Bump version

This was showing as uncovered in CI for some reason

.gitea/workflows/ci.yml

@@ -0,0 +1,306 @@
+name: CI
+
+on: [push, pull_request]
+
+env:
+  CC: clang
+  CXX: clang++
+
+jobs:
+  pre-commit:
+    runs-on: ubuntu-latest-amd64
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/cache@v4
+        with:
+          path: /var/cache/apt/archives
+          key: apt-amd64-${{ hashFiles('.gitea/workflows/ci.yml') }}
+
+      - name: Install dependencies
+        run: |
+          sudo apt-get update -qq
+          sudo apt-get install -y git nodejs pre-commit
+
+      - uses: actions/cache@v4
+        with:
+          path: ~/.cache/pre-commit
+          key: pre-commit-${{ hashFiles('.pre-commit-config.yaml') }}
+
+      - name: Run pre-commit
+        run: |
+          git config --global --add safe.directory "$PWD"
+          pre-commit run --all-files --show-diff-on-failure
+
+  test:
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - name: 64-bit-versions
+            cmake_args: -DCMAKE_CXX_FLAGS=-DUSE_64_BIT=1
+          - name: debug
+            cmake_args: -DCMAKE_BUILD_TYPE=Debug
+          - name: simd-fallback
+            cmake_args: -DUSE_SIMD_FALLBACK=ON
+          - name: gcc
+            cmake_args: -DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++
+    runs-on: ubuntu-latest-amd64
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/cache@v4
+        with:
+          path: /var/cache/apt/archives
+          key: apt-amd64-${{ hashFiles('.gitea/workflows/ci.yml') }}
+
+      - 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-${VERSION_CODENAME}-21 main" | sudo tee /etc/apt/sources.list.d/llvm.list
+          sudo apt-get update -qq
+          sudo apt-get install -y \
+            clang-21 llvm-21 lld-21 mold
+          for tool in clang clang++ llvm-ar llvm-nm llvm-ranlib llvm-objcopy llvm-cov llvm-symbolizer lld ld.lld; do
+            sudo update-alternatives --install /usr/bin/${tool} ${tool} /usr/bin/${tool}-21 100
+          done
+
+      - name: Build
+        run: |
+          export CCACHE_DIR="$GITHUB_WORKSPACE/.ccache"
+          rm -rf build
+          cmake -S . -B build -G Ninja -DCMAKE_CXX_COMPILER_LAUNCHER=ccache ${{ matrix.cmake_args }}
+          ninja -C build
+          ccache -s
+
+      - name: Test
+        run: |
+          cd build
+          ctest --no-compress-output --test-output-size-passed 100000 --test-output-size-failed 100000 -T Test -j "$(nproc)" --timeout 90 > /dev/null
+
+      - name: Upload test results to MinIO
+        if: always()
+        env:
+          MINIO_ACCESS_KEY: ${{ secrets.MINIO_ACCESS_KEY }}
+          MC_HOST_minio: https://${{ secrets.MINIO_ACCESS_KEY }}:${{ secrets.MINIO_SECRET_KEY }}@minio.weaselab.dev
+        run: |
+          if [ -z "$MINIO_ACCESS_KEY" ]; then
+            echo "MinIO credentials not configured; skipping upload"
+            exit 0
+          fi
+          zstd build/Testing/*/Test.xml
+          mc cp build/Testing/*/Test.xml.zst "minio/jenkins/conflict-set/${{ gitea.run_number }}/${{ matrix.name }}/"
+
+      - name: Test summary
+        if: always()
+        run: |
+          python3 ctest_summary.py build/Testing/*/Test.xml \
+            --link "https://minio.weaselab.dev/jenkins/conflict-set/${{ gitea.run_number }}/${{ matrix.name }}/Test.xml.zst" \
+            | tee -a "$GITHUB_STEP_SUMMARY"
+
+  release:
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - runner: ubuntu-latest-amd64
+            arch: amd64
+          - runner: ubuntu-latest-arm64
+            arch: arm64
+    runs-on: ${{ matrix.runner }}
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/cache@v4
+        with:
+          path: /var/cache/apt/archives
+          key: apt-${{ matrix.arch }}-${{ hashFiles('.gitea/workflows/ci.yml') }}
+
+      - name: Install dependencies
+        run: |
+          . /etc/os-release
+          wget -qO- https://apt.llvm.org/llvm-snapshot.gpg.key | sudo tee /etc/apt/trusted.gpg.d/apt.llvm.org.asc
+          echo "deb http://apt.llvm.org/${VERSION_CODENAME}/ llvm-toolchain-${VERSION_CODENAME}-21 main" | sudo tee /etc/apt/sources.list.d/llvm.list
+          sudo apt-get update -qq
+          sudo apt-get install -y \
+            build-essential ccache clang-21 cmake devscripts \
+            libc6-dbg llvm-21 lld-21 mold ninja-build rpm \
+            valgrind zstd
+          sudo curl -Ls "https://minio.weaselab.dev/public/$(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
+          key: ccache-release-${{ matrix.arch }}-${{ gitea.sha }}
+          restore-keys: |
+            ccache-release-${{ matrix.arch }}-
+
+      - name: Build
+        run: |
+          export CCACHE_DIR="$GITHUB_WORKSPACE/.ccache"
+          rm -rf build
+          cmake -S . -B build -G Ninja -DCMAKE_CXX_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_FLAGS=-DNVALGRIND
+          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
+
+      - name: Package
+        run: |
+          cd build
+          cpack -G DEB
+          cpack -G RPM
+
+      - name: Build paper
+        if: matrix.arch == 'amd64'
+        run: |
+          cd paper
+          make
+
+      - name: Upload artifacts to MinIO
+        if: always()
+        shell: bash
+        env:
+          MINIO_ACCESS_KEY: ${{ secrets.MINIO_ACCESS_KEY }}
+          MC_HOST_minio: https://${{ secrets.MINIO_ACCESS_KEY }}:${{ secrets.MINIO_SECRET_KEY }}@minio.weaselab.dev
+        run: |
+          if [ -z "$MINIO_ACCESS_KEY" ]; then
+            echo "MinIO credentials not configured; skipping upload"
+            exit 0
+          fi
+          dest="minio/jenkins/conflict-set/${{ gitea.run_number }}/release-${{ matrix.arch }}/"
+          zstd build/Testing/*/Test.xml
+          mc cp build/Testing/*/Test.xml.zst "$dest"
+          if compgen -G "build/*.deb" > /dev/null; then
+            mc cp build/*.deb "$dest"
+          fi
+          if compgen -G "build/*.rpm" > /dev/null; then
+            mc cp build/*.rpm "$dest"
+          fi
+          if compgen -G "paper/*.pdf" > /dev/null; then
+            mc cp paper/*.pdf "$dest"
+          fi
+
+      - name: Test summary
+        if: always()
+        run: |
+          python3 ctest_summary.py build/Testing/*/Test.xml \
+            --link "https://minio.weaselab.dev/jenkins/conflict-set/${{ gitea.run_number }}/release-${{ matrix.arch }}/Test.xml.zst" \
+            | tee -a "$GITHUB_STEP_SUMMARY"
+
+  coverage:
+    runs-on: ubuntu-latest-amd64
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/cache@v4
+        with:
+          path: /var/cache/apt/archives
+          key: apt-amd64-${{ hashFiles('.gitea/workflows/ci.yml') }}
+
+      - 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 \
+            build-essential ccache clang-21 cmake gcovr \
+            libc6-dbg llvm-21 lld-21 mold ninja-build python3 valgrind zstd
+          sudo curl -Ls "https://minio.weaselab.dev/public/$(uname -m)/mc.RELEASE.2025-08-13T08-35-41Z" \
+            -o /usr/local/bin/mc && sudo chmod +x /usr/local/bin/mc
+          for tool in clang clang++ llvm-ar llvm-nm llvm-ranlib llvm-objcopy llvm-cov llvm-symbolizer lld ld.lld; do
+            sudo update-alternatives --install /usr/bin/${tool} ${tool} /usr/bin/${tool}-21 100
+          done
+
+      - uses: actions/cache@v4
+        with:
+          path: .ccache
+          key: ccache-coverage-${{ gitea.sha }}
+          restore-keys: |
+            ccache-coverage-
+
+      - name: Build
+        run: |
+          export CCACHE_DIR="$GITHUB_WORKSPACE/.ccache"
+          rm -rf build
+          cmake -S . -B build -G Ninja -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_C_FLAGS=--coverage -DCMAKE_CXX_FLAGS=--coverage \
+            -DCMAKE_BUILD_TYPE=Debug -DDISABLE_TSAN=ON
+          ninja -C build
+          ccache -s
+
+      - name: Test
+        run: |
+          cd build
+          ctest --no-compress-output --test-output-size-passed 100000 --test-output-size-failed 100000 -T Test -j "$(nproc)" --timeout 90 > /dev/null
+
+      - name: Coverage report
+        shell: bash
+        run: |
+          gcov_args=(-f ConflictSet.cpp -f LongestCommonPrefix.h -f Metrics.h
+                     --gcov-executable "llvm-cov gcov" --exclude-noncode-lines)
+          gcovr "${gcov_args[@]}" --cobertura > build/coverage.xml
+          gcovr "${gcov_args[@]}"
+          mkdir -p build/coverage_html
+          gcovr "${gcov_args[@]}" --html-details build/coverage_html/index.html
+          gcovr "${gcov_args[@]}" --fail-under-line 100 > /dev/null
+
+      - name: Upload results to MinIO
+        if: always()
+        env:
+          MINIO_ACCESS_KEY: ${{ secrets.MINIO_ACCESS_KEY }}
+          MC_HOST_minio: https://${{ secrets.MINIO_ACCESS_KEY }}:${{ secrets.MINIO_SECRET_KEY }}@minio.weaselab.dev
+        run: |
+          if [ -z "$MINIO_ACCESS_KEY" ]; then
+            echo "MinIO credentials not configured; skipping upload"
+            exit 0
+          fi
+          dest="minio/jenkins/conflict-set/${{ gitea.run_number }}/coverage/"
+          zstd build/Testing/*/Test.xml
+          mc cp build/Testing/*/Test.xml.zst "$dest"
+          if [ -e build/coverage.xml ]; then
+            mc cp build/coverage.xml "$dest"
+          fi
+          if [ -d build/coverage_html ]; then
+            mc cp -r build/coverage_html "$dest"
+          fi
+
+      - name: Test summary
+        if: always()
+        run: |
+          python3 ctest_summary.py build/Testing/*/Test.xml \
+            --link "https://minio.weaselab.dev/jenkins/conflict-set/${{ gitea.run_number }}/coverage/Test.xml.zst" \
+            | tee -a "$GITHUB_STEP_SUMMARY"
+          echo "" | tee -a "$GITHUB_STEP_SUMMARY"
+          echo "📊 [Coverage report](https://minio.weaselab.dev/jenkins/conflict-set/${{ gitea.run_number }}/coverage/coverage_html/index.html)" | tee -a "$GITHUB_STEP_SUMMARY"

