Skip sorting if already sorted for skip list

This works around a weird error I was getting when trying to link a
translation unit that included Internal.h with libconflict-set-static.a
with clang + gnu objcopy

.clangd

@@ -1,2 +1,2 @@
 CompileFlags:
-  Add: [-DENABLE_MAIN, -UNDEBUG, -DENABLE_FUZZ, -fexceptions]
+  Add: [-DENABLE_MAIN, -UNDEBUG, -DENABLE_FUZZ, -DTHREAD_TEST, -fexceptions]

.gitignore

@@ -1,2 +1,3 @@
 .cache
+__pycache__
 build

.pre-commit-config.yaml

@@ -1,11 +1,11 @@
 repos:
   - repo: https://github.com/pre-commit/mirrors-clang-format
-    rev: b111689e7b5cba60be3c62d5db2bd1357f4d36ca
+    rev: 6d365699efc33b1b432eab5b4ae331a19e1857de # frozen: v18.1.2
     hooks:
       - id: clang-format
         exclude: ".*third_party/.*"
   - repo: https://github.com/cheshirekow/cmake-format-precommit
-    rev: e2c2116d86a80e72e7146a06e68b7c228afc6319
+    rev: e2c2116d86a80e72e7146a06e68b7c228afc6319 # frozen: v0.6.13
     hooks:
       - id: cmake-format
   - repo: local
@@ -13,6 +13,22 @@ repos:
       - id: debug verbose check
         name: disallow checking in DEBUG_VERBOSE=1
         description: disallow checking in DEBUG_VERBOSE=1
-        entry: '^#define DEBUG_VERBOSE 1$'
+        entry: "^#define DEBUG_VERBOSE 1$"
         language: pygrep
-        types: [c++]
+        types: [c++]
+  - repo: local
+    hooks:
+      - id: debug verbose check
+        name: disallow checking in SHOW_MEMORY=1
+        description: disallow checking in SHOW_MEMORY=1
+        entry: "^#define SHOW_MEMORY 1$"
+        language: pygrep
+        types: [c++]
+  - repo: https://github.com/shellcheck-py/shellcheck-py
+    rev: a23f6b85d0fdd5bb9d564e2579e678033debbdff # frozen: v0.10.0.1
+    hooks:
+      - id: shellcheck
+  - repo: https://github.com/psf/black
+    rev: 552baf822992936134cbd31a38f69c8cfe7c0f05 # frozen: 24.3.0
+    hooks:
+      - id: black

.vscode/settings.json

@@ -0,0 +1,9 @@
+{
+  "files.associations": {
+    "*.tikz": "latex"
+  },
+  "latex-workshop.view.pdf.invertMode.enabled": "compat",
+  "latex-workshop.view.pdf.invert": 1,
+  "latex-workshop.view.pdf.invertMode.sepia": 1,
+  "latex-workshop.view.pdf.invertMode.grayscale": 0.5
+}

Bench.cpp

@@ -2,622 +2,41 @@
 #include "Internal.h"
 #include <cstdint>
 #include <cstring>
-#include <string>
 
-#define ANKERL_NANOBENCH_IMPLEMENT
+#if SHOW_MEMORY
+void showMemory(const ConflictSet &cs);
+#endif
+
 #include "third_party/nanobench.h"
 
-std::span<const uint8_t> keyAfter(Arena &arena, std::span<const uint8_t> key) {
+constexpr int kNumKeys = 1000000;
-  auto result =
+
-      std::span<uint8_t>(new (arena) uint8_t[key.size() + 1], key.size() + 1);
+constexpr int kOpsPerTx = 100;
-  memcpy(result.data(), key.data(), key.size());
+
-  result[result.size() - 1] = 0;
+constexpr int kPrefixLen = 0;
+
+constexpr int kMvccWindow = 100000;
+
+TrivialSpan makeKey(Arena &arena, int index) {
+
+  uint8_t *buf = new (arena) uint8_t[4 + kPrefixLen];
+  auto result = TrivialSpan{buf, 4 + kPrefixLen};
+  index = __builtin_bswap32(index);
+  memset(buf, 0, kPrefixLen);
+  memcpy(buf, &index, 4);
+
   return result;
 }
 
