Run valgrind in arm64 release in CI

2026-08-02 21:57:06 -04:00
16 changed files with 323 additions and 733 deletions
@@ -38,43 +38,21 @@ 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
-            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++
-          - 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 }}
+    runs-on: ubuntu-latest-amd64
    steps:
      - uses: actions/checkout@v4

      - uses: actions/cache@v4
        with:
          path: /var/cache/apt/archives
-          key: apt-${{ matrix.arch }}-${{ hashFiles('.gitea/workflows/ci.yml') }}
+          key: apt-amd64-${{ hashFiles('.gitea/workflows/ci.yml') }}

      - name: Install common dependencies
        run: |
@@ -90,18 +68,18 @@ jobs:
        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}-${{ matrix.llvm_version }} main" | sudo tee /etc/apt/sources.list.d/llvm.list
+          echo "deb http://apt.llvm.org/${VERSION_CODENAME}/ llvm-toolchain-${VERSION_CODENAME}-21 main" | sudo tee /etc/apt/sources.list.d/llvm.list
          sudo apt-get update -qq
          sudo apt-get install -y \
-            clang-${{ matrix.llvm_version }} llvm-${{ matrix.llvm_version }} lld-${{ matrix.llvm_version }} mold
+            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}-${{ matrix.llvm_version }} 100
+            sudo update-alternatives --install /usr/bin/${tool} ${tool} /usr/bin/${tool}-21 100
          done

      - name: Download MSan toolchain
-        if: matrix.msan_url != ''
+        if: matrix.name == 'debug'
        run: |
-          curl -Ls "${{ matrix.msan_url }}" -o /tmp/msan-toolchain.tar.zst
+          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

@@ -5,7 +5,7 @@ project(
  DESCRIPTION
    "A data structure for optimistic concurrency control on ranges of bitwise-lexicographically-ordered keys."
  HOMEPAGE_URL "https://git.weaselab.dev/weaselab/conflict-set"
-  LANGUAGES C CXX)
+  LANGUAGES C CXX ASM)
 set(CMAKE_CXX_STANDARD 20)

 file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/version.txt ${PROJECT_VERSION})
@@ -130,7 +130,15 @@ endif()

 set(CMAKE_CXX_IMPLICIT_LINK_LIBRARIES "")

-add_library(${PROJECT_NAME}-object OBJECT ConflictSet.cpp)
+# Architecture-specific SIMD assembly. These functions operate on
+# potentially-indeterminate memory, which is UB in C++ but well-defined in
+# assembly.
+set(SIMD_ASM_FILES)
+if(CMAKE_SYSTEM_PROCESSOR STREQUAL x86_64 AND NOT USE_SIMD_FALLBACK)
+  set(SIMD_ASM_FILES ${CMAKE_CURRENT_SOURCE_DIR}/simd_x86_64.S)
+endif()
+
+add_library(${PROJECT_NAME}-object OBJECT ConflictSet.cpp ${SIMD_ASM_FILES})
 target_compile_options(${PROJECT_NAME}-object PRIVATE -fno-exceptions
                                                      -fvisibility=hidden)
 target_include_directories(${PROJECT_NAME}-object
@@ -182,8 +190,6 @@ endif()

 include(CTest)

-option(TEST_SKIP_LIST "Run the slow skip list cross-check tests" OFF)
-
 # disable tests if this is being used through e.g. FetchContent
 if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)

@@ -226,9 +232,8 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
  target_compile_options(driver_skip_list PRIVATE ${TEST_FLAGS})
  target_link_libraries(driver_skip_list PRIVATE skip_list)

-  # Optional cross-check of SkipList.cpp against ReferenceImpl. Opt-in via
-  # -DTEST_SKIP_LIST=ON, since these are slow.
-  if(TEST_SKIP_LIST)
+  # enable to test skip list
+  if(0)
    foreach(TEST ${CORPUS_TESTS})
      get_filename_component(hash ${TEST} NAME)
      add_test(NAME skip_list_${hash} COMMAND driver_skip_list ${TEST})
@@ -236,7 +241,7 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
  endif()

  # ad hoc testing
-  add_executable(conflict_set_main ConflictSet.cpp)
+  add_executable(conflict_set_main ConflictSet.cpp ${SIMD_ASM_FILES})
  target_include_directories(conflict_set_main
                             PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include)
  target_compile_definitions(conflict_set_main PRIVATE ENABLE_MAIN)
@@ -252,7 +257,7 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
    cmake_pop_check_state()

    if(HAS_LIB_FUZZER)
-      add_executable(conflict_set_fuzz_test ConflictSet.cpp)
+      add_executable(conflict_set_fuzz_test ConflictSet.cpp ${SIMD_ASM_FILES})
      target_include_directories(conflict_set_fuzz_test
                                 PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include)
      target_compile_definitions(conflict_set_fuzz_test PRIVATE ENABLE_FUZZ)
@@ -264,7 +269,8 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
  endif()

  # whitebox tests asan+ubsan
-  add_executable(fuzz_driver ConflictSet.cpp FuzzTestDriver.cpp)
+  add_executable(fuzz_driver ConflictSet.cpp FuzzTestDriver.cpp
+                             ${SIMD_ASM_FILES})
  target_compile_options(fuzz_driver PRIVATE ${TEST_FLAGS})
  if(NOT CMAKE_CROSSCOMPILING)
    target_compile_options(fuzz_driver PRIVATE -fsanitize=address,undefined)
@@ -280,7 +286,8 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)

  # whitebox tests msan
  if(MSAN_TOOLCHAIN_PATH)
-    add_executable(fuzz_driver_msan ConflictSet.cpp FuzzTestDriver.cpp)
+    add_executable(fuzz_driver_msan ConflictSet.cpp FuzzTestDriver.cpp
+                                    ${SIMD_ASM_FILES})
    target_compile_options(fuzz_driver_msan PRIVATE ${TEST_FLAGS})
    if(NOT CMAKE_CROSSCOMPILING)
      target_compile_options(
@@ -308,7 +315,8 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)

  # tsan tests
  if(NOT CMAKE_CROSSCOMPILING AND NOT DISABLE_TSAN)
-    add_executable(tsan_driver ConflictSet.cpp FuzzTestDriver.cpp)
+    add_executable(tsan_driver ConflictSet.cpp FuzzTestDriver.cpp
+                               ${SIMD_ASM_FILES})
    target_compile_options(tsan_driver PRIVATE ${TEST_FLAGS} -fsanitize=thread)
    target_link_options(tsan_driver PRIVATE -fsanitize=thread)
    target_compile_definitions(tsan_driver PRIVATE ENABLE_FUZZ THREAD_TEST)
@@ -393,24 +401,6 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
                        PROPERTIES CXX_STANDARD_REQUIRED ON)
  add_test(NAME conflict_set_cxx_api_test COMMAND conflict_set_cxx_api_test)

-  # SHOW_MEMORY instrumentation, which the rest of the test suite doesn't
-  # compile with
-  add_executable(conflict_set_show_memory_test ConflictSet.cpp
-                                               ShowMemoryTest.cpp)
-  target_compile_options(conflict_set_show_memory_test PRIVATE ${TEST_FLAGS})
-  target_compile_definitions(conflict_set_show_memory_test
-                             PRIVATE SHOW_MEMORY=1)
-  target_include_directories(conflict_set_show_memory_test
-                             PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include)
-  if(NOT CMAKE_CROSSCOMPILING)
-    target_compile_options(conflict_set_show_memory_test
-                           PRIVATE -fsanitize=address,undefined)
-    target_link_options(conflict_set_show_memory_test PRIVATE
-                        -fsanitize=address,undefined)
-  endif()
-  add_test(NAME conflict_set_show_memory_test
-           COMMAND conflict_set_show_memory_test)
-
  # symbol visibility tests
  if(NOT CMAKE_BUILD_TYPE STREQUAL Debug)
    if(APPLE)
@@ -28,6 +28,7 @@ limitations under the License.
 #include "Internal.h"
 #include "LongestCommonPrefix.h"
 #include "Metrics.h"
+#include "simd.h"

 #include <algorithm>
 #include <bit>
@@ -80,19 +81,17 @@ using namespace weaselab;
 #define assume assert
 #endif