CMakeLists.txt

@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.18)
 project(
   conflict-set
-  VERSION 0.0.12
+  VERSION 0.0.14
   DESCRIPTION
     "A data structure for optimistic concurrency control on ranges of bitwise-lexicographically-ordered keys."
   HOMEPAGE_URL "https://git.weaselab.dev/weaselab/conflict-set"
@@ -276,9 +276,15 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
 
   find_program(VALGRIND_EXE valgrind)
   if(VALGRIND_EXE AND NOT CMAKE_CROSSCOMPILING)
-    add_test(NAME conflict_set_blackbox_valgrind
+    list(LENGTH CORPUS_TESTS len)
+    math(EXPR last "${len} - 1")
+    set(partition_size 100)
+    foreach(i RANGE 0 ${last} ${partition_size})
+      list(SUBLIST CORPUS_TESTS ${i} ${partition_size} partition)
+      add_test(NAME conflict_set_blackbox_valgrind_${i}
                COMMAND ${VALGRIND_EXE} --error-exitcode=99 --
-                     $<TARGET_FILE:driver> ${CORPUS_TESTS})
+                       $<TARGET_FILE:driver> ${partition})
+    endforeach()
   endif()
 
   # api smoke tests

ConflictSet.cpp

@@ -87,22 +87,42 @@ constexpr int64_t kMaxCorrectVersionWindow =
     std::numeric_limits<int32_t>::max();
 static_assert(kNominalVersionWindow <= kMaxCorrectVersionWindow);
 
+#ifndef USE_64_BIT
+#define USE_64_BIT 0
+#endif
+
 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 {
+#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;
+#if USE_64_BIT
+  static const InternalVersionT zero;
+#else
   static thread_local InternalVersionT zero;
+#endif
 
 private:
+#if USE_64_BIT
+  int64_t value;
+#else
   uint32_t value;
+#endif
 };
+#if USE_64_BIT
+const InternalVersionT InternalVersionT::zero{0};
+#else
 thread_local InternalVersionT InternalVersionT::zero;
+#endif
 
 struct Entry {
   InternalVersionT pointVersion;
@@ -197,7 +217,6 @@ struct Node {
   /* end section that's copied to the next node */
 
   uint8_t *partialKey();
-
   Type getType() const { return type; }
   int32_t getCapacity() const { return partialKeyCapacity; }
 
@@ -221,84 +240,83 @@ constexpr int kNodeCopySize =
 
 struct Node0 : Node {
   constexpr static auto kType = Type_Node0;
-  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
 
+  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
   void copyChildrenAndKeyFrom(const Node0 &other);
   void copyChildrenAndKeyFrom(const struct Node3 &other);
-
   size_t size() const { return sizeof(Node0) + getCapacity(); }
 };
 
 struct Node3 : Node {
   constexpr static auto kMaxNodes = 3;
   constexpr static auto kType = Type_Node3;
-  // Sorted
-  uint8_t index[kMaxNodes];
+
   Node *children[kMaxNodes];
   InternalVersionT childMaxVersion[kMaxNodes];
-  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
+  // Sorted
+  uint8_t index[kMaxNodes];
 
+  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
   void copyChildrenAndKeyFrom(const Node0 &other);
   void copyChildrenAndKeyFrom(const Node3 &other);
   void copyChildrenAndKeyFrom(const struct Node16 &other);
-
   size_t size() const { return sizeof(Node3) + getCapacity(); }
 };
 
 struct Node16 : Node {
   constexpr static auto kType = Type_Node16;
   constexpr static auto kMaxNodes = 16;
-  // Sorted
-  uint8_t index[kMaxNodes];
+
   Node *children[kMaxNodes];
   InternalVersionT childMaxVersion[kMaxNodes];
-  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
+  // Sorted
+  uint8_t index[kMaxNodes];
 
+  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
   void copyChildrenAndKeyFrom(const Node3 &other);
   void copyChildrenAndKeyFrom(const Node16 &other);
   void copyChildrenAndKeyFrom(const struct Node48 &other);
-
   size_t size() const { return sizeof(Node16) + getCapacity(); }
 };
 
 struct Node48 : Node {
   constexpr static auto kType = Type_Node48;
   constexpr static auto kMaxNodes = 48;
-  BitSet bitSet;
-  int8_t nextFree;
-  int8_t index[256];
-  Node *children[kMaxNodes];
-  InternalVersionT 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;
+
+  BitSet bitSet;
+  Node *children[kMaxNodes];
+  InternalVersionT childMaxVersion[kMaxNodes];
   InternalVersionT maxOfMax[kMaxOfMaxTotalPages];
+  uint8_t reverseIndex[kMaxNodes];
+  int8_t index[256];
 
   uint8_t *partialKey() { return (uint8_t *)(this + 1); }
-
   void copyChildrenAndKeyFrom(const Node16 &other);
   void copyChildrenAndKeyFrom(const Node48 &other);
   void copyChildrenAndKeyFrom(const struct Node256 &other);
-
   size_t size() const { return sizeof(Node48) + getCapacity(); }
 };
 
 struct Node256 : Node {
   constexpr static auto kType = Type_Node256;
-  BitSet bitSet;
-  Node *children[256];
-  InternalVersionT childMaxVersion[256];
+  constexpr static auto kMaxNodes = 256;
   constexpr static int kMaxOfMaxPageSize = 16;
   constexpr static int kMaxOfMaxShift =
       std::countr_zero(uint32_t(kMaxOfMaxPageSize));
-  constexpr static int kMaxOfMaxTotalPages = 256 / kMaxOfMaxPageSize;
+  constexpr static int kMaxOfMaxTotalPages = kMaxNodes / kMaxOfMaxPageSize;
+
+  BitSet bitSet;
+  Node *children[kMaxNodes];
+  InternalVersionT childMaxVersion[kMaxNodes];
   InternalVersionT maxOfMax[kMaxOfMaxTotalPages];
+
   uint8_t *partialKey() { return (uint8_t *)(this + 1); }
   void copyChildrenAndKeyFrom(const Node48 &other);
   void copyChildrenAndKeyFrom(const Node256 &other);
-
   size_t size() const { return sizeof(Node256) + getCapacity(); }
 };
 