-namespace {
+ConflictSet::ReadRange singleton(Arena &arena, TrivialSpan key) {
-
+  uint8_t *buf = new (arena) uint8_t[key.size() + 1];
-using Version = int64_t;
+  auto r = TrivialSpan(buf, key.size() + 1);
-#define force_inline __attribute__((always_inline))
+  memcpy(buf, key.data(), key.size());
-using StringRef = std::span<const uint8_t>;
+  buf[key.size()] = 0;
-
+  return {{key.data(), int(key.size())}, {r.data(), int(r.size())}, 0};
-struct KeyRangeRef {
-  StringRef begin;
-  StringRef end;
-  KeyRangeRef() {}
-  KeyRangeRef(StringRef begin, StringRef end) : begin(begin), end(end) {}
-  KeyRangeRef(Arena &arena, StringRef begin)
-      : begin(begin), end(keyAfter(arena, begin)) {}
-};
-
-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) {
+ConflictSet::ReadRange prefixRange(Arena &arena, TrivialSpan key) {
-  int c = memcmp(a.data(), b.data(), std::min(a.size(), b.size()));
-  if (c < 0)
-    return -1;
-  if (c > 0)
-    return +1;
-  if (a.size() < b.size())
-    return -1;
-  if (a.size() == b.size())
-    return 0;
-  return +1;
-}
-
-struct ReadConflictRange {
-  StringRef begin, end;
-  Version version;
-
-  ReadConflictRange() {}
-  ReadConflictRange(StringRef begin, StringRef end, Version version)
-      : begin(begin), end(end), version(version) {}
-  bool operator<(const ReadConflictRange &rhs) const {
-    return compare(begin, rhs.begin) < 0;
-  }
-};
-
-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;
-  }
-
-  // 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
-  // level.
-  struct Node {
-    int level() const { return nPointers - 1; }
-    uint8_t *value() {
-      return end() + nPointers * (sizeof(Node *) + sizeof(Version));
-    }
-    int length() const { return valueLength; }
-
-    // Returns the next node pointer at the given level.
-    Node *getNext(int level) { return *((Node **)end() + level); }
-    // Sets the next node pointer at the given level.
-    void setNext(int level, Node *n) { *((Node **)end() + level) = n; }
-
-    // Returns the max version at the given level.
-    Version getMaxVersion(int i) const {
-      return ((Version *)(end() + nPointers * sizeof(Node *)))[i];
-    }
-    // Sets the max version at the given level.
-    void setMaxVersion(int i, Version v) {
-      ((Version *)(end() + nPointers * sizeof(Node *)))[i] = v;
-    }
-
-    // Return a node with initialized value but uninitialized pointers
-    // Memory layout: *this, (level+1) Node*, (level+1) Version, value
-    static Node *create(const StringRef &value, int level) {
-      int nodeSize = sizeof(Node) + value.size() +
-                     (level + 1) * (sizeof(Node *) + sizeof(Version));
-
-      Node *n;
-      n = (Node *)new char[nodeSize];
-
-      n->nPointers = level + 1;
-
-      n->valueLength = value.size();
-      if (value.size() > 0) {
-        memcpy(n->value(), value.data(), value.size());
-      }
-      return n;
-    }
-
-    // pre: level>0, all lower level nodes between this and getNext(level) have
-    // correct maxversions
-    void calcVersionForLevel(int level) {
-      Node *end = getNext(level);
-      Version v = getMaxVersion(level - 1);
-      for (Node *x = getNext(level - 1); x != end; x = x->getNext(level - 1))
-        v = std::max(v, x->getMaxVersion(level - 1));
-      setMaxVersion(level, v);
-    }
-
-    void destroy() { delete[](char *) this; }
-
-  private:
-    int getNodeSize() const {
-      return sizeof(Node) + valueLength +
-             nPointers * (sizeof(Node *) + sizeof(Version));
-    }
-    // Returns the first Node* pointer
-    uint8_t *end() { return (uint8_t *)(this + 1); }
-    uint8_t const *end() const { return (uint8_t const *)(this + 1); }
-    int nPointers, valueLength;
-  };
-
-  static force_inline bool less(const uint8_t *a, int aLen, const uint8_t *b,
-                                int bLen) {
-    int c = memcmp(a, b, std::min(aLen, bLen));
-    if (c < 0)
-      return true;
-    if (c > 0)
-      return false;
-    return aLen < bLen;
-  }
-
-  Node *header;
-
-  void destroy() {
-    Node *next, *x;
-    for (x = header; x; x = next) {
-      next = x->getNext(0);
-      x->destroy();
-    }
-  }
-
-public:
-  // Points the location (i.e., Node*) that value would appear in the SkipList.
-  // If the "value" is in the list, then finger[0] points to that exact node;
-  // otherwise, the finger points to Nodes that the value should be inserted
-  // before. Note the SkipList organizes all nodes at level 0, higher levels
-  // contain jump pointers.
-  struct Finger {
-    Node *finger[MaxLevels]; // valid for levels >= level
-    int level = MaxLevels;
-    Node *x = nullptr;
-    Node *alreadyChecked = nullptr;
-    StringRef value;
-
-    Finger() = default;
-    Finger(Node *header, const StringRef &ptr) : x(header), value(ptr) {}
-
-    void init(const StringRef &value, Node *header) {
-      this->value = value;
-      x = header;
-      alreadyChecked = nullptr;
-      level = MaxLevels;
-    }
-
-    // pre: !finished()
-    force_inline void prefetch() {
-      Node *next = x->getNext(0);
-      __builtin_prefetch(next);
-    }
-
-    // pre: !finished()
-    // Advances the pointer at the current level to a Node that's >= finger's
-    // value if possible; or move to the next level (i.e., level--). Returns
-    // true if we have advanced to the next level
-    force_inline bool advance() {
-      Node *next = x->getNext(level - 1);
-
-      if (next == alreadyChecked ||
-          !less(next->value(), next->length(), value.data(), value.size())) {
-        alreadyChecked = next;
-        level--;
-        finger[level] = x;
-        return true;
-      } else {
-        x = next;
-        return false;
-      }
-    }
-
-    // pre: !finished()
-    force_inline void nextLevel() {
-      while (!advance())
-        ;
-    }
-
-    force_inline bool finished() const { return level == 0; }
-
-    // Returns if the finger value is found in the SkipList.
-    force_inline Node *found() const {
-      // valid after finished returns true
-      Node *n = finger[0]->getNext(
-          0); // or alreadyChecked, but that is more easily invalidated
-      if (n && n->length() == value.size() &&
-          !memcmp(n->value(), value.data(), value.size()))
-        return n;
-      else
-        return nullptr;
-    }
-
-    StringRef getValue() const {
-      Node *n = finger[0]->getNext(0);
-      return n ? StringRef(n->value(), n->length()) : StringRef();
-    }
-  };
-
-  // Returns the total number of nodes in the list.
-  int count() const {
-    int count = 0;
-    Node *x = header->getNext(0);
-    while (x) {
-      x = x->getNext(0);
-      count++;
-    }
-    return count;
-  }
-
-  explicit SkipList(Version version = 0) {
-    header = Node::create(StringRef(), MaxLevels - 1);
-    for (int l = 0; l < MaxLevels; l++) {
-      header->setNext(l, nullptr);
-      header->setMaxVersion(l, version);
-    }
-  }
-  ~SkipList() { destroy(); }
-  SkipList(SkipList &&other) noexcept : header(other.header) {
-    other.header = nullptr;
-  }
-  void operator=(SkipList &&other) noexcept {
-    destroy();
-    header = other.header;
-    other.header = nullptr;
-  }
-  void swap(SkipList &other) { std::swap(header, other.header); }
-
-  void addConflictRanges(const Finger *fingers, int rangeCount,
-                         Version *version) {
-    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)
-        insert(endF, endF.finger[0]->getMaxVersion(0));
-
-      remove(startF, endF);
-      insert(startF, version[r]);
-    }
-  }
-
-  void detectConflicts(ReadConflictRange *ranges, int count,
-                       ConflictSet::Result *transactionConflictStatus) const {
-    const int M = 16;
-    int nextJob[M];
-    CheckMax inProgress[M];
-    if (!count)
-      return;
-
-    int started = std::min(M, count);
-    for (int i = 0; i < started; i++) {
-      inProgress[i].init(ranges[i], header, transactionConflictStatus + i);
-      nextJob[i] = i + 1;
-    }
-    nextJob[started - 1] = 0;
-
-    int prevJob = started - 1;
-    int job = 0;
-    // vtune: 340 parts
-    while (true) {
-      if (inProgress[job].advance()) {
-        if (started == count) {
-          if (prevJob == job)
-            break;
-          nextJob[prevJob] = nextJob[job];
-          job = prevJob;
-        } else {
-          int temp = started++;
-          inProgress[job].init(ranges[temp], header,
-                               transactionConflictStatus + temp);
-        }
-      }
-      prevJob = job;
-      job = nextJob[job];
-    }
-  }
-
-  void find(const StringRef *values, Finger *results, int *temp, int count) {
-    // Relying on the ordering of values, descend until the values aren't all in
-    // the same part of the tree
-
-    // vtune: 11 parts
-    results[0].init(values[0], header);
-    const StringRef &endValue = values[count - 1];
-    while (results[0].level > 1) {
-      results[0].nextLevel();
-      Node *ac = results[0].alreadyChecked;
-      if (ac &&
-          less(ac->value(), ac->length(), endValue.data(), endValue.size()))
-        break;
-    }
-
-    // Init all the other fingers to start descending where we stopped
-    //   the first one
-
-    // SOMEDAY: this loop showed up on vtune, could be faster?
-    // vtune: 8 parts
-    int startLevel = results[0].level + 1;
-    Node *x = startLevel < MaxLevels ? results[0].finger[startLevel] : header;
-    for (int i = 1; i < count; i++) {
-      results[i].level = startLevel;
-      results[i].x = x;
-      results[i].alreadyChecked = nullptr;
-      results[i].value = values[i];
-      for (int j = startLevel; j < MaxLevels; j++)
-        results[i].finger[j] = results[0].finger[j];
-    }
-
-    int *nextJob = temp;
-    for (int i = 0; i < count - 1; i++)
-      nextJob[i] = i + 1;
-    nextJob[count - 1] = 0;
-
-    int prevJob = count - 1;
-    int job = 0;
-
-    // vtune: 225 parts
-    while (true) {
-      Finger *f = &results[job];
-      f->advance();
-      if (f->finished()) {
-        if (prevJob == job)
-          break;
-        nextJob[prevJob] = nextJob[job];
-      } else {
-        f->prefetch();
-        prevJob = job;
-      }
-      job = nextJob[job];
-    }
-  }
-
-  int removeBefore(Version v, Finger &f, int nodeCount) {
-    // f.x, f.alreadyChecked?
-
-    int removedCount = 0;
-    bool wasAbove = true;
-    while (nodeCount--) {
-      Node *x = f.finger[0]->getNext(0);
-      if (!x)
-        break;
-
-      // double prefetch gives +25% speed (single threaded)
-      Node *next = x->getNext(0);
-      __builtin_prefetch(next);
-      next = x->getNext(1);
-      __builtin_prefetch(next);
-
-      bool isAbove = x->getMaxVersion(0) >= v;
-      if (isAbove || wasAbove) { // f.nextItem
-        for (int l = 0; l <= x->level(); l++)
-          f.finger[l] = x;
-      } else { // f.eraseItem
-        removedCount++;
-        for (int l = 0; l <= x->level(); l++)
-          f.finger[l]->setNext(l, x->getNext(l));
-        for (int i = 1; i <= x->level(); i++)
-          f.finger[i]->setMaxVersion(
-              i, std::max(f.finger[i]->getMaxVersion(i), x->getMaxVersion(i)));
-        x->destroy();
-      }
-      wasAbove = isAbove;
-    }
-
-    return removedCount;
-  }
-
-private:
-  void remove(const Finger &start, const Finger &end) {
-    if (start.finger[0] == end.finger[0])
-      return;
-
-    Node *x = start.finger[0]->getNext(0);
-
-    // vtune says: this loop is the expensive parts (6 parts)
-    for (int i = 0; i < MaxLevels; i++)
-      if (start.finger[i] != end.finger[i])
-        start.finger[i]->setNext(i, end.finger[i]->getNext(i));
-
-    while (true) {
-      Node *next = x->getNext(0);
-      x->destroy();
-      if (x == end.finger[0])
-        break;
-      x = next;
-    }
-  }
-
-  void insert(const Finger &f, Version version) {
-    int level = randomLevel();
-    // std::cout << std::string((const char*)value,length) << " level: " <<
-    // level << std::endl;
-    Node *x = Node::create(f.value, level);
-    x->setMaxVersion(0, version);
-    for (int i = 0; i <= level; i++) {
-      x->setNext(i, f.finger[i]->getNext(i));
-      f.finger[i]->setNext(i, x);
-    }
-    // vtune says: this loop is the costly part of this function
-    for (int i = 1; i <= level; i++) {
-      f.finger[i]->calcVersionForLevel(i);
-      x->calcVersionForLevel(i);
-    }
-    for (int i = level + 1; i < MaxLevels; i++) {
-      Version v = f.finger[i]->getMaxVersion(i);
-      if (v >= version)
-        break;
-      f.finger[i]->setMaxVersion(i, version);
-    }
-  }
-
-  struct CheckMax {
-    Finger start, end;
-    Version version;
-    ConflictSet::Result *result;
-    int state;
-
-    void init(const ReadConflictRange &r, Node *header,
-              ConflictSet::Result *result) {
-      this->start.init(r.begin, header);
-      this->end.init(r.end, header);
-      this->version = r.version;
-      this->state = 0;
-      this->result = result;
-    }
-
-    bool noConflict() const { return true; }
-    bool conflict() {
-      *result = ConflictSet::Conflict;
-      return true;
-    }
-
-    // Return true if finished
-    force_inline bool advance() {
-      if (*result == ConflictSet::TooOld) {
-        return true;
-      }
-      switch (state) {
-      case 0:
-        // find where start and end fingers diverge
-        while (true) {
-          if (!start.advance()) {
-            start.prefetch();
-            return false;
-          }
-          end.x = start.x;
-          while (!end.advance())
-            ;
-
-          int l = start.level;
-          if (start.finger[l] != end.finger[l])
-            break;
-          // accept if the range spans the check range, but does not have a
-          // greater version
-          if (start.finger[l]->getMaxVersion(l) <= version)
-            return noConflict();
-          if (l == 0)
-            return conflict();
-        }
-        state = 1;
-      case 1: {
-        // check the end side of the pyramid
-        Node *e = end.finger[end.level];
-        while (e->getMaxVersion(end.level) > version) {
-          if (end.finished())
-            return conflict();
-          end.nextLevel();
-          Node *f = end.finger[end.level];
-          while (e != f) {
-            if (e->getMaxVersion(end.level) > version)
-              return conflict();
-            e = e->getNext(end.level);
-          }
-        }
-
-        // check the start side of the pyramid
-        Node *s = end.finger[start.level];
-        while (true) {
-          Node *nextS = start.finger[start.level]->getNext(start.level);
-          Node *p = nextS;
-          while (p != s) {
-            if (p->getMaxVersion(start.level) > version)
-              return conflict();
-            p = p->getNext(start.level);
-          }
-          if (start.finger[start.level]->getMaxVersion(start.level) <= version)
-            return noConflict();
-          s = nextS;
-          if (start.finished()) {
-            if (nextS->length() == start.value.size() &&
-                !memcmp(nextS->value(), start.value.data(), start.value.size()))
-              return noConflict();
-            else
-              return conflict();
-          }
-          start.nextLevel();
-        }
-      }
-      default:
-        __builtin_unreachable();
-      }
-    }
-  };
-};
-
-struct SkipListConflictSet {
-  SkipListConflictSet(int64_t oldestVersion)
-      : oldestVersion(oldestVersion), skipList(oldestVersion) {}
-  void check(const ConflictSet::ReadRange *reads, ConflictSet::Result *results,
-             int count) const {
-    Arena arena;
-    auto *ranges = new (arena) ReadConflictRange[count];
-    for (int i = 0; i < count; ++i) {
-      ranges[i].begin = {reads[i].begin.p, size_t(reads[i].begin.len)};
-      ranges[i].end = {reads[i].end.p,
-                       size_t(reads[i].end.len == 0 ? reads[i].begin.len + 1
-                                                    : reads[i].end.len)};
-      ranges[i].version = reads[i].readVersion;
-      if (reads[i].readVersion < oldestVersion) {
-        results[i] = ConflictSet::TooOld;
-      } else {
-        results[i] = ConflictSet::Commit;
-      }
-    }
-    skipList.detectConflicts(ranges, count, results);
-  }
-
-  void addWrites(const ConflictSet::WriteRange *writes, int count) {
-    Arena arena;
-    const int stringCount = count * 2;
-
-    const int stripeSize = 16;
-    SkipList::Finger fingers[stripeSize];
-    int temp[stripeSize];
-    int stripes = (stringCount + stripeSize - 1) / stripeSize;
-    StringRef values[stripeSize];
-    int64_t writeVersions[stripeSize / 2];
-    int ss = stringCount - (stripes - 1) * stripeSize;
-    for (int s = stripes - 1; s >= 0; s--) {
-      for (int i = 0; i * 2 < ss; ++i) {
-        const auto &w = writes[s * stripeSize / 2 + i];
-        values[i * 2] = {w.begin.p, size_t(w.begin.len)};
-        values[i * 2 + 1] = {
-            w.end.p, size_t(w.end.len == 0 ? w.begin.len + 1 : w.end.len)};
-        writeVersions[i] = w.writeVersion;
-        keyUpdates += 2;
-      }
-      skipList.find(values, fingers, temp, ss);
-      skipList.addConflictRanges(fingers, ss / 2, writeVersions);
-      ss = stripeSize;
-    }
-  }
-
-  void setOldestVersion(int64_t 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, 0));
-    removalKey = std::basic_string<uint8_t>(finger.getValue().data(),
-                                            finger.getValue().size());
-  }
-
-private:
-  int64_t keyUpdates = 0;
-  std::basic_string<uint8_t> removalKey;
-  int64_t oldestVersion;
-  SkipList skipList;
-};
-
-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;
-  return {key.data(), int(key.size()), r.data(), int(r.size())};
-}
-
-ConflictSet::ReadRange prefixRange(Arena &arena, std::span<const uint8_t> key) {
   int index;
   for (index = key.size() - 1; index >= 0; index--)
     if ((key[index]) != 255)
@@ -629,39 +48,18 @@ 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);
-  return {key.data(), int(key.size()), r.data(), int(r.size())};
+  buf[r.size() - 1]++;
+  return {{key.data(), int(key.size())}, {r.data(), int(r.size())}, 0};
 }
 
