Remove unnecessary branch for interleaved range writes

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

Bench.cpp

@@ -17,26 +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 =
+  uint8_t *buf = new (arena) uint8_t[4 + kPrefixLen];
-      std::span<uint8_t>{new (arena) uint8_t[4 + kPrefixLen], 4 + kPrefixLen};
+  auto result = TrivialSpan{buf, 4 + kPrefixLen};
   index = __builtin_bswap32(index);
-  memset(result.data(), 0, kPrefixLen);
+  memset(buf, 0, kPrefixLen);
-  memcpy(result.data() + kPrefixLen, &index, 4);
+  memcpy(buf, &index, 4);
 
   return result;
 }
 
-ConflictSet::ReadRange singleton(Arena &arena, std::span<const uint8_t> key) {
+ConflictSet::ReadRange singleton(Arena &arena, TrivialSpan key) {
-  auto r =
+  uint8_t *buf = new (arena) uint8_t[key.size() + 1];
-      std::span<uint8_t>(new (arena) uint8_t[key.size() + 1], key.size() + 1);
+  auto r = TrivialSpan(buf, key.size() + 1);
-  memcpy(r.data(), key.data(), key.size());
+  memcpy(buf, key.data(), key.size());
-  r[key.size()] = 0;
+  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)
@@ -48,14 +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);
+  uint8_t *buf = new (arena) uint8_t[index + 1];
-  memcpy(r.data(), key.data(), index + 1);
+  auto r = TrivialSpan(buf, index + 1);
-  r[r.size() - 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);
@@ -81,14 +83,7 @@ void benchConflictSet() {
     ++version;
   }
 
-  // I don't know why std::less didn't work /shrug
+  auto points = set<TrivialSpan, std::less<>>(arena);
-  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);
 
   while (points.size() < kOpsPerTx * 2 + 1) {
     // TODO don't use rand?

CMakeLists.txt

@@ -31,8 +31,30 @@ if(NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
                                                "MinSizeRel" "RelWithDebInfo")
 endif()
 
-add_compile_options(-fdata-sections -ffunction-sections -Wswitch-enum
+add_compile_options(
-                    -Werror=switch-enum -fPIC)
+  -Werror=switch-enum
+  -Wswitch-enum
+  -Wunused-variable
+  -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)
@@ -47,6 +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()
@@ -64,27 +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(NOT USE_SIMD_FALLBACK)
+if(USE_SIMD_FALLBACK)
-  cmake_push_check_state()
+  add_compile_definitions(USE_SIMD_FALLBACK)
-  list(APPEND CMAKE_REQUIRED_FLAGS -mavx)
+else()
-  check_include_file_cxx("immintrin.h" HAS_AVX)
+  if(CMAKE_SYSTEM_PROCESSOR STREQUAL x86_64)
-  if(HAS_AVX)
     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()
 
@@ -347,6 +375,15 @@ 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} nanobench)

Dockerfile

@@ -11,23 +11,24 @@ RUN TZ=America/Los_Angeles DEBIAN_FRONTEND=noninteractive apt-get install -y \
     binutils-aarch64-linux-gnu \
     build-essential \
     ccache \
-    clang \
     cmake \
     curl \
-    doxygen \
+    devscripts \
-    file \
     g++-aarch64-linux-gnu \
     gcovr \
     git \
-    gperf \
+    gnupg \
-    graphviz \
     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
@@ -43,6 +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++

Internal.h

@@ -18,7 +18,6 @@ using namespace weaselab;
 #include <span>
 #include <string>
 #include <thread>
-#include <unordered_set>
 #include <utility>
 
 #include <callgrind.h>
@@ -26,9 +25,38 @@ using namespace weaselab;
 #define DEBUG_VERBOSE 0
 #define SHOW_MEMORY 0
 
-[[nodiscard]] inline auto
+// std::span is not trivially constructible. We want a span that leaves its
-operator<=>(const std::span<const uint8_t> &lhs,
+// members uninitialized for performance reasons.
-            const std::span<const uint8_t> &rhs) noexcept {
+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) {
@@ -38,7 +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) {
@@ -46,7 +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
@@ -328,23 +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;
@@ -569,7 +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()));
 }
 