@@ -323,7 +341,7 @@ inline void Node3::copyChildrenAndKeyFrom(const Node0 &other) {
 inline void Node3::copyChildrenAndKeyFrom(const Node3 &other) {
   memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
          kNodeCopySize);
-  memcpy(index, other.index, sizeof(*this) - sizeof(Node));
+  memcpy(children, other.children, sizeof(*this) - sizeof(Node));
   memcpy(partialKey(), &other + 1, partialKeyLen);
   for (int i = 0; i < numChildren; ++i) {
     assert(children[i]->parent == &other);
@@ -404,7 +422,6 @@ inline void Node48::copyChildrenAndKeyFrom(const Node16 &other) {
   }
   memcpy(partialKey(), &other + 1, partialKeyLen);
   bitSet.init();
-  nextFree = Node16::kMaxNodes;
   int i = 0;
   for (auto x : other.index) {
     bitSet.set(x);
@@ -424,7 +441,6 @@ inline void Node48::copyChildrenAndKeyFrom(const Node48 &other) {
   memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
          kNodeCopySize);
   bitSet = other.bitSet;
-  nextFree = other.nextFree;
   memcpy(index, other.index, sizeof(index));
   memset(children, 0, sizeof(children));
   const auto z = InternalVersionT::zero;
@@ -451,7 +467,6 @@ inline void Node48::copyChildrenAndKeyFrom(const Node256 &other) {
   for (auto &v : childMaxVersion) {
     v = z;
   }
-  nextFree = other.numChildren;
   bitSet = other.bitSet;
   int i = 0;
   bitSet.forEachSet([&](int c) {
@@ -523,8 +538,13 @@ std::string getSearchPath(Node *n);
 // Each node with an entry present gets a budget of kBytesPerKey. Node0 always
 // has an entry present.
 // Induction hypothesis is that each node's surplus is >= kMinNodeSurplus
+#if USE_64_BIT
+constexpr int kBytesPerKey = 144;
+constexpr int kMinNodeSurplus = 104;
+#else
 constexpr int kBytesPerKey = 112;
 constexpr int kMinNodeSurplus = 80;
+#endif
 // Cound the entry itself as a child
 constexpr int kMinChildrenNode0 = 1;
 constexpr int kMinChildrenNode3 = 2;
@@ -729,9 +749,13 @@ struct WriteContext {
     int64_t write_bytes;
   } accum;
 
+#if USE_64_BIT
+  static constexpr InternalVersionT zero{0};
+#else
   // Cache a copy of InternalVersionT::zero, so we don't need to do the TLS
   // lookup as often.
   InternalVersionT zero;
+#endif
 
   WriteContext() { memset(&accum, 0, sizeof(accum)); }
 
@@ -773,14 +797,7 @@ private:
   BoundedFreeListAllocator<Node256> node256;
 };
 
-template <class NodeT> int getNodeIndex(NodeT *self, uint8_t index) {
-  static_assert(std::is_same_v<NodeT, Node3> || std::is_same_v<NodeT, Node16>);
-
-  // cachegrind says the plain loop is fewer instructions and more mis-predicted
-  // branches. Microbenchmark says plain loop is faster. It's written in this
-  // weird "generic" way though in case someday we can use the simd
-  // implementation easily if we want.
-  if constexpr (std::is_same_v<NodeT, Node3>) {
+int getNodeIndex(Node3 *self, uint8_t index) {
   Node3 *n = (Node3 *)self;
   for (int i = 0; i < n->numChildren; ++i) {
     if (n->index[i] == index) {
@@ -790,6 +807,8 @@ template <class NodeT> int getNodeIndex(NodeT *self, uint8_t index) {
   return -1;
 }
 
+int getNodeIndex(Node16 *self, uint8_t index) {
+
 #ifdef HAS_AVX
   // Based on https://www.the-paper-trail.org/post/art-paper-notes/
 
@@ -801,7 +820,7 @@ template <class NodeT> int getNodeIndex(NodeT *self, uint8_t index) {
   // keys aren't valid, we'll mask the results to only consider the valid ones
   // below.
   __m128i indices;
-  memcpy(&indices, self->index, NodeT::kMaxNodes);
+  memcpy(&indices, self->index, Node16::kMaxNodes);
   __m128i results = _mm_cmpeq_epi8(key_vec, indices);
 
   // Build a mask to select only the first node->num_children values from the
@@ -824,12 +843,11 @@ template <class NodeT> int getNodeIndex(NodeT *self, uint8_t index) {
   // https://community.arm.com/arm-community-blogs/b/infrastructure-solutions-blog/posts/porting-x86-vector-bitmask-optimizations-to-arm-neon
 
   uint8x16_t indices;
-  memcpy(&indices, self->index, NodeT::kMaxNodes);
+  memcpy(&indices, self->index, Node16::kMaxNodes);
   // 0xff for each match
   uint16x8_t results =
       vreinterpretq_u16_u8(vceqq_u8(vdupq_n_u8(index), indices));
-  static_assert(NodeT::kMaxNodes <= 16);
-  assume(self->numChildren <= NodeT::kMaxNodes);
+  assume(self->numChildren <= Node16::kMaxNodes);
   uint64_t mask = self->numChildren == 16
                       ? uint64_t(-1)
                       : (uint64_t(1) << (self->numChildren * 4)) - 1;
@@ -1082,15 +1100,9 @@ ChildAndMaxVersion getChildAndMaxVersion(Node *self, uint8_t index) {
   }
 }
 
-template <class NodeT> Node *getChildGeqSimd(NodeT *self, int child) {
-  static_assert(std::is_same_v<NodeT, Node3> || std::is_same_v<NodeT, Node16>);
+Node *getChildGeq(Node0 *, int) { return nullptr; }
 
-  // cachegrind says the plain loop is fewer instructions and more mis-predicted
-  // branches. Microbenchmark says plain loop is faster. It's written in this
-  // weird "generic" way though so that someday we can use the simd
-  // implementation easily if we want.
-  if constexpr (std::is_same_v<NodeT, Node3>) {
-    Node3 *n = (Node3 *)self;
+Node *getChildGeq(Node3 *n, int child) {
   for (int i = 0; i < n->numChildren; ++i) {
     if (n->index[i] >= child) {
       return n->children[i];
@@ -1098,6 +1110,8 @@ template <class NodeT> Node *getChildGeqSimd(NodeT *self, int child) {
   }
   return nullptr;
 }
+
+Node *getChildGeq(Node16 *self, int child) {
   if (child > 255) {
     return nullptr;
   }
@@ -1105,7 +1119,7 @@ template <class NodeT> Node *getChildGeqSimd(NodeT *self, int child) {
 #ifdef HAS_AVX
   __m128i key_vec = _mm_set1_epi8(child);
   __m128i indices;
-  memcpy(&indices, self->index, NodeT::kMaxNodes);
+  memcpy(&indices, self->index, Node16::kMaxNodes);
   __m128i results = _mm_cmpeq_epi8(key_vec, _mm_min_epu8(key_vec, indices));
   int mask = (1 << self->numChildren) - 1;
   uint32_t bitfield = _mm_movemask_epi8(results) & mask;
@@ -1115,8 +1129,7 @@ template <class NodeT> Node *getChildGeqSimd(NodeT *self, int child) {
   memcpy(&indices, self->index, sizeof(self->index));
   // 0xff for each leq
   auto results = vcleq_u8(vdupq_n_u8(child), indices);
-  static_assert(NodeT::kMaxNodes <= 16);
-  assume(self->numChildren <= NodeT::kMaxNodes);
+  assume(self->numChildren <= Node16::kMaxNodes);
   uint64_t mask = self->numChildren == 16
                       ? uint64_t(-1)
                       : (uint64_t(1) << (self->numChildren * 4)) - 1;
@@ -1141,13 +1154,6 @@ template <class NodeT> Node *getChildGeqSimd(NodeT *self, int child) {
 #endif
 }
 