-} // namespace
+void benchConflictSet() {
-
-constexpr int kNumKeys = 1000000;
-
-constexpr int kOpsPerTx = 100;
-
-constexpr int kPrefixLen = 0;
-
-constexpr int kMvccWindow = 100000;
-
-std::span<const uint8_t> makeKey(Arena &arena, int index) {
-
-  auto result =
-      std::span<uint8_t>{new (arena) uint8_t[4 + kPrefixLen], 4 + kPrefixLen};
-  index = __builtin_bswap32(index);
-  memset(result.data(), 0, kPrefixLen);
-  memcpy(result.data() + kPrefixLen, &index, 4);
-
-  return result;
-}
-
-template <class ConflictSet_> void benchConflictSet(const std::string &name) {
   ankerl::nanobench::Bench bench;
-  ConflictSet_ cs{0};
+  ConflictSet cs{0};
 
   bench.batch(kOpsPerTx);
-  bench.minEpochIterations(2000);
 
   int64_t version = 0;
 
@@ -678,21 +76,13 @@ template <class ConflictSet_> void benchConflictSet(const std::string &name) {
       w.begin.len = r.begin.len;
       w.end.p = r.end.p;
       w.end.len = 0;
-      w.writeVersion = version + 1;
       writes.push_back(w);
     }
-    cs.addWrites(writes.data(), writes.size());
+    cs.addWrites(writes.data(), writes.size(), version + 1);
     ++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?
@@ -712,11 +102,10 @@ template <class ConflictSet_> void benchConflictSet(const std::string &name) {
       w.begin.len = begin.size();
       w.end.p = end.data();
       w.end.len = end.size();
-      w.writeVersion = version + 1;
       writes.push_back(w);
     }
+    cs.addWrites(writes.data(), kOpsPerTx, version + 1);
     ++version;
-    cs.addWrites(writes.data(), kOpsPerTx);
   }
 
   {
@@ -732,7 +121,7 @@ template <class ConflictSet_> void benchConflictSet(const std::string &name) {
 
     auto *results = new (arena) ConflictSet::Result[kOpsPerTx];
 
-    bench.run(name + " (point reads)",
+    bench.run("point reads",
               [&]() { cs.check(reads.data(), results, kOpsPerTx); });
   }
 
