Remove unnecessary branch for interleaved range writes

Do it in a simpler, more robust/type safe way

Bench.cpp

@@ -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)
@@ -49,14 +48,16 @@ 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};
 }
 
 void benchConflictSet() {
   ankerl::nanobench::Bench bench;
+  bench.minEpochIterations(10000);
   ConflictSet cs{0};
 
   bench.batch(kOpsPerTx);
@@ -82,14 +83,7 @@ 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 +327,22 @@ 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();
+        }
       }
     }
   });

CMakeLists.txt

@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.18)
 project(
   conflict-set
-  VERSION 0.0.10
+  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"
@@ -32,13 +32,33 @@ if(NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
 endif()
 
 add_compile_options(
+  -Werror=switch-enum
+  -Wswitch-enum
+  -Wunused-variable
+  -fPIC
   -fdata-sections
   -ffunction-sections
-  -Wswitch-enum
-  -Werror=switch-enum
-  -fPIC
-  -g
-  -fno-omit-frame-pointer)
+  -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 +69,22 @@ if(HAS_FULL_RELRO)
 endif()
 cmake_pop_check_state()
 
+if(CMAKE_SYSTEM_PROCESSOR STREQUAL aarch64 OR CMAKE_SYSTEM_PROCESSOR STREQUAL
+                                              arm64)
+  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()
@@ -66,33 +102,17 @@ option(DISABLE_TSAN "Disable TSAN" OFF)
 # https://valgrind.org/docs/manual/manual-core-adv.html#manual-core-adv.clientreq
 include_directories(SYSTEM ${CMAKE_CURRENT_SOURCE_DIR}/third_party/valgrind)
 
-add_compile_options($<$<COMPILE_LANGUAGE:CXX>:-Wno-invalid-offsetof>)
-
 if(APPLE)
   add_link_options(-Wl,-dead_strip)
 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 -mavx)
-  check_include_file_cxx("immintrin.h" HAS_AVX)
-  if(HAS_AVX)
+if(USE_SIMD_FALLBACK)
+  add_compile_definitions(USE_SIMD_FALLBACK)
+else()
+  if(CMAKE_SYSTEM_PROCESSOR STREQUAL x86_64)
     add_compile_options(-mavx)
-    add_compile_definitions(HAS_AVX)
-  endif()
-  cmake_pop_check_state()
-
-  check_include_file_cxx("arm_neon.h" HAS_ARM_NEON)
-  if(HAS_ARM_NEON)
-    add_compile_definitions(HAS_ARM_NEON)
   endif()
 endif()
 
@@ -103,12 +123,23 @@ 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()
 
@@ -118,19 +149,13 @@ 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
@@ -146,6 +171,8 @@ 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
@@ -183,16 +210,20 @@ 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)
 
-  foreach(TEST ${CORPUS_TESTS})
-    get_filename_component(hash ${TEST} NAME)
-    add_test(NAME skip_list_${hash} COMMAND driver_skip_list ${TEST})
-  endforeach()
+  # 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})
+    endforeach()
+  endif()
 
   # ad hoc testing
   add_executable(conflict_set_main ConflictSet.cpp)
   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
@@ -253,6 +284,19 @@ 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)
@@ -273,16 +317,17 @@ 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
@@ -330,9 +375,18 @@ 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})
@@ -347,6 +401,14 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
   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

Dockerfile

@@ -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++

InterleavingTest.cpp

@@ -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;
+  }
+}

Internal.h

@@ -18,18 +18,45 @@ using namespace weaselab;
 #include <span>
 #include <string>
 #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 +66,7 @@ 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 +74,18 @@ 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 +311,16 @@ 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; }
@@ -319,23 +367,6 @@ template <class T, class C = std::less<T>> auto set(Arena &arena) {
   return Set<T, C>(ArenaAlloc<T>(&arena));
 }
 
-template <class T> struct MyHash;
-
-template <class T> struct MyHash<T *> {
-  size_t operator()(const T *t) const noexcept {
-    size_t result;
-    memcpy(&result, &t, sizeof(result));
-    return result;
-  }
-};
-
-template <class T>
-using HashSet =
-    std::unordered_set<T, MyHash<T>, std::equal_to<T>, ArenaAlloc<T>>;
-template <class T> auto hashSet(Arena &arena) {
-  return HashSet<T>(ArenaAlloc<T>(&arena));
-}
-
 template <class T, class U>
 bool operator==(const ArenaAlloc<T> &lhs, const ArenaAlloc<U> &rhs) {
   return lhs.arena == rhs.arena;
@@ -529,7 +560,7 @@ 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 +591,7 @@ 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()));
 }
 