@@ -677,10 +699,8 @@ struct TestDriver {
             arbitrary->randomBytes(begin + prefixLen, keyLen - prefixLen);
             writes[i].end.len = keyLen;
             writes[i].end.p = begin;
-            auto c =
+            auto c = TrivialSpan(writes[i].begin.p, writes[i].begin.len) <=>
-                std::span<const uint8_t>(writes[i].begin.p,
+                     TrivialSpan(writes[i].end.p, writes[i].end.len);
-                                         writes[i].begin.len) <=>
-                std::span<const uint8_t>(writes[i].end.p, writes[i].end.len);
             if (c > 0) {
               using std::swap;
               swap(writes[i].begin, writes[i].end);

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}/'
 }
 
@@ -36,18 +36,6 @@ pipeline {
                 sh 'pre-commit run --all-files --show-diff-on-failure'
             }
         }
-        stage('Clang') {
-            agent {
-                dockerfile {
-                    args '-v /home/jenkins/ccache:/ccache'
-                    reuseNode true
-                }
-            }
-            steps {
-                CleanBuildAndTest("")
-                recordIssues(tools: [clang()])
-            }
-        }
         stage('64 bit versions') {
             agent {
                 dockerfile {
@@ -81,7 +69,7 @@ pipeline {
                 CleanBuildAndTest("-DUSE_SIMD_FALLBACK=ON")
             }
         }
-        stage('Release [gcc]') {
+        stage('Release [clang]') {
             agent {
                 dockerfile {
                     args '-v /home/jenkins/ccache:/ccache'
@@ -89,8 +77,8 @@ pipeline {
                 }
             }
             steps {
-                CleanBuildAndTest("-DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ -DCMAKE_CXX_FLAGS=-DNVALGRIND")
+                CleanBuildAndTest("-DCMAKE_CXX_FLAGS=-DNVALGRIND")
-                recordIssues(tools: [gcc()])
+                recordIssues(tools: [clang()])
                 sh '''
                 cd build
                 cpack -G DEB
@@ -103,7 +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'
@@ -129,16 +129,16 @@ pipeline {
             }
             steps {
                 script {
-                    filter_args = "-f ConflictSet.cpp -f LongestCommonPrefix.h -f Metrics.h"
+                    gcov_args = "-f ConflictSet.cpp -f LongestCommonPrefix.h -f Metrics.h --gcov-executable 'llvm-cov gcov' --exclude-noncode-lines"
                 }
-                CleanBuildAndTest("-DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ -DCMAKE_C_FLAGS=--coverage -DCMAKE_CXX_FLAGS=--coverage -DCMAKE_BUILD_TYPE=Debug -DDISABLE_TSAN=ON")
+                CleanBuildAndTest("-DCMAKE_C_FLAGS=--coverage -DCMAKE_CXX_FLAGS=--coverage -DCMAKE_BUILD_TYPE=Debug -DDISABLE_TSAN=ON")
                 sh """
-                gcovr ${filter_args} --cobertura > build/coverage.xml
+                gcovr ${gcov_args} --cobertura > build/coverage.xml
                 """
                 recordCoverage qualityGates: [[criticality: 'NOTE', metric: 'MODULE']], tools: [[parser: 'COBERTURA', pattern: 'build/coverage.xml']]
                 sh """
-                gcovr ${filter_args}
+                gcovr ${gcov_args}
-                gcovr ${filter_args} --fail-under-line 100 > /dev/null
+                gcovr ${gcov_args} --fail-under-line 100 > /dev/null
                 """
             }
         }

LongestCommonPrefix.h

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

README.md

@@ -2,7 +2,16 @@ A data structure for optimistic concurrency control on ranges of bitwise-lexicog
 
 Intended as an alternative to FoundationDB's skip list.
 