-// The SHOW_MEMORY instrumentation. Declared unconditionally so that each of
-// these has a single definition in every configuration (see the definitions at
-// the end of this file), rather than a no-op stub standing in for the real
-// thing.
+#if SHOW_MEMORY
 void addNode(struct Node *);
 void removeNode(struct Node *);
 void addKey(struct Node *);
 void removeKey(struct Node *);
-// Like `removeKey`, but with a caller-provided search path length. Used when
-// the node's ancestors may already have been released, which makes walking
-// `parent` pointers invalid.
-void removeKey(struct Node *, int64_t searchPathLength);
-int64_t getSearchPathLength(struct Node *);
+#else
+constexpr void addNode(struct Node *) {}
+constexpr void removeNode(struct Node *) {}
+constexpr void addKey(struct Node *) {}
+constexpr void removeKey(struct Node *) {}
+#endif

 // ==================== BEGIN IMPLEMENTATION ====================

@@ -118,40 +117,13 @@ struct InternalVersionT {
  constexpr InternalVersionT() = default;
  constexpr explicit InternalVersionT(int64_t value) : value(value) {}
  constexpr int64_t toInt64() const { return value; } // GCOVR_EXCL_LINE
-  constexpr bool operator<(const InternalVersionT &rhs) const {
+  constexpr auto operator<=>(const InternalVersionT &rhs) const {
 #if USE_64_BIT
-    return value < rhs.value;
+    return value <=> rhs.value;
 #else
    // Maintains ordering after overflow, as long as the full-precision versions
    // are within `kMaxCorrectVersionWindow` of eachother.
-    return int32_t(value - rhs.value) < 0;
-#endif
-  }
-  constexpr bool operator<=(const InternalVersionT &rhs) const {
-#if USE_64_BIT
-    return value <= rhs.value;
-#else
-    // Maintains ordering after overflow, as long as the full-precision versions
-    // are within `kMaxCorrectVersionWindow` of eachother.
-    return int32_t(value - rhs.value) <= 0;
-#endif
-  }
-  constexpr bool operator>=(const InternalVersionT &rhs) const {
-#if USE_64_BIT
-    return value >= rhs.value;
-#else
-    // Maintains ordering after overflow, as long as the full-precision versions
-    // are within `kMaxCorrectVersionWindow` of eachother.
-    return int32_t(value - rhs.value) >= 0;
-#endif
-  }
-  constexpr bool operator>(const InternalVersionT &rhs) const {
-#if USE_64_BIT
-    return value > rhs.value;
-#else
-    // Maintains ordering after overflow, as long as the full-precision versions
-    // are within `kMaxCorrectVersionWindow` of eachother.
-    return int32_t(value - rhs.value) > 0;
+    return int32_t(value - rhs.value) <=> 0;
 #endif
  }
  constexpr bool operator==(const InternalVersionT &) const = default;
@@ -336,9 +308,7 @@ struct Node {
  uint8_t *partialKey();
  Type getType() const { return type; }
  int32_t getCapacity() const {
-    // `partialKeyCapacity` is not repurposed by `deferRelease`, so this is also
-    // valid for nodes with `releaseDeferred` set. It's needed to account for
-    // the capacity of a deferred node when it's finally released.
+    assert(!releaseDeferred);
    return partialKeyCapacity;
  }

@@ -519,7 +489,6 @@ 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
@@ -531,12 +500,6 @@ 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) {
@@ -553,7 +516,6 @@ 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
@@ -567,12 +529,6 @@ 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) {
@@ -580,7 +536,6 @@ 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;
@@ -600,9 +555,6 @@ inline void Node48::copyChildrenAndKeyFrom(const Node16 &other) {
        std::max(maxOfMax[i >> Node48::kMaxOfMaxShift], childMaxVersion[i]);
    ++i;
  }
-  // The reverseIndex bytes beyond numChildren are initialized so that the
-  // SIMD loads of a full 16-byte page in scan16 are well-defined (reading
-  // indeterminate bytes is UB in C++).
 }

 inline void Node48::copyChildrenAndKeyFrom(const Node48 &other) {
@@ -629,7 +581,6 @@ 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;
@@ -651,9 +602,6 @@ inline void Node48::copyChildrenAndKeyFrom(const Node256 &other) {
    ++i;
  });
  memcpy(partialKey(), &other + 1, partialKeyLen);
-  // The reverseIndex bytes beyond numChildren are initialized so that the
-  // SIMD loads of a full 16-byte page in scan16 are well-defined (reading
-  // indeterminate bytes is UB in C++).
 }

 inline void Node256::copyChildrenAndKeyFrom(const Node48 &other) {
@@ -963,34 +911,11 @@ int getNodeIndexExists(Node3 *self, uint8_t index) {

 int getNodeIndex(Node16 *self, uint8_t index) {

-#ifdef HAS_AVX
-  // Based on https://www.the-paper-trail.org/post/art-paper-notes/
-
-  // key_vec is 16 repeated copies of the searched-for byte, one for every
-  // possible position in child_keys that needs to be searched.
-  __m128i key_vec = _mm_set1_epi8(index);
-
-  // Compare all child_keys to 'index' in parallel. Don't worry if some of the
-  // keys aren't valid, we'll mask the results to only consider the valid ones
-  // below.
-  __m128i indices;
-  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
-  // comparison (because the other values are meaningless)
-  uint32_t mask = (1 << self->numChildren) - 1;
-
-  // Change the results of the comparison into a bitfield, masking off any
-  // invalid comparisons.
-  uint32_t bitfield = _mm_movemask_epi8(results) & mask;
-
-  // No match if there are no '1's in the bitfield.
+#if defined(__x86_64__) && !defined(USE_SIMD_FALLBACK)
+  uint32_t bitfield =
+      find_eq_16(self->index, index) & ((1 << self->numChildren) - 1);
  if (bitfield == 0)
    return -1;
-
-  // Find the index of the first '1' in the bitfield by counting the leading
-  // zeros.
  return std::countr_zero(bitfield);
 #elif defined(HAS_ARM_NEON)
  // Based on
@@ -1023,13 +948,9 @@ int getNodeIndex(Node16 *self, uint8_t index) {

 int getNodeIndexExists(Node16 *self, uint8_t index) {

-#ifdef HAS_AVX
-  __m128i key_vec = _mm_set1_epi8(index);
-  __m128i indices;
-  memcpy(&indices, self->index, Node16::kMaxNodes);
-  __m128i results = _mm_cmpeq_epi8(key_vec, indices);
-  uint32_t mask = (1 << self->numChildren) - 1;
-  uint32_t bitfield = _mm_movemask_epi8(results) & mask;
+#if defined(__x86_64__) && !defined(USE_SIMD_FALLBACK)
+  uint32_t bitfield =
+      find_eq_16(self->index, index) & ((1 << self->numChildren) - 1);
  assume(bitfield != 0);
  return std::countr_zero(bitfield);
 #elif defined(HAS_ARM_NEON)
@@ -1319,13 +1240,9 @@ TaggedNodePointer getChildGeq(Node16 *self, int child) {
    return nullptr;
  }

-#ifdef HAS_AVX
-  __m128i key_vec = _mm_set1_epi8(child);
-  __m128i indices;
-  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;
+#if defined(__x86_64__) && !defined(USE_SIMD_FALLBACK)
+  uint32_t bitfield =
+      find_ge_16(self->index, child) & ((1 << self->numChildren) - 1);
  return bitfield == 0 ? nullptr : self->children[std::countr_zero(bitfield)];
 #elif defined(HAS_ARM_NEON)
  uint8x16_t indices;
@@ -2045,6 +1962,7 @@ Node *erase(Node *self, WriteContext *writeContext, bool logical) {
      parent48->index[parentIndex] = toRemoveChildrenIndex;
      parent48->reverseIndex[toRemoveChildrenIndex] = parentIndex;
    }
+    parent48->childMaxVersion[lastChildrenIndex] = writeContext->zero;
    VALGRIND_MAKE_MEM_UNDEFINED(parent48->children + lastChildrenIndex,
                                sizeof(parent48->children[0]));

@@ -2139,53 +2057,6 @@ compare16_avx512(const InternalVersionT *vs, InternalVersionT rv) {
 }
 #endif