@@ -668,10 +699,8 @@ struct TestDriver {
             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);
@@ -738,7 +767,10 @@ struct TestDriver {
       fprintf(stderr, "%p Set oldest version: %" PRId64 "\n", this,
               oldestVersion);
 #endif
+      CALLGRIND_START_INSTRUMENTATION;
       cs.setOldestVersion(oldestVersion);
+      CALLGRIND_STOP_INSTRUMENTATION;
+
       if constexpr (kEnableAssertions) {
         refImpl.setOldestVersion(oldestVersion);
       }

Jenkinsfile

@@ -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('gcc') {
+            agent {
+                dockerfile {
+                    args '-v /home/jenkins/ccache:/ccache'
+                    reuseNode true
+                }
+            }
+            steps {
+                CleanBuildAndTest("-DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++")
+                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 '''
-                gcovr -f ConflictSet.cpp
-                gcovr -f ConflictSet.cpp --fail-under-line 100 > /dev/null
-                '''
+                sh """
+                gcovr ${gcov_args}
+                gcovr ${gcov_args} --fail-under-line 100 > /dev/null
+                """
             }
         }
     }

LongestCommonPrefix.h

@@ -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;
+}

Metrics.h

@@ -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;
+}

README.md

@@ -2,7 +2,16 @@ A data structure for optimistic concurrency control on ranges of bitwise-lexicog
 
 Intended as an alternative to FoundationDB's skip list.
 
-Hardware for all benchmarks is an AMD Ryzen 9 7900 with (2x32GB) 5600MT/s CL28-34-34-89 1.35V RAM
+Hardware for all benchmarks is an AMD Ryzen 9 7900 with (2x32GB) 5600MT/s CL28-34-34-89 1.35V RAM.
+
+```
+$ clang++ --version
+
+Ubuntu clang version 20.0.0 (++20241120082228+86734c857724-1~exp1~20241120202359.554)
+Target: x86_64-pc-linux-gnu
+Thread model: posix
+InstalledDir: /usr/lib/llvm-20/bin
+```
 
 # Microbenchmark
 
@@ -10,44 +19,45 @@ Hardware for all benchmarks is an AMD Ryzen 9 7900 with (2x32GB) 5600MT/s CL28-3
 
 |               ns/op |                op/s |    err% |          ins/op |          cyc/op |    IPC |         bra/op |   miss% |     total | benchmark
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|              172.03 |        5,812,791.77 |    0.4% |        3,130.62 |          879.00 |  3.562 |         509.23 |    0.0% |      0.01 | `point reads`
-|              167.44 |        5,972,130.71 |    0.2% |        3,065.14 |          862.27 |  3.555 |         494.30 |    0.0% |      0.01 | `prefix reads`
-|              238.77 |        4,188,130.84 |    0.9% |        3,589.93 |        1,259.30 |  2.851 |         637.12 |    0.0% |      0.01 | `range reads`
-|              424.01 |        2,358,426.70 |    0.2% |        5,620.05 |        2,242.35 |  2.506 |         854.80 |    1.7% |      0.01 | `point writes`
-|              418.45 |        2,389,780.56 |    0.4% |        5,525.07 |        2,211.05 |  2.499 |         831.71 |    1.7% |      0.01 | `prefix writes`
-|              254.87 |        3,923,568.88 |    2.6% |        3,187.01 |        1,366.50 |  2.332 |         529.11 |    2.7% |      0.02 | `range writes`
-|              675.96 |        1,479,374.50 |    3.3% |        7,735.41 |        3,468.60 |  2.230 |       1,386.02 |    1.8% |      0.01 | `monotonic increasing point writes`
-|          137,986.20 |            7,247.10 |    0.6% |      789,752.33 |      699,462.00 |  1.129 |     144,824.14 |    0.0% |      0.01 | `worst case for radix tree`
-|               21.63 |       46,231,564.03 |    1.0% |          448.00 |          107.14 |  4.181 |          84.00 |    0.0% |      0.01 | `create and destroy`
+|              161.29 |        6,200,056.17 |    0.1% |        3,014.03 |          831.04 |  3.627 |         504.59 |    0.0% |      1.93 | `point reads`
+|              158.32 |        6,316,160.64 |    0.1% |        2,954.16 |          815.80 |  3.621 |         490.17 |    0.0% |      1.89 | `prefix reads`
+|              237.39 |        4,212,409.50 |    0.2% |        3,592.41 |        1,233.96 |  2.911 |         629.31 |    0.0% |      2.84 | `range reads`
+|              442.11 |        2,261,878.94 |    0.0% |        4,450.57 |        2,314.25 |  1.923 |         707.92 |    2.1% |      5.28 | `point writes`
+|              439.89 |        2,273,308.53 |    0.1% |        4,410.22 |        2,302.29 |  1.916 |         694.74 |    2.1% |      5.25 | `prefix writes`
+|              290.96 |        3,436,936.78 |    0.0% |        2,315.38 |        1,528.68 |  1.515 |         396.69 |    3.3% |      3.49 | `range writes`
+|              476.93 |        2,096,762.02 |    0.6% |        6,999.33 |        2,484.94 |  2.817 |       1,251.73 |    1.3% |      0.06 | `monotonic increasing point writes`
+|          131,736.57 |            7,590.91 |    1.1% |      807,444.50 |      704,941.71 |  1.145 |     144,584.60 |    0.9% |      0.01 | `worst case for radix tree`
+|               45.50 |       21,978,369.95 |    1.1% |          902.00 |          232.36 |  3.882 |         132.00 |    0.0% |      0.01 | `create and destroy`
 
