Skip sorting if already sorted for skip list

Point read/write workload for server_bench
Require musttail and preserve_none for interleaved
2024-11-12 13:25:29 -08:00 · 2024-11-12 10:50:03 -08:00 · 2024-11-11 21:40:05 -08:00 · 2024-11-11 16:56:28 -08:00 · 2024-11-11 16:17:57 -08:00 · 2024-11-11 12:09:36 -08:00
3037 changed files with 5847 additions and 2367 deletions
@@ -7,7 +7,6 @@
 void showMemory(const ConflictSet &cs);
 #endif

-#define ANKERL_NANOBENCH_IMPLEMENT
 #include "third_party/nanobench.h"

 constexpr int kNumKeys = 1000000;
@@ -18,26 +17,26 @@ constexpr int kPrefixLen = 0;

 constexpr int kMvccWindow = 100000;

-std::span<const uint8_t> makeKey(Arena &arena, int index) {
+TrivialSpan makeKey(Arena &arena, int index) {

-  auto result =
-      std::span<uint8_t>{new (arena) uint8_t[4 + kPrefixLen], 4 + kPrefixLen};
+  uint8_t *buf = new (arena) uint8_t[4 + kPrefixLen];
+  auto result = TrivialSpan{buf, 4 + kPrefixLen};
  index = __builtin_bswap32(index);
-  memset(result.data(), 0, kPrefixLen);
-  memcpy(result.data() + kPrefixLen, &index, 4);
+  memset(buf, 0, kPrefixLen);
+  memcpy(buf, &index, 4);

  return result;
 }

-ConflictSet::ReadRange singleton(Arena &arena, std::span<const uint8_t> key) {
-  auto r =
-      std::span<uint8_t>(new (arena) uint8_t[key.size() + 1], key.size() + 1);
-  memcpy(r.data(), key.data(), key.size());
-  r[key.size()] = 0;
+ConflictSet::ReadRange singleton(Arena &arena, TrivialSpan key) {
+  uint8_t *buf = new (arena) uint8_t[key.size() + 1];
+  auto r = TrivialSpan(buf, key.size() + 1);
+  memcpy(buf, key.data(), key.size());
+  buf[key.size()] = 0;
  return {{key.data(), int(key.size())}, {r.data(), int(r.size())}, 0};
 }

-ConflictSet::ReadRange prefixRange(Arena &arena, std::span<const uint8_t> key) {
+ConflictSet::ReadRange prefixRange(Arena &arena, TrivialSpan key) {
  int index;
  for (index = key.size() - 1; index >= 0; index--)
    if ((key[index]) != 255)
@@ -49,9 +48,10 @@ ConflictSet::ReadRange prefixRange(Arena &arena, std::span<const uint8_t> key) {
    assert(false);
  }

-  auto r = std::span<uint8_t>(new (arena) uint8_t[index + 1], index + 1);
-  memcpy(r.data(), key.data(), index + 1);
-  r[r.size() - 1]++;
+  uint8_t *buf = new (arena) uint8_t[index + 1];
+  auto r = TrivialSpan(buf, index + 1);
+  memcpy(buf, key.data(), index + 1);
+  buf[r.size() - 1]++;
  return {{key.data(), int(key.size())}, {r.data(), int(r.size())}, 0};
 }

@@ -82,14 +82,7 @@ void benchConflictSet() {
    ++version;
  }

-  // I don't know why std::less didn't work /shrug
-  struct Less {
-    bool operator()(const std::span<const uint8_t> &lhs,
-                    const std::span<const uint8_t> &rhs) const {
-      return lhs < rhs;
-    }
-  };
-  auto points = set<std::span<const uint8_t>, Less>(arena);
+  auto points = set<TrivialSpan, std::less<>>(arena);

  while (points.size() < kOpsPerTx * 2 + 1) {
    // TODO don't use rand?
@@ -333,16 +326,22 @@ void benchWorstCaseForRadixRangeRead() {
  auto end = std::vector<uint8_t>(kKeyLenForWorstCase - 1, 255);
  end.push_back(254);

-  weaselab::ConflictSet::Result result;
-  weaselab::ConflictSet::ReadRange r{
-      {begin.data(), int(begin.size())}, {end.data(), int(end.size())}, 0};
+  weaselab::ConflictSet::ReadRange r[] = {
+      {{begin.data(), int(begin.size())}, {end.data(), int(end.size())}, 0},
+  };
+  weaselab::ConflictSet::Result results[sizeof(r) / sizeof(r[0])];
+  for (auto &result : results) {
+    result = weaselab::ConflictSet::TooOld;
+  }
+  bench.batch(sizeof(r) / sizeof(r[0]));

  bench.run("worst case for radix tree", [&]() {
    for (int i = 0; i < 256; ++i) {
-      result = weaselab::ConflictSet::TooOld;
-      cs[i]->check(&r, &result, 1);
-      if (result != weaselab::ConflictSet::Commit) {
-        abort();
+      cs[i]->check(r, results, sizeof(r) / sizeof(r[0]));
+      for (auto result : results) {
+        if (result != weaselab::ConflictSet::Commit) {
+          abort();
+        }
      }
    }
  });
@@ -358,21 +357,6 @@ void benchWorstCaseForRadixRangeRead() {
  // }
 }

-void benchMetrics() {
-  ankerl::nanobench::Bench bench;
-  ConflictSet cs{0};
-
-  int count;
-  ConflictSet::MetricsV1 *m;
-  cs.getMetricsV1(&m, &count);
-  bench.batch(count);
-  bench.run("fetch metric", [&]() {
-    for (int i = 0; i < count; ++i) {
-      m[i].getValue();
-    }
-  });
-}
-
 void benchCreateAndDestroy() {
  ankerl::nanobench::Bench bench;

@@ -382,6 +366,5 @@ void benchCreateAndDestroy() {
 int main(void) {
  benchConflictSet();
  benchWorstCaseForRadixRangeRead();
-  benchMetrics();
  benchCreateAndDestroy();
 }
@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.18)
 project(
  conflict-set
-  VERSION 0.0.9
+  VERSION 0.0.14
  DESCRIPTION
    "A data structure for optimistic concurrency control on ranges of bitwise-lexicographically-ordered keys."
  HOMEPAGE_URL "https://git.weaselab.dev/weaselab/conflict-set"
@@ -19,12 +19,6 @@ include(CheckCXXSourceCompiles)

 set(DEFAULT_BUILD_TYPE "Release")

-if(EMSCRIPTEN OR CMAKE_SYSTEM_NAME STREQUAL WASI)
-  set(WASM ON)
-else()
-  set(WASM OFF)
-endif()
-
 if(NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
  message(
    STATUS
@@ -37,8 +31,29 @@ if(NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
                                               "MinSizeRel" "RelWithDebInfo")
 endif()

-add_compile_options(-fdata-sections -ffunction-sections -Wswitch-enum
-                    -Werror=switch-enum -fPIC)
+add_compile_options(
+  -Werror=switch-enum -Wswitch-enum -fPIC -fdata-sections -ffunction-sections
+  -fno-jump-tables # https://github.com/llvm/llvm-project/issues/54247
+)
+
+if(CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+  add_link_options("-Wno-unused-command-line-argument")
+  find_program(LLVM_OBJCOPY llvm-objcopy)
+  if(LLVM_OBJCOPY)
+    set(CMAKE_OBJCOPY
+        ${LLVM_OBJCOPY}
+        CACHE FILEPATH "path to objcopy binary" FORCE)
+  endif()
+endif()
+
+if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
+  add_compile_options("-Wno-maybe-uninitialized")
+endif()
+
+if(NOT APPLE)
+  # This causes some versions of clang to crash on macos
+  add_compile_options(-g -fno-omit-frame-pointer)
+endif()

 set(full_relro_flags "-pie;LINKER:-z,relro,-z,now,-z,noexecstack")
 cmake_push_check_state()
@@ -49,6 +64,21 @@ if(HAS_FULL_RELRO)
 endif()
 cmake_pop_check_state()

+if(CMAKE_SYSTEM_PROCESSOR STREQUAL aarch64)
+  add_compile_options(-mbranch-protection=standard)
+else()
+  add_compile_options(-fcf-protection)
+  set(rewrite_endbr_flags "-fuse-ld=mold;LINKER:-z,rewrite-endbr")
+  cmake_push_check_state()
+  list(APPEND CMAKE_REQUIRED_LINK_OPTIONS ${rewrite_endbr_flags})
+  check_cxx_source_compiles("int main(){}" HAS_REWRITE_ENDBR FAIL_REGEX
+                            "warning:")
+  if(HAS_REWRITE_ENDBR)
+    add_link_options(${rewrite_endbr_flags})
+  endif()
+  cmake_pop_check_state()
+endif()
+
 set(version_script_flags
    LINKER:--version-script=${CMAKE_CURRENT_SOURCE_DIR}/linker.map)
 cmake_push_check_state()
@@ -74,22 +104,7 @@ else()
  add_link_options(-Wl,--gc-sections)
 endif()

-if(EMSCRIPTEN)
-  # https://github.com/emscripten-core/emscripten/issues/15377#issuecomment-1285167486
-  add_link_options(-lnodefs.js -lnoderawfs.js)
-  add_link_options(-s ALLOW_MEMORY_GROWTH)
-endif()
-
 if(NOT USE_SIMD_FALLBACK)
-  cmake_push_check_state()
-  list(APPEND CMAKE_REQUIRED_FLAGS -msimd128)
-  check_include_file_cxx("wasm_simd128.h" HAS_WASM_SIMD)
-  if(HAS_WASM_SIMD)
-    add_compile_options(-msimd128)
-    add_compile_definitions(HAS_WASM_SIMD)
-  endif()
-  cmake_pop_check_state()
-
  cmake_push_check_state()
  list(APPEND CMAKE_REQUIRED_FLAGS -mavx)
  check_include_file_cxx("immintrin.h" HAS_AVX)
@@ -112,12 +127,23 @@ target_compile_options(${PROJECT_NAME}-object PRIVATE -fno-exceptions
                                                      -fvisibility=hidden)
 target_include_directories(${PROJECT_NAME}-object
                           PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include)
+if(NOT LD_EXE)
+  set(LD_EXE ld)
+endif()
+add_custom_command(
+  OUTPUT ${CMAKE_BINARY_DIR}/${PROJECT_NAME}.o
+  COMMAND ${LD_EXE} -r $<TARGET_OBJECTS:${PROJECT_NAME}-object> -o
+          ${CMAKE_BINARY_DIR}/${PROJECT_NAME}.o
+  DEPENDS $<TARGET_OBJECTS:${PROJECT_NAME}-object>
+  COMMAND_EXPAND_LISTS)

-add_library(${PROJECT_NAME} SHARED $<TARGET_OBJECTS:${PROJECT_NAME}-object>)
+add_library(${PROJECT_NAME} SHARED ${CMAKE_BINARY_DIR}/${PROJECT_NAME}.o)
 set_target_properties(
  ${PROJECT_NAME} PROPERTIES LIBRARY_OUTPUT_DIRECTORY
                             "${CMAKE_CURRENT_BINARY_DIR}/radix_tree")
-if(NOT CMAKE_BUILD_TYPE STREQUAL Debug)
+if(CMAKE_BUILD_TYPE STREQUAL Debug)
+  set_target_properties(${PROJECT_NAME} PROPERTIES LINKER_LANGUAGE CXX)
+else()
  set_target_properties(${PROJECT_NAME} PROPERTIES LINKER_LANGUAGE C)
 endif()

@@ -127,19 +153,13 @@ if(HAS_VERSION_SCRIPT)
    LINKER:--version-script=${CMAKE_CURRENT_SOURCE_DIR}/linker.map)
 endif()

-add_library(${PROJECT_NAME}-static STATIC
-            $<TARGET_OBJECTS:${PROJECT_NAME}-object>)
-if(NOT CMAKE_BUILD_TYPE STREQUAL Debug)
+add_library(${PROJECT_NAME}-static STATIC ${CMAKE_BINARY_DIR}/${PROJECT_NAME}.o)
+if(CMAKE_BUILD_TYPE STREQUAL Debug)
+  set_target_properties(${PROJECT_NAME}-static PROPERTIES LINKER_LANGUAGE CXX)
+else()
  set_target_properties(${PROJECT_NAME}-static PROPERTIES LINKER_LANGUAGE C)
 endif()
-
-if(APPLE)
-  add_custom_command(
-    TARGET ${PROJECT_NAME}-static
-    PRE_LINK
-    COMMAND ${CMAKE_CURRENT_SOURCE_DIR}/privatize_symbols_macos.sh
-            $<TARGET_OBJECTS:${PROJECT_NAME}-object>)
-else()
+if(NOT APPLE)
  add_custom_command(
    TARGET ${PROJECT_NAME}-static
    POST_BUILD
@@ -155,6 +175,8 @@ include(CTest)
 # disable tests if this is being used through e.g. FetchContent
 if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)