-#ifdef HAS_ARM_NEON
-// Returns a uint8x16_t whose byte i is 0xff if vs[i] > readVersion and 0
-// otherwise. When USE_64_BIT is set the comparison uses the full 64-bit value
-// (via vcgtq_s64); otherwise it falls back to the 32-bit subtract-and-compare
-// used by the original NEON fast path. The caller narrows this 16-byte vector
-// into the nibble-packed (4 bits per version) mask expected by scan16 and
-// checkMaxBetweenExclusiveImpl.
-inline uint8x16_t conflictMask16(const InternalVersionT *vs,
-                                 InternalVersionT readVersion) {
-#if USE_64_BIT
-  int64_t rv;
-  memcpy(&rv, &readVersion, sizeof(rv));
-  const auto rvVec = vdupq_n_s64(rv);
-  int32x2_t r32[8];
-  const auto *vsp = reinterpret_cast<const int64_t *>(vs);
-  for (int j = 0; j < 8; ++j) {
-    r32[j] = vmovn_s64(
-        vreinterpretq_s64_u64(vcgtq_s64(vld1q_s64(vsp + 2 * j), rvVec)));
-  }
-  uint32x4_t w4[4];
-  for (int k = 0; k < 4; ++k) {
-    w4[k] = vreinterpretq_u32_s32(vcombine_s32(r32[2 * k], r32[2 * k + 1]));
-  }
-  return vcombine_u8(
-      vmovn_u16(vcombine_u16(vmovn_u32(w4[0]), vmovn_u32(w4[1]))),
-      vmovn_u16(vcombine_u16(vmovn_u32(w4[2]), vmovn_u32(w4[3]))));
-#else
-  uint32x4_t w4[4];
-  memcpy(w4, vs, sizeof(w4));
-  uint32_t rv;
-  memcpy(&rv, &readVersion, sizeof(rv));
-  const auto rvVec = vdupq_n_u32(rv);
-
-  int32x4_t z;
-  memset(&z, 0, sizeof(z));
-
-  uint16x4_t conflicting[4];
-  for (int i = 0; i < 4; ++i) {
-    conflicting[i] =
-        vmovn_u32(vcgtq_s32(vreinterpretq_s32_u32(vsubq_u32(w4[i], rvVec)), z));
-  }
-  return vcombine_u8(vmovn_u16(vcombine_u16(conflicting[0], conflicting[1])),
-                     vmovn_u16(vcombine_u16(conflicting[2], conflicting[3])));
-#endif
-}
-#endif
-
 // Returns true if v[i] <= readVersion for all i such that begin <= is[i] < end
 // Preconditions: begin <= end, end - begin < 256
 template <bool kAVX512>
@@ -2206,20 +2077,32 @@ 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(conflictMask16(vs, readVersion)), 4)),
-      0);
+      vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(combined), 4)), 0);

  return !(compared & mask);

-#elif defined(HAS_AVX)
+#elif defined(__x86_64__) && !defined(USE_SIMD_FALLBACK)

-  __m128i indices;
-  memcpy(&indices, is, 16);
-  indices = _mm_sub_epi8(indices, _mm_set1_epi8(begin));
-  uint32_t mask = ~_mm_movemask_epi8(_mm_cmpeq_epi8(
-      indices, _mm_max_epu8(indices, _mm_set1_epi8(end - begin))));
+  uint32_t mask = mask_in_range_16(is, begin, end);

  uint32_t compared = 0;
  if constexpr (kAVX512) {
@@ -2259,10 +2142,26 @@ 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(conflictMask16(vs, readVersion)), 4)),
-      0);
+      vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(combined), 4)), 0);

  conflict &= end == 16 ? -1 : (uint64_t(1) << (end << 2)) - 1;
  conflict >>= begin << 2;
@@ -2311,8 +2210,6 @@ bool checkMaxBetweenExclusiveImpl(Node3 *n, int begin, int end,
  assert(!(begin == -1 && end == 256));

  auto *self = static_cast<Node3 *>(n);
-  assume(1 <= self->numChildren);
-  assume(self->numChildren <= 3);

  ++begin;

@@ -2378,23 +2275,33 @@ 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(conflictMask16(
-                                          self->childMaxVersion, readVersion)),
-                                      4)),
-      0);
+      vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(combined), 4)), 0);

  return !(compared & mask) && firstRangeOk;

-#elif defined(HAS_AVX)
+#elif defined(__x86_64__) && !defined(USE_SIMD_FALLBACK)

-  __m128i indices;
-  memcpy(&indices, self->index, 16);
-  indices = _mm_sub_epi8(indices, _mm_set1_epi8(begin));
-  uint32_t mask =
-      0xffff & ~_mm_movemask_epi8(_mm_cmpeq_epi8(
-                   indices, _mm_max_epu8(indices, _mm_set1_epi8(end - begin))));
-  mask &= (1 << self->numChildren) - 1;
+  uint32_t mask = mask_in_range_16(self->index, begin, end) &
+                  ((1 << self->numChildren) - 1);
  if (!mask) {
    return true;
  }
@@ -2786,39 +2693,10 @@ checkMaxBetweenExclusiveImpl<true>(Node256 *n, int begin, int end,
  return self;
 }

-#if SHOW_MEMORY
-// Length of the key leading to `child`, given the length of the key leading to
-// its parent. Consistent with `getSearchPathLength`.
-int64_t childSearchPathLength(int64_t parentSearchPathLength, Node *child) {
-  return parentSearchPathLength + 1 + child->partialKeyLen;
-}
-#endif
-
 void eraseTree(Node *root, WriteContext *writeContext) {
  Arena arena;
  auto toFree = vector<Node *>(arena);
  toFree.push_back(root);
-  // Length of the key leading to each node in `toFree`, on a parallel stack.
-  // A node is released before its children are visited, so the `parent`
-  // pointers that `getSearchPathLength` walks are dangling by the time a queued
-  // node is popped. Carry the lengths along instead. Only the SHOW_MEMORY
-  // accounting needs them.
-#if SHOW_MEMORY
-  auto searchPathLengths = vector<int64_t>(arena);
-  searchPathLengths.push_back(getSearchPathLength(root));
-#endif
-
-  // Queue `child` for release, where `parentSearchPathLength` is the length of
-  // the key leading to its parent.
-  auto queueChild = [&](Node *child, int64_t parentSearchPathLength) {
-    toFree.push_back(child);
-#if SHOW_MEMORY
-    searchPathLengths.push_back(
-        childSearchPathLength(parentSearchPathLength, child));
-#else
-    (void)parentSearchPathLength;
-#endif
-  };

  while (toFree.size() > 0) {
    auto *n = toFree.back();
@@ -2826,13 +2704,7 @@ void eraseTree(Node *root, WriteContext *writeContext) {
    writeContext->accum.entries_erased += n->entryPresent;
    ++writeContext->accum.nodes_released;

-#if SHOW_MEMORY
-    const int64_t searchPathLength = searchPathLengths.back();
-    searchPathLengths.pop_back();
-#else
-    const int64_t searchPathLength = 0;
-#endif
-    removeKey(n, searchPathLength);
+    removeKey(n);

    switch (n->getType()) {
    case Type_Node0: {
@@ -2842,39 +2714,28 @@ void eraseTree(Node *root, WriteContext *writeContext) {
    case Type_Node3: {
      auto *n3 = static_cast<Node3 *>(n);
      for (int i = 0; i < n3->numChildren; ++i) {
-        queueChild(n3->children[i], searchPathLength);
+        toFree.push_back(n3->children[i]);
      }
      writeContext->release(n3);
    } break;
    case Type_Node16: {
      auto *n16 = static_cast<Node16 *>(n);
      for (int i = 0; i < n16->numChildren; ++i) {
-        queueChild(n16->children[i], searchPathLength);
+        toFree.push_back(n16->children[i]);
      }
      writeContext->release(n16);
    } break;
    case Type_Node48: {
      auto *n48 = static_cast<Node48 *>(n);
      for (int i = 0; i < n48->numChildren; ++i) {
-        queueChild(n48->children[i], searchPathLength);
+        toFree.push_back(n48->children[i]);
      }
      writeContext->release(n48);
    } break;
    case Type_Node256: {
      auto *n256 = static_cast<Node256 *>(n);
      auto *out = toFree.unsafePrepareAppend(n256->numChildren).data();
-#if SHOW_MEMORY
-      auto *outSearchPathLengths =
-          searchPathLengths.unsafePrepareAppend(n256->numChildren).data();
-#endif
-      n256->bitSet.forEachSet([&](int i) {
-        Node *child = n256->children[i];
-        *out++ = child;
-#if SHOW_MEMORY
-        *outSearchPathLengths++ =
-            childSearchPathLength(searchPathLength, child);
-#endif
-      });
+      n256->bitSet.forEachSet([&](int i) { *out++ = n256->children[i]; });
      assert(out == toFree.end());
      writeContext->release(n256);
    } break;
@@ -2971,11 +2832,7 @@ horizontalMax16(InternalVersionT *vs, InternalVersionT) {
  return max;
 }

