage collection erases entries that can only fail with TooOld anyway.
+    // Doing it one version at a time walks the tree, erasing, downsizing and
+    // releasing nodes as it goes.
+    for (int i = 0; i < 256; ++i) {
+      cs.setOldestVersion(i + 1);
+    }
+    key[0] = '\0';
+    assert(checkPoint(cs, key, 0) == ConflictSet::TooOld);
+  }
+  checkFullyReleased();
+}

SkipList.cpp

@@ -38,6 +38,9 @@ std::span<const uint8_t> keyAfter(Arena &arena, std::span<const uint8_t> key) {
 
 std::span<const uint8_t> copyToArena(Arena &arena,
                                      std::span<const uint8_t> key) {
+  if (key.size() == 0) {
+    return std::span<const uint8_t>();
+  }
   auto result = std::span<uint8_t>(new (arena) uint8_t[key.size()], key.size());
   memcpy(result.data(), key.data(), key.size());
   return result;
@@ -97,7 +100,7 @@ force_inline bool getCharacter(const KeyInfo &ki, int character,
 
 bool operator<(const KeyInfo &lhs, const KeyInfo &rhs) {
   int i = std::min(lhs.key.size(), rhs.key.size());
-  int c = memcmp(lhs.key.data(), rhs.key.data(), i);
+  int c = i > 0 ? memcmp(lhs.key.data(), rhs.key.data(), i) : 0;
   if (c != 0)
     return c < 0;
 
@@ -287,7 +290,8 @@ private:
 
   static force_inline bool less(const uint8_t *a, int aLen, const uint8_t *b,
                                 int bLen) {
-    int c = memcmp(a, b, std::min(aLen, bLen));
+    int n = std::min(aLen, bLen);
+    int c = n > 0 ? memcmp(a, b, n) : 0;
     if (c < 0)
       return true;
     if (c > 0)
@@ -754,7 +758,8 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
 
   void addWrites(const ConflictSet::WriteRange *writes, int count,
                  int64_t writeVersion) {
-    auto points = std::vector<KeyInfo>(count * 2);
+    auto points = std::vector<KeyInfo>();
+    points.reserve(count * 2);
     Arena arena;
 
     for (int r = 0; r < count; r++) {
@@ -814,10 +819,17 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   }
 
   void setOldestVersion(int64_t oldestVersion) {
+    // oldestVersion becomes the max of its previous value and the argument, so
+    // an older version is a no-op. We still fall through in the equal case so
+    // that garbage collection can make progress without advancing the oldest
+    // version.
+    if (oldestVersion < this->oldestVersion) {
+      return;
+    }
+
     // This isn't 100% accurate. It overcounts if you hit the end
     gc_iterations_total.add(keyUpdates);
 
-    assert(oldestVersion >= this->oldestVersion);
     this->oldestVersion = oldestVersion;
     SkipList::Finger finger;
     int temp;

TestDriver.cpp

@@ -35,36 +35,5 @@ int main(int argc, char **argv) {
         break;
       }
     }
-
-    {
-      ConflictSet::MetricsV1 *metrics;
-      int metricsCount;
-      driver1.cs.getMetricsV1(&metrics, &metricsCount);
-      printf("#################### METRICS for ConflictSet 1 for %s "
-             "####################\n",
-             argv[i]);
-      for (int i = 0; i < metricsCount; ++i) {
-        printf("# HELP %s %s\n", metrics[i].name, metrics[i].help);
-        printf("# TYPE %s %s\n", metrics[i].name,
-               metrics[i].type == metrics[i].Counter ? "counter" : "gauge");
-        printf("%s %g\n", metrics[i].name, metrics[i].getValue());
-      }
-      puts("");
-    }
-    {
-      ConflictSet::MetricsV1 *metrics;
-      int metricsCount;
-      driver2.cs.getMetricsV1(&metrics, &metricsCount);
-      printf("#################### METRICS for ConflictSet 2 for %s "
-             "####################\n",
-             argv[i]);
-      for (int i = 0; i < metricsCount; ++i) {
-        printf("# HELP %s %s\n", metrics[i].name, metrics[i].help);
-        printf("# TYPE %s %s\n", metrics[i].name,
-               metrics[i].type == metrics[i].Counter ? "counter" : "gauge");
-        printf("%s %g\n", metrics[i].name, metrics[i].getValue());
-      }
-      puts("");
-    }
   }
 }