+  add_library(nanobench ${CMAKE_CURRENT_SOURCE_DIR}/nanobench.cpp)
+
  set(TEST_FLAGS -Wall -Wextra -Wunreachable-code -Wpedantic -UNDEBUG)

  # corpus tests, which are tests curated by libfuzzer. The goal is to get broad
@@ -162,42 +184,38 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)

  file(GLOB CORPUS_TESTS ${CMAKE_CURRENT_SOURCE_DIR}/corpus/*)

-  # extra testing that relies on shared libraries, which aren't available with
-  # wasm
-  if(NOT WASM)
-    # Shared library version of FoundationDB's skip list implementation
-    add_library(skip_list SHARED SkipList.cpp)
-    target_compile_options(skip_list PRIVATE -fno-exceptions
-                                             -fvisibility=hidden)
-    target_include_directories(skip_list
-                               PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
-    set_target_properties(
-      skip_list PROPERTIES LIBRARY_OUTPUT_DIRECTORY
-                           "${CMAKE_CURRENT_BINARY_DIR}/skip_list")
-    set_target_properties(skip_list PROPERTIES OUTPUT_NAME ${PROJECT_NAME})
-    set_target_properties(
-      skip_list PROPERTIES VERSION ${PROJECT_VERSION} SOVERSION
-                                                      ${PROJECT_VERSION_MAJOR})
+  # Shared library version of FoundationDB's skip list implementation
+  add_library(skip_list SHARED SkipList.cpp)
+  target_compile_options(skip_list PRIVATE -fno-exceptions -fvisibility=hidden)
+  target_include_directories(skip_list
+                             PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
+  set_target_properties(
+    skip_list PROPERTIES LIBRARY_OUTPUT_DIRECTORY
+                         "${CMAKE_CURRENT_BINARY_DIR}/skip_list")
+  set_target_properties(skip_list PROPERTIES OUTPUT_NAME ${PROJECT_NAME})
+  set_target_properties(skip_list PROPERTIES VERSION ${PROJECT_VERSION}
+                                             SOVERSION ${PROJECT_VERSION_MAJOR})

-    # Shared library version of a std::unordered_map-based conflict set (point
-    # queries only)
-    add_library(hash_table SHARED HashTable.cpp)
-    target_compile_options(hash_table PRIVATE -fno-exceptions
-                                              -fvisibility=hidden)
-    target_include_directories(hash_table
-                               PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
-    set_target_properties(
-      hash_table PROPERTIES LIBRARY_OUTPUT_DIRECTORY
-                            "${CMAKE_CURRENT_BINARY_DIR}/hash_table")
-    set_target_properties(hash_table PROPERTIES OUTPUT_NAME ${PROJECT_NAME})
-    set_target_properties(
-      hash_table PROPERTIES VERSION ${PROJECT_VERSION} SOVERSION
-                                                       ${PROJECT_VERSION_MAJOR})
+  # Shared library version of a std::unordered_map-based conflict set (point
+  # queries only)
+  add_library(hash_table SHARED HashTable.cpp)
+  target_compile_options(hash_table PRIVATE -fno-exceptions -fvisibility=hidden)
+  target_include_directories(hash_table
+                             PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
+  set_target_properties(
+    hash_table PROPERTIES LIBRARY_OUTPUT_DIRECTORY
+                          "${CMAKE_CURRENT_BINARY_DIR}/hash_table")
+  set_target_properties(hash_table PROPERTIES OUTPUT_NAME ${PROJECT_NAME})
+  set_target_properties(
+    hash_table PROPERTIES VERSION ${PROJECT_VERSION} SOVERSION
+                                                     ${PROJECT_VERSION_MAJOR})

-    add_executable(driver_skip_list TestDriver.cpp)
-    target_compile_options(driver_skip_list PRIVATE ${TEST_FLAGS})
-    target_link_libraries(driver_skip_list PRIVATE skip_list)
+  add_executable(driver_skip_list TestDriver.cpp)
+  target_compile_options(driver_skip_list PRIVATE ${TEST_FLAGS})
+  target_link_libraries(driver_skip_list PRIVATE skip_list)

+  # enable to test skip list
+  if(0)
    foreach(TEST ${CORPUS_TESTS})
      get_filename_component(hash ${TEST} NAME)
      add_test(NAME skip_list_${hash} COMMAND driver_skip_list ${TEST})
@@ -209,6 +227,7 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
  target_include_directories(conflict_set_main
                             PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include)
  target_compile_definitions(conflict_set_main PRIVATE ENABLE_MAIN)
+  target_link_libraries(conflict_set_main PRIVATE nanobench)

  if(NOT APPLE)
    # libfuzzer target, to generate/manage corpus
@@ -269,6 +288,19 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
    add_test(NAME conflict_set_blackbox_${hash} COMMAND driver ${TEST})
  endforeach()

+  find_program(VALGRIND_EXE valgrind)
+  if(VALGRIND_EXE AND NOT CMAKE_CROSSCOMPILING)
+    list(LENGTH CORPUS_TESTS len)
+    math(EXPR last "${len} - 1")
+    set(partition_size 100)
+    foreach(i RANGE 0 ${last} ${partition_size})
+      list(SUBLIST CORPUS_TESTS ${i} ${partition_size} partition)
+      add_test(NAME conflict_set_blackbox_valgrind_${i}
+               COMMAND ${VALGRIND_EXE} --error-exitcode=99 --
+                       $<TARGET_FILE:driver> ${partition})
+    endforeach()
+  endif()
+
  # scripted tests. Written manually to fill in anything libfuzzer couldn't
  # find.
  if(NOT CMAKE_CROSSCOMPILING)
@@ -289,16 +321,17 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
          ${Python3_EXECUTABLE}
          ${CMAKE_CURRENT_SOURCE_DIR}/test_conflict_set.py test ${TEST}
          --build-dir ${CMAKE_CURRENT_BINARY_DIR})
+      if(VALGRIND_EXE AND NOT CMAKE_CROSSCOMPILING)
+        add_test(
+          NAME script_test_${TEST}_valgrind
+          COMMAND
+            ${VALGRIND_EXE} ${Python3_EXECUTABLE}
+            ${CMAKE_CURRENT_SOURCE_DIR}/test_conflict_set.py test ${TEST}
+            --build-dir ${CMAKE_CURRENT_BINARY_DIR})
+      endif()
    endforeach()
  endif()

-  find_program(VALGRIND_EXE valgrind)
-  if(VALGRIND_EXE AND NOT CMAKE_CROSSCOMPILING)
-    add_test(NAME conflict_set_blackbox_valgrind
-             COMMAND ${VALGRIND_EXE} --error-exitcode=99 --
-                     $<TARGET_FILE:driver> ${CORPUS_TESTS})
-  endif()
-
  # api smoke tests

  # c90
@@ -313,14 +346,15 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
  # c++98
  add_executable(conflict_set_cxx_api_test conflict_set_cxx_api_test.cpp)
  target_compile_options(conflict_set_cxx_api_test PRIVATE ${TEST_FLAGS})
-  target_link_libraries(conflict_set_cxx_api_test PRIVATE ${PROJECT_NAME})
+  target_link_libraries(conflict_set_cxx_api_test
+                        PRIVATE ${PROJECT_NAME}-static)
  set_target_properties(conflict_set_cxx_api_test PROPERTIES CXX_STANDARD 98)
  set_target_properties(conflict_set_cxx_api_test
                        PROPERTIES CXX_STANDARD_REQUIRED ON)
  add_test(NAME conflict_set_cxx_api_test COMMAND conflict_set_cxx_api_test)

  # symbol visibility tests
-  if(NOT WASM AND NOT CMAKE_BUILD_TYPE STREQUAL Debug)
+  if(NOT CMAKE_BUILD_TYPE STREQUAL Debug)
    if(APPLE)
      set(symbol_exports ${CMAKE_CURRENT_SOURCE_DIR}/apple-symbol-exports.txt)
      set(symbol_imports ${CMAKE_CURRENT_SOURCE_DIR}/apple-symbol-imports.txt)
@@ -346,13 +380,40 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
        ${symbol_imports})
  endif()

+  if(NOT CMAKE_CROSSCOMPILING)
+    find_program(HARDENING_CHECK hardening-check)
+    if(HARDENING_CHECK)
+      add_test(NAME hardening_check
+               COMMAND ${HARDENING_CHECK} $<TARGET_FILE:${PROJECT_NAME}>
+                       --nofortify --nostackprotector)
+    endif()
+  endif()
+
  # bench
  add_executable(conflict_set_bench Bench.cpp)
-  target_link_libraries(conflict_set_bench PRIVATE ${PROJECT_NAME})
+  target_link_libraries(conflict_set_bench PRIVATE ${PROJECT_NAME} nanobench)
  set_target_properties(conflict_set_bench PROPERTIES SKIP_BUILD_RPATH ON)
  add_executable(real_data_bench RealDataBench.cpp)
  target_link_libraries(real_data_bench PRIVATE ${PROJECT_NAME})
  set_target_properties(real_data_bench PROPERTIES SKIP_BUILD_RPATH ON)
+
+  # fuzzer-based perf
+  add_executable(driver_perf TestDriver.cpp)
+  target_compile_definitions(driver_perf PRIVATE PERF_TEST=1)
+  target_link_libraries(driver_perf PRIVATE ${PROJECT_NAME})
+
+  # server bench
+  add_executable(server_bench ServerBench.cpp)
+  target_link_libraries(server_bench PRIVATE ${PROJECT_NAME})
+  set_target_properties(server_bench PROPERTIES SKIP_BUILD_RPATH ON)
+
+  add_executable(interleaving_test InterleavingTest.cpp)
+  # work around lack of musttail for gcc
+  if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND CMAKE_BUILD_TYPE STREQUAL "Debug")
+    target_compile_options(interleaving_test PRIVATE -Og
+                                                     -foptimize-sibling-calls)
+  endif()
+  target_link_libraries(interleaving_test PRIVATE nanobench)
 endif()

 # packaging
@@ -8,25 +8,27 @@ RUN chmod -R 777 /tmp
 RUN apt-get update
 RUN apt-get upgrade -y
 RUN TZ=America/Los_Angeles DEBIAN_FRONTEND=noninteractive apt-get install -y \
+    binutils-aarch64-linux-gnu \
    build-essential \
    ccache \
-    clang \
    cmake \
    curl \
-    doxygen \
-    file \
+    devscripts \
    g++-aarch64-linux-gnu \
    gcovr \
    git \
-    gperf \
-    graphviz \
+    gnupg \
    libc6-dbg \
+    lsb-release \
+    mold \
    ninja-build \
    pre-commit \
    python3-requests \
    qemu-user \
    rpm \
+    software-properties-common \
    texlive-full \
+    wget \
    zstd

 # Install recent valgrind from source
@@ -42,6 +44,11 @@ RUN curl -Ls https://sourceware.org/pub/valgrind/valgrind-3.22.0.tar.bz2 -o valg
    cd .. && \
    rm -rf /tmp/*

+# Recent clang
+RUN wget https://apt.llvm.org/llvm.sh && chmod +x ./llvm.sh && ./llvm.sh 20
+
+RUN apt-get -y install clang llvm
+
 # Set after building valgrind, which doesn't build with clang for some reason
 ENV CC=clang
 ENV CXX=clang++
@@ -98,6 +98,13 @@ void ConflictSet::setOldestVersion(int64_t oldestVersion) {

 int64_t ConflictSet::getBytes() const { return -1; }

+void ConflictSet::getMetricsV1(MetricsV1 **metrics, int *count) const {
+  *metrics = nullptr;
+  *count = 0;
+}
+
+double ConflictSet::MetricsV1::getValue() const { return 0; }
+
 ConflictSet::ConflictSet(int64_t oldestVersion)
    : impl(new(safe_malloc(sizeof(Impl))) Impl{oldestVersion}) {}