-/** @pre `node` is not the root, `node->entryPresent`, and `writeContext !=
- * nullptr`.
- * @post `maxVersion(node)` is the max of `writeContext->zero` and the max
- * version of all logical keys starting with the search path of `node`.
- */
+// Precondition: `node->entryPresent`, and node is not the root
 void fixupMaxVersion(Node *node, WriteContext *writeContext) {
  assert(node->parent);
  InternalVersionT max;
@@ -2986,13 +2843,9 @@ void fixupMaxVersion(Node *node, WriteContext *writeContext) {
    break;
  case Type_Node3: {
    auto *self3 = static_cast<Node3 *>(node);
-    assume(1 <= self3->numChildren);
-    assume(self3->numChildren <= 3);
-    InternalVersionT m = self3->childMaxVersion[0];
-    for (int i = 1; i < self3->numChildren; ++i) {
-      m = std::max(m, self3->childMaxVersion[i]);
-    }
-    max = std::max(max, m);
+    max = std::max(max,
+                   horizontalMaxUpTo16(self3->childMaxVersion,
+                                       writeContext->zero, self3->numChildren));
  } break;
  case Type_Node16: {
    auto *self16 = static_cast<Node16 *>(node);
@@ -3008,7 +2861,8 @@ void fixupMaxVersion(Node *node, WriteContext *writeContext) {
  } break;
  case Type_Node256: {
    auto *self256 = static_cast<Node256 *>(node);
-    max = std::max(max, horizontalMax16(self256->maxOfMax, writeContext->zero));
+    max = std::max(
+        max, horizontalMax16(self256->childMaxVersion, writeContext->zero));
  } break;
  default:                   // GCOVR_EXCL_LINE
    __builtin_unreachable(); // GCOVR_EXCL_LINE
@@ -4955,7 +4809,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
        auto *results2 = new (arena) Result[count];
        check::Context context2;
        useSequential(reads, results2, count, context2);
-        assert(memcmp(result, results2, count * sizeof(Result)) == 0);
+        assert(memcmp(result, results2, count) == 0);
        assert(context.readContext == context2.readContext);
 #endif
      }
@@ -5056,18 +4910,16 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
        }
        if (e != nullptr) {
          while (e->releaseDeferred) {
-            e = e->forwardTo;
+            e = b->forwardTo;
          }
        }
        fprintf(stderr, "search path: %s, begin: %s\n",
                getSearchPathPrintable(b).c_str(),
                printable(context.results[i].remaining).c_str());
-        if (e != nullptr) {
        fprintf(stderr, "search path: %s, end: %s\n",
                getSearchPathPrintable(e).c_str(),
                printable(context.results[i].endRemaining).c_str());
      }
-      }
 #endif

      while (context.results[i].insertionPoint->releaseDeferred) {
@@ -5168,7 +5020,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {

    if (oldestExtantVersion < writeVersion - kMaxCorrectVersionWindow)
        [[unlikely]] {
-      if (writeVersion - kNominalVersionWindow > newestVersionFullPrecision) {
+      if (writeVersion > newestVersionFullPrecision + kNominalVersionWindow) {
        eraseTree(rootParent->children[0], &writeContext);
        init(writeVersion - kNominalVersionWindow);
      }
@@ -5546,9 +5398,8 @@ void internal_addWrites(ConflictSet::Impl *impl,
  impl->totalBytes += mallocBytesDelta;
  impl->memory_bytes.set(impl->totalBytes);
 #if SHOW_MEMORY
-  // Testing-only cross-check of the accounting.
  if (impl->totalBytes != mallocBytes) {
-    abort(); // GCOVR_EXCL_LINE
+    abort();
  }
 #endif
 }
@@ -5559,9 +5410,8 @@ void internal_setOldestVersion(ConflictSet::Impl *impl, int64_t oldestVersion) {
  impl->totalBytes += mallocBytesDelta;
  impl->memory_bytes.set(impl->totalBytes);
 #if SHOW_MEMORY
-  // Testing-only cross-check of the accounting.
  if (impl->totalBytes != mallocBytes) {
-    abort(); // GCOVR_EXCL_LINE
+    abort();
  }
 #endif
 }
@@ -6027,14 +5877,6 @@ checkMaxVersion(Node *root, Node *node, InternalVersionT oldestVersion,

 } // namespace

-// The SHOW_MEMORY instrumentation. Each of these functions is defined exactly
-// once, with the accounting compiled out when SHOW_MEMORY is disabled. Defining
-// them once (rather than defining no-op stubs when SHOW_MEMORY is disabled)
-// keeps every configuration of this file reporting each function on the same
-// line, which merging coverage from objects compiled with different options
-// requires: gcovr, merging the SHOW_MEMORY test with the rest of the build,
-// otherwise aborts with "Got function ... on multiple lines". The optimizer
-// removes the resulting empty calls.
 #if SHOW_MEMORY

 int64_t nodeBytes = 0;
@@ -6063,10 +5905,7 @@ int64_t getNodeSize(struct Node *n) {
  }
 }

-#endif
-
 int64_t getSearchPathLength(Node *n) {
-#if SHOW_MEMORY
  assert(n != nullptr);
  int64_t result = 0;
  for (;;) {
@@ -6078,14 +5917,9 @@ int64_t getSearchPathLength(Node *n) {
    n = n->parent;
  }
  return result;
-#else
-  (void)n;
-  return 0;
-#endif
 }

 void addNode(Node *n) {
-#if SHOW_MEMORY
  nodeBytes += getNodeSize(n);
  partialCapacityBytes += n->getCapacity();
  if (nodeBytes > peakNodeBytes) {
@@ -6094,22 +5928,14 @@ void addNode(Node *n) {
  if (partialCapacityBytes > peakPartialCapacityBytes) {
    peakPartialCapacityBytes = partialCapacityBytes;
  }
-#else
-  (void)n;
-#endif
 }

 void removeNode(Node *n) {
-#if SHOW_MEMORY
  nodeBytes -= getNodeSize(n);
  partialCapacityBytes -= n->getCapacity();
-#else
-  (void)n;
-#endif
 }

 void addKey(Node *n) {
-#if SHOW_MEMORY
  if (!n->entryPresent) {
    ++totalKeys;
    keyBytes += getSearchPathLength(n);
@@ -6120,36 +5946,15 @@ void addKey(Node *n) {
      peakKeyBytes = keyBytes;
    }
  }
-#else
-  (void)n;
-#endif
 }

 void removeKey(Node *n) {
-#if SHOW_MEMORY
  if (n->entryPresent) {
    --totalKeys;
    keyBytes -= getSearchPathLength(n);
  }
-#else
-  (void)n;
-#endif
 }

-void removeKey(Node *n, int64_t searchPathLength) {
-#if SHOW_MEMORY
-  if (n->entryPresent) {
-    --totalKeys;
-    keyBytes -= searchPathLength;
-  }
-#else
-  (void)n;
-  (void)searchPathLength;
-#endif
-}
-
-#if SHOW_MEMORY
-
 struct __attribute__((visibility("default"))) PeakPrinter {
  ~PeakPrinter() {
    printf("--- radix_tree ---\n");
@@ -59,12 +59,6 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
    while (keyUpdates > 0) {
      if (iter == map.end()) {
        iter = map.begin();
-        // The map is empty, so there are no more entries to garbage-collect.
-        // Break rather than spinning forever (keyUpdates is charged at 2x
-        // the entry count, so it may still be > 0 after emptying the map).
-        if (iter == map.end()) {
-          break;
-        }
      }
      for (; iter != map.end(); --keyUpdates) {
        if (iter->second <= oldestVersion) {
@@ -23,9 +23,7 @@ using namespace weaselab;
 #include <callgrind.h>

