Merge pull request 'Replace SIMD assembly with initialized Node16::index + valgrind client request' (#77 ) from client-request-instead-of-asm into main

Reviewed-on: #77
Poison Node48 reverseIndex slot on erase
2026-08-04 00:07:19 +00:00 · 2026-08-03 17:38:21 -04:00 · 2026-08-03 17:32:41 -04:00 · 2026-08-03 13:42:45 +00:00 · 2026-08-03 09:17:28 -04:00 · 2026-08-03 12:27:49 +00:00
94 changed files with 2094 additions and 194 deletions
@@ -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
@@ -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)
@@ -253,7 +260,7 @@ 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 ${TEST_FLAGS})
  if(NOT CMAKE_CROSSCOMPILING)
@@ -268,6 +275,34 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
    add_test(NAME conflict_set_fuzz_${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)
@@ -488,6 +488,7 @@ 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 +500,12 @@ 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 +522,7 @@ 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 +536,12 @@ 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 +549,7 @@ 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 +569,13 @@ 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++). valgrind then treats them as undefined
+  // so that it can still flag any read that relies on them.
+  VALGRIND_MAKE_MEM_UNDEFINED(reverseIndex + numChildren,
+                              sizeof(reverseIndex) -
+                                  numChildren * sizeof(reverseIndex[0]));
 }

 inline void Node48::copyChildrenAndKeyFrom(const Node48 &other) {
@@ -580,6 +602,7 @@ 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 +624,13 @@ 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++). valgrind then treats them as undefined
+  // so that it can still flag any read that relies on them.
+  VALGRIND_MAKE_MEM_UNDEFINED(reverseIndex + numChildren,
+                              sizeof(reverseIndex) -
+                                  numChildren * sizeof(reverseIndex[0]));
 }

 inline void Node256::copyChildrenAndKeyFrom(const Node48 &other) {
@@ -1332,8 +1362,8 @@ 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) {
@@ -1995,6 +2025,8 @@ Node *erase(Node *self, WriteContext *writeContext, bool logical) {
    parent48->childMaxVersion[lastChildrenIndex] = writeContext->zero;
    VALGRIND_MAKE_MEM_UNDEFINED(parent48->children + lastChildrenIndex,
                                sizeof(parent48->children[0]));
+    VALGRIND_MAKE_MEM_UNDEFINED(parent48->reverseIndex + lastChildrenIndex,
+                                sizeof(parent48->reverseIndex[0]));

    if (needsDownsize(parent48)) {
      downsize(parent48, writeContext);
@@ -2087,6 +2119,52 @@ 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(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 +2185,10 @@ 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 +2215,14 @@ 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 +2238,10 @@ 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;
@@ -2250,17 +2298,9 @@ 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 +2308,11 @@ 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 +2355,11 @@ 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 +2394,9 @@ 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 +2404,7 @@ 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;
@@ -3864,17 +3882,17 @@ 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;
@@ -5046,8 +5064,8 @@ 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 {
@@ -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
@@ -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`
+|              164.29 |        6,086,873.38 |    0.0% |        3,107.03 |          604.19 |  5.142 |         558.59 |    0.0% |      1.96 | `point reads`
+|              161.05 |        6,209,395.38 |    0.1% |        3,036.76 |          592.21 |  5.128 |         539.35 |    0.0% |      1.93 | `prefix reads`
+|              239.55 |        4,174,539.38 |    0.1% |        3,722.71 |          880.68 |  4.227 |         692.00 |    0.0% |      2.86 | `range reads`
+|              354.75 |        2,818,919.14 |    0.7% |        4,523.64 |        1,304.75 |  3.467 |         720.22 |    2.0% |      4.23 | `point writes`
+|              345.32 |        2,895,878.47 |    0.1% |        4,484.57 |        1,270.31 |  3.530 |         705.00 |    1.8% |      4.12 | `prefix writes`
+|              193.48 |        5,168,547.42 |    0.1% |        2,224.10 |          711.72 |  3.125 |         377.17 |    3.3% |      2.32 | `range writes`
+|              404.89 |        2,469,777.50 |    2.4% |        6,855.96 |        1,489.70 |  4.602 |       1,227.82 |    1.3% |      0.05 | `monotonic increasing point writes`
+|          134,231.80 |            7,449.80 |    1.9% |      812,045.25 |      495,770.40 |  1.638 |     151,246.50 |    0.9% |      0.01 | `worst case for radix tree`
+|               37.80 |       26,454,311.17 |    0.4% |          701.00 |          139.14 |  5.038 |         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`
+|               12.89 |       77,565,115.56 |    0.1% |          244.55 |           47.43 |  5.155 |          34.21 |    0.6% |      0.15 | `point reads`
+|               15.11 |       66,162,047.76 |    0.1% |          297.79 |           55.60 |  5.356 |          43.23 |    0.4% |      0.18 | `prefix reads`
+|               36.29 |       27,559,358.29 |    0.1% |          783.16 |          133.44 |  5.869 |         109.52 |    0.2% |      0.43 | `range reads`
+|               20.53 |       48,719,405.55 |    0.1% |          381.81 |           75.51 |  5.057 |          51.04 |    0.5% |      0.25 | `point writes`
+|               39.37 |       25,402,042.40 |    0.1% |          685.00 |          144.83 |  4.730 |         106.72 |    0.3% |      0.47 | `prefix writes`
+|               43.78 |       22,843,841.63 |    0.1% |          800.40 |          161.06 |  4.970 |         127.36 |    0.1% |      0.53 | `range writes`
+|               78.37 |       12,760,008.75 |    1.0% |        1,452.61 |          288.24 |  5.040 |         278.69 |    0.1% |      0.01 | `monotonic increasing point writes`
+|          322,885.50 |            3,097.07 |    1.5% |    4,362,382.00 |    1,183,852.00 |  3.685 |     765,301.00 |    0.1% |      0.01 | `worst case for radix tree`
+|               99.99 |       10,000,718.79 |    0.4% |        1,775.00 |          367.93 |  4.824 |         288.00 |    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: 4.62967 seconds, 352.195 MB/s, Add: 3.34177 seconds, 167.771 MB/s, Gc ratio: 37.9399%, 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: 1.00477 seconds, 1622.8 MB/s, Add: 1.21142 seconds, 462.808 MB/s, Gc ratio: 39.4716%, 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: 0.854254 seconds, 1908.74 MB/s, Add: 0.632626 seconds, 886.232 MB/s, Gc ratio: 41.0827%, Peak idle memory: 0
 ```

@@ -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}"