-Node *getChildGeq(Node0 *, int) { return nullptr; }
-Node *getChildGeq(Node3 *self, int child) {
-  return getChildGeqSimd(self, child);
-}
-Node *getChildGeq(Node16 *self, int child) {
-  return getChildGeqSimd(self, child);
-}
 Node *getChildGeq(Node48 *self, int child) {
   int c = self->bitSet.firstSetGeq(child);
   if (c < 0) {
@@ -1360,7 +1366,7 @@ Node *&getOrCreateChild(Node *&self, std::span<const uint8_t> &key,
     auto *self3 = static_cast<Node3 *>(self);
     int i = self->numChildren - 1;
     for (; i >= 0; --i) {
-      if (int(self3->index[i]) < int(index)) {
+      if (self3->index[i] < index) {
         break;
       }
       self3->index[i + 1] = self3->index[i];
@@ -1390,7 +1396,7 @@ Node *&getOrCreateChild(Node *&self, std::span<const uint8_t> &key,
     auto *self16 = static_cast<Node16 *>(self);
     int i = self->numChildren - 1;
     for (; i >= 0; --i) {
-      if (int(self16->index[i]) < int(index)) {
+      if (self16->index[i] < index) {
         break;
       }
       self16->index[i + 1] = self16->index[i];
@@ -1419,9 +1425,7 @@ Node *&getOrCreateChild(Node *&self, std::span<const uint8_t> &key,
   insert48:
     auto *self48 = static_cast<Node48 *>(self);
     self48->bitSet.set(index);
-    ++self->numChildren;
-    assert(self48->nextFree < 48);
-    int nextFree = self48->nextFree++;
+    auto nextFree = self48->numChildren++;
     self48->index[index] = nextFree;
     self48->reverseIndex[nextFree] = index;
     auto &result = self48->children[nextFree];
@@ -1569,8 +1573,6 @@ void maybeDecreaseCapacity(Node *&self, WriteContext *tls,
 }
 
 #if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
-// This gets covered in local development
-// GCOVR_EXCL_START
 __attribute__((target("avx512f"))) void rezero16(InternalVersionT *vs,
                                                  InternalVersionT zero) {
   uint32_t z;
@@ -1580,7 +1582,6 @@ __attribute__((target("avx512f"))) void rezero16(InternalVersionT *vs,
       _mm512_sub_epi32(_mm512_loadu_epi32(vs), zvec), _mm512_setzero_epi32());
   _mm512_mask_storeu_epi32(vs, m, zvec);
 }
-// GCOVR_EXCL_STOP
 
 __attribute__((target("default")))
 #endif
@@ -1591,6 +1592,9 @@ void rezero16(InternalVersionT *vs, InternalVersionT zero) {
   }
 }
 
+#if USE_64_BIT
+void rezero(Node *, InternalVersionT) {}
+#else
 void rezero(Node *n, InternalVersionT z) {
 #if DEBUG_VERBOSE && !defined(NDEBUG)
   fprintf(stderr, "rezero to %" PRId64 ": %s\n", z.toInt64(),
@@ -1633,6 +1637,7 @@ void rezero(Node *n, InternalVersionT z) {
     __builtin_unreachable(); // GCOVR_EXCL_LINE
   }
 }
+#endif
 
 void mergeWithChild(Node *&self, WriteContext *tls, ConflictSet::Impl *impl,
                     Node *&dontInvalidate, Node3 *self3) {
@@ -1812,7 +1817,7 @@ Node *erase(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
     parent48->bitSet.reset(parentsIndex);
     int8_t toRemoveChildrenIndex =
         std::exchange(parent48->index[parentsIndex], -1);
-    int8_t lastChildrenIndex = --parent48->nextFree;
+    auto lastChildrenIndex = --parent48->numChildren;
     assert(toRemoveChildrenIndex >= 0);
     assert(lastChildrenIndex >= 0);
     if (toRemoveChildrenIndex != lastChildrenIndex) {
@@ -1831,8 +1836,6 @@ Node *erase(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
     }
     parent48->childMaxVersion[lastChildrenIndex] = tls->zero;
 
-    --parent->numChildren;
-
     if (needsDownsize(parent48)) {
       downsize(parent48, tls, impl, result);
     }
@@ -2017,11 +2020,18 @@ downLeftSpine:
 }
 
 #ifdef HAS_AVX
-uint32_t compare16_32bit(const InternalVersionT *vs, InternalVersionT rv) {
+uint32_t compare16(const InternalVersionT *vs, InternalVersionT rv) {
+#if USE_64_BIT
   uint32_t compared = 0;
-  __m128i w[4];
+  for (int i = 0; i < 16; ++i) {
+    compared |= (vs[i] > rv) << i;
+  }
+  return compared;
+#else
+  uint32_t compared = 0;
+  __m128i w[4]; // GCOVR_EXCL_LINE
   memcpy(w, vs, sizeof(w));
-  uint32_t r;
+  uint32_t r; // GCOVR_EXCL_LINE
   memcpy(&r, &rv, sizeof(r));
   const auto rvVec = _mm_set1_epi32(r);
   const auto zero = _mm_setzero_si128();
@@ -2031,20 +2041,28 @@ uint32_t compare16_32bit(const InternalVersionT *vs, InternalVersionT rv) {
                 << (i * 4);
   }
   return compared;
+#endif
 }
 
-// This gets covered in local development
-// GCOVR_EXCL_START
 __attribute__((target("avx512f"))) uint32_t
-compare16_32bit_avx512(const InternalVersionT *vs, InternalVersionT rv) {
+compare16_avx512(const InternalVersionT *vs, InternalVersionT rv) {
+#if USE_64_BIT
+  int64_t r;
+  memcpy(&r, &rv, sizeof(r));
+  uint32_t low =
+      _mm512_cmpgt_epi64_mask(_mm512_loadu_epi64(vs), _mm512_set1_epi64(r));
+  uint32_t high =
+      _mm512_cmpgt_epi64_mask(_mm512_loadu_epi64(vs + 8), _mm512_set1_epi64(r));
+  return low | (high << 8);
+#else
   uint32_t r;
   memcpy(&r, &rv, sizeof(r));
   return _mm512_cmpgt_epi32_mask(
       _mm512_sub_epi32(_mm512_loadu_epi32(vs), _mm512_set1_epi32(r)),
       _mm512_setzero_epi32());
+#endif
 }
 #endif
-// GCOVR_EXCL_STOP
 
 // Returns true if v[i] <= readVersion for all i such that begin <= is[i] < end
 // Preconditions: begin <= end, end - begin < 256
@@ -2099,9 +2117,9 @@ bool scan16(const InternalVersionT *vs, const uint8_t *is, int begin, int end,
 
   uint32_t compared = 0;
   if constexpr (kAVX512) {
-    compared = compare16_32bit_avx512(vs, readVersion); // GCOVR_EXCL_LINE
+    compared = compare16_avx512(vs, readVersion);
   } else {
-    compared = compare16_32bit(vs, readVersion); // GCOVR_EXCL_LINE
+    compared = compare16(vs, readVersion);
   }
   return !(compared & mask);
 
@@ -2160,9 +2178,9 @@ bool scan16(const InternalVersionT *vs, int begin, int end,
 #elif defined(HAS_AVX)
   uint32_t conflict;
   if constexpr (kAVX512) {
-    conflict = compare16_32bit_avx512(vs, readVersion); // GCOVR_EXCL_LINE
+    conflict = compare16_avx512(vs, readVersion);
   } else {
-    conflict = compare16_32bit(vs, readVersion); // GCOVR_EXCL_LINE
+    conflict = compare16(vs, readVersion);
   }
   conflict &= (1 << end) - 1;
   conflict >>= begin;
@@ -2297,12 +2315,9 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
 
     uint32_t compared = 0;
     if constexpr (kAVX512) {
-      compared =                                        // GCOVR_EXCL_LINE
-          compare16_32bit_avx512(self->childMaxVersion, // GCOVR_EXCL_LINE
-                                 readVersion);          // GCOVR_EXCL_LINE
+      compared = compare16_avx512(self->childMaxVersion, readVersion);
     } else {
-      compared = compare16_32bit(self->childMaxVersion,
-                                 readVersion); // GCOVR_EXCL_LINE
+      compared = compare16(self->childMaxVersion, readVersion);
     }
     return !(compared & mask) && firstRangeOk;
 
@@ -2340,10 +2355,8 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
       int c = self->bitSet.firstSetGeq(begin + 1);
       if (c >= 0 && c < end) {
         auto *child = self->children[self->index[c]];
-        if (child->entryPresent) {
-          if (!(child->entry.rangeVersion <= readVersion)) {
+        if (child->entryPresent && child->entry.rangeVersion > readVersion) {
           return false;
-          };
         }
         begin = c;
       } else {
@@ -2376,10 +2389,8 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
       int c = self->bitSet.firstSetGeq(begin + 1);
       if (c >= 0 && c < end) {
         auto *child = self->children[c];
-        if (child->entryPresent) {
-          if (!(child->entry.rangeVersion <= readVersion)) {
+        if (child->entryPresent && child->entry.rangeVersion > readVersion) {
           return false;
-          };
         }
         begin = c;
       } else {
@@ -2422,9 +2433,7 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
       }
     }
     // Check inner pages
-    const int innerPageBegin = (begin >> Node256::kMaxOfMaxShift) + 1;
-    const int innerPageEnd = (end - 1) >> Node256::kMaxOfMaxShift;
-    return scan16<kAVX512>(self->maxOfMax, innerPageBegin, innerPageEnd,
+    return scan16<kAVX512>(self->maxOfMax, firstPage + 1, lastPage,
                            readVersion);
   }
   default:                   // GCOVR_EXCL_LINE
@@ -2433,14 +2442,11 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
 }
 