 ## Radix tree (this implementation)
 
+
 |               ns/op |                op/s |    err% |          ins/op |          cyc/op |    IPC |         bra/op |   miss% |     total | benchmark
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|               12.42 |       80,500,398.66 |    0.8% |          180.38 |           61.57 |  2.930 |          41.51 |    0.4% |      0.01 | `point reads`
-|               15.17 |       65,917,580.99 |    0.2% |          279.47 |           74.95 |  3.729 |          55.54 |    0.3% |      0.01 | `prefix reads`
-|               38.16 |       26,202,393.91 |    0.1% |          803.07 |          189.13 |  4.246 |         141.68 |    0.2% |      0.01 | `range reads`
-|               20.20 |       49,504,615.44 |    0.4% |          363.00 |          100.35 |  3.617 |          49.81 |    0.3% |      0.01 | `point writes`
-|               41.99 |       23,816,559.99 |    0.3% |          799.27 |          209.63 |  3.813 |         154.32 |    0.1% |      0.01 | `prefix writes`
-|               46.28 |       21,607,605.88 |    1.5% |          953.79 |          231.47 |  4.121 |         168.34 |    0.0% |      0.01 | `range writes`
-|               80.99 |       12,347,449.98 |    0.9% |        1,501.97 |          406.50 |  3.695 |         281.89 |    0.1% |      0.01 | `monotonic increasing point writes`
-|          318,010.00 |            3,144.56 |    1.0% |    3,994,511.50 |    1,657,831.50 |  2.409 |     805,969.50 |    0.0% |      0.01 | `worst case for radix tree`
-|               75.85 |       13,183,612.56 |    0.5% |        1,590.01 |          385.64 |  4.123 |         258.00 |    0.0% |      0.01 | `create and destroy`
+|               12.36 |       80,885,626.43 |    0.2% |          243.56 |           63.62 |  3.828 |          31.07 |    0.6% |      0.15 | `point reads`
+|               14.18 |       70,502,196.81 |    0.1% |          297.72 |           73.13 |  4.071 |          40.31 |    0.5% |      0.17 | `prefix reads`
+|               33.44 |       29,901,623.04 |    0.1% |          767.90 |          172.42 |  4.454 |         101.32 |    0.2% |      0.40 | `range reads`
+|               19.48 |       51,342,564.70 |    0.3% |          374.45 |          100.43 |  3.728 |          48.92 |    0.5% |      0.23 | `point writes`
+|               37.46 |       26,694,471.44 |    0.1% |          672.00 |          193.14 |  3.479 |         101.28 |    0.3% |      0.45 | `prefix writes`
+|               38.78 |       25,784,784.34 |    0.0% |          738.26 |          199.93 |  3.693 |         111.59 |    0.1% |      0.47 | `range writes`
+|               76.05 |       13,148,995.74 |    0.7% |        1,450.77 |          397.16 |  3.653 |         275.72 |    0.0% |      0.01 | `monotonic increasing point writes`
+|          286,920.33 |            3,485.29 |    0.4% |    4,117,948.00 |    1,521,352.00 |  2.707 |     714,833.00 |    0.1% |      0.01 | `worst case for radix tree`
+|               95.66 |       10,453,798.72 |    0.5% |        1,986.00 |          495.04 |  4.012 |         315.00 |    0.0% |      0.01 | `create and destroy`
 