 #define DEBUG_VERBOSE 0
-#ifndef SHOW_MEMORY
 #define SHOW_MEMORY 0
-#endif

 // std::span is not trivially constructible. We want a span that leaves its
 // members uninitialized for performance reasons.
@@ -98,11 +96,6 @@ static_assert(std::is_trivial_v<TrivialSpan>);
 #if SHOW_MEMORY
 inline int64_t mallocBytes = 0;
 inline int64_t peakMallocBytes = 0;
-// Defined in ConflictSet.cpp
-extern int64_t nodeBytes;
-extern int64_t partialCapacityBytes;
-extern int64_t totalKeys;
-extern int64_t keyBytes;
 #endif

 inline thread_local int64_t mallocBytesDelta = 0;
@@ -96,7 +96,7 @@ inline int firstNeqStride(const uint8_t *ap, const uint8_t *bp) {
 // GCOVR_EXCL_START
 #if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
 __attribute__((target("avx512f,avx512bw"))) inline int
-longestCommonPrefixLong(const uint8_t *ap, const uint8_t *bp, int cl) {
+longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
  int i = 0;
  int end = cl & ~63;
  while (i < end) {
@@ -123,7 +123,7 @@ __attribute__((target("default")))
 // GCOVR_EXCL_STOP

 inline int
-longestCommonPrefixLong(const uint8_t *ap, const uint8_t *bp, int cl) {
+longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
  if (!(cl >= 0)) {
    __builtin_unreachable(); // GCOVR_EXCL_LINE
  }
@@ -183,71 +183,3 @@ longestCommonPrefixLong(const uint8_t *ap, const uint8_t *bp, int cl) {

  return i;
 }
-// Use in-bounds native-word loads for short common lengths. The second load
-// overlaps the first; because the first load matched, any mismatch found in
-// the overlap is impossible and the returned absolute index is still correct.
-// Precondition: cl <= 2 * sizeof(uint64_t)
-__attribute__((always_inline)) static inline int
-longestCommonPrefixShort(const uint8_t *ap, const uint8_t *bp, int cl) {
-  assert(cl <= 2 * int(sizeof(uint64_t)));
-  if (cl >= 8) {
-    uint64_t a0;
-    uint64_t b0;
-    memcpy(&a0, ap, sizeof(a0));
-    memcpy(&b0, bp, sizeof(b0));
-    if (uint64_t mismatched = a0 ^ b0) {
-      return std::countr_zero(mismatched) / 8;
-    }
-    if (cl == 8) {
-      return 8;
-    }
-
-    uint64_t a1;
-    uint64_t b1;
-    memcpy(&a1, ap + cl - 8, sizeof(a1));
-    memcpy(&b1, bp + cl - 8, sizeof(b1));
-    if (uint64_t mismatched = a1 ^ b1) {
-      return cl - 8 + std::countr_zero(mismatched) / 8;
-    }
-    return cl;
-  }
-
-  if (cl >= 4) {
-    uint32_t a0;
-    uint32_t b0;
-    memcpy(&a0, ap, sizeof(a0));
-    memcpy(&b0, bp, sizeof(b0));
-    if (uint32_t mismatched = a0 ^ b0) {
-      return std::countr_zero(mismatched) / 8;
-    }
-    if (cl == 4) {
-      return 4;
-    }
-
-    uint32_t a1;
-    uint32_t b1;
-    memcpy(&a1, ap + cl - 4, sizeof(a1));
-    memcpy(&b1, bp + cl - 4, sizeof(b1));
-    if (uint32_t mismatched = a1 ^ b1) {
-      return cl - 4 + std::countr_zero(mismatched) / 8;
-    }
-    return cl;
-  }
-
-  int i = 0;
-  while (i < cl && ap[i] == bp[i]) {
-    ++i;
-  }
-  return i;
-}
-
-__attribute__((always_inline)) static inline int
-longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
-  if (!(cl >= 0)) {
-    __builtin_unreachable(); // GCOVR_EXCL_LINE
-  }
-  if (cl <= 2 * int(sizeof(uint64_t))) {
-    return longestCommonPrefixShort(ap, bp, cl);
-  }
-  return longestCommonPrefixLong(ap, bp, cl);
-}
@@ -19,30 +19,30 @@ InstalledDir: /usr/lib/llvm-21/bin