 #if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
-// This gets covered in local development
-// GCOVR_EXCL_START
 __attribute__((target("avx512f"))) bool
 checkMaxBetweenExclusive(Node *n, int begin, int end,
                          InternalVersionT readVersion, ReadContext *tls) {
   return checkMaxBetweenExclusiveImpl<true>(n, begin, end, readVersion, tls);
 }
-// GCOVR_EXCL_STOP
 __attribute__((target("default")))
 #endif
 bool checkMaxBetweenExclusive(Node *n, int begin, int end,
@@ -2906,8 +2912,7 @@ checkMaxBetweenExclusiveImpl<true>(Node *n, int begin, int end,
 // of the result will have `maxVersion` set to `writeVersion` as a
 // postcondition. Nodes along the search path may be invalidated. Callers must
 // ensure that the max version of the self argument is updated.
-[[nodiscard]]
-Node **insert(Node **self, std::span<const uint8_t> key,
+[[nodiscard]] Node **insert(Node **self, std::span<const uint8_t> key,
                             InternalVersionT writeVersion, WriteContext *tls) {
 
   for (; key.size() != 0; ++tls->accum.insert_iterations) {
@@ -2927,7 +2932,6 @@ void eraseTree(Node *root, WriteContext *tls) {
     tls->accum.entries_erased += n->entryPresent;
     ++tls->accum.nodes_released;
 
-    removeNode(n);
     removeKey(n);
 
     switch (n->getType()) {
@@ -3144,6 +3148,8 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     tls.impl = this;
     int64_t check_byte_accum = 0;
     for (int i = 0; i < count; ++i) {
+      assert(reads[i].readVersion >= 0);
+      assert(reads[i].readVersion <= newestVersionFullPrecision);
       const auto &r = reads[i];
       check_byte_accum += r.begin.len + r.end.len;
       auto begin = std::span<const uint8_t>(r.begin.p, r.begin.len);
@@ -3180,10 +3186,12 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   }
 
   void addWrites(const WriteRange *writes, int count, int64_t writeVersion) {
+#if !USE_64_BIT
     // There could be other conflict sets in the same thread. We need
     // InternalVersionT::zero to be correct for this conflict set for the
     // lifetime of the current call frame.
     InternalVersionT::zero = tls.zero = oldestVersion;
+#endif
 
     assert(writeVersion >= newestVersionFullPrecision);
     assert(tls.accum.entries_erased == 0);
@@ -3198,7 +3206,10 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
 
       newestVersionFullPrecision = writeVersion;
       newest_version.set(newestVersionFullPrecision);
+      if (newestVersionFullPrecision - kNominalVersionWindow >
+          oldestVersionFullPrecision) {
         setOldestVersion(newestVersionFullPrecision - kNominalVersionWindow);
+      }
       while (oldestExtantVersion <
              newestVersionFullPrecision - kMaxCorrectVersionWindow) {
         gcScanStep(1000);
@@ -3206,8 +3217,11 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     } else {
       newestVersionFullPrecision = writeVersion;
       newest_version.set(newestVersionFullPrecision);
+      if (newestVersionFullPrecision - kNominalVersionWindow >
+          oldestVersionFullPrecision) {
         setOldestVersion(newestVersionFullPrecision - kNominalVersionWindow);
       }
+    }
 
     for (int i = 0; i < count; ++i) {
       const auto &w = writes[i];
@@ -3291,14 +3305,22 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     return fuel;
   }
 
-  void setOldestVersion(int64_t o) {
-    if (o <= oldestVersionFullPrecision) {
+  void setOldestVersion(int64_t newOldestVersion) {
+    assert(newOldestVersion >= 0);
+    assert(newOldestVersion <= newestVersionFullPrecision);
+    // If addWrites advances oldestVersion to keep within valid window, a
+    // subsequent setOldestVersion can be legitimately called with a version
+    // older than `oldestVersionFullPrecision`. < instead of <= so that we can
+    // do garbage collection work without advancing the oldest version.
+    if (newOldestVersion < oldestVersionFullPrecision) {
       return;
     }
-    InternalVersionT oldestVersion{o};
-    this->oldestVersionFullPrecision = o;
+    InternalVersionT oldestVersion{newOldestVersion};
+    this->oldestVersionFullPrecision = newOldestVersion;
     this->oldestVersion = oldestVersion;
+#if !USE_64_BIT
     InternalVersionT::zero = tls.zero = oldestVersion;
+#endif
 #ifdef NDEBUG
     // This is here for performance reasons, since we want to amortize the cost
     // of storing the search path as a string. In tests, we want to exercise the
@@ -3347,12 +3369,15 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     root->entry.pointVersion = this->oldestVersion;
     root->entry.rangeVersion = this->oldestVersion;
 
+#if !USE_64_BIT
     InternalVersionT::zero = tls.zero = this->oldestVersion;
+#endif
 
     // Intentionally not resetting totalBytes
   }
 
   explicit Impl(int64_t oldestVersion) {
+    assert(oldestVersion >= 0);
     init(oldestVersion);
     initMetrics();
   }
@@ -3623,7 +3648,12 @@ ConflictSet::ConflictSet(ConflictSet &&other) noexcept
     : impl(std::exchange(other.impl, nullptr)) {}
 
 ConflictSet &ConflictSet::operator=(ConflictSet &&other) noexcept {
+  if (this != &other) {
+    if (impl) {
+      internal_destroy(impl);
+    }
     impl = std::exchange(other.impl, nullptr);
+  }
   return *this;
 }
 
@@ -3747,13 +3777,13 @@ std::string getSearchPath(Node *n) {
         fprintf(file,
                 " k_%p [label=\"m=%" PRId64 " p=%" PRId64 " r=%" PRId64
                 "\n%s\", pos=\"%d,%d!\"];\n",
-                (void *)n, maxVersion(n).toInt64(),
+                (void *)n, n->parent == nullptr ? -1 : maxVersion(n).toInt64(),
                 n->entry.pointVersion.toInt64(),
                 n->entry.rangeVersion.toInt64(),
                 getPartialKeyPrintable(n).c_str(), x, y);
       } else {
         fprintf(file, " k_%p [label=\"m=%" PRId64 "\n%s\", pos=\"%d,%d!\"];\n",
-                (void *)n, maxVersion(n).toInt64(),
+                (void *)n, n->parent == nullptr ? -1 : maxVersion(n).toInt64(),
                 getPartialKeyPrintable(n).c_str(), x, y);
       }
       x += kSeparation;
@@ -3794,6 +3824,9 @@ Node *firstGeq(Node *n, std::string_view key) {
       n, std::span<const uint8_t>((const uint8_t *)key.data(), key.size()));
 }
 
+#if USE_64_BIT
+void checkVersionsGeqOldestExtant(Node *, InternalVersionT) {}
+#else
 void checkVersionsGeqOldestExtant(Node *n,
                                   InternalVersionT oldestExtantVersion) {
   if (n->entryPresent) {
@@ -3837,6 +3870,7 @@ void checkVersionsGeqOldestExtant(Node *n,
     abort();
   }
 }
+#endif
 
 [[maybe_unused]] InternalVersionT
 checkMaxVersion(Node *root, Node *node, InternalVersionT oldestVersion,

HashTable.cpp

@@ -119,7 +119,13 @@ ConflictSet::ConflictSet(ConflictSet &&other) noexcept
     : impl(std::exchange(other.impl, nullptr)) {}
 
 ConflictSet &ConflictSet::operator=(ConflictSet &&other) noexcept {
+  if (this != &other) {
+    if (impl) {
+      impl->~Impl();
+      safe_free(impl, sizeof(Impl));
+    }
     impl = std::exchange(other.impl, nullptr);
+  }
   return *this;
 }
 

Internal.h

@@ -20,7 +20,6 @@ using namespace weaselab;
 #include <thread>
 #include <unordered_set>
 #include <utility>
-#include <vector>
 