@@ -0,0 +1,256 @@
+#include <alloca.h>
+#include <cassert>
+#ifdef __x86_64__
+#include <immintrin.h>
+#endif
+
+#include "third_party/nanobench.h"
+
+struct Job {
+  int *input;
+  // Returned void* is a function pointer to the next continuation. We have to
+  // use void* because otherwise the type would be recursive.
+  typedef void *(*continuation)(Job *);
+  continuation next;
+};
+
+void *stepJob(Job *j) {
+  auto done = --(*j->input) == 0;
+#ifdef __x86_64__
+  _mm_clflush(j->input);
+#endif
+  return done ? nullptr : (void *)stepJob;
+}
+
+void sequential(Job **jobs, int count) {
+  for (int i = 0; i < count; ++i) {
+    do {
+      jobs[i]->next = (Job::continuation)jobs[i]->next(jobs[i]);
+    } while (jobs[i]->next);
+  }
+}
+
+void sequentialNoFuncPtr(Job **jobs, int count) {
+  for (int i = 0; i < count; ++i) {
+    while (stepJob(jobs[i]))
+      ;
+  }
+}
+
+void interleaveSwapping(Job **jobs, int remaining) {
+  int current = 0;
+  while (remaining > 0) {
+    auto next = (Job::continuation)jobs[current]->next(jobs[current]);
+    jobs[current]->next = next;
+    if (next == nullptr) {
+      jobs[current] = jobs[remaining - 1];
+      --remaining;
+    } else {
+      ++current;
+    }
+    if (current == remaining) {
+      current = 0;
+    }
+  }
+}
+
+void interleaveBoundedCyclicList(Job **jobs, int count) {
+  if (count == 0) {
+    return;
+  }
+
+  constexpr int kConcurrent = 32;
+  Job *inProgress[kConcurrent];
+  int nextJob[kConcurrent];
+
+  int started = std::min(kConcurrent, count);
+  for (int i = 0; i < started; i++) {
+    inProgress[i] = jobs[i];
+    nextJob[i] = i + 1;
+  }
+  nextJob[started - 1] = 0;
+
+  int prevJob = started - 1;
+  int job = 0;
+  for (;;) {
+    auto next = (Job::continuation)inProgress[job]->next(inProgress[job]);
+    inProgress[job]->next = next;
+    if (next == nullptr) {
+      if (started == count) {
+        if (prevJob == job)
+          break;
+        nextJob[prevJob] = nextJob[job];
+        job = prevJob;
+      } else {
+        int temp = started++;
+        inProgress[job] = jobs[temp];
+      }
+    }
+    prevJob = job;
+    job = nextJob[job];
+  }
+}
+
+#ifndef __has_attribute
+#define __has_attribute(x) 0
+#endif
+
+#if __has_attribute(musttail)
+#define MUSTTAIL __attribute__((musttail))
+#else
+#define MUSTTAIL
+#endif
+
+struct Context {
+  constexpr static int kConcurrent = 32;
+  Job **jobs;
+  Job *inProgress[kConcurrent];
+  void (*continuation[kConcurrent])(Context *, int64_t prevJob, int64_t job,
+                                    int64_t started, int64_t count);
+  int nextJob[kConcurrent];
+};
+
+void keepGoing(Context *context, int64_t prevJob, int64_t job, int64_t started,
+               int64_t count) {
+  prevJob = job;
+  job = context->nextJob[job];
+  MUSTTAIL return context->continuation[job](context, prevJob, job, started,
+                                             count);
+}
+
+void stepJobTailCall(Context *context, int64_t prevJob, int64_t job,
+                     int64_t started, int64_t count);
+
+void complete(Context *context, int64_t prevJob, int64_t job, int64_t started,
+              int64_t count) {
+  if (started == count) {
+    if (prevJob == job) {
+      return;
+    }
+    context->nextJob[prevJob] = context->nextJob[job];
+    job = prevJob;
+  } else {
+    context->inProgress[job] = context->jobs[started++];
+    context->continuation[job] = stepJobTailCall;
+  }
+  prevJob = job;
+  job = context->nextJob[job];
+  MUSTTAIL return context->continuation[job](context, prevJob, job, started,
+                                             count);
+}
+
+void stepJobTailCall(Context *context, int64_t prevJob, int64_t job,
+                     int64_t started, int64_t count) {
+  auto *j = context->inProgress[job];
+  auto done = --(*j->input) == 0;
+#ifdef __x86_64__
+  _mm_clflush(j->input);
+#endif
+  if (done) {
+    MUSTTAIL return complete(context, prevJob, job, started, count);
+  } else {
+    context->continuation[job] = stepJobTailCall;
+    MUSTTAIL return keepGoing(context, prevJob, job, started, count);
+  }
+}
+
+void useTailCalls(Job **jobs, int count) {
+  if (count == 0) {
+    return;
+  }
+  Context context;
+  context.jobs = jobs;
+  int64_t started = std::min(Context::kConcurrent, count);
+  for (int i = 0; i < started; i++) {
+    context.inProgress[i] = jobs[i];
+    context.nextJob[i] = i + 1;
+    context.continuation[i] = stepJobTailCall;
+  }
+  context.nextJob[started - 1] = 0;
+  int prevJob = started - 1;
+  int job = 0;
+  return context.continuation[job](&context, prevJob, job, started, count);
+}
+
+void interleaveCyclicList(Job **jobs, int count) {
+  auto *nextJob = (int *)alloca(sizeof(int) * count);
+
+  for (int i = 0; i < count - 1; ++i) {
+    nextJob[i] = i + 1;
+  }
+  nextJob[count - 1] = 0;
+
+  int prevJob = count - 1;
+  int job = 0;
+  for (;;) {
+    auto next = (Job::continuation)jobs[job]->next(jobs[job]);
+    jobs[job]->next = next;
+    if (next == nullptr) {
+      if (prevJob == job)
+        break;
+      nextJob[prevJob] = nextJob[job];
+      job = prevJob;
+    }
+    prevJob = job;
+    job = nextJob[job];
+  }
+}
+
+int main() {
+  ankerl::nanobench::Bench bench;
+
+  constexpr int kNumJobs = 10000;
+  bench.relative(true);
+
+  Job jobs[kNumJobs];
+  Job jobsCopy[kNumJobs];
+  int iters = 0;
+  int originalInput[kNumJobs];
+  for (int i = 0; i < kNumJobs; ++i) {
+    originalInput[i] = rand() % 5 + 3;
+    jobs[i].input = new int{originalInput[i]};
+    jobs[i].next = stepJob;
+    iters += *jobs[i].input;
+  }
+  bench.batch(iters);
+
+  for (auto [scheduler, name] :
+       {std::make_pair(sequentialNoFuncPtr, "sequentialNoFuncPtr"),
+        std::make_pair(sequential, "sequential"),
+        std::make_pair(useTailCalls, "useTailCalls"),
+        std::make_pair(interleaveSwapping, "interleavingSwapping"),
+        std::make_pair(interleaveBoundedCyclicList,
+                       "interleaveBoundedCyclicList"),
+        std::make_pair(interleaveCyclicList, "interleaveCyclicList")}) {
+    for (int i = 0; i < kNumJobs; ++i) {
+      *jobs[i].input = originalInput[i];
+    }
+    memcpy(jobsCopy, jobs, sizeof(jobs));
+    Job *ps[kNumJobs];
+    for (int i = 0; i < kNumJobs; ++i) {
+      ps[i] = jobsCopy + i;
+    }
+    scheduler(ps, kNumJobs);
+    for (int i = 0; i < kNumJobs; ++i) {
+      if (*jobsCopy[i].input != 0) {
+        fprintf(stderr, "%s failed\n", name);
+        abort();
+      }
+    }
+
+    bench.run(name, [&]() {
+      for (int i = 0; i < kNumJobs; ++i) {
+        *jobs[i].input = originalInput[i];
+      }
+      memcpy(jobsCopy, jobs, sizeof(jobs));
+      Job *ps[kNumJobs];
+      for (int i = 0; i < kNumJobs; ++i) {
+        ps[i] = jobsCopy + i;
+      }
+      scheduler(ps, kNumJobs);
+    });
+  }
+  for (int i = 0; i < kNumJobs; ++i) {
+    delete jobs[i].input;
+  }
+}
@@ -20,16 +20,44 @@ using namespace weaselab;
 #include <thread>
 #include <unordered_set>
 #include <utility>
-#include <vector>

 #include <callgrind.h>

 #define DEBUG_VERBOSE 0
 #define SHOW_MEMORY 0

-[[nodiscard]] inline auto
-operator<=>(const std::span<const uint8_t> &lhs,
-            const std::span<const uint8_t> &rhs) noexcept {
+// std::span is not trivially constructible. We want a span that leaves its
+// members uninitialized for performance reasons.
+struct TrivialSpan {
+  TrivialSpan() = default;
+  TrivialSpan(const uint8_t *begin, int len) : begin(begin), len(len) {}
+
+  uint8_t back() const {
+    assert(len > 0);
+    return begin[len - 1];
+  }
+  uint8_t front() const {
+    assert(len > 0);
+    return begin[0];
+  }
+  uint8_t operator[](int i) const {
+    assert(0 <= i);
+    assert(i < len);
+    return begin[i];
+  }
+  int size() const { return len; }
+  TrivialSpan subspan(int offset, int len) { return {begin + offset, len}; }
+  const uint8_t *data() const { return begin; }
+
+private:
+  const uint8_t *begin;
+  int len;
+};
+
+static_assert(std::is_trivial_v<TrivialSpan>);
+
+[[nodiscard]] inline auto operator<=>(const TrivialSpan &lhs,
+                                      const TrivialSpan &rhs) noexcept {
  int cl = std::min<int>(lhs.size(), rhs.size());
  if (cl > 0) {
    if (auto c = memcmp(lhs.data(), rhs.data(), cl) <=> 0; c != 0) {
@@ -39,7 +67,7 @@ operator<=>(const std::span<const uint8_t> &lhs,
  return lhs.size() <=> rhs.size();
 }

-[[nodiscard]] inline auto operator<=>(const std::span<const uint8_t> &lhs,
+[[nodiscard]] inline auto operator<=>(const TrivialSpan &lhs,
                                      const ConflictSet::Key &rhs) noexcept {
  int cl = std::min<int>(lhs.size(), rhs.len);
  if (cl > 0) {
@@ -47,7 +75,18 @@ operator<=>(const std::span<const uint8_t> &lhs,
      return c;
    }
  }
-  return lhs.size() <=> size_t(rhs.len);
+  return lhs.size() <=> rhs.len;
+}
+
+[[nodiscard]] inline auto operator<=>(const ConflictSet::Key &lhs,
+                                      const ConflictSet::Key &rhs) noexcept {
+  int cl = std::min<int>(lhs.len, rhs.len);
+  if (cl > 0) {
+    if (auto c = memcmp(lhs.p, rhs.p, cl) <=> 0; c != 0) {
+      return c;
+    }
+  }
+  return lhs.len <=> rhs.len;
 }

 // This header contains code that we want to reuse outside of ConflictSet.cpp or
@@ -273,6 +312,16 @@ template <class T> struct Vector {
    size_ += slice.size();
  }

+  // Caller must write to the returned slice
+  std::span<T> unsafePrepareAppend(int appendSize) {
+    if (size_ + appendSize > capacity) {
+      grow(std::max<int>(size_ + appendSize, capacity * 2));
+    }
+    auto result = std::span<T>(t + size_, appendSize);
+    size_ += appendSize;
+    return result;
+  }
+
  void push_back(const T &t) { append(std::span<const T>(&t, 1)); }

  T *begin() { return t; }
@@ -529,7 +578,7 @@ struct ReferenceImpl {

 using Key = ConflictSet::Key;

-inline Key operator"" _s(const char *str, size_t size) {
+inline Key operator""_s(const char *str, size_t size) {
  return {reinterpret_cast<const uint8_t *>(str), int(size)};
 }

@@ -560,7 +609,7 @@ inline std::string printable(const Key &key) {
  return printable(std::string_view((const char *)key.p, key.len));
 }

-inline std::string printable(std::span<const uint8_t> key) {
+inline std::string printable(TrivialSpan key) {
  return printable(std::string_view((const char *)key.data(), key.size()));
 }

@@ -578,12 +627,16 @@ inline const char *resultToStr(ConflictSet::Result r) {

 namespace {

-template <class ConflictSetImpl> struct TestDriver {
-  Arbitrary arbitrary;
-  explicit TestDriver(const uint8_t *data, size_t size)
-      : arbitrary({data, size}) {}
+template <class ConflictSetImpl, bool kEnableAssertions = true>
+struct TestDriver {
+  Arbitrary *arbitrary;
+  explicit TestDriver(Arbitrary &a) : arbitrary(&a) {
+#if DEBUG_VERBOSE && !defined(NDEBUG)
+    fprintf(stderr, "%p Initial version: {%" PRId64 "}\n", this, writeVersion);
+#endif
+  }

-  int64_t oldestVersion = arbitrary.next();
+  int64_t oldestVersion = arbitrary->next();
  int64_t writeVersion = oldestVersion;
  ConflictSetImpl cs{oldestVersion};
  ReferenceImpl refImpl{oldestVersion};
@@ -592,33 +645,34 @@ template <class ConflictSetImpl> struct TestDriver {

  bool ok = true;

-  const int prefixLen = arbitrary.bounded(512);
-  const int prefixByte = arbitrary.randT<uint8_t>();
+  const int prefixLen = arbitrary->bounded(512);
+  const int prefixByte = arbitrary->randT<uint8_t>();