 |               ns/op |                op/s |    err% |          ins/op |          cyc/op |    IPC |         bra/op |   miss% |     total | benchmark
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|              165.93 |        6,026,520.04 |    0.1% |        3,107.03 |          610.42 |  5.090 |         558.59 |    0.0% |      1.98 | `point reads`
-|              163.08 |        6,132,046.57 |    0.1% |        3,036.76 |          599.68 |  5.064 |         539.35 |    0.0% |      1.95 | `prefix reads`
-|              250.26 |        3,995,834.21 |    0.1% |        3,722.71 |          920.25 |  4.045 |         692.00 |    0.0% |      2.99 | `range reads`
-|              359.64 |        2,780,581.63 |    0.1% |        4,523.64 |        1,322.63 |  3.420 |         720.22 |    2.1% |      4.30 | `point writes`
-|              358.20 |        2,791,726.93 |    0.1% |        4,484.57 |        1,317.23 |  3.405 |         705.00 |    2.0% |      4.28 | `prefix writes`
-|              193.91 |        5,157,081.29 |    0.1% |        2,224.10 |          713.33 |  3.118 |         377.17 |    3.3% |      2.32 | `range writes`
-|              422.03 |        2,369,513.38 |    1.1% |        6,855.96 |        1,551.55 |  4.419 |       1,227.82 |    1.4% |      0.05 | `monotonic increasing point writes`
-|          142,467.00 |            7,019.17 |    3.8% |      812,047.33 |      508,535.40 |  1.597 |     151,246.50 |    1.0% |      0.01 | `worst case for radix tree`
-|               37.66 |       26,554,217.74 |    0.8% |          701.00 |          138.80 |  5.050 |         102.00 |    0.0% |      0.01 | `create and destroy`
+|              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
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|               12.25 |       81,664,669.29 |    0.4% |          233.49 |           44.94 |  5.196 |          31.23 |    0.7% |      0.15 | `point reads`
-|               13.35 |       74,903,950.49 |    0.1% |          271.75 |           49.10 |  5.534 |          40.21 |    0.4% |      0.16 | `prefix reads`
-|               31.37 |       31,874,583.62 |    0.1% |          700.39 |          115.43 |  6.067 |         100.83 |    0.2% |      0.37 | `range reads`
-|               19.55 |       51,153,057.80 |    0.1% |          366.51 |           71.90 |  5.098 |          47.92 |    0.5% |      0.23 | `point writes`
-|               36.92 |       27,088,344.17 |    0.1% |          645.11 |          135.72 |  4.753 |         100.31 |    0.4% |      0.44 | `prefix writes`
-|               36.84 |       27,143,553.08 |    0.1% |          691.04 |          135.42 |  5.103 |         114.01 |    0.1% |      0.45 | `range writes`
-|               74.62 |       13,400,467.72 |    0.9% |        1,418.37 |          274.20 |  5.173 |         273.39 |    0.1% |      0.01 | `monotonic increasing point writes`
-|          352,671.00 |            2,835.50 |   15.3% |    3,687,535.00 |    1,287,119.00 |  2.865 |     712,030.00 |    0.1% |      0.01 | :wavy_dash: `worst case for radix tree` (Unstable with ~1.0 iters. Increase `minEpochIterations` to e.g. 10)
-|               96.63 |       10,348,729.42 |    1.2% |        1,766.00 |          355.53 |  4.967 |         283.00 |    0.0% |      0.01 | `create and destroy`
+|               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: 4.69875 seconds, 347.017 MB/s, Add: 3.39475 seconds, 165.153 MB/s, Gc ratio: 38.7414%, Peak idle memory: 5.51852e+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: 0.976745 seconds, 1669.37 MB/s, Add: 1.20464 seconds, 465.411 MB/s, Gc ratio: 39.4158%, Peak idle memory: 2.0226e+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: 0.976745 seconds, 1669.37 MB/s, Add: 1.20464 seconds, 465.411 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: 0.864589 seconds, 1885.92 MB/s, Add: 0.634604 seconds, 883.47 MB/s, Gc ratio: 41.184%, 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
 ```

@@ -1,115 +0,0 @@
-// Exercises the SHOW_MEMORY instrumentation, which the other tests don't
-// compile with (this target passes -DSHOW_MEMORY=1). The instrumentation has
-// its own consistency checks, which abort on failure, and asserts are enabled
-// by TEST_FLAGS. Only one ConflictSet may exist at a time, since the
-// accounting is process-wide.
-#include "ConflictSet.h"
-#include "Internal.h"
-
-#include <cassert>
-#include <cstdint>
-#include <string>
-
-using namespace weaselab;
-
-namespace {
-
-void writePoint(ConflictSet &cs, const std::string &key, int64_t writeVersion) {
-  ConflictSet::WriteRange w{{(const uint8_t *)key.data(), int(key.size())},
-                            {nullptr, 0}};
-  cs.addWrites(&w, 1, writeVersion);
-}
-
-ConflictSet::Result checkPoint(const ConflictSet &cs, const std::string &key,
-                               int64_t readVersion) {
-  ConflictSet::ReadRange r{{(const uint8_t *)key.data(), int(key.size())},
-                           {nullptr, 0},
-                           readVersion};
-  ConflictSet::Result result;
-  cs.check(&r, &result, 1);
-  return result;
-}
-
-// Everything a ConflictSet allocated must have been accounted for and released
-// by the time it's destroyed.
-void checkFullyReleased() {
-  assert(nodeBytes == 0);
-  assert(partialCapacityBytes == 0);
-  assert(totalKeys == 0);
-  assert(keyBytes == 0);
-  assert(mallocBytes == 0);
-}
-
-} // namespace
-
-int main() {
-  // Splitting a node replaces it, which defers the release of the old node.
-  // Releasing a deferred node must account for its capacity, and the key it
-  // held must be accounted for without walking `parent` pointers of nodes that
-  // have already been released.
-  {
-    ConflictSet cs{0};
-    writePoint(cs, "a", 1);
-    writePoint(cs, "ab", 2);
-    assert(checkPoint(cs, "a", 0) == ConflictSet::Conflict);
-    assert(checkPoint(cs, "ab", 0) == ConflictSet::Conflict);
-    assert(checkPoint(cs, "b", 0) == ConflictSet::Commit);
-    assert(nodeBytes > 0);
-    assert(keyBytes > 0);
-  }
-  checkFullyReleased();
-
-  // Exceeding the version window erases the tree mid-workload.
-  {
-    ConflictSet cs{0};
-    writePoint(cs, "abc", 1);
-    const int64_t version = int64_t(1) << 40;
-    writePoint(cs, "abc", version);
-    assert(checkPoint(cs, "abc", 0) == ConflictSet::TooOld);
-    assert(checkPoint(cs, "abc", version) == ConflictSet::Commit);
-  }
-  checkFullyReleased();
-
-  // Erase a tree with many nodes, both mid-workload and at destruction.
-  {
-    ConflictSet cs{0};
-    for (int i = 0; i < 1000; ++i) {
-      writePoint(cs, "prefix/" + std::to_string(i), i + 1);
-    }
-    const int64_t version = int64_t(1) << 40;
-    writePoint(cs, "prefix/1", version);
-    assert(checkPoint(cs, "prefix/1", version) == ConflictSet::Commit);
-
-    for (int i = 0; i < 1000; ++i) {
-      writePoint(cs, "other/" + std::to_string(i), version + i + 1);
-    }
-    assert(checkPoint(cs, "other/999", version) == ConflictSet::Conflict);
-  }
-  checkFullyReleased();
-
-  // Grow a node through every node type, so that the accounting for each node
-  // type's size is exercised, then erase the entries one at a time so that
-  // nodes are released individually rather than as part of a whole tree.
-  {
-    ConflictSet cs{0};
-    std::string key(1, '\0');
-    for (int i = 0; i < 256; ++i) {
-      key[0] = char(i);
-      writePoint(cs, key, i + 1);
-    }
-    for (int i = 0; i < 256; ++i) {
-      key[0] = char(i);
-      assert(checkPoint(cs, key, i) == ConflictSet::Conflict);
-    }
-
-    // Garbage collection erases entries that can only fail with TooOld anyway.
-    // Doing it one version at a time walks the tree, erasing, downsizing and
-    // releasing nodes as it goes.
-    for (int i = 0; i < 256; ++i) {
-      cs.setOldestVersion(i + 1);
-    }
-    key[0] = '\0';
-    assert(checkPoint(cs, key, 0) == ConflictSet::TooOld);
-  }
-  checkFullyReleased();
-}
@@ -819,17 +819,10 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
  }

  void setOldestVersion(int64_t oldestVersion) {
-    // oldestVersion becomes the max of its previous value and the argument, so
-    // an older version is a no-op. We still fall through in the equal case so
-    // that garbage collection can make progress without advancing the oldest
-    // version.
-    if (oldestVersion < this->oldestVersion) {
-      return;
-    }
-
    // This isn't 100% accurate. It overcounts if you hit the end
    gc_iterations_total.add(keyUpdates);

+    assert(oldestVersion >= this->oldestVersion);
    this->oldestVersion = oldestVersion;
    SkipList::Finger finger;
    int temp;
@@ -35,5 +35,36 @@ int main(int argc, char **argv) {
        break;
      }
    }
+
+    {
+      ConflictSet::MetricsV1 *metrics;
+      int metricsCount;
+      driver1.cs.getMetricsV1(&metrics, &metricsCount);
+      printf("#################### METRICS for ConflictSet 1 for %s "
+             "####################\n",
+             argv[i]);
+      for (int i = 0; i < metricsCount; ++i) {
+        printf("# HELP %s %s\n", metrics[i].name, metrics[i].help);
+        printf("# TYPE %s %s\n", metrics[i].name,
+               metrics[i].type == metrics[i].Counter ? "counter" : "gauge");
+        printf("%s %g\n", metrics[i].name, metrics[i].getValue());
+      }
+      puts("");
+    }
+    {
+      ConflictSet::MetricsV1 *metrics;
+      int metricsCount;
+      driver2.cs.getMetricsV1(&metrics, &metricsCount);
+      printf("#################### METRICS for ConflictSet 2 for %s "
+             "####################\n",
+             argv[i]);