 #include <callgrind.h>
 
@@ -748,7 +747,10 @@ struct TestDriver {
       fprintf(stderr, "%p Set oldest version: %" PRId64 "\n", this,
               oldestVersion);
 #endif
+      CALLGRIND_START_INSTRUMENTATION;
       cs.setOldestVersion(oldestVersion);
+      CALLGRIND_STOP_INSTRUMENTATION;
+
       if constexpr (kEnableAssertions) {
         refImpl.setOldestVersion(oldestVersion);
       }

Jenkinsfile

@@ -48,6 +48,17 @@ pipeline {
                 recordIssues(tools: [clang()])
             }
         }
+        stage('64 bit versions') {
+            agent {
+                dockerfile {
+                    args '-v /home/jenkins/ccache:/ccache'
+                    reuseNode true
+                }
+            }
+            steps {
+                CleanBuildAndTest("-DCMAKE_CXX_FLAGS=-DUSE_64_BIT=1")
+            }
+        }
         stage('Debug') {
             agent {
                 dockerfile {

LongestCommonPrefix.h

@@ -129,7 +129,7 @@ longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
   }
 
   int i = 0;
-  int end;
+  int end; // GCOVR_EXCL_LINE
 
   // kStride * kUnrollCount at a time
   end = cl & ~(kStride * kUnrollFactor - 1);

README.md

@@ -2,7 +2,16 @@ A data structure for optimistic concurrency control on ranges of bitwise-lexicog
 
 Intended as an alternative to FoundationDB's skip list.
 
-Hardware for all benchmarks is an AMD Ryzen 9 7900 with (2x32GB) 5600MT/s CL28-34-34-89 1.35V RAM
+Hardware for all benchmarks is an AMD Ryzen 9 7900 with (2x32GB) 5600MT/s CL28-34-34-89 1.35V RAM.
+
+```
+$ clang++ --version
+
+Ubuntu clang version 21.1.8 (6ubuntu1)
+Target: x86_64-pc-linux-gnu
+Thread model: posix
+InstalledDir: /usr/lib/llvm-21/bin
+```
 
 # Microbenchmark
 
@@ -10,44 +19,45 @@ Hardware for all benchmarks is an AMD Ryzen 9 7900 with (2x32GB) 5600MT/s CL28-3
 
 |               ns/op |                op/s |    err% |          ins/op |          cyc/op |    IPC |         bra/op |   miss% |     total | benchmark
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|              172.03 |        5,812,791.77 |    0.4% |        3,130.62 |          879.00 |  3.562 |         509.23 |    0.0% |      0.01 | `point reads`
-|              167.44 |        5,972,130.71 |    0.2% |        3,065.14 |          862.27 |  3.555 |         494.30 |    0.0% |      0.01 | `prefix reads`
-|              238.77 |        4,188,130.84 |    0.9% |        3,589.93 |        1,259.30 |  2.851 |         637.12 |    0.0% |      0.01 | `range reads`
-|              424.01 |        2,358,426.70 |    0.2% |        5,620.05 |        2,242.35 |  2.506 |         854.80 |    1.7% |      0.01 | `point writes`
-|              418.45 |        2,389,780.56 |    0.4% |        5,525.07 |        2,211.05 |  2.499 |         831.71 |    1.7% |      0.01 | `prefix writes`
-|              254.87 |        3,923,568.88 |    2.6% |        3,187.01 |        1,366.50 |  2.332 |         529.11 |    2.7% |      0.02 | `range writes`
-|              675.96 |        1,479,374.50 |    3.3% |        7,735.41 |        3,468.60 |  2.230 |       1,386.02 |    1.8% |      0.01 | `monotonic increasing point writes`
-|          137,986.20 |            7,247.10 |    0.6% |      789,752.33 |      699,462.00 |  1.129 |     144,824.14 |    0.0% |      0.01 | `worst case for radix tree`
-|               21.63 |       46,231,564.03 |    1.0% |          448.00 |          107.14 |  4.181 |          84.00 |    0.0% |      0.01 | `create and destroy`
+|              171.37 |        5,835,467.28 |    0.5% |        2,915.46 |          629.31 |  4.633 |         502.91 |    0.0% |      0.01 | `point reads`
+|              168.16 |        5,946,774.62 |    0.5% |        2,859.73 |          619.03 |  4.620 |         488.59 |    0.0% |      0.01 | `prefix reads`
+|              245.86 |        4,067,294.63 |    0.3% |        3,508.97 |          904.82 |  3.878 |         627.73 |    0.0% |      0.01 | `range reads`
+|              381.14 |        2,623,691.76 |    0.2% |        5,127.16 |        1,402.60 |  3.655 |         830.59 |    1.7% |      0.01 | `point writes`
+|              376.24 |        2,657,887.07 |    0.3% |        5,062.85 |        1,381.07 |  3.666 |         811.68 |    1.6% |      0.01 | `prefix writes`
+|              257.09 |        3,889,688.44 |    3.8% |        3,010.09 |          949.42 |  3.170 |         521.34 |    2.9% |      0.01 | `range writes`
+|              629.71 |        1,588,026.28 |    4.0% |        6,809.60 |        2,320.32 |  2.935 |       1,247.08 |    1.3% |      0.01 | `monotonic increasing point writes`
+|          126,164.50 |            7,926.16 |    1.7% |      772,460.25 |      463,264.67 |  1.667 |     141,512.00 |    0.2% |      0.01 | `worst case for radix tree`
+|               15.33 |       65,217,669.23 |    0.4% |          299.00 |           56.45 |  5.297 |          64.00 |    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
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|               11.04 |       90,614,308.12 |    0.8% |          180.38 |           55.13 |  3.272 |          41.51 |    0.4% |      0.01 | `point reads`
-|               14.96 |       66,843,629.12 |    0.4% |          274.41 |           74.73 |  3.672 |          55.05 |    0.3% |      0.01 | `prefix reads`
-|               37.06 |       26,982,847.61 |    0.2% |          791.04 |          185.28 |  4.269 |         142.67 |    0.2% |      0.01 | `range reads`
-|               17.89 |       55,887,365.73 |    0.6% |          335.54 |           89.79 |  3.737 |          43.84 |    0.4% |      0.01 | `point writes`
-|               31.85 |       31,394,336.65 |    0.3% |          615.32 |          159.63 |  3.855 |          87.69 |    0.2% |      0.01 | `prefix writes`
-|               36.17 |       27,647,221.45 |    0.6% |          705.11 |          182.80 |  3.857 |         100.62 |    0.1% |      0.01 | `range writes`
-|               79.01 |       12,656,457.78 |    0.7% |        1,498.35 |          402.46 |  3.723 |         270.50 |    0.1% |      0.01 | `monotonic increasing point writes`
-|          303,667.50 |            3,293.08 |    1.1% |    3,931,273.00 |    1,612,702.50 |  2.438 |     806,223.33 |    0.0% |      0.01 | `worst case for radix tree`
-|               83.70 |       11,947,443.83 |    0.7% |        1,738.03 |          429.06 |  4.051 |         270.01 |    0.0% |      0.01 | `create and destroy`
+|               12.81 |       78,088,759.03 |    0.7% |          187.38 |           47.15 |  3.974 |          33.70 |    0.5% |      0.01 | `point reads`
+|               16.77 |       59,626,501.14 |    0.4% |          277.61 |           61.74 |  4.497 |          44.76 |    0.4% |      0.01 | `prefix reads`
+|               42.45 |       23,556,818.87 |    0.4% |          856.96 |          156.09 |  5.490 |         126.22 |    0.2% |      0.01 | `range reads`
+|               15.76 |       63,453,196.94 |    0.3% |          249.89 |           58.04 |  4.306 |          33.21 |    0.5% |      0.01 | `point writes`
+|               29.71 |       33,662,499.39 |    0.1% |          509.14 |          109.33 |  4.657 |          70.26 |    0.5% |      0.01 | `prefix writes`
+|               35.77 |       27,956,388.03 |    0.8% |          614.18 |          131.72 |  4.663 |          85.57 |    0.1% |      0.01 | `range writes`
+|               80.13 |       12,480,456.47 |    1.5% |        1,279.04 |          294.90 |  4.337 |         234.68 |    0.1% |      0.01 | `monotonic increasing point writes`
+|          360,748.00 |            2,772.02 |    2.6% |    4,401,835.00 |    1,318,754.00 |  3.338 |     728,537.00 |    0.0% |      0.01 | `worst case for radix tree`
+|               90.69 |       11,027,103.45 |    1.0% |        1,686.00 |          333.80 |  5.051 |         284.00 |    0.0% |      0.01 | `create and destroy`
 
 # "Real data" test
 
-Point queries only, best of three runs. Gc ratio is the ratio of time spent doing garbage collection to time spent adding writes or doing garbage collection. Lower is better.
+Point queries only. Gc ratio is the ratio of time spent doing garbage collection to time spent adding writes or doing garbage collection. Lower is better.
 