  // Call until it returns true, for "done". Check internal invariants etc
  // between calls to next.
  bool next() {
    assert(cs.getBytes() >= 0);
-    if (!arbitrary.hasEntropy()) {
+    if (!arbitrary->hasEntropy()) {
      return true;
    }
    Arena arena;
    {
-      int numPointWrites = arbitrary.bounded(100);
-      int numRangeWrites = arbitrary.bounded(100);
-      int64_t v = (writeVersion += arbitrary.bounded(10) ? arbitrary.bounded(10)
-                                                         : arbitrary.next());
+      int numPointWrites = arbitrary->bounded(100);
+      int numRangeWrites = arbitrary->bounded(100);
+      int64_t v =
+          (writeVersion +=
+           arbitrary->bounded(10) ? arbitrary->bounded(10) : arbitrary->next());
      auto *writes =
          new (arena) ConflictSet::WriteRange[numPointWrites + numRangeWrites];
      auto keys = set<std::string_view>(arena);
      while (int(keys.size()) < numPointWrites + numRangeWrites * 2) {
-        if (!arbitrary.hasEntropy()) {
+        if (!arbitrary->hasEntropy()) {
          return true;
        }
-        int keyLen = prefixLen + arbitrary.bounded(kMaxKeySuffixLen);
+        int keyLen = prefixLen + arbitrary->bounded(kMaxKeySuffixLen);
        auto *begin = new (arena) uint8_t[keyLen];
        memset(begin, prefixByte, prefixLen);
-        arbitrary.randomBytes(begin + prefixLen, keyLen - prefixLen);
+        arbitrary->randomBytes(begin + prefixLen, keyLen - prefixLen);
        keys.insert(std::string_view((const char *)begin, keyLen));
      }

@@ -628,7 +682,7 @@ template <class ConflictSetImpl> struct TestDriver {
               rangesRemaining = numRangeWrites;
           pointsRemaining > 0 || rangesRemaining > 0; ++i) {
        bool pointRead = pointsRemaining > 0 && rangesRemaining > 0
-                             ? bool(arbitrary.bounded(2))
+                             ? bool(arbitrary->bounded(2))
                             : pointsRemaining > 0;
        if (pointRead) {
          assert(pointsRemaining > 0);
@@ -647,39 +701,24 @@ template <class ConflictSetImpl> struct TestDriver {
          ++iter;
          --rangesRemaining;
        }
-#if DEBUG_VERBOSE && !defined(NDEBUG)
-        if (writes[i].end.len == 0) {
-          fprintf(stderr, "Write: {%s}\n", printable(writes[i].begin).c_str());
-        } else {
-          fprintf(stderr, "Write: [%s, %s)\n",
-                  printable(writes[i].begin).c_str(),
-                  printable(writes[i].end).c_str());
-        }
-#endif
      }
      assert(iter == keys.end());
      assert(i == numPointWrites + numRangeWrites);

-#if DEBUG_VERBOSE && !defined(NDEBUG)
-      fprintf(stderr, "Write @ %" PRId64 "\n", v);
-#endif
-
      // Test non-canonical writes
      if (numPointWrites > 0) {
-        int overlaps = arbitrary.bounded(numPointWrites);
+        int overlaps = arbitrary->bounded(numPointWrites);
        for (int i = 0; i < numPointWrites + numRangeWrites && overlaps > 0;
             ++i) {
          if (writes[i].end.len == 0) {
-            int keyLen = prefixLen + arbitrary.bounded(kMaxKeySuffixLen);
+            int keyLen = prefixLen + arbitrary->bounded(kMaxKeySuffixLen);
            auto *begin = new (arena) uint8_t[keyLen];
            memset(begin, prefixByte, prefixLen);
-            arbitrary.randomBytes(begin + prefixLen, keyLen - prefixLen);
+            arbitrary->randomBytes(begin + prefixLen, keyLen - prefixLen);
            writes[i].end.len = keyLen;
            writes[i].end.p = begin;
-            auto c =
-                std::span<const uint8_t>(writes[i].begin.p,
-                                         writes[i].begin.len) <=>
-                std::span<const uint8_t>(writes[i].end.p, writes[i].end.len);
+            auto c = TrivialSpan(writes[i].begin.p, writes[i].begin.len) <=>
+                     TrivialSpan(writes[i].end.p, writes[i].end.len);
            if (c > 0) {
              using std::swap;
              swap(writes[i].begin, writes[i].end);
@@ -691,10 +730,10 @@ template <class ConflictSetImpl> struct TestDriver {
          }
        }
      }
-      if (arbitrary.bounded(2)) {
+      if (arbitrary->bounded(2)) {
        // Shuffle writes
        for (int i = numPointWrites + numRangeWrites - 1; i > 0; --i) {
-          int j = arbitrary.bounded(i + 1);
+          int j = arbitrary->bounded(i + 1);
          if (i != j) {
            using std::swap;
            swap(writes[i], writes[j]);
@@ -703,7 +742,7 @@ template <class ConflictSetImpl> struct TestDriver {
      }

      oldestVersion +=
-          arbitrary.bounded(10) ? arbitrary.bounded(10) : arbitrary.next();
+          arbitrary->bounded(10) ? arbitrary->bounded(10) : arbitrary->next();
      oldestVersion = std::min(oldestVersion, writeVersion);

 #ifdef THREAD_TEST
@@ -720,38 +759,63 @@ template <class ConflictSetImpl> struct TestDriver {
      ready.wait();
 #endif

+#if DEBUG_VERBOSE && !defined(NDEBUG)
+      for (int i = 0; i < numPointWrites + numRangeWrites; ++i) {
+        if (writes[i].end.len == 0) {
+          fprintf(stderr, "%p Write: {%s}\n", this,
+                  printable(writes[i].begin).c_str());
+        } else {
+          fprintf(stderr, "%p Write: [%s, %s)\n", this,
+                  printable(writes[i].begin).c_str(),
+                  printable(writes[i].end).c_str());
+        }
+      }
+      fprintf(stderr, "%p Write @ %" PRId64 "\n", this, v);
+#endif
+
      CALLGRIND_START_INSTRUMENTATION;
      cs.addWrites(writes, numPointWrites + numRangeWrites, v);
      CALLGRIND_STOP_INSTRUMENTATION;

-      refImpl.addWrites(writes, numPointWrites + numRangeWrites, v);
+      if constexpr (kEnableAssertions) {
+        refImpl.addWrites(writes, numPointWrites + numRangeWrites, v);
+      }

+#if DEBUG_VERBOSE && !defined(NDEBUG)
+      fprintf(stderr, "%p Set oldest version: %" PRId64 "\n", this,
+              oldestVersion);
+#endif
+      CALLGRIND_START_INSTRUMENTATION;
      cs.setOldestVersion(oldestVersion);
-      refImpl.setOldestVersion(oldestVersion);
+      CALLGRIND_STOP_INSTRUMENTATION;
+
+      if constexpr (kEnableAssertions) {
+        refImpl.setOldestVersion(oldestVersion);
+      }

 #ifdef THREAD_TEST
      thread2.join();
 #endif
    }
    {
-      int numPointReads = arbitrary.bounded(100);
-      int numRangeReads = arbitrary.bounded(100);
+      int numPointReads = arbitrary->bounded(100);
+      int numRangeReads = arbitrary->bounded(100);

-      int64_t v = std::max<int64_t>(writeVersion - (arbitrary.bounded(10)
-                                                        ? arbitrary.bounded(10)
-                                                        : arbitrary.next()),
+      int64_t v = std::max<int64_t>(writeVersion - (arbitrary->bounded(10)
+                                                        ? arbitrary->bounded(10)
+                                                        : arbitrary->next()),
                                    0);
      auto *reads =
          new (arena) ConflictSet::ReadRange[numPointReads + numRangeReads];
      auto keys = set<std::string_view>(arena);
      while (int(keys.size()) < numPointReads + numRangeReads * 2) {
-        if (!arbitrary.hasEntropy()) {
+        if (!arbitrary->hasEntropy()) {
          return true;
        }
-        int keyLen = prefixLen + arbitrary.bounded(kMaxKeySuffixLen);
+        int keyLen = prefixLen + arbitrary->bounded(kMaxKeySuffixLen);
        auto *begin = new (arena) uint8_t[keyLen];
        memset(begin, prefixByte, prefixLen);
-        arbitrary.randomBytes(begin + prefixLen, keyLen - prefixLen);
+        arbitrary->randomBytes(begin + prefixLen, keyLen - prefixLen);
        keys.insert(std::string_view((const char *)begin, keyLen));
      }

@@ -760,7 +824,7 @@ template <class ConflictSetImpl> struct TestDriver {
      for (int pointsRemaining = numPointReads, rangesRemaining = numRangeReads;
           pointsRemaining > 0 || rangesRemaining > 0; ++i) {
        bool pointRead = pointsRemaining > 0 && rangesRemaining > 0
-                             ? bool(arbitrary.bounded(2))
+                             ? bool(arbitrary->bounded(2))
                             : pointsRemaining > 0;
        if (pointRead) {
          assert(pointsRemaining > 0);
@@ -782,10 +846,10 @@ template <class ConflictSetImpl> struct TestDriver {
        reads[i].readVersion = v;
 #if DEBUG_VERBOSE && !defined(NDEBUG)
        if (reads[i].end.len == 0) {
-          fprintf(stderr, "Read: {%s} @ %" PRId64 "\n",
+          fprintf(stderr, "%p Read: {%s} @ %" PRId64 "\n", this,
                  printable(reads[i].begin).c_str(), reads[i].readVersion);
        } else {
-          fprintf(stderr, "Read: [%s, %s) @ %" PRId64 "\n",
+          fprintf(stderr, "%p Read: [%s, %s) @ %" PRId64 "\n", this,
                  printable(reads[i].begin).c_str(),
                  printable(reads[i].end).c_str(), reads[i].readVersion);
        }
@@ -821,35 +885,40 @@ template <class ConflictSetImpl> struct TestDriver {
      cs.check(reads, results1, numPointReads + numRangeReads);
      CALLGRIND_STOP_INSTRUMENTATION;

-      // Call remaining const methods
-      cs.getBytes();
-      ConflictSet::MetricsV1 *m;
-      int count;
-      cs.getMetricsV1(&m, &count);
-      for (int i = 0; i < count; ++i) {
-        m[i].getValue();
+      if constexpr (kEnableAssertions) {
+        // Call remaining const methods
+        cs.getBytes();
+        ConflictSet::MetricsV1 *m;
+        int count;
+        cs.getMetricsV1(&m, &count);
+        for (int i = 0; i < count; ++i) {
+          m[i].getValue();
+        }
+
+        refImpl.check(reads, results2, numPointReads + numRangeReads);
      }

-      refImpl.check(reads, results2, numPointReads + numRangeReads);
-
-      auto compareResults = [reads](ConflictSet::Result *results1,
-                                    ConflictSet::Result *results2, int count) {
+      auto compareResults = [reads, this](ConflictSet::Result *results1,
+                                          ConflictSet::Result *results2,
+                                          int count) {
        for (int i = 0; i < count; ++i) {
          if (results1[i] != results2[i]) {
            if (reads[i].end.len == 0) {
-              fprintf(stderr,
-                      "Expected %s, got %s for read of {%s} at version %" PRId64
-                      "\n",
-                      resultToStr(results2[i]), resultToStr(results1[i]),
-                      printable(reads[i].begin).c_str(), reads[i].readVersion);
-            } else {
              fprintf(
                  stderr,
-                  "Expected %s, got %s for read of [%s, %s) at version %" PRId64
+                  "%p Expected %s, got %s for read of {%s} at version %" PRId64
                  "\n",
-                  resultToStr(results2[i]), resultToStr(results1[i]),
-                  printable(reads[i].begin).c_str(),
-                  printable(reads[i].end).c_str(), reads[i].readVersion);
+                  (void *)this, resultToStr(results2[i]),
+                  resultToStr(results1[i]), printable(reads[i].begin).c_str(),
+                  reads[i].readVersion);
+            } else {
+              fprintf(stderr,
+                      "%p Expected %s, got %s for read of [%s, %s) at version "
+                      "%" PRId64 "\n",
+                      (void *)this, resultToStr(results2[i]),
+                      resultToStr(results1[i]),
+                      printable(reads[i].begin).c_str(),
+                      printable(reads[i].end).c_str(), reads[i].readVersion);
            }
            return false;
          }
@@ -857,9 +926,12 @@ template <class ConflictSetImpl> struct TestDriver {
        return true;
      };

-      if (!compareResults(results1, results2, numPointReads + numRangeReads)) {
-        ok = false;
-        return true;
+      if constexpr (kEnableAssertions) {
+        if (!compareResults(results1, results2,
+                            numPointReads + numRangeReads)) {
+          ok = false;
+          return true;
+        }
      }