+      for (int i = 0; i < metricsCount; ++i) {
+        printf("# HELP %s %s\n", metrics[i].name, metrics[i].help);
+        printf("# TYPE %s %s\n", metrics[i].name,
+               metrics[i].type == metrics[i].Counter ? "counter" : "gauge");
+        printf("%s %g\n", metrics[i].name, metrics[i].getValue());
+      }
+      puts("");
+    }
  }
 }
@@ -2,7 +2,6 @@

 #include <cassert>
 #include <cstdio>
-#include <limits>

 using namespace weaselab;

@@ -24,28 +23,6 @@ int main(void) {
  int64_t bytes = cs.getBytes();
  assert(bytes > 0);

-  // Regression: a write just below INT64_MAX must not overflow the nominal
-  // version window comparison and erase entries that reads still observe.
-  // Versions are only required to be >= 0, so this is valid input.
-  {
-    const int64_t imax = std::numeric_limits<int64_t>::max();
-    const int64_t v0 = imax - int64_t(1000000000);
-    ConflictSet nearMax(0);
-    ConflictSet::WriteRange w2;
-    w2.begin.p = (const uint8_t *)"key";
-    w2.begin.len = 3;
-    w2.end.len = 0;
-    nearMax.addWrites(&w2, 1, v0);
-    nearMax.addWrites(&w2, 0, imax);
-    ConflictSet::ReadRange r2;
-    r2.begin.p = (const uint8_t *)"key";
-    r2.begin.len = 3;
-    r2.end.len = 0;
-    r2.readVersion = v0 - 1;
-    nearMax.check(&r2, &result, 1);
-    assert(result == ConflictSet::Conflict);
-  }
-
  ConflictSet::MetricsV1 *metrics;
  int metricsCount;
  cs.getMetricsV1(&metrics, &metricsCount);
@@ -21,12 +21,7 @@ limitations under the License.
 #ifdef __cplusplus
 namespace weaselab {
 /** A data structure for optimistic concurrency control on ranges of
- * bitwise-lexicographically-ordered keys.
- *
- * All versions are >= 0. `oldestVersion` is the smallest version a subsequent
- * `check` can read and `newestVersion` is the largest version passed to
- * `addWrites` or the constructor. `oldestVersion <= newestVersion`, and
- * `oldestVersion >= newestVersion - two billion`.
+ * bitwise-lexicographically-ordered keys. All versions must be >= 0.
 *
 * Thread safety:
 *   - It's safe to operate on two different ConflictSets in two different
@@ -41,7 +36,8 @@ struct __attribute__((__visibility__("default"))) ConflictSet {
    /** The result of a check which intersects a write at a version >
       readVersion */
    Conflict,
-    /** The result of a check with a readVersion older than `oldestVersion` */
+    /** The result of a check with a readVersion older than the highest call to
+       `setOldestVersion` */
    TooOld
  };

@@ -58,6 +54,9 @@ struct __attribute__((__visibility__("default"))) ConflictSet {
    /** `end` having length 0 denotes that this range is the single key {begin}.
     * Otherwise this denotes the range [begin, end), and begin must be < end */
    Key end;
+    /** `readVersion` older than the oldestVersion or the version of the
+     * latest call to `addWrites` minus two billion will result in `TooOld`.
+     * Must be <= the version of the latest call to `addWrites` */
    int64_t readVersion;
  };

@@ -69,38 +68,23 @@ struct __attribute__((__visibility__("default"))) ConflictSet {
    Key end;
  };

-  /** The result of checking reads[i] is written in results[i].
-   * @pre `reads` and `results` each hold at least `count` elements; each
-   * `reads[i]` is a valid `ReadRange` (see its field documentation); each
-   * `reads[i].readVersion <= newestVersion`.
-   * @post For every i < count, `results[i]` is `TooOld` if
-   * `reads[i].readVersion < oldestVersion`; `Conflict` if any write at a
-   * version greater than `reads[i].readVersion` intersects the range
-   * `reads[i]`; and `Commit` otherwise. */
+  /** The result of checking reads[i] is written in results[i] */
  void check(const ReadRange *reads, Result *results, int count) const;

-  /** Records writes at `writeVersion`, making them visible to `check`.
-   * @pre `writes` holds at least `count` elements; every `writes[i]` is a valid
-   * `WriteRange` (see its field documentation); `count >= 0`; `writeVersion >=
-   * newestVersion`.
-   * @post Writes (writes[i], writeVersion) are recorded for every i < count,
-   * and `newestVersion` becomes `writeVersion`. `oldestVersion` advances only
-   * as far as necessary: it becomes the larger of its previous value and
-   * `newestVersion - two billion`. Calling with `count == 0` adds no write;
-   * `newestVersion` (and possibly `oldestVersion`) still advance. */
+  /** Reads intersecting writes where readVersion < `writeVersion` will result
+   * in `Conflict` (or `TooOld`, eventually).  `writeVersion` must be greater
+   * than or equal to all previous write versions. Call `addWrites` with `count`
+   * zero to only advance the version. */
  void addWrites(const WriteRange *writes, int count, int64_t writeVersion);

-  /** Reads where readVersion < `oldestVersion` will result in `TooOld`.
-   * @pre `oldestVersion >= 0` and `oldestVersion <= newestVersion`.
-   * @post `oldestVersion` becomes the maximum of its previous value and the
-   * argument; `newestVersion` is unchanged. */
+  /** Reads where readVersion < `oldestVersion` will result in `TooOld`. Must be
+   * greater than or equal to all previous oldest versions. Must be <= the
+   * version of the latest call to `addWrites` */
  void setOldestVersion(int64_t oldestVersion);

-  /** Reads where readVersion < `oldestVersion` will result in `TooOld`.
-   * @pre `version >= 0`.
-   * @post No writes are recorded, and `oldestVersion == newestVersion ==
-   * version`. */
-  explicit ConflictSet(int64_t version);
+  /** Reads where readVersion < oldestVersion will result in `TooOld`. There are
+   * no writes initially */
+  explicit ConflictSet(int64_t oldestVersion);

  ~ConflictSet();

@@ -162,12 +146,7 @@ private:
 #else

 /** A data structure for optimistic concurrency control on ranges of
- * bitwise-lexicographically-ordered keys.
- *
- * All versions are >= 0. `oldestVersion` is the smallest version a subsequent
- * `check` can read and `newestVersion` is the largest version passed to
- * `addWrites` or the constructor. `oldestVersion <= newestVersion`, and
- * `oldestVersion >= newestVersion - two billion`.
+ * bitwise-lexicographically-ordered keys. All versions must be >= 0.
 *
 * Thread safety:
 *   - It's safe to operate on two different ConflictSets in two different
@@ -181,10 +160,11 @@ typedef struct ConflictSet ConflictSet;
 typedef enum {
  /** The result of a check which does not intersect any conflicting writes */
  ConflictSet_Commit,
-  /** The result of a check which intersects a write at a version >
-     readVersion */
+  /** The result of a check which intersects a write at a version > readVersion
+   */
  ConflictSet_Conflict,
-  /** The result of a check with a readVersion older than `oldestVersion` */
+  /** The result of a check with a readVersion older than the highest call to
+     `setOldestVersion` */
  ConflictSet_TooOld
 } ConflictSet_Result;

@@ -201,6 +181,9 @@ typedef struct {
  /** `end` having length 0 denotes that this range is the single key {begin}.
   * Otherwise this denotes the range [begin, end), and begin must be < end */
  ConflictSet_Key end;
+  /** `readVersion` older than the oldestVersion or the version of the
+   * latest call to `addWrites` minus two billion will result in `TooOld`.
+   * Must be <= the version of the latest call to `addWrites` */
  int64_t readVersion;
 } ConflictSet_ReadRange;

@@ -212,43 +195,27 @@ typedef struct {
  ConflictSet_Key end;
 } ConflictSet_WriteRange;

-/** The result of checking reads[i] is written in results[i].
- * @pre `reads` and `results` each hold at least `count` elements; each
- * `reads[i]` is a valid `ConflictSet_ReadRange` (see its field documentation);
- * each `reads[i].readVersion <= newestVersion`.
- * @post For every i < count, `results[i]` is `ConflictSet_TooOld` if
- * `reads[i].readVersion < oldestVersion`; `ConflictSet_Conflict` if any write
- * at a version greater than `reads[i].readVersion` intersects the range
- * `reads[i]`; and `ConflictSet_Commit` otherwise. */
+/** The result of checking reads[i] is written in results[i] */
 void ConflictSet_check(const ConflictSet *cs,
                       const ConflictSet_ReadRange *reads,
                       ConflictSet_Result *results, int count);

-/** Records writes at `writeVersion`, making them visible to
- * `ConflictSet_check`.
- * @pre `writes` holds at least `count` elements; every `writes[i]` is a valid
- * `ConflictSet_WriteRange` (see its field documentation); `count >= 0`;
- * `writeVersion >= newestVersion`.
- * @post Writes (writes[i], writeVersion) are recorded for every i < count,
- * and `newestVersion` becomes `writeVersion`. `oldestVersion` advances only
- * as far as necessary: it becomes the larger of its previous value and
- * `newestVersion - two billion`. Calling with `count == 0` adds no write;
- * `newestVersion` (and possibly `oldestVersion`) still advance. */
+/** Reads intersecting writes where readVersion < `writeVersion` will result
+ * in `Conflict` (or `TooOld`, eventually).  `writeVersion` must be greater
+ * than or equal to all previous write versions. Call `addWrites` with `count`
+ * zero to only advance the version. */
 void ConflictSet_addWrites(ConflictSet *cs,
                           const ConflictSet_WriteRange *writes, int count,
                           int64_t writeVersion);

-/** Reads where readVersion < `oldestVersion` will result in `TooOld`.
- * @pre `oldestVersion >= 0` and `oldestVersion <= newestVersion`.
- * @post `oldestVersion` becomes the maximum of its previous value and the