 ## skip list
 
 ```
-Check: 4.47891 seconds, 364.05 MB/s, Add: 4.55599 seconds, 123.058 MB/s, Gc ratio: 37.1145%
+Check: 4.42779 seconds, 368.253 MB/s, Add: 4.64942 seconds, 120.586 MB/s, Gc ratio: 28.8784%, Peak idle memory: 5.51735e+06
 ```
 
 ## radix tree
 
 ```
-Check: 0.958985 seconds, 1700.28 MB/s, Add: 1.35083 seconds, 415.044 MB/s, Gc ratio: 44.4768%, Peak idle memory: 2.33588e+06
+Check: 0.891163 seconds, 1829.69 MB/s, Add: 1.23064 seconds, 455.579 MB/s, Gc ratio: 42.9417%, Peak idle memory: 2.28333e+06
 ```
 
 ## hash table
@@ -55,5 +65,6 @@ Check: 0.958985 seconds, 1700.28 MB/s, Add: 1.35083 seconds, 415.044 MB/s, Gc ra
 (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: 0.804094 seconds, 2027.81 MB/s, Add: 0.652952 seconds, 858.645 MB/s, Gc ratio: 35.3885%
+Check: 0.858484 seconds, 1899.33 MB/s, Add: 0.632886 seconds, 885.868 MB/s, Gc ratio: 41.252%, Peak idle memory: 0
 ```
+

ServerBench.cpp

@@ -6,11 +6,15 @@
 #include <string.h>
 #include <string>
 #include <string_view>
+#include <sys/ioctl.h>
 #include <sys/resource.h>
 #include <sys/socket.h>
+#include <sys/types.h>
 #include <sys/uio.h>
 #include <thread>
 #include <unistd.h>
+#include <utility>
+#include <vector>
 
 #include "ConflictSet.h"
 #include "third_party/nadeau.h"
@@ -20,8 +24,8 @@ std::atomic<int64_t> transactions;
 constexpr int kBaseSearchDepth = 32;
 constexpr int kWindowSize = 10000000;
 
-std::basic_string<uint8_t> numToKey(int64_t num) {
-  std::basic_string<uint8_t> result;
+std::string numToKey(int64_t num) {
+  std::string result;
   result.resize(kBaseSearchDepth + sizeof(int64_t));
   memset(result.data(), 0, kBaseSearchDepth);
   int64_t be = __builtin_bswap64(num);
@@ -41,13 +45,13 @@ void workload(weaselab::ConflictSet *cs) {
       auto pointK = numToKey(pointRv);
       weaselab::ConflictSet::ReadRange reads[] = {
           {
-              {pointK.data(), int(pointK.size())},
+              {(const uint8_t *)pointK.data(), int(pointK.size())},
               {nullptr, 0},
               pointRv,
           },
           {
-              {beginK.data(), int(beginK.size())},
-              {endK.data(), int(endK.size())},
+              {(const uint8_t *)beginK.data(), int(beginK.size())},
+              {(const uint8_t *)endK.data(), int(endK.size())},
               version - 2,
           },
       };
@@ -66,7 +70,7 @@ void workload(weaselab::ConflictSet *cs) {
     {
       weaselab::ConflictSet::WriteRange w;
       auto k = numToKey(version);
-      w.begin.p = k.data();
+      w.begin.p = (const uint8_t *)k.data();
       w.end.len = 0;
       if (version % (kWindowSize / 2) == 0) {
         for (int l = 0; l <= k.size(); ++l) {
@@ -79,9 +83,9 @@ void workload(weaselab::ConflictSet *cs) {
         int64_t beginN = version - kWindowSize + rand() % kWindowSize;
         auto b = numToKey(beginN);
         auto e = numToKey(beginN + 1000);
-        w.begin.p = b.data();
+        w.begin.p = (const uint8_t *)b.data();
         w.begin.len = b.size();
-        w.end.p = e.data();
+        w.end.p = (const uint8_t *)e.data();
         w.end.len = e.size();
         cs->addWrites(&w, 1, version);
       }
@@ -164,36 +168,29 @@ double toSeconds(timeval t) {
   return double(t.tv_sec) + double(t.tv_usec) * 1e-6;
 }
 
-#include <linux/perf_event.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-#include <sys/ioctl.h>
-#include <sys/syscall.h>
-#include <sys/types.h>
-#include <unistd.h>
-
 #ifdef __linux__
+#include <linux/perf_event.h>
 struct PerfCounter {
-  explicit PerfCounter(int event) {
+  PerfCounter(int type, int config, const std::string &labels = {},
+              int groupLeaderFd = -1)
+      : labels(labels) {
     struct perf_event_attr pe;
 
     memset(&pe, 0, sizeof(pe));
-    pe.type = PERF_TYPE_HARDWARE;
+    pe.type = type;
     pe.size = sizeof(pe);
-    pe.config = event;
+    pe.config = config;
     pe.inherit = 1;
     pe.exclude_kernel = 1;
     pe.exclude_hv = 1;
 
-    fd = perf_event_open(&pe, 0, -1, -1, 0);
-    if (fd == -1) {
-      fprintf(stderr, "Error opening leader %llx\n", pe.config);
-      exit(EXIT_FAILURE);
+    fd = perf_event_open(&pe, 0, -1, groupLeaderFd, 0);
+    if (fd < 0 && errno != ENOENT && errno != EINVAL) {
+      perror(labels.c_str());
     }
   }
 
-  int64_t total() {
+  int64_t total() const {
     int64_t count;
     if (read(fd, &count, sizeof(count)) != sizeof(count)) {
       perror("read instructions from perf");
@@ -202,10 +199,27 @@ struct PerfCounter {
     return count;
   }
 
-  ~PerfCounter() { close(fd); }
+  PerfCounter(PerfCounter &&other)
+      : fd(std::exchange(other.fd, -1)), labels(std::move(other.labels)) {}
+  PerfCounter &operator=(PerfCounter &&other) {
+    fd = std::exchange(other.fd, -1);
+    labels = std::move(other.labels);
+    return *this;
+  }
+
+  ~PerfCounter() {
+    if (fd >= 0) {
+      close(fd);
+    }
+  }
+
+  bool ok() const { return fd >= 0; }
+  const std::string &getLabels() const { return labels; }
+  int getFd() const { return fd; }
 
 private:
   int fd;
+  std::string labels;
   static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
                               int cpu, int group_fd, unsigned long flags) {
     int ret;
@@ -214,11 +228,6 @@ private:
     return ret;
   }
 };
-#else
-struct PerfCounter {
-  explicit PerPerfCounter(int) {}
-  int64_t total() { return 0; }
-};
 #endif
 
 int main(int argc, char **argv) {
@@ -233,8 +242,50 @@ int main(int argc, char **argv) {
     int metricsCount;
     cs.getMetricsV1(&metrics, &metricsCount);
 
-    PerfCounter instructions{PERF_COUNT_HW_INSTRUCTIONS};
-    PerfCounter cycles{PERF_COUNT_HW_CPU_CYCLES};
+#ifdef __linux__
+    PerfCounter instructions{PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS};
+    PerfCounter cycles{PERF_TYPE_HARDWARE, PERF_COUNT_HW_CPU_CYCLES, "",
+                       instructions.getFd()};
+
+    std::vector<PerfCounter> cacheCounters;
+    for (auto [id, idStr] : std::initializer_list<std::pair<int, std::string>>{
+             {PERF_COUNT_HW_CACHE_L1D, "l1d"},
+             {PERF_COUNT_HW_CACHE_L1I, "l1i"},
+             {PERF_COUNT_HW_CACHE_LL, "ll"},
+             {PERF_COUNT_HW_CACHE_DTLB, "dtlb"},
+             {PERF_COUNT_HW_CACHE_ITLB, "itlb"},
+             {PERF_COUNT_HW_CACHE_BPU, "bpu"},
+             {PERF_COUNT_HW_CACHE_NODE, "node"},
+         }) {
+      for (auto [op, opStr] :
+           std::initializer_list<std::pair<int, std::string>>{
+               {PERF_COUNT_HW_CACHE_OP_READ, "read"},
+               {PERF_COUNT_HW_CACHE_OP_WRITE, "write"},
+               {PERF_COUNT_HW_CACHE_OP_PREFETCH, "prefetch"},
+           }) {
+        int groupLeaderFd = -1;
+        for (auto [result, resultStr] :
+             std::initializer_list<std::pair<int, std::string>>{
+                 {PERF_COUNT_HW_CACHE_RESULT_MISS, "miss"},
+                 {PERF_COUNT_HW_CACHE_RESULT_ACCESS, "access"},
+             }) {
+          auto labels = "{id=\"" + idStr + "\", op=\"" + opStr +
+                        "\", result=\"" + resultStr + "\"}";
+          cacheCounters.emplace_back(PERF_TYPE_HW_CACHE,
+                                     id | (op << 8) | (result << 16), labels,