 #ifdef THREAD_TEST
@@ -11,11 +11,11 @@ def CleanBuildAndTest(String cmakeArgs) {
    catchError {
        sh '''
        cd build
-        ctest --no-compress-output --test-output-size-passed 100000 --test-output-size-failed 100000 -T Test -j `nproc` --timeout 90
+        ctest --no-compress-output --test-output-size-passed 100000 --test-output-size-failed 100000 -T Test -j `nproc` --timeout 90 > /dev/null
        zstd Testing/*/Test.xml
        '''
    }
-    xunit tools: [CTest(pattern: 'build/Testing/*/Test.xml')], reduceLog: false, skipPublishingChecks: false
+    xunit tools: [CTest(pattern: 'build/Testing/*/Test.xml')], skipPublishingChecks: false
    minio bucket: 'jenkins', credentialsId: 'jenkins-minio', excludes: '', host: 'minio.weaselab.dev', includes: 'build/Testing/*/Test.xml.zst', targetFolder: '${JOB_NAME}/${BUILD_NUMBER}/${STAGE_NAME}/'
 }

@@ -36,7 +36,7 @@ pipeline {
                sh 'pre-commit run --all-files --show-diff-on-failure'
            }
        }
-        stage('Clang') {
+        stage('64 bit versions') {
            agent {
                dockerfile {
                    args '-v /home/jenkins/ccache:/ccache'
@@ -44,8 +44,7 @@ pipeline {
                }
            }
            steps {
-                CleanBuildAndTest("")
-                recordIssues(tools: [clang()])
+                CleanBuildAndTest("-DCMAKE_CXX_FLAGS=-DUSE_64_BIT=1")
            }
        }
        stage('Debug') {
@@ -70,7 +69,7 @@ pipeline {
                CleanBuildAndTest("-DUSE_SIMD_FALLBACK=ON")
            }
        }
-        stage('Release [gcc]') {
+        stage('Release [clang]') {
            agent {
                dockerfile {
                    args '-v /home/jenkins/ccache:/ccache'
@@ -78,8 +77,8 @@ pipeline {
                }
            }
            steps {
-                CleanBuildAndTest("-DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ -DCMAKE_CXX_FLAGS=-DNVALGRIND")
-                recordIssues(tools: [gcc()])
+                CleanBuildAndTest("-DCMAKE_CXX_FLAGS=-DNVALGRIND")
+                recordIssues(tools: [clang()])
                sh '''
                cd build
                cpack -G DEB
@@ -92,7 +91,19 @@ pipeline {
                minio bucket: 'jenkins', credentialsId: 'jenkins-minio', excludes: '', host: 'minio.weaselab.dev', includes: 'build/*.deb,build/*.rpm,paper/*.pdf', targetFolder: '${JOB_NAME}/${BUILD_NUMBER}/${STAGE_NAME}/'
            }
        }
-        stage('Release [gcc,aarch64]') {
+        stage('Release [gcc]') {
+            agent {
+                dockerfile {
+                    args '-v /home/jenkins/ccache:/ccache'
+                    reuseNode true
+                }
+            }
+            steps {
+                CleanBuildAndTest("-DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ -DCMAKE_CXX_FLAGS=-DNVALGRIND")
+                recordIssues(tools: [gcc()])
+            }
+        }
+        stage('Release [clang,aarch64]') {
            agent {
                dockerfile {
                    args '-v /home/jenkins/ccache:/ccache'
@@ -117,15 +128,18 @@ pipeline {
                }
            }
            steps {
-                CleanBuildAndTest("-DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ -DCMAKE_C_FLAGS=--coverage -DCMAKE_CXX_FLAGS=--coverage -DCMAKE_BUILD_TYPE=Debug -DDISABLE_TSAN=ON")
-                sh '''
-                gcovr -f ConflictSet.cpp --cobertura > build/coverage.xml
-                '''
+                script {
+                    gcov_args = "-f ConflictSet.cpp -f LongestCommonPrefix.h -f Metrics.h --gcov-executable 'llvm-cov gcov' --exclude-noncode-lines"
+                }
+                CleanBuildAndTest("-DCMAKE_C_FLAGS=--coverage -DCMAKE_CXX_FLAGS=--coverage -DCMAKE_BUILD_TYPE=Debug -DDISABLE_TSAN=ON")
+                sh """
+                gcovr ${gcov_args} --cobertura > build/coverage.xml
+                """
                recordCoverage qualityGates: [[criticality: 'NOTE', metric: 'MODULE']], tools: [[parser: 'COBERTURA', pattern: 'build/coverage.xml']]
-                sh '''
-                # Suppress again, because we haven't dealt with function multi-versioning for x86 yet
-                # gcovr -f ConflictSet.cpp --fail-under-line 100 > /dev/null
-                '''
+                sh """
+                gcovr ${gcov_args}
+                gcovr ${gcov_args} --fail-under-line 100 > /dev/null
+                """
            }
        }
    }
@@ -0,0 +1,185 @@
+#pragma once
+
+#include <assert.h>
+#include <bit>
+#include <stdint.h>
+#include <string.h>
+
+#ifdef HAS_AVX
+#include <immintrin.h>
+#elif HAS_ARM_NEON
+#include <arm_neon.h>
+#endif
+
+#ifndef __SANITIZE_THREAD__
+#if defined(__has_feature)
+#if __has_feature(thread_sanitizer)
+#define __SANITIZE_THREAD__
+#endif
+#endif
+#endif
+
+#if defined(HAS_AVX) || defined(HAS_ARM_NEON)
+constexpr int kStride = 64;
+#else
+constexpr int kStride = 16;
+#endif
+
+constexpr int kUnrollFactor = 4;
+
+inline bool compareStride(const uint8_t *ap, const uint8_t *bp) {
+#if defined(HAS_ARM_NEON)
+  static_assert(kStride == 64);
+  uint8x16_t x[4]; // GCOVR_EXCL_LINE
+  for (int i = 0; i < 4; ++i) {
+    x[i] = vceqq_u8(vld1q_u8(ap + i * 16), vld1q_u8(bp + i * 16));
+  }
+  auto results = vreinterpretq_u16_u8(
+      vandq_u8(vandq_u8(x[0], x[1]), vandq_u8(x[2], x[3])));
+  bool eq = vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(results, 4)), 0) ==
+            uint64_t(-1);
+#elif defined(HAS_AVX)
+  static_assert(kStride == 64);
+  __m128i x[4]; // GCOVR_EXCL_LINE
+  for (int i = 0; i < 4; ++i) {
+    x[i] = _mm_cmpeq_epi8(_mm_loadu_si128((__m128i *)(ap + i * 16)),
+                          _mm_loadu_si128((__m128i *)(bp + i * 16)));
+  }
+  auto eq =
+      _mm_movemask_epi8(_mm_and_si128(_mm_and_si128(x[0], x[1]),
+                                      _mm_and_si128(x[2], x[3]))) == 0xffff;
+#else
+  // Hope it gets vectorized
+  auto eq = memcmp(ap, bp, kStride) == 0;
+#endif
+  return eq;
+}
+
+// Precondition: ap[:kStride] != bp[:kStride]
+inline int firstNeqStride(const uint8_t *ap, const uint8_t *bp) {
+#if defined(HAS_AVX)
+  static_assert(kStride == 64);
+  uint64_t c[kStride / 16]; // GCOVR_EXCL_LINE
+  for (int i = 0; i < kStride; i += 16) {
+    const auto a = _mm_loadu_si128((__m128i *)(ap + i));
+    const auto b = _mm_loadu_si128((__m128i *)(bp + i));
+    const auto compared = _mm_cmpeq_epi8(a, b);
+    c[i / 16] = _mm_movemask_epi8(compared) & 0xffff;
+  }
+  return std::countr_zero(~(c[0] | c[1] << 16 | c[2] << 32 | c[3] << 48));
+#elif defined(HAS_ARM_NEON)
+  static_assert(kStride == 64);
+  for (int i = 0; i < kStride; i += 16) {
+    // 0xff for each match
+    uint16x8_t results =
+        vreinterpretq_u16_u8(vceqq_u8(vld1q_u8(ap + i), vld1q_u8(bp + i)));
+    // 0xf for each mismatch
+    uint64_t bitfield =
+        ~vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(results, 4)), 0);
+    if (bitfield) {
+      return i + (std::countr_zero(bitfield) >> 2);
+    }
+  }
+  __builtin_unreachable(); // GCOVR_EXCL_LINE
+#else
+  int i = 0;
+  for (; i < kStride - 1; ++i) {
+    if (*ap++ != *bp++) {
+      break;
+    }
+  }
+  return i;
+#endif
+}
+
+// This gets covered in local development
+// GCOVR_EXCL_START
+#if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
+__attribute__((target("avx512f,avx512bw"))) inline int
+longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
+  int i = 0;
+  int end = cl & ~63;
+  while (i < end) {
+    const uint64_t eq =
+        _mm512_cmpeq_epi8_mask(_mm512_loadu_epi8(ap), _mm512_loadu_epi8(bp));
+    if (eq != uint64_t(-1)) {
+      return i + std::countr_one(eq);
+    }
+    i += 64;
+    ap += 64;
+    bp += 64;
+  }
+  if (i < cl) {
+    const uint64_t mask = (uint64_t(1) << (cl - i)) - 1;
+    const uint64_t eq = _mm512_cmpeq_epi8_mask(
+        _mm512_maskz_loadu_epi8(mask, ap), _mm512_maskz_loadu_epi8(mask, bp));
+    return i + std::countr_one(eq & mask);
+  }
+  assert(i == cl);
+  return i;
+}
+__attribute__((target("default")))
+#endif
+// GCOVR_EXCL_STOP
+
+inline int
+longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
+  if (!(cl >= 0)) {
+    __builtin_unreachable(); // GCOVR_EXCL_LINE
+  }
+
+  int i = 0;
+  int end; // GCOVR_EXCL_LINE
+
+  // kStride * kUnrollCount at a time
+  end = cl & ~(kStride * kUnrollFactor - 1);
+  while (i < end) {
+    for (int j = 0; j < kUnrollFactor; ++j) {
+      if (!compareStride(ap, bp)) {
+        return i + firstNeqStride(ap, bp);
+      }
+      i += kStride;
+      ap += kStride;
+      bp += kStride;
+    }
+  }
+
+  // kStride at a time
+  end = cl & ~(kStride - 1);
+  while (i < end) {
+    if (!compareStride(ap, bp)) {
+      return i + firstNeqStride(ap, bp);
+    }
+    i += kStride;
+    ap += kStride;
+    bp += kStride;
+  }
+
+  // word at a time
+  end = cl & ~(sizeof(uint64_t) - 1);
+  while (i < end) {
+    uint64_t a; // GCOVR_EXCL_LINE
+    uint64_t b; // GCOVR_EXCL_LINE
+    memcpy(&a, ap, 8);
+    memcpy(&b, bp, 8);
+    const auto mismatched = a ^ b;
+    if (mismatched) {
+      return i + std::countr_zero(mismatched) / 8;
+    }
+    i += 8;
+    ap += 8;
+    bp += 8;
+  }
+
+  // byte at a time
+  while (i < cl) {
+    if (*ap != *bp) {
+      break;
+    }
+    ++ap;
+    ++bp;
+    ++i;
+  }
+
+  return i;
+}
@@ -0,0 +1,64 @@
+#pragma once
+
+#include "ConflictSet.h"
+#include "Internal.h"
+
+#include <assert.h>
+#include <atomic>
+#include <tuple>
+
+struct Metric {
+  Metric *prev;
+  const char *name;
+  const char *help;
+  weaselab::ConflictSet::MetricsV1::Type type;
+  std::atomic<int64_t> value;
+
+protected:
+  Metric(Metric *&metricList, int &metricsCount, const char *name,
+         const char *help, weaselab::ConflictSet::MetricsV1::Type type)
+      : prev(std::exchange(metricList, this)), name(name), help(help),
+        type(type), value(0) {
+    ++metricsCount;
+  }
+};
+
+struct Gauge : private Metric {
+  Gauge(Metric *&metricList, int &metricsCount, const char *name,
+        const char *help)
+      : Metric(metricList, metricsCount, name, help,
+               weaselab::ConflictSet::MetricsV1::Gauge) {}
+
+  void set(int64_t value) {
+    this->value.store(value, std::memory_order_relaxed);
+  }
+};
+
+struct Counter : private Metric {
+  Counter(Metric *&metricList, int &metricsCount, const char *name,
+          const char *help)
+      : Metric(metricList, metricsCount, name, help,
+               weaselab::ConflictSet::MetricsV1::Counter) {}
+  // Expensive. Accumulate locally and then call add instead of repeatedly
+  // calling add.
+  void add(int64_t value) {
+    assert(value >= 0);
+    static_assert(std::atomic<int64_t>::is_always_lock_free);
+    this->value.fetch_add(value, std::memory_order_relaxed);
+  }
+};
+
+inline weaselab::ConflictSet::MetricsV1 *initMetrics(Metric *metricsList,
+                                                     int metricsCount) {
+  weaselab::ConflictSet::MetricsV1 *metrics =
+      (weaselab::ConflictSet::MetricsV1 *)safe_malloc(metricsCount *
+                                                      sizeof(metrics[0]));
+  for (auto [i, m] = std::make_tuple(metricsCount - 1, metricsList); i >= 0;
+       --i, m = m->prev) {
+    metrics[i].name = m->name;
+    metrics[i].help = m->help;
+    metrics[i].p = m;
+    metrics[i].type = m->type;
+  }
+  return metrics;
+}
@@ -1,86 +1,40 @@
 A data structure for optimistic concurrency control on ranges of bitwise-lexicographically-ordered keys.