@@ -748,7 +137,7 @@ template <class ConflictSet_> void benchConflictSet(const std::string &name) {
 
     auto *results = new (arena) ConflictSet::Result[kOpsPerTx];
 
-    bench.run(name + " (prefix reads)",
+    bench.run("prefix reads",
               [&]() { cs.check(reads.data(), results, kOpsPerTx); });
   }
 
@@ -769,7 +158,7 @@ template <class ConflictSet_> void benchConflictSet(const std::string &name) {
 
     auto *results = new (arena) ConflictSet::Result[kOpsPerTx];
 
-    bench.run(name + " (range reads)",
+    bench.run("range reads",
               [&]() { cs.check(reads.data(), results, kOpsPerTx); });
   }
 
@@ -787,13 +176,14 @@ template <class ConflictSet_> void benchConflictSet(const std::string &name) {
       ++iter;
     }
 
-    bench.run(name + " (point writes)", [&]() {
+    while (version < kMvccWindow) {
       auto v = ++version;
-      for (auto &w : writes) {
+      cs.addWrites(writes.data(), 1, v);
-        w.writeVersion = v;
+    }
-      }
+
-      cs.addWrites(writes.data(), writes.size());
+    bench.run("point writes", [&]() {
-      cs.setOldestVersion(std::max<int64_t>(version - kMvccWindow, 0));
+      auto v = ++version;
+      cs.addWrites(writes.data(), writes.size(), v);
     });
   }
 
@@ -811,13 +201,9 @@ template <class ConflictSet_> void benchConflictSet(const std::string &name) {
       ++iter;
     }
 
-    bench.run(name + " (prefix writes)", [&]() {
+    bench.run("prefix writes", [&]() {
       auto v = ++version;
-      for (auto &w : writes) {
+      cs.addWrites(writes.data(), writes.size(), v);
-        w.writeVersion = v;
-      }
-      cs.addWrites(writes.data(), writes.size());
-      cs.setOldestVersion(std::max<int64_t>(version - kMvccWindow, 0));
     });
   }
 
@@ -835,45 +221,150 @@ template <class ConflictSet_> void benchConflictSet(const std::string &name) {
       writes.push_back(w);
     }
 
-    bench.run(name + " (range writes)", [&]() {
+    bench.run("range writes", [&]() {
       auto v = ++version;
-      for (auto &w : writes) {
+      cs.addWrites(writes.data(), writes.size(), v);
-        w.writeVersion = v;
+    });
-      }
+  }
-      cs.addWrites(writes.data(), writes.size());
+
-      cs.setOldestVersion(std::max<int64_t>(version - kMvccWindow, 0));
+  bench.batch(1);
+
+  bench.warmup(10000);
+
+  {
+    bench.run("monotonic increasing point writes", [&]() {
+      auto v = ++version;
+      ConflictSet::WriteRange w;
+
+      uint8_t b[9];
+      b[8] = 0;
+      auto x = __builtin_bswap64(version);
+      memcpy(b, &x, 8);
+
+      w.begin.p = b;
+      w.begin.len = 8;
+      w.end.len = 0;
+      w.end.p = b;
+      cs.addWrites(&w, 1, v);
+      cs.setOldestVersion(version - kMvccWindow);
     });
   }
 }
 
-int main(void) {
+constexpr int kKeyLenForWorstCase = 50;
-  benchConflictSet<SkipListConflictSet>("skip list");
+
-  benchConflictSet<ConflictSet>("radix tree");
+ConflictSet worstCaseConflictSetForRadixRangeRead(int cardinality) {
+  ConflictSet cs{0};
+
+  for (int i = 0; i < kKeyLenForWorstCase; ++i) {
+    for (int j = 0; j < cardinality; ++j) {
+      auto b = std::vector<uint8_t>(i, 0);
+      b.push_back(j);
+      auto e = std::vector<uint8_t>(i, 255);
+      e.push_back(255 - j);
+      weaselab::ConflictSet::WriteRange w[] = {{
+                                                   {b.data(), int(b.size())},
+                                                   {nullptr, 0},
+                                               },
+                                               {
+                                                   {e.data(), int(e.size())},
+                                                   {nullptr, 0},
+                                               }};
+      std::sort(std::begin(w), std::end(w),
+                [](const auto &lhs, const auto &rhs) {
+                  int cl = std::min(lhs.begin.len, rhs.begin.len);
+                  if (cl > 0) {
+                    int c = memcmp(lhs.begin.p, rhs.begin.p, cl);
+                    if (c != 0) {
+                      return c < 0;
+                    }
+                  }
+                  return lhs.begin.len < rhs.begin.len;
+                });
+      cs.addWrites(w, sizeof(w) / sizeof(w[0]), 0);
+    }
+  }
+
+  // Defeat short-circuiting on the left
+  {
+    auto k = std::vector<uint8_t>(kKeyLenForWorstCase, 0);
+    weaselab::ConflictSet::WriteRange w[] = {
+        {
+            {k.data(), int(k.size())},
+            {nullptr, 0},
+        },
+    };
+    cs.addWrites(w, sizeof(w) / sizeof(w[0]), 1);
+  }
+
+  // Defeat short-circuiting on the right
+  {
+    auto k = std::vector<uint8_t>(kKeyLenForWorstCase, 255);
+    weaselab::ConflictSet::WriteRange w[] = {
+        {
+            {k.data(), int(k.size())},
+            {nullptr, 0},
+        },
+    };
+    cs.addWrites(w, sizeof(w) / sizeof(w[0]), 1);
+  }
+
+  return cs;
 }
 
-// extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+void benchWorstCaseForRadixRangeRead() {
-//   TestDriver<SkipListConflictSet> driver{data, size};
+  ankerl::nanobench::Bench bench;
 
-//   for (;;) {
+  std::unique_ptr<ConflictSet> cs[256];
-//     bool done = driver.next();
+  for (int i = 0; i < 256; ++i) {
-//     if (!driver.ok) {
+    cs[i] =
-//       // debugPrintDot(stdout, driver.cs.root);
+        std::make_unique<ConflictSet>(worstCaseConflictSetForRadixRangeRead(i));
-//       // fflush(stdout);
+  }
-//       abort();
-//     }
-// #if DEBUG_VERBOSE && !defined(NDEBUG)
-//     fprintf(stderr, "Check correctness\n");
-// #endif
-//     // bool success = checkCorrectness(driver.cs.root);
-//     // if (!success) {
-//     //   debugPrintDot(stdout, driver.cs.root);
-//     //   fflush(stdout);
-//     //   abort();
-//     // }
-//     if (done) {
-//       break;
-//     }
-//   }
 
-//   return 0;
+  auto begin = std::vector<uint8_t>(kKeyLenForWorstCase - 1, 0);
-// }
+  begin.push_back(1);
+  auto end = std::vector<uint8_t>(kKeyLenForWorstCase - 1, 255);
+  end.push_back(254);
+
+  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) {
+      cs[i]->check(r, results, sizeof(r) / sizeof(r[0]));
+      for (auto result : results) {
+        if (result != weaselab::ConflictSet::Commit) {
+          abort();
+        }
+      }
+    }
+  });
+
+  // for (int i = 0; i < 256; ++i) {
+  //   bench.run("worst case for radix tree, span " + std::to_string(i), [&]() {
+  //     result = weaselab::ConflictSet::TooOld;
+  //     cs[i]->check(&r, &result, 1);
+  //     if (result != weaselab::ConflictSet::Commit) {
+  //       abort();
+  //     }
+  //   });
+  // }
+}
+
+void benchCreateAndDestroy() {
+  ankerl::nanobench::Bench bench;
+
+  bench.run("create and destroy", [&]() { ConflictSet cs{0}; });
+}
+
+int main(void) {
+  benchConflictSet();
+  benchWorstCaseForRadixRangeRead();
+  benchCreateAndDestroy();
+}

CMakeLists.txt

@@ -1,12 +1,22 @@
 cmake_minimum_required(VERSION 3.18)
 project(
-  conflict_set
+  conflict-set
-  VERSION 0.0.1
+  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"
   LANGUAGES C CXX)
 set(CMAKE_CXX_STANDARD 20)
 
+file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/version.txt ${PROJECT_VERSION})
+configure_file(${CMAKE_CURRENT_SOURCE_DIR}/version.txt.in
+               ${CMAKE_CURRENT_SOURCE_DIR}/paper/version.txt)
+
+include(CMakePushCheckState)
+include(CheckCXXCompilerFlag)
+include(CheckIncludeFileCXX)
+include(CheckCXXSourceCompiles)
+
 set(DEFAULT_BUILD_TYPE "Release")
 
 if(NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
@@ -21,16 +31,72 @@ if(NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
                                                "MinSizeRel" "RelWithDebInfo")
 endif()
 