+                                     groupLeaderFd);
+          if (!cacheCounters.back().ok()) {
+            cacheCounters.pop_back();
+          } else {
+            if (groupLeaderFd == -1) {
+              groupLeaderFd = cacheCounters.back().getFd();
+            }
+          }
+        }
+      }
+    }
+#endif
+
     auto w = std::thread{workload, &cs};
 
     for (;;) {
@@ -262,6 +313,7 @@ int main(int argc, char **argv) {
               "transactions_total ";
       body += std::to_string(transactions.load(std::memory_order_relaxed));
       body += "\n";
+#ifdef __linux__
       body += "# HELP instructions_total Total number of instructions\n"
               "# TYPE instructions_total counter\n"
               "instructions_total ";
@@ -272,6 +324,13 @@ int main(int argc, char **argv) {
               "cycles_total ";
       body += std::to_string(cycles.total());
       body += "\n";
+      body += "# HELP cache_event_total Total number of cache events\n"
+              "# TYPE cache_event_total counter\n";
+      for (const auto &counter : cacheCounters) {
+        body += "cache_event_total" + counter.getLabels() + " " +
+                std::to_string(counter.total()) + "\n";
+      }
+#endif
 
       for (int i = 0; i < metricsCount; ++i) {
         body += "# HELP ";

SkipList.cpp

@@ -25,6 +25,7 @@
 
 #include <algorithm>
 #include <span>
+#include <vector>
 
 std::span<const uint8_t> keyAfter(Arena &arena, std::span<const uint8_t> key) {
   auto result =
@@ -115,15 +116,6 @@ bool operator==(const KeyInfo &lhs, const KeyInfo &rhs) {
   return !(lhs < rhs || rhs < lhs);
 }
 
-void swapSort(std::vector<KeyInfo> &points, int a, int b) {
-  if (points[b] < points[a]) {
-    KeyInfo temp;
-    temp = points[a];
-    points[a] = points[b];
-    points[b] = temp;
-  }
-}
-
 struct SortTask {
   int begin;
   int size;
@@ -183,13 +175,6 @@ void sortPoints(std::vector<KeyInfo> &points) {
   }
 }
 
-static thread_local uint32_t g_seed = 0;
-
-static inline int skfastrand() {
-  g_seed = g_seed * 1664525L + 1013904223L;
-  return g_seed;
-}
-
 static int compare(const StringRef &a, const StringRef &b) {
   int c = memcmp(a.data(), b.data(), std::min(a.size(), b.size()));
   if (c < 0)
@@ -215,21 +200,25 @@ struct ReadConflictRange {
   }
 };
 
-class SkipList {
-private:
 static constexpr int MaxLevels = 26;
 
-  int randomLevel() const {
-    uint32_t i = uint32_t(skfastrand()) >> (32 - (MaxLevels - 1));
-    int level = 0;
-    while (i & 1) {
-      i >>= 1;
-      level++;
-    }
-    assert(level < MaxLevels);
-    return level;
+struct RandomLevel {
+  explicit RandomLevel(uint32_t seed) : seed(seed) {}
+
+  int next() {
+    int result = __builtin_clz(seed | (uint32_t(-1) >> (MaxLevels - 1)));
+    seed = seed * 1664525L + 1013904223L;
+    return result;
   }
 
+private:
+  uint32_t seed;
+};
+
+class SkipList {
+private:
+  RandomLevel randomLevel{0};
+
   // Represent a node in the SkipList. The node has multiple (i.e., level)
   // pointers to other nodes, and keeps a record of the max versions for each
   // level.
@@ -426,18 +415,23 @@ public:
   }
   void swap(SkipList &other) { std::swap(header, other.header); }
 
-  void addConflictRanges(const Finger *fingers, int rangeCount,
+  // Returns the change in the number of entries
+  int64_t addConflictRanges(const Finger *fingers, int rangeCount,
                             Version version) {
+    int64_t result = rangeCount;
     for (int r = rangeCount - 1; r >= 0; r--) {
       const Finger &startF = fingers[r * 2];
       const Finger &endF = fingers[r * 2 + 1];
 
-      if (endF.found() == nullptr)
+      if (endF.found() == nullptr) {
+        ++result;
         insert(endF, endF.finger[0]->getMaxVersion(0));
+      }
 
-      remove(startF, endF);
+      result -= remove(startF, endF);
       insert(startF, version);
     }
+    return result;
   }
 
   void detectConflicts(ReadConflictRange *ranges, int count,
@@ -567,9 +561,10 @@ public:
   }
 
 private:
-  void remove(const Finger &start, const Finger &end) {
+  // Returns the number of entries removed
+  int64_t remove(const Finger &start, const Finger &end) {
     if (start.finger[0] == end.finger[0])
-      return;
+      return 0;
 
     Node *x = start.finger[0]->getNext(0);
 
@@ -578,17 +573,20 @@ private:
       if (start.finger[i] != end.finger[i])
         start.finger[i]->setNext(i, end.finger[i]->getNext(i));
 
+    int64_t result = 0;
     while (true) {
       Node *next = x->getNext(0);
       x->destroy();
+      ++result;
       if (x == end.finger[0])
         break;
       x = next;
     }
+    return result;
   }
 
   void insert(const Finger &f, Version version) {
-    int level = randomLevel();
+    int level = randomLevel.next();
     // std::cout << std::string((const char*)value,length) << " level: " <<
     // level << std::endl;
     Node *x = Node::create(f.value, level);
@@ -704,8 +702,6 @@ private:
   };
 };
 
-struct SkipListConflictSet {};
-
 struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   Impl(int64_t oldestVersion)
       : oldestVersion(oldestVersion), newestVersion(oldestVersion),
@@ -775,17 +771,20 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     StringRef values[stripeSize];
     int64_t writeVersions[stripeSize / 2];
     int ss = stringCount - (stripes - 1) * stripeSize;
+    int64_t entryDelta = 0;
     for (int s = stripes - 1; s >= 0; s--) {
       for (int i = 0; i * 2 < ss; ++i) {
         const auto &w = combinedWriteConflictRanges[s * stripeSize / 2 + i];
         values[i * 2] = w.first;
         values[i * 2 + 1] = w.second;
-        keyUpdates += 3;
       }
       skipList.find(values, fingers, temp, ss);
-      skipList.addConflictRanges(fingers, ss / 2, writeVersion);
+      entryDelta += skipList.addConflictRanges(fingers, ss / 2, writeVersion);
       ss = stripeSize;
     }
+
+    // Run gc at least 200% the rate we're inserting entries
+    keyUpdates += std::max<int64_t>(entryDelta, 0) * 2;
   }
 
   void setOldestVersion(int64_t oldestVersion) {
@@ -795,7 +794,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     int temp;
     std::span<const uint8_t> key = removalKey;
     skipList.find(&key, &finger, &temp, 1);
-    skipList.removeBefore(oldestVersion, finger, std::exchange(keyUpdates, 10));
+    skipList.removeBefore(oldestVersion, finger, std::exchange(keyUpdates, 0));
     removalArena = Arena();
     removalKey = copyToArena(
         removalArena, {finger.getValue().data(), finger.getValue().size()});
@@ -804,7 +803,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   int64_t totalBytes = 0;
 
 private:
-  int64_t keyUpdates = 10;
+  int64_t keyUpdates = 0;
   Arena removalArena;
   std::span<const uint8_t> removalKey;
   int64_t oldestVersion;
@@ -904,7 +903,12 @@ ConflictSet::ConflictSet(ConflictSet &&other) noexcept
     : impl(std::exchange(other.impl, nullptr)) {}
 
 ConflictSet &ConflictSet::operator=(ConflictSet &&other) noexcept {
+  if (this != &other) {
+    if (impl) {
+      internal_destroy(impl);
+    }
     impl = std::exchange(other.impl, nullptr);
+  }
   return *this;
 }