-Intended to replace FoundationDB's skip list.
+Intended as an alternative to FoundationDB's skip list.

-Hardware for all benchmarks is a mac m1 2020.
+Hardware for all benchmarks is an AMD Ryzen 9 7900 with (2x32GB) 5600MT/s CL28-34-34-89 1.35V RAM.

-# FoundationDB's benchmark
+Compiler is `Ubuntu clang version 20.0.0 (++20241029082144+7544d3af0e28-1~exp1~20241029082307.506)`.
+
+# Microbenchmark

 ## Skip list

-```
-New conflict set: 1.957 sec
-                  0.639 Mtransactions/sec
-                  2.555 Mkeys/sec
-Detect only:      1.845 sec
-                  0.678 Mtransactions/sec
-                  2.710 Mkeys/sec
-Skiplist only:    1.263 sec
-                  0.990 Mtransactions/sec
-                  3.960 Mkeys/sec
-Performance counters:
-               Build: 0.0546
-                 Add: 0.0563
-              Detect: 1.84
-              D.Sort: 0.412
-           D.Combine: 0.0141
-         D.CheckRead: 0.671
-   D.CheckIntraBatch: 0.0068
-        D.MergeWrite: 0.592
-      D.RemoveBefore: 0.146
-```
+|               ns/op |                op/s |    err% |          ins/op |          cyc/op |    IPC |         bra/op |   miss% |     total | benchmark
+|--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
+|              159.65 |        6,263,576.52 |    1.6% |        2,972.36 |          820.37 |  3.623 |         504.59 |    0.0% |      0.01 | `point reads`
+|              156.32 |        6,397,320.65 |    0.7% |        2,913.62 |          806.87 |  3.611 |         490.19 |    0.0% |      0.01 | `prefix reads`
+|              229.18 |        4,363,293.65 |    1.2% |        3,541.05 |        1,219.75 |  2.903 |         629.33 |    0.0% |      0.01 | `range reads`
+|              363.37 |        2,752,026.30 |    0.3% |        5,273.63 |        1,951.54 |  2.702 |         851.66 |    1.7% |      0.01 | `point writes`
+|              364.99 |        2,739,787.02 |    0.3% |        5,250.92 |        1,958.54 |  2.681 |         839.24 |    1.7% |      0.01 | `prefix writes`
+|              242.26 |        4,127,796.58 |    2.9% |        3,117.33 |        1,304.41 |  2.390 |         541.07 |    2.8% |      0.02 | `range writes`
+|              562.48 |        1,777,855.27 |    0.8% |        7,305.21 |        3,034.34 |  2.408 |       1,329.30 |    1.3% |      0.01 | `monotonic increasing point writes`
+|          122,688.57 |            8,150.72 |    0.7% |      798,766.00 |      666,842.00 |  1.198 |     144,584.50 |    0.1% |      0.01 | `worst case for radix tree`
+|               41.71 |       23,976,459.34 |    1.7% |          885.00 |          219.17 |  4.038 |         132.00 |    0.0% |      0.01 | `create and destroy`

 ## Radix tree (this implementation)

-```
-New conflict set: 1.366 sec
-                  0.915 Mtransactions/sec
-                  3.660 Mkeys/sec
-Detect only:      1.248 sec
-                  1.002 Mtransactions/sec
-                  4.007 Mkeys/sec
-Skiplist only:    0.573 sec
-                  2.182 Mtransactions/sec
-                  8.730 Mkeys/sec
-Performance counters:
-               Build: 0.0594
-                 Add: 0.0572
-              Detect: 1.25
-              D.Sort: 0.418
-           D.Combine: 0.0149
-         D.CheckRead: 0.232
-   D.CheckIntraBatch: 0.0067
-        D.MergeWrite: 0.341
-      D.RemoveBefore: 0.232
-```
-
-# Our benchmark
-
-## Skip list
-
-|               ns/op |                op/s |    err% |     total | benchmark
-|--------------------:|--------------------:|--------:|----------:|:----------
-|              245.99 |        4,065,232.81 |    0.3% |      0.01 | `point reads`
-|              265.93 |        3,760,430.49 |    0.2% |      0.01 | `prefix reads`
-|              485.30 |        2,060,569.50 |    0.2% |      0.01 | `range reads`
-|              449.60 |        2,224,195.17 |    0.4% |      0.01 | `point writes`
-|              441.76 |        2,263,688.18 |    1.1% |      0.01 | `prefix writes`
-|              245.42 |        4,074,647.54 |    2.4% |      0.02 | `range writes`
-|              572.80 |        1,745,810.06 |    1.3% |      0.01 | `monotonic increasing point writes`
-|          154,819.33 |            6,459.14 |    0.9% |      0.01 | `worst case for radix tree`
-
-## Radix tree (this implementation)
-
-|               ns/op |                op/s |    err% |     total | benchmark
-|--------------------:|--------------------:|--------:|----------:|:----------
-|               19.17 |       52,163,930.66 |    0.1% |      0.01 | `point reads`
-|               23.68 |       42,224,388.21 |    0.7% |      0.01 | `prefix reads`
-|               63.30 |       15,797,506.06 |    0.9% |      0.01 | `range reads`
-|               29.66 |       33,720,994.74 |    0.3% |      0.01 | `point writes`
-|               43.50 |       22,987,781.25 |    1.0% |      0.01 | `prefix writes`
-|               50.00 |       20,000,000.00 |    0.8% |      0.01 | `range writes`
-|              103.25 |        9,684,786.47 |    2.9% |      0.01 | `monotonic increasing point writes`
-|        1,181,500.00 |              846.38 |    2.3% |      0.01 | `worst case for radix tree`
+|               ns/op |                op/s |    err% |          ins/op |          cyc/op |    IPC |         bra/op |   miss% |     total | benchmark
+|--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
+|               12.63 |       79,186,868.18 |    1.4% |          241.61 |           64.76 |  3.731 |          31.64 |    0.8% |      0.01 | `point reads`
+|               14.48 |       69,078,073.40 |    0.3% |          292.42 |           74.69 |  3.915 |          41.49 |    0.5% |      0.01 | `prefix reads`
+|               34.37 |       29,094,694.11 |    0.2% |          759.53 |          179.77 |  4.225 |         100.38 |    0.2% |      0.01 | `range reads`
+|               19.34 |       51,713,896.36 |    0.7% |          369.70 |          101.81 |  3.631 |          47.88 |    0.6% |      0.01 | `point writes`
+|               39.16 |       25,538,968.61 |    0.2% |          653.16 |          206.77 |  3.159 |          89.62 |    0.8% |      0.01 | `prefix writes`
+|               40.58 |       24,642,681.12 |    4.7% |          718.44 |          216.44 |  3.319 |          99.28 |    0.6% |      0.01 | `range writes`
+|               78.77 |       12,694,520.69 |    3.8% |        1,395.55 |          421.73 |  3.309 |         249.81 |    0.1% |      0.01 | `monotonic increasing point writes`
+|          287,760.50 |            3,475.11 |    0.5% |    3,929,266.50 |    1,550,225.50 |  2.535 |     639,064.00 |    0.0% |      0.01 | `worst case for radix tree`
+|              104.76 |        9,545,250.65 |    3.1% |        2,000.00 |          552.82 |  3.618 |         342.00 |    0.0% |      0.01 | `create and destroy`

 # "Real data" test

@@ -89,13 +43,13 @@ Point queries only, best of three runs. Gc ratio is the ratio of time spent doin
 ## skip list

 ```
-Check: 11.3385 seconds, 329.718 MB/s, Add: 5.35612 seconds, 131.072 MB/s, Gc ratio: 45.7173%
+Check: 4.39702 seconds, 370.83 MB/s, Add: 4.50025 seconds, 124.583 MB/s, Gc ratio: 29.1333%, Peak idle memory: 5.51852e+06
 ```

 ## radix tree

 ```
-Check: 2.48583 seconds, 1503.93 MB/s, Add: 2.12768 seconds, 329.954 MB/s, Gc ratio: 41.7943%
+Check: 0.987757 seconds, 1650.76 MB/s, Add: 1.24815 seconds, 449.186 MB/s, Gc ratio: 41.4675%, Peak idle memory: 2.02872e+06
 ```