 # "Real data" test
 
-Point queries only, best of three runs. Gc ratio is the ratio of time spent doing garbage collection to time spent adding writes or doing garbage collection. Lower is better.
+Point queries only. Gc ratio is the ratio of time spent doing garbage collection to time spent adding writes or doing garbage collection. Lower is better.
 
 ## skip list
 
 ```
-Check: 4.47891 seconds, 364.05 MB/s, Add: 4.55599 seconds, 123.058 MB/s, Gc ratio: 37.1145%
+Check: 4.53508 seconds, 371.81 MB/s, Add: 3.81222 seconds, 150.919 MB/s, Gc ratio: 33.66%, Peak idle memory: 5.61007e+06
 ```
 
 ## radix tree
 
 ```
-Check: 0.963721 seconds, 1691.93 MB/s, Add: 1.3288 seconds, 421.924 MB/s, Gc ratio: 42.8819%
+Check: 0.957735 seconds, 1760.6 MB/s, Add: 1.19942 seconds, 479.678 MB/s, Gc ratio: 38.6069%, Peak idle memory: 2.05667e+06
 ```
 
 ## hash table
@@ -55,5 +65,6 @@ Check: 0.963721 seconds, 1691.93 MB/s, Add: 1.3288 seconds, 421.924 MB/s, Gc rat
 (The hash table implementation doesn't work on range queries, and its purpose is to provide an idea of how fast point queries can be)
 
 ```
-Check: 0.804094 seconds, 2027.81 MB/s, Add: 0.652952 seconds, 858.645 MB/s, Gc ratio: 35.3885%
+Check: 0.804598 seconds, 2095.69 MB/s, Add: 0.671221 seconds, 857.147 MB/s, Gc ratio: 35.0034%, Peak idle memory: 0
 ```
+

RealDataBench.cpp

@@ -5,7 +5,7 @@
 #include <cstdio>
 #include <cstring>
 #include <fcntl.h>
-#include <string_view>
+#include <span>
 #include <sys/mman.h>
 #include <sys/stat.h>
 #include <unistd.h>
@@ -64,7 +64,7 @@ int main(int argc, const char **argv) {
     auto *const mapOriginal = begin;
     const auto sizeOriginal = size;
 
-    using StringView = std::basic_string_view<uint8_t>;
+    using StringView = std::span<const uint8_t>;
 
     StringView write;
     std::vector<StringView> reads;
@@ -78,9 +78,9 @@ int main(int argc, const char **argv) {
       end = (uint8_t *)memchr(begin, '\n', size);
 
       if (line.size() > 0 && line[0] == 'P') {
-        write = line.substr(2, line.size());
+        write = line.subspan(2, line.size());
       } else if (line.size() > 0 && line[0] == 'L') {
-        reads.push_back(line.substr(2, line.size()));
+        reads.push_back(line.subspan(2, line.size()));
       } else if (line.empty()) {
         {
           readRanges.resize(reads.size());
@@ -133,10 +133,10 @@ int main(int argc, const char **argv) {
   int metricsCount;
   cs.getMetricsV1(&metrics, &metricsCount);
   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());
+    fprintf(stderr, "# HELP %s %s\n", metrics[i].name, metrics[i].help);
+    fprintf(stderr, "# TYPE %s %s\n", metrics[i].name,
+            metrics[i].type == metrics[i].Counter ? "counter" : "gauge");
+    fprintf(stderr, "%s %g\n", metrics[i].name, metrics[i].getValue());
   }
 
   printf("Check: %g seconds, %g MB/s, Add: %g seconds, %g MB/s, Gc ratio: "

ServerBench.cpp

@@ -1,4 +1,7 @@
+#include <algorithm>
 #include <atomic>
+#include <cstdint>
+#include <cstdlib>
 #include <errno.h>
 #include <netdb.h>
 #include <stdio.h>
@@ -6,83 +9,221 @@
 #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"
 