-add_compile_options(-fdata-sections -ffunction-sections)
+add_compile_options(
+  -Werror=switch-enum -Wswitch-enum -fPIC -fdata-sections -ffunction-sections
+  -fno-jump-tables # https://github.com/llvm/llvm-project/issues/54247
+)
+
+if(CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+  add_link_options("-Wno-unused-command-line-argument")
+  find_program(LLVM_OBJCOPY llvm-objcopy)
+  if(LLVM_OBJCOPY)
+    set(CMAKE_OBJCOPY
+        ${LLVM_OBJCOPY}
+        CACHE FILEPATH "path to objcopy binary" FORCE)
+  endif()
+endif()
+
+if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
+  add_compile_options("-Wno-maybe-uninitialized")
+endif()
+
+if(NOT APPLE)
+  # This causes some versions of clang to crash on macos
+  add_compile_options(-g -fno-omit-frame-pointer)
+endif()
+
+set(full_relro_flags "-pie;LINKER:-z,relro,-z,now,-z,noexecstack")
+cmake_push_check_state()
+list(APPEND CMAKE_REQUIRED_LINK_OPTIONS ${full_relro_flags})
+check_cxx_source_compiles("int main(){}" HAS_FULL_RELRO FAIL_REGEX "warning:")
+if(HAS_FULL_RELRO)
+  add_link_options(${full_relro_flags})
+endif()
+cmake_pop_check_state()
+
+if(CMAKE_SYSTEM_PROCESSOR STREQUAL aarch64)
+  add_compile_options(-mbranch-protection=standard)
+else()
+  add_compile_options(-fcf-protection)
+  set(rewrite_endbr_flags "-fuse-ld=mold;LINKER:-z,rewrite-endbr")
+  cmake_push_check_state()
+  list(APPEND CMAKE_REQUIRED_LINK_OPTIONS ${rewrite_endbr_flags})
+  check_cxx_source_compiles("int main(){}" HAS_REWRITE_ENDBR FAIL_REGEX
+                            "warning:")
+  if(HAS_REWRITE_ENDBR)
+    add_link_options(${rewrite_endbr_flags})
+  endif()
+  cmake_pop_check_state()
+endif()
+
+set(version_script_flags
+    LINKER:--version-script=${CMAKE_CURRENT_SOURCE_DIR}/linker.map)
+cmake_push_check_state()
+list(APPEND CMAKE_REQUIRED_LINK_OPTIONS ${version_script_flags})
+check_cxx_source_compiles("int main(){}" HAS_VERSION_SCRIPT FAIL_REGEX
+                          "warning:")
+cmake_pop_check_state()
+
+option(USE_SIMD_FALLBACK
+       "Use fallback implementations of functions that use SIMD" OFF)
+
+option(DISABLE_TSAN "Disable TSAN" OFF)
 
 # This is encouraged according to
 # https://valgrind.org/docs/manual/manual-core-adv.html#manual-core-adv.clientreq
-include_directories(SYSTEM ${CMAKE_SOURCE_DIR}/third_party/valgrind)
+include_directories(SYSTEM ${CMAKE_CURRENT_SOURCE_DIR}/third_party/valgrind)
 
-if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
+add_compile_options($<$<COMPILE_LANGUAGE:CXX>:-Wno-invalid-offsetof>)
-  add_compile_options(-Wno-maybe-uninitialized
-                      $<$<COMPILE_LANGUAGE:CXX>:-Wno-invalid-offsetof>)
-endif()
 
 if(APPLE)
   add_link_options(-Wl,-dead_strip)
@@ -38,66 +104,130 @@ else()
   add_link_options(-Wl,--gc-sections)
 endif()
 
-include(CheckIncludeFileCXX)
+if(NOT USE_SIMD_FALLBACK)
-include(CMakePushCheckState)
+  cmake_push_check_state()
+  list(APPEND CMAKE_REQUIRED_FLAGS -mavx)
+  check_include_file_cxx("immintrin.h" HAS_AVX)
+  if(HAS_AVX)
+    add_compile_options(-mavx)
+    add_compile_definitions(HAS_AVX)
+  endif()
+  cmake_pop_check_state()
 
-cmake_push_check_state()
+  check_include_file_cxx("arm_neon.h" HAS_ARM_NEON)
-list(APPEND CMAKE_REQUIRED_FLAGS -mavx)
+  if(HAS_ARM_NEON)
-check_include_file_cxx("immintrin.h" HAS_AVX)
+    add_compile_definitions(HAS_ARM_NEON)
-if(HAS_AVX)
+  endif()
-  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()
 
 set(CMAKE_CXX_IMPLICIT_LINK_LIBRARIES "")
 
-add_library(${PROJECT_NAME}_object OBJECT ConflictSet.cpp)
+add_library(${PROJECT_NAME}-object OBJECT ConflictSet.cpp)
-target_compile_options(${PROJECT_NAME}_object PRIVATE -fPIC -fno-exceptions
+target_compile_options(${PROJECT_NAME}-object PRIVATE -fno-exceptions
                                                       -fvisibility=hidden)
-target_include_directories(${PROJECT_NAME}_object
+target_include_directories(${PROJECT_NAME}-object
-                           PRIVATE ${CMAKE_SOURCE_DIR}/include)
+                           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)
-if(NOT CMAKE_BUILD_TYPE STREQUAL Debug)
+set_target_properties(
+  ${PROJECT_NAME} PROPERTIES LIBRARY_OUTPUT_DIRECTORY
+                             "${CMAKE_CURRENT_BINARY_DIR}/radix_tree")
+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()
 
+if(HAS_VERSION_SCRIPT)
+  target_link_options(
+    ${PROJECT_NAME} PRIVATE
+    LINKER:--version-script=${CMAKE_CURRENT_SOURCE_DIR}/linker.map)
+endif()
+
+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(NOT APPLE)
-  target_link_options(${PROJECT_NAME} PRIVATE
-                      LINKER:--version-script=${CMAKE_SOURCE_DIR}/linker.map)
-endif()
-
-add_library(${PROJECT_NAME}_static STATIC
-            $<TARGET_OBJECTS:${PROJECT_NAME}_object>)
-if(NOT CMAKE_BUILD_TYPE STREQUAL Debug)
-  set_target_properties(${PROJECT_NAME}_static PROPERTIES LINKER_LANGUAGE C)
-endif()
-
-if(NOT APPLE AND CMAKE_OBJCOPY)
   add_custom_command(
-    TARGET conflict_set_static
+    TARGET ${PROJECT_NAME}-static
     POST_BUILD
     COMMAND
-      ${CMAKE_OBJCOPY} --keep-global-symbols=${CMAKE_SOURCE_DIR}/symbols.txt
+      ${CMAKE_OBJCOPY}
-      $<TARGET_FILE:${PROJECT_NAME}_static>)
+      --keep-global-symbols=${CMAKE_CURRENT_SOURCE_DIR}/symbol-exports.txt
+      $<TARGET_FILE:${PROJECT_NAME}-static> || echo
+      "Proceeding with all symbols global in static library")
 endif()
 
-set(TEST_FLAGS -Wall -Wextra -Wpedantic -Wunreachable-code -UNDEBUG)
-
 include(CTest)
 