 ## hash table
@@ -103,5 +57,5 @@ Check: 2.48583 seconds, 1503.93 MB/s, Add: 2.12768 seconds, 329.954 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: 1.83386 seconds, 2038.6 MB/s, Add: 0.601411 seconds, 1167.32 MB/s, Gc ratio: 48.9776%
+Check: 0.84256 seconds, 1935.23 MB/s, Add: 0.697204 seconds, 804.146 MB/s, Gc ratio: 35.4091%
 ```
@@ -0,0 +1,383 @@
+#include <algorithm>
+#include <atomic>
+#include <cstdint>
+#include <cstdlib>
+#include <errno.h>
+#include <netdb.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <string>
+#include <string_view>
+#include <sys/ioctl.h>
+#include <sys/resource.h>
+#include <sys/socket.h>
+#include <sys/types.h>
+#include <sys/uio.h>
+#include <thread>
+#include <unistd.h>
+#include <utility>
+#include <vector>
+
+#include "ConflictSet.h"
+#include "Internal.h"
+#include "third_party/nadeau.h"
+
+std::atomic<int64_t> transactions;
+
+int64_t safeUnaryMinus(int64_t x) {
+  return x == std::numeric_limits<int64_t>::min() ? x : -x;
+}
+
+void tupleAppend(std::string &output, int64_t value) {
+  if (value == 0) {
+    output.push_back(0x14);
+    return;
+  }
+  uint32_t size = 8 - __builtin_clrsbll(value) / 8;
+  int typeCode = 0x14 + (value < 0 ? -1 : 1) * size;
+  output.push_back(typeCode);
+  if (value < 0) {
+    value = ~safeUnaryMinus(value);
+  }
+  uint64_t swap = __builtin_bswap64(value);
+  output.insert(output.end(), (uint8_t *)&swap + 8 - size,
+                (uint8_t *)&swap + 8);
+}
+
+void tupleAppend(std::string &output, std::string_view value) {
+  output.push_back('\x02');
+  for (auto c : value) {
+    if (c == '\x00') {
+      output.push_back('\x00');
+      output.push_back('\xff');
+    } else {
+      output.push_back(c);
+    }
+  }
+  output.push_back('\x00');
+}
+
+template <class... Ts> std::string tupleKey(const Ts &...ts) {
+  std::string result;
+  (tupleAppend(result, ts), ...);
+  return result;
+}
+
+constexpr int kTotalKeyRange = 1'000'000'000;
+constexpr int kWindowSize = 1'000'000;
+constexpr int kNumKeys = 10;
+
+void workload(weaselab::ConflictSet *cs) {
+  int64_t version = kWindowSize;
+  for (;; transactions.fetch_add(1, std::memory_order_relaxed)) {
+    std::vector<int64_t> keyIndices;
+    for (int i = 0; i < kNumKeys; ++i) {
+      keyIndices.push_back(rand() % kTotalKeyRange);
+    }
+    std::sort(keyIndices.begin(), keyIndices.end());
+    std::vector<std::string> keys;
+    constexpr std::string_view fullString =
+        "this is a string, where a prefix of it is used as an element of the "
+        "tuple forming the key";
+    for (int i = 0; i < kNumKeys; ++i) {
+      keys.push_back(
+          tupleKey(0x100, keyIndices[i] / fullString.size(),
+                   fullString.substr(0, keyIndices[i] % fullString.size())));
+      // printf("%s\n", printable(keys.back()).c_str());
+    }
+
+    std::vector<weaselab::ConflictSet::ReadRange> reads;
+    std::vector<weaselab::ConflictSet::WriteRange> writes;
+    std::vector<weaselab::ConflictSet::Result> results;
+    for (int i = 0; i < kNumKeys; ++i) {
+      writes.push_back({{(const uint8_t *)keys[i].data(), int(keys[i].size())},
+                        {nullptr, 0}});
+      reads.push_back({{(const uint8_t *)keys[i].data(), int(keys[i].size())},
+                       {nullptr, 0},
+                       version - kWindowSize});
+    }
+    results.resize(reads.size());
+
+    cs->check(reads.data(), results.data(), reads.size());
+    bool ok = true;
+    for (auto result : results) {
+      ok &= result == weaselab::ConflictSet::Commit;
+    }
+    cs->addWrites(writes.data(), ok ? writes.size() : 0, version);
+    cs->setOldestVersion(version - kWindowSize);
+    ++version;
+  }
+}
+
+// Adapted from getaddrinfo man page
+int getListenFd(const char *node, const char *service) {
+
+  struct addrinfo hints;
+  struct addrinfo *result, *rp;
+  int sfd, s;
+
+  memset(&hints, 0, sizeof(hints));
+  hints.ai_family = AF_UNSPEC;     /* Allow IPv4 or IPv6 */
+  hints.ai_socktype = SOCK_STREAM; /* stream socket */
+  hints.ai_flags = AI_PASSIVE;     /* For wildcard IP address */
+  hints.ai_protocol = 0;           /* Any protocol */
+  hints.ai_canonname = nullptr;
+  hints.ai_addr = nullptr;
+  hints.ai_next = nullptr;
+
+  s = getaddrinfo(node, service, &hints, &result);
+  if (s != 0) {
+    fprintf(stderr, "getaddrinfo: %s\n", gai_strerror(s));
+    abort();
+  }
+
+  /* getaddrinfo() returns a list of address structures.
+     Try each address until we successfully bind(2).
+     If socket(2) (or bind(2)) fails, we (close the socket
+     and) try the next address. */
+
+  for (rp = result; rp != nullptr; rp = rp->ai_next) {
+    sfd = socket(rp->ai_family, rp->ai_socktype, rp->ai_protocol);
+    if (sfd == -1) {
+      continue;
+    }
+
+    int val = 1;
+    setsockopt(sfd, SOL_SOCKET, SO_REUSEADDR, &val, sizeof(val));
+
+    if (bind(sfd, rp->ai_addr, rp->ai_addrlen) == 0) {
+      break; /* Success */
+    }
+
+    close(sfd);
+  }
+
+  freeaddrinfo(result); /* No longer needed */
+
+  if (rp == nullptr) { /* No address succeeded */
+    fprintf(stderr, "Could not bind\n");
+    abort();
+  }
+
+  int rv = listen(sfd, SOMAXCONN);
+  if (rv) {
+    perror("listen()");
+    abort();
+  }
+
+  return sfd;
+}
+
+// HTTP response
+//
+std::string_view part1 =
+    "HTTP/1.1 200 OK \r\nContent-type: text/plain; version=0.0.4; "
+    "charset=utf-8; escaping=values\r\nContent-Length: ";
+// Decimal content length
+std::string_view part2 = "\r\n\r\n";
+// Body
+
+double toSeconds(timeval t) {
+  return double(t.tv_sec) + double(t.tv_usec) * 1e-6;
+}
+
+#ifdef __linux__
+#include <linux/perf_event.h>
+struct PerfCounter {
+  PerfCounter(int type, int config, const std::string &labels = {},
+              int groupLeaderFd = -1)
+      : labels(labels) {
+    struct perf_event_attr pe;
+
+    memset(&pe, 0, sizeof(pe));
+    pe.type = type;
+    pe.size = sizeof(pe);
+    pe.config = config;
+    pe.inherit = 1;
+    pe.exclude_kernel = 1;
+    pe.exclude_hv = 1;
+
+    fd = perf_event_open(&pe, 0, -1, groupLeaderFd, 0);
+    if (fd < 0 && errno != ENOENT && errno != EINVAL) {
+      perror(labels.c_str());
+    }
+  }
+
+  int64_t total() const {
+    int64_t count;
+    if (read(fd, &count, sizeof(count)) != sizeof(count)) {
+      perror("read instructions from perf");
+      abort();
+    }
+    return count;
+  }
+
+  PerfCounter(PerfCounter &&other)
+      : fd(std::exchange(other.fd, -1)), labels(std::move(other.labels)) {}
+  PerfCounter &operator=(PerfCounter &&other) {
+    fd = std::exchange(other.fd, -1);
+    labels = std::move(other.labels);
+    return *this;
+  }
+
+  ~PerfCounter() {
+    if (fd >= 0) {
+      close(fd);
+    }
+  }
+
+  bool ok() const { return fd >= 0; }
+  const std::string &getLabels() const { return labels; }
+  int getFd() const { return fd; }
+
+private:
+  int fd;
+  std::string labels;
+  static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
+                              int cpu, int group_fd, unsigned long flags) {
+    int ret;
+
+    ret = syscall(SYS_perf_event_open, hw_event, pid, cpu, group_fd, flags);
+    return ret;
+  }
+};
+#endif
+
+int main(int argc, char **argv) {
+  if (argc != 3) {
+    goto fail;
+  }
+  {
+    int listenFd = getListenFd(argv[1], argv[2]);
+
+    weaselab::ConflictSet cs{0};
+    weaselab::ConflictSet::MetricsV1 *metrics;
+    int metricsCount;
+    cs.getMetricsV1(&metrics, &metricsCount);
+
+#ifdef __linux__
+    PerfCounter instructions{PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS};
+    PerfCounter cycles{PERF_TYPE_HARDWARE, PERF_COUNT_HW_CPU_CYCLES, "",
+                       instructions.getFd()};
+
+    std::vector<PerfCounter> cacheCounters;
+    for (auto [id, idStr] : std::initializer_list<std::pair<int, std::string>>{
+             {PERF_COUNT_HW_CACHE_L1D, "l1d"},
+             {PERF_COUNT_HW_CACHE_L1I, "l1i"},
+             {PERF_COUNT_HW_CACHE_LL, "ll"},
+             {PERF_COUNT_HW_CACHE_DTLB, "dtlb"},
+             {PERF_COUNT_HW_CACHE_ITLB, "itlb"},
+             {PERF_COUNT_HW_CACHE_BPU, "bpu"},
+             {PERF_COUNT_HW_CACHE_NODE, "node"},
+         }) {
+      for (auto [op, opStr] :
+           std::initializer_list<std::pair<int, std::string>>{
+               {PERF_COUNT_HW_CACHE_OP_READ, "read"},
+               {PERF_COUNT_HW_CACHE_OP_WRITE, "write"},
+               {PERF_COUNT_HW_CACHE_OP_PREFETCH, "prefetch"},
+           }) {
+        int groupLeaderFd = -1;
+        for (auto [result, resultStr] :
+             std::initializer_list<std::pair<int, std::string>>{
+                 {PERF_COUNT_HW_CACHE_RESULT_MISS, "miss"},
+                 {PERF_COUNT_HW_CACHE_RESULT_ACCESS, "access"},
+             }) {
+          auto labels = "{id=\"" + idStr + "\", op=\"" + opStr +
+                        "\", result=\"" + resultStr + "\"}";
+          cacheCounters.emplace_back(PERF_TYPE_HW_CACHE,
+                                     id | (op << 8) | (result << 16), labels,
+                                     groupLeaderFd);
+          if (!cacheCounters.back().ok()) {
+            cacheCounters.pop_back();
+          } else {
+            if (groupLeaderFd == -1) {
+              groupLeaderFd = cacheCounters.back().getFd();
+            }
+          }
+        }
+      }
+    }
+#endif
+
+    auto w = std::thread{workload, &cs};
+
+    for (;;) {
+      struct sockaddr_storage peer_addr = {};
+      socklen_t peer_addr_len = sizeof(peer_addr);
+      const int connfd =
+          accept(listenFd, (struct sockaddr *)&peer_addr, &peer_addr_len);
+
+      std::string body;
+
+      rusage r;
+      getrusage(RUSAGE_SELF, &r);
+      body += "# HELP process_cpu_seconds_total Total user and system CPU time "
+              "spent in seconds.\n# TYPE process_cpu_seconds_total counter\n"
+              "process_cpu_seconds_total ";
+      body += std::to_string(toSeconds(r.ru_utime) + toSeconds(r.ru_stime));
+      body += "\n";
+      body += "# HELP process_resident_memory_bytes Resident memory size in "
+              "bytes.\n# TYPE process_resident_memory_bytes gauge\n"
+              "process_resident_memory_bytes ";
+      body += std::to_string(getCurrentRSS());
+      body += "\n";
+      body += "# HELP transactions_total Total number of transactions\n"
+              "# TYPE transactions_total counter\n"
+              "transactions_total ";
+      body += std::to_string(transactions.load(std::memory_order_relaxed));
+      body += "\n";
+#ifdef __linux__
+      body += "# HELP instructions_total Total number of instructions\n"
+              "# TYPE instructions_total counter\n"
+              "instructions_total ";
+      body += std::to_string(instructions.total());
+      body += "\n";
+      body += "# HELP cycles_total Total number of cycles\n"
+              "# TYPE cycles_total counter\n"
+              "cycles_total ";
+      body += std::to_string(cycles.total());
+      body += "\n";
+      body += "# HELP cache_event_total Total number of cache events\n"
+              "# TYPE cache_event_total counter\n";
+      for (const auto &counter : cacheCounters) {
+        body += "cache_event_total" + counter.getLabels() + " " +
+                std::to_string(counter.total()) + "\n";
+      }
+#endif
+
+      for (int i = 0; i < metricsCount; ++i) {
+        body += "# HELP ";
+        body += metrics[i].name;
+        body += " ";
+        body += metrics[i].help;
+        body += "\n";
+        body += "# TYPE ";
+        body += metrics[i].name;
+        body += " ";
+        body += metrics[i].type == metrics[i].Counter ? "counter" : "gauge";
+        body += "\n";
+        body += metrics[i].name;
+        body += " ";
+        body += std::to_string(metrics[i].getValue());
+        body += "\n";
+      }
+
+      auto len = std::to_string(body.size());
+      iovec iov[] = {
+          {(void *)part1.data(), part1.size()},
+          {(void *)len.data(), len.size()},
+          {(void *)part2.data(), part2.size()},
+          {(void *)body.data(), body.size()},
+      };
+      int written;
+      do {
+        written = writev(connfd, iov, sizeof(iov) / sizeof(iov[0]));
+      } while (written < 0 && errno == EINTR);
+      close(connfd);
+    }
+  }
+fail:
+  fprintf(stderr, "Expected ./%s <host> <port>\n", argv[0]);
+  return 1;
+}
@@ -22,9 +22,11 @@

 #include "ConflictSet.h"
 #include "Internal.h"
+#include "Metrics.h"

 #include <algorithm>
 #include <span>
+#include <vector>

 std::span<const uint8_t> keyAfter(Arena &arena, std::span<const uint8_t> key) {
  auto result =
@@ -115,15 +117,6 @@ bool operator==(const KeyInfo &lhs, const KeyInfo &rhs) {
  return !(lhs < rhs || rhs < lhs);
 }

-void swapSort(std::vector<KeyInfo> &points, int a, int b) {
-  if (points[b] < points[a]) {
-    KeyInfo temp;
-    temp = points[a];
-    points[a] = points[b];
-    points[b] = temp;
-  }
-}
-
 struct SortTask {
  int begin;
  int size;
@@ -183,13 +176,6 @@ void sortPoints(std::vector<KeyInfo> &points) {
  }
 }

-static thread_local uint32_t g_seed = 0;
-
-static inline int skfastrand() {
-  g_seed = g_seed * 1664525L + 1013904223L;
-  return g_seed;
-}
-
 static int compare(const StringRef &a, const StringRef &b) {