+constexpr int kCacheLine = 64; // TODO mac m1 is 128
+
+template <class T> struct TxQueue {
+
+  explicit TxQueue(int lgSlotCount)
+      : slotCount(1 << lgSlotCount), slotCountMask(slotCount - 1),
+        slots(new T[slotCount]) {
+    // Otherwise we can't tell the difference between full and empty.
+    assert(!(slotCountMask & 0x80000000));
+  }
+
+  /// Call from producer thread, after ensuring consumer is no longer accessing
+  /// it somehow
+  ~TxQueue() { delete[] slots; }
+
+  /// Must be called from the producer thread
+  void push(T t) {
+    if (wouldBlock()) {
+      // Wait for pops to change and try again
+      consumer.pops.wait(producer.lastPopRead, std::memory_order_relaxed);
+      producer.lastPopRead = consumer.pops.load(std::memory_order_acquire);
+    }
+    slots[producer.pushesNonAtomic++ & slotCountMask] = std::move(t);
+    // seq_cst so that the notify can't be ordered before the store
+    producer.pushes.store(producer.pushesNonAtomic, std::memory_order_seq_cst);
+    // We have to notify every time, since we don't know if this is the last
+    // push ever
+    producer.pushes.notify_one();
+  }
+
+  /// Must be called from the producer thread
+  uint32_t outstanding() {
+    return producer.pushesNonAtomic -
+           consumer.pops.load(std::memory_order_relaxed);
+  }
+
+  /// Returns true if a call to push might block. Must be called from the
+  /// producer thread.
+  bool wouldBlock() {
+    // See if we can determine that overflow won't happen entirely from state
+    // local to the producer
+    if (producer.pushesNonAtomic - producer.lastPopRead == slotCount - 1) {
+      // Re-read pops with memory order
+      producer.lastPopRead = consumer.pops.load(std::memory_order_acquire);
+      return producer.pushesNonAtomic - producer.lastPopRead == slotCount - 1;
+    }
+    return false;
+  }
+
+  /// Valid until the next pop, or until this queue is destroyed.
+  T *pop() {
+    // See if we can determine that there's an entry we can pop entirely from
+    // state local to the consumer
+    if (consumer.lastPushRead - consumer.popsNonAtomic == 0) {
+      // Re-read pushes with memory order and try again
+      consumer.lastPushRead = producer.pushes.load(std::memory_order_acquire);
+      if (consumer.lastPushRead - consumer.popsNonAtomic == 0) {
+        // Wait for pushes to change and try again
+        producer.pushes.wait(consumer.lastPushRead, std::memory_order_relaxed);
+        consumer.lastPushRead = producer.pushes.load(std::memory_order_acquire);
+      }
+    }
+    auto result = &slots[consumer.popsNonAtomic++ & slotCountMask];
+    // We only have to write pops with memory order if we've run out of items.
+    // We know that we'll eventually run out.
+    if (consumer.lastPushRead - consumer.popsNonAtomic == 0) {
+      // seq_cst so that the notify can't be ordered before the store
+      consumer.pops.store(consumer.popsNonAtomic, std::memory_order_seq_cst);
+      consumer.pops.notify_one();
+    }
+    return result;
+  }
+
+private:
+  const uint32_t slotCount;
+  const uint32_t slotCountMask;
+  T *slots;
+  struct alignas(kCacheLine) ProducerState {
+    std::atomic<uint32_t> pushes{0};
+    uint32_t pushesNonAtomic{0};
+    uint32_t lastPopRead{0};
+  };
+  struct alignas(kCacheLine) ConsumerState {
+    std::atomic<uint32_t> pops{0};
+    uint32_t popsNonAtomic{0};
+    uint32_t lastPushRead{0};
+  };
+  ProducerState producer;
+  ConsumerState consumer;
+};
+
 std::atomic<int64_t> transactions;
 
-constexpr int kBaseSearchDepth = 32;
-constexpr int kWindowSize = 10000000;
+int64_t safeUnaryMinus(int64_t x) {
+  return x == std::numeric_limits<int64_t>::min() ? x : -x;
+}
 