-if(BUILD_TESTING)
+# 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
+  # coverage with a small number of tests.
+
+  file(GLOB CORPUS_TESTS ${CMAKE_CURRENT_SOURCE_DIR}/corpus/*)
+
+  # Shared library version of FoundationDB's skip list implementation
+  add_library(skip_list SHARED SkipList.cpp)
+  target_compile_options(skip_list PRIVATE -fno-exceptions -fvisibility=hidden)
+  target_include_directories(skip_list
+                             PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
+  set_target_properties(
+    skip_list PROPERTIES LIBRARY_OUTPUT_DIRECTORY
+                         "${CMAKE_CURRENT_BINARY_DIR}/skip_list")
+  set_target_properties(skip_list PROPERTIES OUTPUT_NAME ${PROJECT_NAME})
+  set_target_properties(skip_list PROPERTIES VERSION ${PROJECT_VERSION}
+                                             SOVERSION ${PROJECT_VERSION_MAJOR})
+
+  # Shared library version of a std::unordered_map-based conflict set (point
+  # queries only)
+  add_library(hash_table SHARED HashTable.cpp)
+  target_compile_options(hash_table PRIVATE -fno-exceptions -fvisibility=hidden)
+  target_include_directories(hash_table
+                             PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
+  set_target_properties(
+    hash_table PROPERTIES LIBRARY_OUTPUT_DIRECTORY
+                          "${CMAKE_CURRENT_BINARY_DIR}/hash_table")
+  set_target_properties(hash_table PROPERTIES OUTPUT_NAME ${PROJECT_NAME})
+  set_target_properties(
+    hash_table PROPERTIES VERSION ${PROJECT_VERSION} SOVERSION
+                                                     ${PROJECT_VERSION_MAJOR})
+
+  add_executable(driver_skip_list TestDriver.cpp)
+  target_compile_options(driver_skip_list PRIVATE ${TEST_FLAGS})
+  target_link_libraries(driver_skip_list PRIVATE skip_list)
+
+  # enable to test skip list
+  if(0)
+    foreach(TEST ${CORPUS_TESTS})
+      get_filename_component(hash ${TEST} NAME)
+      add_test(NAME skip_list_${hash} COMMAND driver_skip_list ${TEST})
+    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
@@ -120,12 +250,13 @@ if(BUILD_TESTING)
     endif()
   endif()
 
-  # corpus tests
+  # whitebox tests
-
-  file(GLOB CORPUS_TESTS ${CMAKE_SOURCE_DIR}/corpus/*)
-
   add_executable(fuzz_driver ConflictSet.cpp FuzzTestDriver.cpp)
   target_compile_options(fuzz_driver PRIVATE ${TEST_FLAGS})
+  if(NOT CMAKE_CROSSCOMPILING)
+    target_compile_options(fuzz_driver PRIVATE -fsanitize=address,undefined)
+    target_link_options(fuzz_driver PRIVATE -fsanitize=address,undefined)
+  endif()
   target_compile_definitions(fuzz_driver PRIVATE ENABLE_FUZZ)
   target_include_directories(fuzz_driver
                              PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include)
@@ -134,66 +265,195 @@ if(BUILD_TESTING)
     add_test(NAME conflict_set_fuzz_${hash} COMMAND fuzz_driver ${TEST})
   endforeach()
 
+  # tsan tests
+  if(NOT CMAKE_CROSSCOMPILING AND NOT DISABLE_TSAN)
+    add_executable(tsan_driver ConflictSet.cpp FuzzTestDriver.cpp)
+    target_compile_options(tsan_driver PRIVATE ${TEST_FLAGS} -fsanitize=thread)
+    target_link_options(tsan_driver PRIVATE -fsanitize=thread)
+    target_compile_definitions(tsan_driver PRIVATE ENABLE_FUZZ THREAD_TEST)
+    target_include_directories(tsan_driver
+                               PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include)
+    foreach(TEST ${CORPUS_TESTS})
+      get_filename_component(hash ${TEST} NAME)
+      add_test(NAME conflict_set_tsan_${hash} COMMAND tsan_driver ${TEST})
+    endforeach()
+  endif()
+
+  # blackbox tests
   add_executable(driver TestDriver.cpp)
   target_compile_options(driver PRIVATE ${TEST_FLAGS})
   target_link_libraries(driver PRIVATE ${PROJECT_NAME})
-
-  find_program(VALGRIND_EXE valgrind)
-  if(VALGRIND_EXE)
-    add_test(NAME conflict_set_blackbox_valgrind
-             COMMAND ${VALGRIND_EXE} --error-exitcode=99 --
-                     $<TARGET_FILE:driver> ${CORPUS_TESTS})
-  endif()
-
   foreach(TEST ${CORPUS_TESTS})
     get_filename_component(hash ${TEST} NAME)
     add_test(NAME conflict_set_blackbox_${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)
+    find_package(Python3 REQUIRED COMPONENTS Interpreter)
+    set_property(
+      DIRECTORY
+      APPEND
+      PROPERTY CMAKE_CONFIGURE_DEPENDS
+               ${CMAKE_CURRENT_SOURCE_DIR}/test_conflict_set.py)
+    execute_process(
+      COMMAND ${Python3_EXECUTABLE}
+              ${CMAKE_CURRENT_SOURCE_DIR}/test_conflict_set.py list
+      OUTPUT_VARIABLE SCRIPT_TESTS)
+    foreach(TEST ${SCRIPT_TESTS})
+      add_test(
+        NAME script_test_${TEST}
+        COMMAND
+          ${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()
+
   # api smoke tests
 
   # c90
   add_executable(conflict_set_c_api_test conflict_set_c_api_test.c)
   target_compile_options(conflict_set_c_api_test PRIVATE ${TEST_FLAGS})
   target_link_libraries(conflict_set_c_api_test PRIVATE ${PROJECT_NAME})
-  set_property(TARGET conflict_set_c_api_test PROPERTY C_STANDARD 90)
+  set_target_properties(conflict_set_c_api_test PROPERTIES C_STANDARD 90)
-  set_property(TARGET conflict_set_c_api_test PROPERTY C_STANDARD_REQUIRED ON)
+  set_target_properties(conflict_set_c_api_test PROPERTIES C_STANDARD_REQUIRED
+                                                           ON)
   add_test(NAME conflict_set_c_api_test COMMAND conflict_set_c_api_test)
 
   # c++98
   add_executable(conflict_set_cxx_api_test conflict_set_cxx_api_test.cpp)
   target_compile_options(conflict_set_cxx_api_test PRIVATE ${TEST_FLAGS})
-  target_link_libraries(conflict_set_cxx_api_test PRIVATE ${PROJECT_NAME})
+  target_link_libraries(conflict_set_cxx_api_test
-  set_property(TARGET conflict_set_cxx_api_test PROPERTY CXX_STANDARD 98)
+                        PRIVATE ${PROJECT_NAME}-static)
-  set_property(TARGET conflict_set_cxx_api_test PROPERTY CXX_STANDARD_REQUIRED
+  set_target_properties(conflict_set_cxx_api_test PROPERTIES CXX_STANDARD 98)
-                                                         ON)
+  set_target_properties(conflict_set_cxx_api_test
+                        PROPERTIES CXX_STANDARD_REQUIRED ON)
   add_test(NAME conflict_set_cxx_api_test COMMAND conflict_set_cxx_api_test)
 
-  if(NOT APPLE AND NOT CMAKE_BUILD_TYPE STREQUAL Debug)
+  # symbol visibility tests
+  if(NOT CMAKE_BUILD_TYPE STREQUAL Debug)
+    if(APPLE)
+      set(symbol_exports ${CMAKE_CURRENT_SOURCE_DIR}/apple-symbol-exports.txt)
+      set(symbol_imports ${CMAKE_CURRENT_SOURCE_DIR}/apple-symbol-imports.txt)
+    else()
+      set(symbol_exports ${CMAKE_CURRENT_SOURCE_DIR}/symbol-exports.txt)
+      if(CMAKE_SYSTEM_PROCESSOR STREQUAL aarch64)
+        set(symbol_imports
+            ${CMAKE_CURRENT_SOURCE_DIR}/aarch64-symbol-imports.txt)
+      else()
+        set(symbol_imports ${CMAKE_CURRENT_SOURCE_DIR}/symbol-imports.txt)
+      endif()
+    endif()
     add_test(
       NAME conflict_set_shared_symbols
-      COMMAND ${CMAKE_SOURCE_DIR}/test_symbols.sh
+      COMMAND
-              $<TARGET_FILE:${PROJECT_NAME}> ${CMAKE_SOURCE_DIR}/symbols.txt)
+        ${CMAKE_CURRENT_SOURCE_DIR}/test_symbols.sh
+        $<TARGET_FILE:${PROJECT_NAME}> ${symbol_exports} ${symbol_imports})
     add_test(
       NAME conflict_set_static_symbols
       COMMAND
-        ${CMAKE_SOURCE_DIR}/test_symbols.sh
+        ${CMAKE_CURRENT_SOURCE_DIR}/test_symbols.sh
-        $<TARGET_FILE:${PROJECT_NAME}_static> ${CMAKE_SOURCE_DIR}/symbols.txt)
+        $<TARGET_FILE:${PROJECT_NAME}-static> ${symbol_exports}
+        ${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)
-  # target_compile_options(conflict_set_bench PRIVATE
+  set_target_properties(conflict_set_bench PROPERTIES SKIP_BUILD_RPATH ON)
-  # "-fsanitize=address,undefined,fuzzer")
+  add_executable(real_data_bench RealDataBench.cpp)
-  # target_link_options(conflict_set_bench PRIVATE
+  target_link_libraries(real_data_bench PRIVATE ${PROJECT_NAME})
-  # "-fsanitize=address,undefined,fuzzer")
+  set_target_properties(real_data_bench PROPERTIES SKIP_BUILD_RPATH ON)
+
+  # fuzzer-based perf
+  add_executable(driver_perf TestDriver.cpp)
+  target_compile_definitions(driver_perf PRIVATE PERF_TEST=1)
+  target_link_libraries(driver_perf PRIVATE ${PROJECT_NAME})
+
+  # server bench
+  add_executable(server_bench ServerBench.cpp)
+  target_link_libraries(server_bench PRIVATE ${PROJECT_NAME})
+  set_target_properties(server_bench PROPERTIES SKIP_BUILD_RPATH ON)
+
+  add_executable(interleaving_test InterleavingTest.cpp)
+  # work around lack of musttail for gcc
+  if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND CMAKE_BUILD_TYPE STREQUAL "Debug")
+    target_compile_options(interleaving_test PRIVATE -Og
+                                                     -foptimize-sibling-calls)
+  endif()
+  target_link_libraries(interleaving_test PRIVATE nanobench)
 endif()
+
 # packaging
 