  int c = memcmp(a.data(), b.data(), std::min(a.size(), b.size()));
  if (c < 0)
@@ -215,20 +201,24 @@ struct ReadConflictRange {
  }
 };

+static constexpr int MaxLevels = 26;
+
+struct RandomLevel {
+  explicit RandomLevel(uint32_t seed) : seed(seed) {}
+
+  int next() {
+    int result = __builtin_clz(seed | (uint32_t(-1) >> (MaxLevels - 1)));
+    seed = seed * 1664525L + 1013904223L;
+    return result;
+  }
+
+private:
+  uint32_t seed;
+};
+
 class SkipList {
 private:
-  static constexpr int MaxLevels = 26;
-
-  int randomLevel() const {
-    uint32_t i = uint32_t(skfastrand()) >> (32 - (MaxLevels - 1));
-    int level = 0;
-    while (i & 1) {
-      i >>= 1;
-      level++;
-    }
-    assert(level < MaxLevels);
-    return level;
-  }
+  RandomLevel randomLevel{0};

  // Represent a node in the SkipList. The node has multiple (i.e., level)
  // pointers to other nodes, and keeps a record of the max versions for each
@@ -426,27 +416,33 @@ public:
  }
  void swap(SkipList &other) { std::swap(header, other.header); }

-  void addConflictRanges(const Finger *fingers, int rangeCount,
-                         Version version) {
+  // Returns the change in the number of entries
+  int64_t addConflictRanges(const Finger *fingers, int rangeCount,
+                            Version version) {
+    int64_t result = rangeCount;
    for (int r = rangeCount - 1; r >= 0; r--) {
      const Finger &startF = fingers[r * 2];
      const Finger &endF = fingers[r * 2 + 1];

-      if (endF.found() == nullptr)
+      if (endF.found() == nullptr) {
+        ++result;
        insert(endF, endF.finger[0]->getMaxVersion(0));
+      }

-      remove(startF, endF);
+      result -= remove(startF, endF);
      insert(startF, version);
    }
+    return result;
  }

-  void detectConflicts(ReadConflictRange *ranges, int count,
-                       ConflictSet::Result *transactionConflictStatus) const {
+  // Return number of iterations of main loop
+  int detectConflicts(ReadConflictRange *ranges, int count,
+                      ConflictSet::Result *transactionConflictStatus) const {
    const int M = 16;
    int nextJob[M];
    CheckMax inProgress[M];
    if (!count)
-      return;
+      return 0;

    int started = std::min(M, count);
    for (int i = 0; i < started; i++) {
@@ -457,8 +453,9 @@ public:

    int prevJob = started - 1;
    int job = 0;
+    int iters = 0;
    // vtune: 340 parts
-    while (true) {
+    for (;; ++iters) {
      if (inProgress[job].advance()) {
        if (started == count) {
          if (prevJob == job)
@@ -474,6 +471,7 @@ public:
      prevJob = job;
      job = nextJob[job];
    }
+    return iters;
  }

  void find(const StringRef *values, Finger *results, int *temp, int count) {
@@ -567,9 +565,10 @@ public:
  }

 private:
-  void remove(const Finger &start, const Finger &end) {
+  // Returns the number of entries removed
+  int64_t remove(const Finger &start, const Finger &end) {
    if (start.finger[0] == end.finger[0])
-      return;
+      return 0;

    Node *x = start.finger[0]->getNext(0);

@@ -578,17 +577,20 @@ private:
      if (start.finger[i] != end.finger[i])
        start.finger[i]->setNext(i, end.finger[i]->getNext(i));

+    int64_t result = 0;
    while (true) {
      Node *next = x->getNext(0);
      x->destroy();
+      ++result;
      if (x == end.finger[0])
        break;
      x = next;
    }
+    return result;
  }

  void insert(const Finger &f, Version version) {
-    int level = randomLevel();
+    int level = randomLevel.next();
    // std::cout << std::string((const char*)value,length) << " level: " <<
    // level << std::endl;
    Node *x = Node::create(f.value, level);
@@ -704,17 +706,27 @@ private:
  };
 };

-struct SkipListConflictSet {};
+struct ReadContext {
+  int64_t commits_accum = 0;
+  int64_t conflicts_accum = 0;
+  int64_t too_olds_accum = 0;
+  int64_t check_bytes_accum = 0;
+};

 struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
  Impl(int64_t oldestVersion)
      : oldestVersion(oldestVersion), newestVersion(oldestVersion),
-        skipList(oldestVersion) {}
+        skipList(oldestVersion) {
+    metrics = initMetrics(metricsList, metricsCount);
+  }
+  ~Impl() { safe_free(metrics, metricsCount * sizeof(metrics[0])); }
  void check(const ConflictSet::ReadRange *reads, ConflictSet::Result *results,
-             int count) const {
+             int count) {
+    ReadContext tls;
    Arena arena;
    auto *ranges = new (arena) ReadConflictRange[count];
    for (int i = 0; i < count; ++i) {
+      tls.check_bytes_accum += reads[i].begin.len + reads[i].end.len;
      ranges[i].begin = {reads[i].begin.p, size_t(reads[i].begin.len)};
      ranges[i].end = reads[i].end.len > 0
                          ? StringRef{reads[i].end.p, size_t(reads[i].end.len)}
@@ -722,13 +734,22 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
      ranges[i].version = reads[i].readVersion;
      results[i] = ConflictSet::Commit;
    }
-    skipList.detectConflicts(ranges, count, results);
+    int iters = skipList.detectConflicts(ranges, count, results);
    for (int i = 0; i < count; ++i) {
      if (reads[i].readVersion < oldestVersion ||
          reads[i].readVersion < newestVersion - 2e9) {
        results[i] = TooOld;
      }
+      tls.commits_accum += results[i] == Commit;
+      tls.conflicts_accum += results[i] == Conflict;
+      tls.too_olds_accum += results[i] == TooOld;
    }
+    range_read_iterations_total.add(iters);
+    range_read_total.add(count);
+    commits_total.add(tls.commits_accum);
+    conflicts_total.add(tls.conflicts_accum);
+    too_olds_total.add(tls.too_olds_accum);
+    check_bytes_total.add(tls.check_bytes_accum);
  }

  void addWrites(const ConflictSet::WriteRange *writes, int count,
@@ -746,7 +767,9 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
          false, true);
    }

-    sortPoints(points);
+    if (!std::is_sorted(points.begin(), points.end())) {
+      sortPoints(points);
+    }

    int activeWriteCount = 0;
    std::vector<std::pair<StringRef, StringRef>> combinedWriteConflictRanges;
@@ -775,31 +798,33 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
    StringRef values[stripeSize];
    int64_t writeVersions[stripeSize / 2];
    int ss = stringCount - (stripes - 1) * stripeSize;
+    int64_t entryDelta = 0;
    for (int s = stripes - 1; s >= 0; s--) {
      for (int i = 0; i * 2 < ss; ++i) {
        const auto &w = combinedWriteConflictRanges[s * stripeSize / 2 + i];
-#if DEBUG_VERBOSE
-        printf("Write begin: %s\n", printable(w.begin).c_str());
-        fflush(stdout);
-#endif
        values[i * 2] = w.first;
        values[i * 2 + 1] = w.second;
-        keyUpdates += 3;
      }
      skipList.find(values, fingers, temp, ss);
-      skipList.addConflictRanges(fingers, ss / 2, writeVersion);
+      entryDelta += skipList.addConflictRanges(fingers, ss / 2, writeVersion);
      ss = stripeSize;
    }
+
+    // Run gc at least 200% the rate we're inserting entries
+    keyUpdates += std::max<int64_t>(entryDelta, 0) * 2;
  }

  void setOldestVersion(int64_t oldestVersion) {
+    // 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;
    std::span<const uint8_t> key = removalKey;
    skipList.find(&key, &finger, &temp, 1);
-    skipList.removeBefore(oldestVersion, finger, std::exchange(keyUpdates, 10));
+    skipList.removeBefore(oldestVersion, finger, std::exchange(keyUpdates, 0));
    removalArena = Arena();
    removalKey = copyToArena(
        removalArena, {finger.getValue().data(), finger.getValue().size()});
@@ -807,8 +832,56 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {

  int64_t totalBytes = 0;

+  MetricsV1 *metrics;
+  int metricsCount = 0;
+  Metric *metricsList = nullptr;
+
+#define GAUGE(name, help)                                                      \
+  Gauge name { metricsList, metricsCount, #name, help }
+#define COUNTER(name, help)                                                    \
+  Counter name { metricsList, metricsCount, #name, help }
+  // ==================== METRICS DEFINITIONS ====================
+  COUNTER(range_read_total, "Total number of range reads checked");
+  COUNTER(range_read_iterations_total,
+          "Total number of iterations of the main loops for range read checks");
+  COUNTER(commits_total,
+          "Total number of checks where the result is \"commit\"");
+  COUNTER(conflicts_total,
+          "Total number of checks where the result is \"conflict\"");
+  COUNTER(too_olds_total,
+          "Total number of checks where the result is \"too old\"");
+  COUNTER(check_bytes_total, "Total number of key bytes checked");
+  GAUGE(memory_bytes, "Total number of bytes in use");
+  COUNTER(nodes_allocated_total,
+          "The total number of physical tree nodes allocated");
+  COUNTER(nodes_released_total,
+          "The total number of physical tree nodes released");
+  COUNTER(insert_iterations_total,
+          "The total number of iterations of the main loop for insertion. "
+          "Includes searches where the entry already existed, and so insertion "
+          "did not take place");
+  COUNTER(entries_inserted_total,
+          "The total number of entries inserted in the tree");
+  COUNTER(entries_erased_total,
+          "The total number of entries erased from the tree");
+  COUNTER(
+      gc_iterations_total,
+      "The total number of iterations of the main loop for garbage collection");
+  COUNTER(write_bytes_total, "Total number of key bytes in calls to addWrites");
+  GAUGE(oldest_version,
+        "The lowest version that doesn't result in \"TooOld\" for checks");
+  GAUGE(newest_version, "The version of the most recent call to addWrites");
+  // ==================== END METRICS DEFINITIONS ====================
+#undef GAUGE
+#undef COUNTER
+
+  void getMetricsV1(MetricsV1 **metrics, int *count) {
+    *metrics = this->metrics;
+    *count = metricsCount;
+  }
+
 private:
-  int64_t keyUpdates = 10;
+  int64_t keyUpdates = 0;
  Arena removalArena;
  std::span<const uint8_t> removalKey;
  int64_t oldestVersion;
@@ -829,6 +902,7 @@ void internal_addWrites(ConflictSet::Impl *impl,
  mallocBytesDelta = 0;
  impl->addWrites(writes, count, writeVersion);
  impl->totalBytes += mallocBytesDelta;
+  impl->memory_bytes.set(impl->totalBytes);
 #if SHOW_MEMORY
  if (impl->totalBytes != mallocBytes) {
    abort();
@@ -840,6 +914,7 @@ void internal_setOldestVersion(ConflictSet::Impl *impl, int64_t oldestVersion) {
  mallocBytesDelta = 0;
  impl->setOldestVersion(oldestVersion);
  impl->totalBytes += mallocBytesDelta;
+  impl->memory_bytes.set(impl->totalBytes);
 #if SHOW_MEMORY
  if (impl->totalBytes != mallocBytes) {
    abort();
@@ -863,12 +938,11 @@ int64_t internal_getBytes(ConflictSet::Impl *impl) { return impl->totalBytes; }

 void internal_getMetricsV1(ConflictSet::Impl *impl,
                           ConflictSet::MetricsV1 **metrics, int *count) {
-  *metrics = nullptr;
-  *count = 0;
+  return impl->getMetricsV1(metrics, count);
 }

 double internal_getMetricValue(const ConflictSet::MetricsV1 *metric) {
-  return 0;
+  return ((Metric *)metric->p)->value.load(std::memory_order_relaxed);
 }

 void ConflictSet::check(const ReadRange *reads, Result *results,
@@ -3,17 +3,68 @@
 #include <fstream>
 #include <sstream>

+#ifndef PERF_TEST
+#define PERF_TEST 0
+#endif
+
 int main(int argc, char **argv) {
  for (int i = 1; i < argc; ++i) {
    std::ifstream t(argv[i], std::ios::binary);
    std::stringstream buffer;
    buffer << t.rdbuf();
    auto str = buffer.str();
-    TestDriver<ConflictSet> driver{(const uint8_t *)str.data(), str.size()};
-    while (!driver.next())
-      ;
-    if (!driver.ok) {
-      abort();
+    Arbitrary arbitrary({(const uint8_t *)str.data(), str.size()});
+    TestDriver<ConflictSet, !PERF_TEST> driver1{arbitrary};
+    TestDriver<ConflictSet, !PERF_TEST> driver2{arbitrary};
+    bool done1 = false;
+    bool done2 = false;
+    for (;;) {
+      if (!done1) {
+        done1 = driver1.next();
+        if (!driver1.ok) {
+          abort();
+        }
+      }
+      if (!done2) {
+        done2 = driver2.next();
+        if (!driver2.ok) {
+          abort();
+        }
+      }
+      if (done1 && done2) {
+        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("");
    }
  }
 }
@@ -1,10 +1,11 @@
-__aarch64_cas8_relax
+__aarch64_ldadd8_relax
 __getauxval@GLIBC_2.17
 __stack_chk_fail@GLIBC_2.17
 __stack_chk_guard@GLIBC_2.17
 abort@GLIBC_2.17
 free@GLIBC_2.17
 malloc@GLIBC_2.17
+memcmp@GLIBC_2.17
 memcpy@GLIBC_2.17
 memmove@GLIBC_2.17
 memset@GLIBC_2.17
@@ -1,7 +1,8 @@
 set(CMAKE_SYSTEM_NAME Linux)
 set(CMAKE_SYSTEM_PROCESSOR aarch64)
-set(CMAKE_C_COMPILER "/usr/bin/aarch64-linux-gnu-gcc")
-set(CMAKE_CXX_COMPILER "/usr/bin/aarch64-linux-gnu-g++")
+set(CMAKE_C_COMPILER "clang;--target=aarch64-linux-gnu")
+set(CMAKE_CXX_COMPILER "clang++;--target=aarch64-linux-gnu")
 set(CMAKE_FIND_ROOT_PATH /usr/aarch64-linux-gnu)
 set(CMAKE_CROSSCOMPILING_EMULATOR "qemu-aarch64;-L;/usr/aarch64-linux-gnu/")
 set(CPACK_DEBIAN_PACKAGE_ARCHITECTURE arm64)
+set(LD_EXE "/usr/bin/aarch64-linux-gnu-ld")
--- a/Show More
+++ b/Show More