-std::basic_string<uint8_t> numToKey(int64_t num) {
-  std::basic_string<uint8_t> result;
-  result.resize(kBaseSearchDepth + sizeof(int64_t));
-  memset(result.data(), 0, kBaseSearchDepth);
-  int64_t be = __builtin_bswap64(num);
-  memcpy(result.data() + kBaseSearchDepth, &be, sizeof(int64_t));
+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');
+  if (memchr(value.data(), '\x00', value.size()) != nullptr) {
+    for (auto c : value) {
+      if (c == '\x00') {
+        output.push_back('\x00');
+        output.push_back('\xff');
+      } else {
+        output.push_back(c);
+      }
+    }
+  } else {
+    output.insert(output.end(), value.begin(), value.end());
+  }
+  output.push_back('\x00');
+}
+
+template <class... Ts> std::string tupleKey(const Ts &...ts) {
+  std::string result;
+  (tupleAppend(result, ts), ...);
   return result;
 }
 
-void workload(weaselab::ConflictSet *cs) {
-  int64_t version = kWindowSize;
-  cs->addWrites(nullptr, 0, version);
-  for (;; transactions.fetch_add(1, std::memory_order_relaxed)) {
-    // Reads
-    {
-      auto beginK = numToKey(version - kWindowSize);
-      auto endK = numToKey(version - 1);
-      auto pointRv = version - kWindowSize + rand() % kWindowSize + 1;
-      auto pointK = numToKey(pointRv);
-      weaselab::ConflictSet::ReadRange reads[] = {
-          {
-              {pointK.data(), int(pointK.size())},
-              {nullptr, 0},
-              pointRv,
-          },
-          {
-              {beginK.data(), int(beginK.size())},
-              {endK.data(), int(endK.size())},
-              version - 2,
-          },
-      };
-      weaselab::ConflictSet::Result result[sizeof(reads) / sizeof(reads[0])];
-      cs->check(reads, result, sizeof(reads) / sizeof(reads[0]));
-      // for (int i = 0; i < sizeof(reads) / sizeof(reads[0]); ++i) {
-      //   if (result[i] != weaselab::ConflictSet::Commit) {
-      //     fprintf(stderr, "Unexpected conflict: [%s, %s) @ %" PRId64 "\n",
-      //             printable(reads[i].begin).c_str(),
-      //             printable(reads[i].end).c_str(), reads[i].readVersion);
-      //     abort();
-      //   }
-      // }
+constexpr int kTotalKeyRange = 1'000'000'000;
+constexpr int kWindowSize = 1'000'000;
+constexpr int kNumReadKeysPerTx = 5;
+constexpr int kNumWriteKeysPerTx = 10;
+
+struct Transaction {
+  std::vector<std::string> keys;
+  std::vector<weaselab::ConflictSet::ReadRange> reads;
+  std::vector<weaselab::ConflictSet::WriteRange> writes;
+  int64_t version;
+  int64_t oldestVersion;
+  Transaction() = default;
+  explicit Transaction(int64_t version)
+      : version(version), oldestVersion(version - kWindowSize) {
+    std::vector<int64_t> keyIndices;
+    for (int i = 0; i < std::max(kNumReadKeysPerTx, kNumWriteKeysPerTx); ++i) {
+      keyIndices.push_back(rand() % kTotalKeyRange);
     }
-    // Writes
-    {
-      weaselab::ConflictSet::WriteRange w;
-      auto k = numToKey(version);
-      w.begin.p = k.data();
-      w.end.len = 0;
-      if (version % (kWindowSize / 2) == 0) {
-        for (int l = 0; l <= k.size(); ++l) {
-          w.begin.len = l;
-          cs->addWrites(&w, 1, version);
-        }
-      } else {
-        w.begin.len = k.size();
-        cs->addWrites(&w, 1, version);
-      }
+    std::sort(keyIndices.begin(), keyIndices.end());
+    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 < int(keyIndices.size()); ++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());
+    }
+    for (int i = 0; i < kNumWriteKeysPerTx; ++i) {
+      writes.push_back({{(const uint8_t *)keys[i].data(), int(keys[i].size())},
+                        {nullptr, 0}});
+    }
+    reads.push_back({{(const uint8_t *)keys[0].data(), int(keys[0].size())},
+                     {(const uint8_t *)keys[1].data(), int(keys[1].size())},
+                     version - std::min(10, kWindowSize)});
+    static_assert(kNumReadKeysPerTx >= 3);
+    for (int i = 2; i < kNumReadKeysPerTx; ++i) {
+      reads.push_back({{(const uint8_t *)keys[i].data(), int(keys[i].size())},
+                       {nullptr, 0},
+                       version - kWindowSize});
     }
-    // GC
-    cs->setOldestVersion(version - kWindowSize);
-    ++version;
   }
-}
+
+  Transaction(Transaction &&) = default;
+  Transaction &operator=(Transaction &&) = default;
+  Transaction(Transaction const &) = delete;
+  Transaction const &operator=(Transaction const &) = delete;
+};
+
+struct Resolver {
+
+  void resolve(const weaselab::ConflictSet::ReadRange *reads, int readCount,
+               const weaselab::ConflictSet::WriteRange *writes, int writeCount,
+               int64_t newVersion, int64_t newOldestVersion) {
+    results.resize(readCount);
+    cs.check(reads, results.data(), readCount);
+    cs.addWrites(writes, writeCount, newVersion);
+    cs.setOldestVersion(newOldestVersion);
+  }
+
+  ConflictSet cs{0};
+
+private:
+  std::vector<weaselab::ConflictSet::Result> results;
+};
 
