Add cycles_total

If numChildren + entryPresent is enough, we don't have to copy even if
it would fit in a smaller node.

If we have to copy, we might as well use the smallest acceptable node
type.

Bench.cpp

@@ -361,7 +361,21 @@ void benchWorstCaseForRadixRangeRead() {
 void benchCreateAndDestroy() {
   ankerl::nanobench::Bench bench;
 
-  bench.run("create and destroy", [&]() { ConflictSet cs{0}; });
+  bench.run("create and destroy", [&]() {
+    ConflictSet cs{0};
+    ConflictSet::WriteRange w;
+    uint8_t b[9];
+    b[8] = 0;
+    for (int64_t i = 0; i < 1000; i += 7) {
+      auto x = __builtin_bswap64(i);
+      memcpy(b, &x, 8);
+      w.begin.p = b;
+      w.begin.len = 8;
+      w.end.len = 0;
+      w.end.p = b;
+      cs.addWrites(&w, 1, 1);
+    }
+  });
 }
 
 int main(void) {

CMakeLists.txt

@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.18)
 project(
   conflict-set
-  VERSION 0.0.10
+  VERSION 0.0.12
   DESCRIPTION
     "A data structure for optimistic concurrency control on ranges of bitwise-lexicographically-ordered keys."
   HOMEPAGE_URL "https://git.weaselab.dev/weaselab/conflict-set"
@@ -31,14 +31,12 @@ if(NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
                                                "MinSizeRel" "RelWithDebInfo")
 endif()
 
-add_compile_options(
-  -fdata-sections
-  -ffunction-sections
-  -Wswitch-enum
-  -Werror=switch-enum
-  -fPIC
-  -g
-  -fno-omit-frame-pointer)
+add_compile_options(-fdata-sections -ffunction-sections -Wswitch-enum
+                    -Werror=switch-enum -fPIC)
+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()

Internal.h

@@ -273,6 +273,16 @@ 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; }

Jenkinsfile

@@ -117,15 +117,18 @@ pipeline {
                 }
             }
             steps {
+                script {
+                    filter_args = "-f ConflictSet.cpp -f LongestCommonPrefix.h"
+                }
                 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
-                '''
+                sh """
+                gcovr ${filter_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 ${filter_args}
+                gcovr ${filter_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;
+
+  // 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;
+}

README.md

@@ -24,15 +24,16 @@ 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
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|               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`
+|               10.80 |       92,600,541.52 |    0.6% |          180.38 |           54.49 |  3.310 |          41.51 |    0.4% |      0.01 | `point reads`
+|               15.00 |       66,687,691.68 |    0.4% |          278.44 |           76.44 |  3.642 |          55.56 |    0.3% |      0.01 | `prefix reads`
+|               36.81 |       27,163,394.61 |    0.4% |          795.06 |          187.91 |  4.231 |         142.67 |    0.2% |      0.01 | `range reads`
+|               18.14 |       55,137,674.01 |    1.2% |          338.19 |           92.86 |  3.642 |          42.81 |    0.4% |      0.01 | `point writes`
+|               33.19 |       30,127,119.71 |    0.1% |          681.03 |          170.05 |  4.005 |          98.68 |    0.2% |      0.01 | `prefix writes`
+|               37.37 |       26,759,432.70 |    1.9% |          779.70 |          195.45 |  3.989 |         114.21 |    0.0% |      0.01 | `range writes`
+|               74.36 |       13,448,582.47 |    1.9% |        1,425.68 |          389.08 |  3.664 |         258.88 |    0.1% |      0.01 | `monotonic increasing point writes`
+|          316,928.00 |            3,155.29 |    1.5% |    3,992,986.00 |    1,699,813.00 |  2.349 |     806,226.50 |    0.0% |      0.01 | `worst case for radix tree`
+|               75.26 |       13,286,517.16 |    0.5% |        1,590.01 |          386.67 |  4.112 |         258.00 |    0.0% |      0.01 | `create and destroy`
+
 
 # "Real data" test
 
@@ -47,7 +48,7 @@ Check: 4.47891 seconds, 364.05 MB/s, Add: 4.55599 seconds, 123.058 MB/s, Gc rati
 ## radix tree
 
 ```
-Check: 0.963721 seconds, 1691.93 MB/s, Add: 1.3288 seconds, 421.924 MB/s, Gc ratio: 42.8819%
+Check: 0.910234 seconds, 1791.35 MB/s, Add: 1.25908 seconds, 445.287 MB/s, Gc ratio: 44.0415%
 ```
 
 ## hash table

ServerBench.cpp

@@ -76,6 +76,14 @@ void workload(weaselab::ConflictSet *cs) {
       } else {
         w.begin.len = k.size();
         cs->addWrites(&w, 1, version);
+        int64_t beginN = version - kWindowSize + rand() % kWindowSize;
+        auto b = numToKey(beginN);
+        auto e = numToKey(beginN + 1000);
+        w.begin.p = b.data();
+        w.begin.len = b.size();
+        w.end.p = e.data();
+        w.end.len = e.size();
+        cs->addWrites(&w, 1, version);
       }
     }
     // GC
@@ -156,6 +164,63 @@ double toSeconds(timeval t) {
   return double(t.tv_sec) + double(t.tv_usec) * 1e-6;
 }
 
+#include <linux/perf_event.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/ioctl.h>
+#include <sys/syscall.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#ifdef __linux__
+struct PerfCounter {
+  explicit PerfCounter(int event) {
+    struct perf_event_attr pe;
+
+    memset(&pe, 0, sizeof(pe));
+    pe.type = PERF_TYPE_HARDWARE;
+    pe.size = sizeof(pe);
+    pe.config = event;
+    pe.inherit = 1;
+    pe.exclude_kernel = 1;
+    pe.exclude_hv = 1;
+
+    fd = perf_event_open(&pe, 0, -1, -1, 0);
+    if (fd == -1) {
+      fprintf(stderr, "Error opening leader %llx\n", pe.config);
+      exit(EXIT_FAILURE);
+    }
+  }
+
+  int64_t total() {
+    int64_t count;
+    if (read(fd, &count, sizeof(count)) != sizeof(count)) {
+      perror("read instructions from perf");
+      abort();
+    }
+    return count;
+  }
+
+  ~PerfCounter() { close(fd); }
+
+private:
+  int fd;
+  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;
+  }
+};
+#else
+struct PerfCounter {
+  explicit PerPerfCounter(int) {}
+  int64_t total() { return 0; }
+};
+#endif
+
 int main(int argc, char **argv) {
   if (argc != 3) {
     goto fail;
@@ -168,6 +233,8 @@ int main(int argc, char **argv) {
     int metricsCount;
     cs.getMetricsV1(&metrics, &metricsCount);
 
+    PerfCounter instructions{PERF_COUNT_HW_INSTRUCTIONS};
+    PerfCounter cycles{PERF_COUNT_HW_CPU_CYCLES};
     auto w = std::thread{workload, &cs};
 
     for (;;) {
@@ -195,6 +262,16 @@ int main(int argc, char **argv) {
               "transactions_total ";
       body += std::to_string(transactions.load(std::memory_order_relaxed));
       body += "\n";
+      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";
 
       for (int i = 0; i < metricsCount; ++i) {
         body += "# HELP ";

corpus/0f5c5f28c31e44ee0ec085fec09adf0981065771

@@ -1,5 +0,0 @@
-
-<EFBFBD><EFBFBD>
-
-
-2

corpus/12bdd00fd4038756cbcf8ecdad1b0cd862603cd8

@@ -1 +0,0 @@
-<EFBFBD>