 set(CPACK_PACKAGE_CONTACT andrew@weaselab.dev)
-set(CPACK_DEBIAN_PACKAGE_SHLIBDEPS ON)
+set(CMAKE_INSTALL_DEFAULT_COMPONENT_NAME all)
+
+set(CPACK_PACKAGE_VENDOR "Weaselab")
+set(CPACK_RESOURCE_FILE_LICENSE "${CMAKE_CURRENT_SOURCE_DIR}/LICENSE")
+set(CPACK_RESOURCE_FILE_README "${CMAKE_CURRENT_SOURCE_DIR}/README.md")
+
+# rpm
+set(CPACK_RPM_PACKAGE_ARCHITECTURE ${CMAKE_SYSTEM_PROCESSOR})
+set(CPACK_RPM_SPEC_INSTALL_POST "/bin/true") # avoid stripping
+set(CPACK_RPM_PACKAGE_LICENSE "Apache 2.0")
+set(CPACK_RPM_FILE_NAME RPM-DEFAULT)
+
+# deb
+set(CPACK_DEBIAN_FILE_NAME DEB-DEFAULT)
+# see *-imports.txt - dependency versions need to be synced with symbol versions
+if(CMAKE_SYSTEM_PROCESSOR STREQUAL aarch64)
+  set(CPACK_DEBIAN_PACKAGE_DEPENDS "libc6 (>= 2.17)")
+else()
+  set(CPACK_DEBIAN_PACKAGE_DEPENDS "libc6 (>= 2.14)")
+endif()
+
+# macos
+set(CMAKE_OSX_DEPLOYMENT_TARGET 11.0)
+if(APPLE)
+  find_program(PANDOC_EXE pandoc)
+  if(PANDOC_EXE)
+    execute_process(COMMAND ${PANDOC_EXE} ${CMAKE_CURRENT_SOURCE_DIR}/README.md
+                            -o ${CMAKE_CURRENT_BINARY_DIR}/README.txt)
+    set(CPACK_RESOURCE_FILE_README ${CMAKE_CURRENT_BINARY_DIR}/README.txt)
+  endif()
+  configure_file(${CMAKE_CURRENT_SOURCE_DIR}/LICENSE
+                 ${CMAKE_CURRENT_BINARY_DIR}/LICENSE.txt COPYONLY)
+  set(CPACK_RESOURCE_FILE_LICENSE ${CMAKE_CURRENT_BINARY_DIR}/LICENSE.txt)
+endif()
+
 include(CPack)
 
 include(GNUInstallDirs)
@@ -204,7 +464,7 @@ target_include_directories(
          $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/${PROJECT_NAME}>)
 
 target_include_directories(
-  ${PROJECT_NAME}_static
+  ${PROJECT_NAME}-static
   PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
          $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/${PROJECT_NAME}>)
 
@@ -213,13 +473,16 @@ set_target_properties(
                              SOVERSION ${PROJECT_VERSION_MAJOR})
 
 install(
-  TARGETS ${PROJECT_NAME} ${PROJECT_NAME}_static
+  TARGETS ${PROJECT_NAME} ${PROJECT_NAME}-static
-  EXPORT ConflictSetConfig
+  EXPORT ${PROJECT_NAME}Config
   ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
   LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
   RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR})
+
 install(DIRECTORY include/
         DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/${PROJECT_NAME})
 
-install(EXPORT ConflictSetConfig
+install(EXPORT ${PROJECT_NAME}Config
-        DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/ConflictSet/cmake)
+        DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/${PROJECT_NAME}/cmake)
+
+cpack_add_component(all)

Dockerfile

@@ -8,83 +8,51 @@ 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 \
+    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
-RUN curl -Ls https://sourceware.org/pub/valgrind/valgrind-3.20.0.tar.bz2 -o valgrind.tar.bz2 && \
+RUN curl -Ls https://sourceware.org/pub/valgrind/valgrind-3.22.0.tar.bz2 -o valgrind.tar.bz2 && \
-    echo "8536c031dbe078d342f121fa881a9ecd205cb5a78e639005ad570011bdb9f3c6  valgrind.tar.bz2" > valgrind-sha.txt && \
+    echo "c811db5add2c5f729944caf47c4e7a65dcaabb9461e472b578765dd7bf6d2d4c  valgrind.tar.bz2" > valgrind-sha.txt && \
     sha256sum --quiet -c valgrind-sha.txt && \
     mkdir valgrind && \
     tar --strip-components 1 --no-same-owner --no-same-permissions --directory valgrind -xjf valgrind.tar.bz2 && \
     cd valgrind && \
     ./configure --enable-only64bit --enable-lto && \
-    make && \
+    make -j`nproc` && \
     make install && \
     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++
 
-# Install recent flatbuffers from source
-RUN curl -Ls https://github.com/google/flatbuffers/archive/refs/tags/v23.3.3.tar.gz -o flatbuffers.tar.gz && \
-    echo "8aff985da30aaab37edf8e5b02fda33ed4cbdd962699a8e2af98fdef306f4e4d  flatbuffers.tar.gz" > flatbuffers-sha.txt && \
-    sha256sum --quiet -c flatbuffers-sha.txt && \
-    mkdir flatbuffers && \
-    tar --strip-components 1 --no-same-owner --no-same-permissions --directory flatbuffers  -xf flatbuffers.tar.gz && \
-    cd flatbuffers && \
-    cmake -S. -B build -G Ninja -DCMAKE_BUILD_TYPE=Release && \
-    ninja -C build install && \
-    rm -rf /tmp/*
-
-# Build msan-instrumented libc++ (llvmorg-16.0.0)
-RUN curl -Ls https://github.com/llvm/llvm-project/releases/download/llvmorg-16.0.0/llvm-project-16.0.0.src.tar.xz -o llvm-project.tar.gz && \
-    echo "9a56d906a2c81f16f06efc493a646d497c53c2f4f28f0cb1f3c8da7f74350254  llvm-project.tar.gz" > llvm-project-sha.txt && \
-    sha256sum --quiet -c llvm-project-sha.txt && \
-    mkdir llvm-project && \
-    tar --strip-components 1 --no-same-owner --no-same-permissions --directory llvm-project  -xf llvm-project.tar.gz && \
-    cmake -Sllvm-project/runtimes -B build -G Ninja \
-        -DCMAKE_BUILD_TYPE=Release \
-        -DLLVM_ENABLE_RUNTIMES="libcxx;libcxxabi" \
-        -DCMAKE_C_COMPILER=clang \
-        -DCMAKE_CXX_COMPILER=clang++ \
-        -DCMAKE_INSTALL_PREFIX=/opt/llvm-msan \
-        -DLLVM_USE_SANITIZER=MemoryWithOrigins && \
-    ninja -C build cxx cxxabi && \
-    ninja -C build install-cxx install-cxxabi && \
-    rm -rf /tmp/*
-
-# Install bloaty from source
-RUN curl -Ls https://github.com/google/bloaty/releases/download/v1.1/bloaty-1.1.tar.bz2 -o bloaty.tar.bz2 && \
-    echo "a308d8369d5812aba45982e55e7c3db2ea4780b7496a5455792fb3dcba9abd6f  bloaty.tar.bz2" > bloaty-sha.txt && \
-    sha256sum --quiet -c bloaty-sha.txt && \
-    mkdir bloaty && \
-    tar --strip-components 1 --no-same-owner --no-same-permissions --directory bloaty  -xf bloaty.tar.bz2 && \
-    cd bloaty && \
-    cmake -S. -B build -G Ninja && \
-    ninja -C build install && \
-    rm -rf /tmp/*
-
-RUN apt-get update
-RUN apt-get upgrade -y
-RUN TZ=America/Los_Angeles DEBIAN_FRONTEND=noninteractive apt-get install -y texlive-full
-
 # Try to have all the pre-commit hooks we'll need already initialized
 COPY .pre-commit-config.yaml /tmp/
 RUN git init && pre-commit install-hooks