-Hardware for all benchmarks is an AMD Ryzen 9 7900 with (2x32GB) 5600MT/s CL28-34-34-89 1.35V RAM
+Hardware for all benchmarks is an AMD Ryzen 9 7900 with (2x32GB) 5600MT/s CL28-34-34-89 1.35V RAM.
+
+```
+$ clang++ --version
+
+Ubuntu clang version 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`
+|              161.29 |        6,200,056.17 |    0.1% |        3,014.03 |          831.04 |  3.627 |         504.59 |    0.0% |      1.93 | `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`
+|              158.32 |        6,316,160.64 |    0.1% |        2,954.16 |          815.80 |  3.621 |         490.17 |    0.0% |      1.89 | `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`
+|              237.39 |        4,212,409.50 |    0.2% |        3,592.41 |        1,233.96 |  2.911 |         629.31 |    0.0% |      2.84 | `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`
+|              442.11 |        2,261,878.94 |    0.0% |        4,450.57 |        2,314.25 |  1.923 |         707.92 |    2.1% |      5.28 | `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`
+|              439.89 |        2,273,308.53 |    0.1% |        4,410.22 |        2,302.29 |  1.916 |         694.74 |    2.1% |      5.25 | `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`
+|              290.96 |        3,436,936.78 |    0.0% |        2,315.38 |        1,528.68 |  1.515 |         396.69 |    3.3% |      3.49 | `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`
+|              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`
-|          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`
+|          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`
-|               21.63 |       46,231,564.03 |    1.0% |          448.00 |          107.14 |  4.181 |          84.00 |    0.0% |      0.01 | `create and destroy`
+|               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
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|               12.88 |       77,653,350.77 |    0.5% |          185.37 |           64.45 |  2.876 |          41.51 |    0.4% |      0.01 | `point reads`
+|               12.36 |       80,885,626.43 |    0.2% |          243.56 |           63.62 |  3.828 |          31.07 |    0.6% |      0.15 | `point reads`
-|               14.67 |       68,179,354.49 |    0.1% |          271.44 |           73.40 |  3.698 |          53.70 |    0.3% |      0.01 | `prefix reads`
+|               14.18 |       70,502,196.81 |    0.1% |          297.72 |           73.13 |  4.071 |          40.31 |    0.5% |      0.17 | `prefix reads`
-|               34.84 |       28,701,444.36 |    0.3% |          715.74 |          175.27 |  4.084 |         127.30 |    0.2% |      0.01 | `range reads`
+|               33.44 |       29,901,623.04 |    0.1% |          767.90 |          172.42 |  4.454 |         101.32 |    0.2% |      0.40 | `range reads`
-|               17.12 |       58,422,988.28 |    0.2% |          314.30 |           86.11 |  3.650 |          39.82 |    0.4% |      0.01 | `point writes`
+|               19.48 |       51,342,564.70 |    0.3% |          374.45 |          100.43 |  3.728 |          48.92 |    0.5% |      0.23 | `point writes`
-|               31.42 |       31,830,804.65 |    0.1% |          591.06 |          158.07 |  3.739 |          82.67 |    0.2% |      0.01 | `prefix writes`
+|               37.46 |       26,694,471.44 |    0.1% |          672.00 |          193.14 |  3.479 |         101.28 |    0.3% |      0.45 | `prefix writes`
-|               37.37 |       26,759,432.70 |    2.2% |          681.98 |          188.95 |  3.609 |          96.10 |    0.1% |      0.01 | `range writes`
+|               38.78 |       25,784,784.34 |    0.0% |          738.26 |          199.93 |  3.693 |         111.59 |    0.1% |      0.47 | `range writes`
-|               76.72 |       13,035,140.63 |    2.3% |        1,421.28 |          387.17 |  3.671 |         257.76 |    0.1% |      0.01 | `monotonic increasing point 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`
-|          297,452.00 |            3,361.89 |    0.9% |    3,508,083.00 |    1,500,834.67 |  2.337 |     727,525.33 |    0.1% |      0.01 | `worst case for radix tree`
+|          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`
-|               87.70 |       11,402,490.60 |    1.0% |        1,795.00 |          442.09 |  4.060 |         297.00 |    0.0% |      0.01 | `create and destroy`
+|               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: 0.953012 seconds, 1710.94 MB/s, Add: 1.30025 seconds, 431.188 MB/s, Gc ratio: 43.9816%, Peak idle memory: 2.28375e+06
+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: 0.953012 seconds, 1710.94 MB/s, Add: 1.30025 seconds, 431.188 MB/s, Gc ra