build_msan_toolchain.sh

@@ -0,0 +1,43 @@
+#!/bin/bash
+
+set -euxo pipefail
+
+LLVM_VERSION="${LLVM_VERSION:-21}"
+MSAN_PREFIX="${MSAN_PREFIX:-$PWD/msan}"
+JOBS="${JOBS:-$(nproc)}"
+
+cd /tmp
+rm -rf libcxx-msan
+mkdir libcxx-msan
+cd libcxx-msan
+
+git clone --depth=1 "https://github.com/llvm/llvm-project.git" -b "release/${LLVM_VERSION}.x"
+
+cmake -S llvm-project/runtimes -B build_msan \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_CXX_COMPILER="clang++-${LLVM_VERSION}" \
+  -DCMAKE_C_COMPILER="clang-${LLVM_VERSION}" \
+  -DCMAKE_INSTALL_PREFIX="${MSAN_PREFIX}" \
+  -DLLVM_ENABLE_RUNTIMES="libcxx;libcxxabi;libunwind" \
+  -DLLVM_USE_SANITIZER=MemoryWithOrigins
+
+cmake --build build_msan -j"${JOBS}"
+cmake --install build_msan
+
+cmake -S llvm-project/runtimes -B build_nomsan \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_CXX_COMPILER="clang++-${LLVM_VERSION}" \
+  -DCMAKE_C_COMPILER="clang-${LLVM_VERSION}" \
+  -DCMAKE_INSTALL_PREFIX="${MSAN_PREFIX}" \
+  -DLLVM_ENABLE_RUNTIMES="libunwind"
+
+cmake --build build_nomsan -j"${JOBS}"
+cmake --install build_nomsan
+
+VERSION="$(clang-"${LLVM_VERSION}" --version | head -n1 | sed -E 's/.*clang version ([0-9.]+).*/\1/')"
+TARBALL="msan-toolchain-${VERSION}.tar.zst"
+
+tar --zstd -cf "${TARBALL}" -C "${MSAN_PREFIX}" .
+
+TARBALL_PATH="$(pwd)/${TARBALL}"
+echo "Created: ${TARBALL_PATH}"

conflict_set.py

@@ -27,23 +27,37 @@ class Result(enum.Enum):
     TOO_OLD = 2
 
 
-def write(begin: bytes, end: Optional[bytes] = None) -> WriteRange:
-    b = (ctypes.c_ubyte * len(begin)).from_buffer(bytearray(begin))
+def _make_key(buf: bytes) -> tuple[_Key, bytearray]:
+    """Create a _Key and a backing bytearray that must be kept alive."""
+    backing = bytearray(buf)
+    array = (ctypes.c_ubyte * len(backing)).from_buffer(backing)
+    return _Key(array, len(array)), backing
 
+
+def write(begin: bytes, end: Optional[bytes] = None) -> WriteRange:
+    begin_key, begin_buf = _make_key(begin)
     if end is None:
-        e = (ctypes.c_ubyte * 0)()
+        end_key = _Key((ctypes.c_ubyte * 0)(), 0)
+        end_buf = None
     else:
-        e = (ctypes.c_ubyte * len(end)).from_buffer(bytearray(end))
-    return WriteRange(_Key(b, len(b)), _Key(e, len(e)))
+        end_key, end_buf = _make_key(end)
+    result = WriteRange(begin_key, end_key)
+    result._begin_buf = begin_buf
+    result._end_buf = end_buf
+    return result
 
 
 def read(version: int, begin: bytes, end: Optional[bytes] = None) -> ReadRange:
-    b = (ctypes.c_ubyte * len(begin)).from_buffer(bytearray(begin))
+    begin_key, begin_buf = _make_key(begin)
     if end is None:
-        e = (ctypes.c_ubyte * 0)()
+        end_key = _Key((ctypes.c_ubyte * 0)(), 0)
+        end_buf = None
     else:
-        e = (ctypes.c_ubyte * len(end)).from_buffer(bytearray(end))
-    return ReadRange(_Key(b, len(b)), _Key(e, len(e)), version)
+        end_key, end_buf = _make_key(end)
+    result = ReadRange(begin_key, end_key, version)
+    result._begin_buf = begin_buf
+    result._end_buf = end_buf
+    return result
 
 
 class ConflictSet:

conflict_set_cxx_api_test.cpp

@@ -2,6 +2,7 @@
 
 #include <cassert>
 #include <cstdio>
+#include <limits>
 
 using namespace weaselab;
 
@@ -23,6 +24,28 @@ int main(void) {
   int64_t bytes = cs.getBytes();
   assert(bytes > 0);
 
+  // Regression: a write just below INT64_MAX must not overflow the nominal
+  // version window comparison and erase entries that reads still observe.
+  // Versions are only required to be >= 0, so this is valid input.
+  {
+    const int64_t imax = std::numeric_limits<int64_t>::max();
+    const int64_t v0 = imax - int64_t(1000000000);
+    ConflictSet nearMax(0);
+    ConflictSet::WriteRange w2;
+    w2.begin.p = (const uint8_t *)"key";
+    w2.begin.len = 3;
+    w2.end.len = 0;
+    nearMax.addWrites(&w2, 1, v0);
+    nearMax.addWrites(&w2, 0, imax);
+    ConflictSet::ReadRange r2;
+    r2.begin.p = (const uint8_t *)"key";
+    r2.begin.len = 3;
+    r2.end.len = 0;
+    r2.readVersion = v0 - 1;
+    nearMax.check(&r2, &result, 1);
+    assert(result == ConflictSet::Conflict);
+  }
+
   ConflictSet::MetricsV1 *metrics;
   int metricsCount;
   cs.getMetricsV1(&metrics, &metricsCount);