- * argument; `newestVersion` is unchanged. */
+/** Reads where readVersion < `oldestVersion` will result in `TooOld`. Must be
+ * greater than or equal to all previous oldest versions. Must be <= the
+ * version of the latest call to `addWrites` */
 void ConflictSet_setOldestVersion(ConflictSet *cs, int64_t oldestVersion);

-/** Reads where readVersion < `oldestVersion` will result in `TooOld`.
- * @pre `version >= 0`.
- * @post No writes are recorded, and `oldestVersion == newestVersion ==
- * version`. */
-ConflictSet *ConflictSet_create(int64_t version);
+/** Reads where readVersion < oldestVersion will result in `TooOld`. There are
+ * no writes initially */
+ConflictSet *ConflictSet_create(int64_t oldestVersion);

 void ConflictSet_destroy(ConflictSet *cs);

@@ -0,0 +1,25 @@
+#pragma once
+
+#include <cstdint>
+
+#if defined(__x86_64__) && !defined(USE_SIMD_FALLBACK)
+
+// SIMD operations on potentially-indeterminate Node16::index[16] bytes.
+// Implemented in file-level assembly (simd_x86_64.S) because loading and
+// operating on indeterminate values is UB in C++ but well-defined in
+// assembly. The caller must mask the returned bitfield to
+// [0, numChildren) before using it.
+//
+// Each function returns a 16-bit bitmask in the low 16 bits of a uint32_t
+// (upper 16 bits are zero). Bit i is set iff the condition holds at index i.
+
+extern "C" {
+// Returns bit i set iff idx[i] == key
+uint32_t find_eq_16(const uint8_t idx[16], uint8_t key);
+// Returns bit i set iff idx[i] >= child
+uint32_t find_ge_16(const uint8_t idx[16], uint8_t child);
+// Returns bit i set iff begin <= idx[i] < end
+uint32_t mask_in_range_16(const uint8_t idx[16], uint8_t begin, uint8_t end);
+}
+
+#endif
@@ -0,0 +1,74 @@
+// SIMD operations on potentially-indeterminate Node16::index[16] bytes.
+// Written in assembly so msan doesn't track the loads. The caller is
+// responsible for masking the returned bitfield to [0, numChildren) before
+// using it.
+//
+// All functions return a 16-bit bitmask in %eax (bit i set = condition true
+// at index i). The upper 16 bits of %eax are zero.
+//
+// System V AMD64 ABI:
+//   %rdi = const uint8_t *idx  (16 bytes)
+//   %esi = uint8_t key         (find_eq_16, find_ge_16)
+//   %sil = uint8_t begin       (mask_in_range_16)
+//   %dl  = uint8_t end         (mask_in_range_16)
+
+    .text
+
+// uint32_t find_eq_16(const uint8_t idx[16], uint8_t key)
+//   Returns bit i set if idx[i] == key
+    .globl find_eq_16
+    .type  find_eq_16, @function
+find_eq_16:
+    vmovd      %esi, %xmm1          // broadcast key
+    vpbroadcastb %xmm1, %xmm1
+    vmovdqu    (%rdi), %xmm0        // load 16 bytes (may contain indeterminate data)
+    vpcmpeqb   %xmm0, %xmm1, %xmm0  // 0xff for each match
+    vpmovmskb  %xmm0, %eax          // 16-bit bitmask
+    movzwl     %ax, %eax            // zero-extend to 32 bits
+    ret
+    .size find_eq_16, .-find_eq_16
+
+// uint32_t find_ge_16(const uint8_t idx[16], uint8_t child)
+//   Returns bit i set if idx[i] >= child
+//   x86 doesn't have a "compare unsigned >=" for bytes directly, so we use:
+//   min(key, idx[i]) == key  iff  idx[i] >= key
+    .globl find_ge_16
+    .type  find_ge_16, @function
+find_ge_16:
+    vmovd      %esi, %xmm1
+    vpbroadcastb %xmm1, %xmm1       // key broadcast
+    vmovdqu    (%rdi), %xmm0        // load 16 bytes
+    vpminub    %xmm0, %xmm1, %xmm2  // min(key, idx[i])
+    vpcmpeqb   %xmm2, %xmm1, %xmm0  // 0xff where min == key, i.e. idx[i] >= key
+    vpmovmskb  %xmm0, %eax
+    movzwl     %ax, %eax
+    ret
+    .size find_ge_16, .-find_ge_16
+
+// uint32_t mask_in_range_16(const uint8_t idx[16], uint8_t begin, uint8_t end)
+//   Returns bit i set if begin <= idx[i] < end
+//   Logic: (idx[i] - begin) < (end - begin)  [unsigned wrapping arithmetic]
+//   Equivalently: idx[i] - begin != max(idx[i] - begin, end - begin)
+//   i.e. idx[i] - begin is NOT equal to the saturated value.
+//   We compute: sub = idx - begin; result = (sub < (end-begin)) for each byte.
+//   Using: sub == max(sub, end-begin) means NOT in range.
+//   So: in_range = ~(movemask(cmpeq(sub, max(sub, range_size))))
+    .globl mask_in_range_16
+    .type  mask_in_range_16, @function
+mask_in_range_16:
+    vmovd      %esi, %xmm1          // begin
+    vpbroadcastb %xmm1, %xmm1
+    vmovd      %edx, %xmm2          // end
+    vpbroadcastb %xmm2, %xmm2
+    vmovdqu    (%rdi), %xmm0        // load 16 bytes
+    vpsubb     %xmm1, %xmm0, %xmm0  // idx - begin (wrapping)
+    vpsubb     %xmm1, %xmm2, %xmm2  // end - begin (range size)
+    vpmaxub    %xmm0, %xmm2, %xmm3  // max(idx-begin, range_size)
+    vpcmpeqb   %xmm3, %xmm0, %xmm0  // 0xff where NOT in range
+    vpmovmskb  %xmm0, %eax
+    not        %eax                 // invert: 1 = in range
+    movzwl     %ax, %eax
+    ret
+    .size mask_in_range_16, .-mask_in_range_16
+
+    .section .note.GNU-stack,"",@progbits
@@ -68,54 +68,6 @@ def test_hash_table_getBytes():
        assert cs.check(read(0, b"key")) == [Result.CONFLICT]


-def test_setOldestVersion_does_not_decrease():
-    # Regression test for issue #87: `setOldestVersion` sets oldestVersion to
-    # the maximum of its previous value and its argument, so passing an older
-    # version is valid input and must be a no-op. skip_list used to assert in
-    # Debug builds and, in Release builds, lower its oldest version, which
-    # resurrected a window of versions that callers were told resulted in
-    # TooOld.
-    with DebugConflictSet() as cs:
-        cs.addWrites(5, write(b"k"))
-        cs.setOldestVersion(5)
-        # Valid: `2 >= 0` and `2 <= newestVersion == 5`. oldestVersion stays 5,
-        # so 3 and 4 are still TooOld.
-        cs.setOldestVersion(2)
-        assert cs.check(read(3, b"k")) == [Result.TOO_OLD]
-        assert cs.check(read(4, b"k")) == [Result.TOO_OLD]
-        assert cs.check(read(5, b"k")) == [Result.COMMIT]
-
-    # Same, but with the entries at the old oldestVersion already garbage
-    # collected, so lowering oldestVersion would report Commit for a version
-    # that the caller was told resulted in TooOld.
-    with DebugConflictSet() as cs:
-        cs.addWrites(1, write(b"a"))
-        cs.addWrites(2)  # No writes; advances newestVersion to 2
-        cs.setOldestVersion(2)  # Garbage collects the write at version 1
-        cs.setOldestVersion(1)
-        assert cs.check(read(1, b"a")) == [Result.TOO_OLD]
-        assert cs.check(read(2, b"a")) == [Result.COMMIT]
-
-
-def test_hash_table_setOldestVersion_empties_map():
-    # Regression test for issue #78: setOldestVersion must terminate when
-    # its garbage-collection pass erases every entry in the map. Each write
-    # charges keyUpdates += 2, so with 50 writes keyUpdates = 100 and GC is
-    # enabled; erasing all entries leaves keyUpdates > 0 while the map is
-    # empty, which previously caused the GC loop to spin forever.
-    with ConflictSet(0, build_dir=build_dir, implementation="hash_table") as cs:
-        for i in range(50):
-            cs.addWrites(1, write(("k%06d" % i).encode()))
-        # All entries have version 1 <= oldestVersion 1 -> all erased.
-        cs.setOldestVersion(1)
-        # If we reach this point, setOldestVersion returned instead of hanging.
-        # The keys were erased, so reads at a valid version commit.
-        assert cs.check(read(1, b"k000000")) == [Result.COMMIT]
-        assert cs.check(read(2, b"k000000")) == [Result.COMMIT]
-        # A read older than oldestVersion is TooOld.
-        assert cs.check(read(0, b"k000000")) == [Result.TOO_OLD]
-
-
 def test_write_read_without_outer_reference():
    # Regression test for issue #42: WriteRange/ReadRange must keep their
    # backing key buffers alive, because the C library reads the pointer
@@ -251,18 +203,13 @@ def test_merge_child_node48():
        cs.addWrites(1, write(b"\x00" * 8, b"\x00" * 10))


-def test_fixup_256_heterogeneous_versions():
-    # Regression test for issue #89: recomputing the maxVersion of a Node256
-    # must consider all of its children, not just those with an index < 16.
-    # Here the highest-version child ("\x10\x50") has an index >= 16, and the
-    # range write ends at the Node256, which erases that child from the range
-    # and recomputes the Node256's maxVersion.
+def test_fixup_256():
    with DebugConflictSet() as cs:
+        cs.addWrites(0, write(bytes([1])))
        for i in range(256):
-            cs.addWrites(1, write(bytes([0x10, i])))
-        cs.addWrites(10, write(bytes([0x10, 0x50])))
-        cs.addWrites(11, write(bytes([0x05]), bytes([0x10])))
-        assert cs.check(read(5, bytes([0x10, 0x50]))) == [Result.CONFLICT]
+            cs.addWrites(1, write(bytes([1, i])))
+        cs.addWrites(2, write(bytes([0]), bytes([1])))
+        cs.check(read(0, bytes([1]), bytes([2])))


 def test_large_removal_buffer():
@@ -1,8 +1,7 @@
 #!/bin/bash

 # Disable frequency scaling
-# shellcheck disable=SC2002
-for i in $(seq "$(nproc)") ; do cat /sys/devices/system/cpu/cpu$((i-1))/cpufreq/scaling_max_freq | sudo tee /sys/devices/system/cpu/cpu$((i-1))/cpufreq/scaling_min_freq >/dev/null ; done
+for i in $(seq "$(nproc)") ; do sudo cat /sys/devices/system/cpu/cpu$((i-1))/cpufreq/scaling_max_freq | sudo tee /sys/devices/system/cpu/cpu$((i-1))/cpufreq/scaling_min_freq >/dev/null ; done

 cmake -S. -Bbuild -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_COMPILER=clang++ -G Ninja -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DCMAKE_CXX_FLAGS=-DNVALGRIND > /dev/null
 cmake --build build > /dev/null