 (The hash table implementation doesn't work on range queries, and its purpose is to provide an idea of how fast point queries can be)
 
 ```
-Check: 0.804094 seconds, 2027.81 MB/s, Add: 0.652952 seconds, 858.645 MB/s, Gc ratio: 35.3885%
+Check: 0.804598 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);
+    fprintf(stderr, "# HELP %s %s\n", metrics[i].name, metrics[i].help);
-    printf("# TYPE %s %s\n", metrics[i].name,
+    fprintf(stderr, "# TYPE %s %s\n", metrics[i].name,
-           metrics[i].type == metrics[i].Counter ? "counter" : "gauge");
+            metrics[i].type == metrics[i].Counter ? "counter" : "gauge");
-    printf("%s %g\n", metrics[i].name, metrics[i].getValue());
+    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

@@ -1,5 +1,7 @@
+#include <algorithm>
 #include <atomic>
 #include <cstdint>
+#include <cstdlib>
 #include <errno.h>
 #include <netdb.h>
 #include <stdio.h>
@@ -18,62 +20,210 @@
 #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 kWindowSize = 10000000;
+int64_t safeUnaryMinus(int64_t x) {
+  return x == std::numeric_limits<int64_t>::min() ? x : -x;
+}
 
-constexpr int kNumPrefixes = 250000;
+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);
+}
 
-std::string makeKey(int64_t num, int suffixLen) {
+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;
-  result.resize(sizeof(int64_t) + suffixLen);
+  (tupleAppend(result, ts), ...);
-  int64_t be = __builtin_bswap64(num);
-  memcpy(result.data(), &be, sizeof(int64_t));
-  memset(result.data() + sizeof(int64_t), 0, suffixLen);
   return result;
 }
 
-void workload(weaselab::ConflictSet *cs) {
+constexpr int kTotalKeyRange = 1'000'000'000;
-  int64_t version = kWindowSize;
+constexpr int kWindowSize = 1'000'000;
-  for (int i = 0; i < kNumPrefixes; ++i) {
+constexpr int kNumReadKeysPerTx = 5;
-    for (int j = 0; j < 50; ++j) {
+constexpr int kNumWriteKeysPerTx = 10;
-      weaselab::ConflictSet::WriteRange wr;
+
-      auto k = makeKey(i, j);
+struct Transaction {
-      wr.begin.p = (const uint8_t *)k.data();
+  std::vector<std::string> keys;
-      wr.begin.len = k.size();
+  std::vector<weaselab::ConflictSet::ReadRange> reads;
-      wr.end.len = 0;
+  std::vector<weaselab::ConflictSet::WriteRange> writes;
-      cs->addWrites(&wr, 1, version);
+  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);
+    }
+    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});
     }
-  }
-  ++version;
-  for (int i = 0; i < kNumPrefixes; ++i) {
-    weaselab::ConflictSet::WriteRange wr;
-    auto k = makeKey(i, 50);
-    wr.begin.p = (const uint8_t *)k.data();
-    wr.begin.len = k.size();
-    wr.end.len = 0;
-    cs->addWrites(&wr, 1, version);
   }
 
-  constexpr int kNumReads = 1;
+  Transaction(Transaction &&) = default;
-  std::vector<weaselab::ConflictSet::Result> results(kNumReads);
+  Transaction &operator=(Transaction &&) = default;
-  for (;; transactions.fetch_add(1, std::memory_order_relaxed)) {
+  Transaction(Transaction const &) = delete;
-    std::vector<std::string> keys(kNumReads);
+  Transaction const &operator=(Transaction const &) = delete;
-    for (auto &k : keys) {
+};
-      k = makeKey(rand() % kNumPrefixes, 49);
+
-    }
+struct Resolver {
-    std::vector<weaselab::ConflictSet::ReadRange> reads(kNumReads);
+
-    for (int i = 0; i < reads.size(); ++i) {
+  void resolve(const weaselab::ConflictSet::ReadRange *reads, int readCount,
-      reads[i].begin.p = (const uint8_t *)(keys[i].data());
+               const weaselab::ConflictSet::WriteRange *writes, int writeCount,
-      reads[i].begin.len = keys[i].size();
+               int64_t newVersion, int64_t newOldestVersion) {
-      reads[i].end.len = 0;
+    results.resize(readCount);
-      reads[i].readVersion = version - 1;
+    cs.check(reads, results.data(), readCount);
-    }
+    cs.addWrites(writes, writeCount, newVersion);
-    cs->check(reads.data(), results.data(), kNumReads);
+    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) {
@@ -216,7 +366,8 @@ 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);
@@ -265,7 +416,22 @@ int main(int argc, char **argv) {
     }
 #endif
 
-    auto w = std::thread{workload, &cs};
+    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 = {};

SkipList.cpp

@@ -767,7 +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;
@@ -794,7 +796,6 @@ 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--) {

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 +1,7 @@
 set(CMAKE_SYSTEM_NAME Linux)
 set(CMAKE_SYSTEM_PROCESSOR aarch64)
-set(CMAKE_C_COMPILER "/usr/bin/aarch64-linux-gnu-gcc")
+set(CMAKE_C_COMPILER "clang;--target=aarch64-linux-gnu")
-set(CMAKE_CXX_COMPILER "/usr/bin/aarch64-linux-gnu-g++")
+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)

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