 // Adapted from getaddrinfo man page
 int getListenFd(const char *node, const char *service) {
@@ -156,6 +297,68 @@ 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;
@@ -163,12 +366,72 @@ int main(int argc, char **argv) {
   {
     int listenFd = getListenFd(argv[1], argv[2]);
 
-    weaselab::ConflictSet cs{0};
+    Resolver resolver;
+    auto &cs = resolver.cs;
     weaselab::ConflictSet::MetricsV1 *metrics;
     int metricsCount;
     cs.getMetricsV1(&metrics, &metricsCount);
 
-    auto w = std::thread{workload, &cs};
+#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
+
+    TxQueue<Transaction> queue{10};
+
+    auto workloadThread = std::thread{[&]() {
+      for (int64_t version = kWindowSize;;
+           ++version, transactions.fetch_add(1, std::memory_order_relaxed)) {
+        queue.push(Transaction(version));
+      }
+    }};
+
+    auto resolverThread = std::thread{[&]() {
+      for (;;) {
+        auto tx = queue.pop();
+        resolver.resolve(tx->reads.data(), tx->reads.size(), tx->writes.data(),
+                         tx->writes.size(), tx->version, tx->oldestVersion);
+      }
+    }};
 
     for (;;) {
       struct sockaddr_storage peer_addr = {};
@@ -195,6 +458,24 @@ int main(int argc, char **argv) {
               "transactions_total ";
       body += std::to_string(transactions.load(std::memory_order_relaxed));
       body += "\n";
+#ifdef __linux__
+      body += "# HELP instructions_total Total number of instructions\n"
+              "# TYPE instructions_total counter\n"
+              "instructions_total ";
+      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 ";

SkipList.cpp

@@ -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;
@@ -773,29 +796,34 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     int temp[stripeSize];
     int stripes = (stringCount + stripeSize - 1) / stripeSize;
     StringRef values[stripeSize];
-    int64_t writeVersions[stripeSize / 2];
     int ss = stringCount - (stripes - 1) * stripeSize;
+    int64_t entryDelta = 0;
     for (int s = stripes - 1; s >= 0; s--) {
       for (int i = 0; i * 2 < ss; ++i) {
         const auto &w = combinedWriteConflictRanges[s * stripeSize / 2 + i];
         values[i * 2] = w.first;
         values[i * 2 + 1] = w.second;
-        keyUpdates += 3;
       }
       skipList.find(values, fingers, temp, ss);
-      skipList.addConflictRanges(fingers, ss / 2, writeVersion);
+      entryDelta += skipList.addConflictRanges(fingers, ss / 2, writeVersion);
       ss = stripeSize;
     }
+
+    // Run gc at least 200% the rate we're inserting entries
+    keyUpdates += std::max<int64_t>(entryDelta, 0) * 2;
   }
 
   void setOldestVersion(int64_t oldestVersion) {
+    // 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()});
@@ -803,8 +831,56 @@ 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;
@@ -825,6 +901,7 @@ 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();
@@ -836,6 +913,7 @@ 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();
@@ -859,12 +937,11 @@ 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,

aarch64-symbol-imports.txt

@@ -5,6 +5,7 @@ __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

aarch64-toolchain.cmake

@@ -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")

apple-symbol-imports.txt

@@ -1,3 +1,4 @@
+___chkstk_darwin
 ___stack_chk_fail
 ___stack_chk_guard
 __tlv_bootstrap
@@ -5,6 +6,7 @@ _abort
 _bzero
 _free
 _malloc
+_memcmp
 _memcpy
 _memmove
 dyld_stub_binder