FuzzTestDriver.cpp

@@ -2,15 +2,25 @@
 #include <cstdint>
 #include <fstream>
 #include <sstream>
+#include <thread>
 
 extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size);
 
 int main(int argc, char **argv) {
-  for (int i = 1; i < argc; ++i) {
+  auto doTest = [&]() {
-    std::ifstream t(argv[i], std::ios::binary);
+    for (int i = 1; i < argc; ++i) {
-    std::stringstream buffer;
+      std::ifstream t(argv[i], std::ios::binary);
-    buffer << t.rdbuf();
+      std::stringstream buffer;
-    auto str = buffer.str();
+      buffer << t.rdbuf();
-    LLVMFuzzerTestOneInput((const uint8_t *)str.data(), str.size());
+      auto str = buffer.str();
-  }
+      LLVMFuzzerTestOneInput((const uint8_t *)str.data(), str.size());
+    }
+  };
+#ifdef THREAD_TEST
+  std::thread thread2{doTest};
+#endif
+  doTest();
+#ifdef THREAD_TEST
+  thread2.join();
+#endif
 }

HashTable.cpp

@@ -0,0 +1,161 @@
+#include "ConflictSet.h"
+#include "Internal.h"
+#include <functional>
+#include <string>
+#include <unordered_map>
+
+// This implementation isn't correct for range queries :). It's just intended as
+// a reference for performance comparison with point queries.
+
+// struct is from "https://www.cppstories.com/2021/heterogeneous-access-cpp20/"
+struct string_hash {
+  using is_transparent = void;
+  [[nodiscard]] size_t operator()(std::string_view txt) const {
+    return std::hash<std::string_view>{}(txt);
+  }
+  [[nodiscard]] size_t operator()(const std::string &txt) const {
+    return std::hash<std::string>{}(txt);
+  }
+};
+
+struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
+  Impl(int64_t oldestVersion) : oldestVersion(oldestVersion) {}
+  void check(const ConflictSet::ReadRange *reads, ConflictSet::Result *results,
+             int count) const {
+    for (int i = 0; i < count; ++i) {
+      auto key =
+          std::string_view((const char *)reads[i].begin.p, reads[i].begin.len);
+      auto version = reads[i].readVersion;
+      if (version < oldestVersion) {
+        results[i] = TooOld;
+        continue;
+      }
+      auto iter = map.find(key);
+      results[i] =
+          iter == map.end() || iter->second <= version ? Commit : Conflict;
+    }
+  }
+
+  void addWrites(const ConflictSet::WriteRange *writes, int count,
+                 int64_t writeVersion) {
+    for (int i = 0; i < count; ++i) {
+      auto &max = map[std::string((const char *)writes[i].begin.p,
+                                  writes[i].begin.len)];
+      assert(writeVersion >= max);
+      max = writeVersion;
+      keyUpdates += 2;
+    }
+  }
+
+  void setOldestVersion(int64_t oldestVersion) {
+    if (oldestVersion <= this->oldestVersion) {
+      return;
+    }
+    this->oldestVersion = oldestVersion;
+    if (keyUpdates < 100) {
+      return;
+    }
+    auto iter = map.find(removalKey);
+    while (keyUpdates > 0) {
+      if (iter == map.end()) {
+        iter = map.begin();
+      }
+      for (; iter != map.end(); --keyUpdates) {
+        if (iter->second <= oldestVersion) {
+          iter = map.erase(iter);
+        } else {
+          ++iter;
+        }
+      }
+    }
+    if (iter == map.end()) {
+      removalKey.clear();
+    } else {
+      removalKey = iter->first;
+    }
+  }
+
+private:
+  int64_t keyUpdates = 0;
+  int64_t oldestVersion;
+  std::unordered_map<std::string, int64_t, string_hash, std::equal_to<>> map;
+  std::string removalKey;
+};
+
+void ConflictSet::check(const ReadRange *reads, Result *results,
+                        int count) const {
+  return impl->check(reads, results, count);
+}
+
+void ConflictSet::addWrites(const WriteRange *writes, int count,
+                            int64_t writeVersion) {
+  return impl->addWrites(writes, count, writeVersion);
+}
+
+void ConflictSet::setOldestVersion(int64_t oldestVersion) {
+  return impl->setOldestVersion(oldestVersion);
+}
+
+int64_t ConflictSet::getBytes() const { return -1; }
+
+void ConflictSet::getMetricsV1(MetricsV1 **metrics, int *count) const {
+  *metrics = nullptr;
+  *count = 0;
+}
+
+double ConflictSet::MetricsV1::getValue() const { return 0; }
+
+ConflictSet::ConflictSet(int64_t oldestVersion)
+    : impl(new(safe_malloc(sizeof(Impl))) Impl{oldestVersion}) {}
+
+ConflictSet::~ConflictSet() {
+  if (impl) {
+    impl->~Impl();
+    safe_free(impl, sizeof(Impl));
+  }
+}
+
+ConflictSet::ConflictSet(ConflictSet &&other) noexcept
+    : impl(std::exchange(other.impl, nullptr)) {}
+
+ConflictSet &ConflictSet::operator=(ConflictSet &&other) noexcept {
+  impl = std::exchange(other.impl, nullptr);
+  return *this;
+}
+
+using ConflictSet_Result = ConflictSet::Result;
+using ConflictSet_Key = ConflictSet::Key;
+using ConflictSet_ReadRange = ConflictSet::ReadRange;
+using ConflictSet_WriteRange = ConflictSet::WriteRange;
+
+extern "C" {
+__attribute__((__visibility__("default"))) void
+ConflictSet_check(void *cs, const ConflictSet_ReadRange *reads,
+                  ConflictSet_Result *results, int count) {
+  ((ConflictSet::Impl *)cs)->check(reads, results, count);
+}
+__attribute__((__visibility__("default"))) void
+ConflictSet_addWrites(void *cs, const ConflictSet_WriteRange *writes, int count,
+                      int64_t writeVersion) {
+  ((ConflictSet::Impl *)cs)->addWrites(writes, count, writeVersion);
+}
+__attribute__((__visibility__("default"))) void
+ConflictSet_setOldestVersion(void *cs, int64_t oldestVersion) {
+  ((ConflictSet::Impl *)cs)->setOldestVersion(oldestVersion);
+}
+__attribute__((__visibility__("default"))) void *
+ConflictSet_create(int64_t oldestVersion) {
+  return new (safe_malloc(sizeof(ConflictSet::Impl)))
+      ConflictSet::Impl{oldestVersion};
+}
+__attribute__((__visibility__("default"))) void ConflictSet_destroy(void *cs) {
+  using Impl = ConflictSet::Impl;
+  ((Impl *)cs)->~Impl();
+  safe_free(cs, sizeof(Impl));
+}
+__attribute__((__visibility__("default"))) int64_t
+ConflictSet_getBytes(void *cs) {
+  using Impl = ConflictSet::Impl;
+  return -1;
+}
+}

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

@@ -2,6 +2,8 @@
 
 #include "ConflictSet.h"
 
+using namespace weaselab;
+
 #include <bit>
 #include <cassert>
 #include <compare>
@@ -10,21 +12,52 @@
 #include <cstdlib>
 #include <cstring>
 #include <inttypes.h>
+#include <latch>
 #include <map>
 #include <set>
 #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
+// 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) {
@@ -34,16 +67,85 @@ operator<=>(const std::span<const uint8_t> &lhs,
   return lhs.size() <=> rhs.size();
 }
 
+[[nodiscard]] inline auto operator<=>(const TrivialSpan &lhs,
+                                      const ConflictSet::Key &rhs) noexcept {
+  int cl = std::min<int>(lhs.size(), rhs.len);
+  if (cl > 0) {
+    if (auto c = memcmp(lhs.data(), rhs.p, cl) <=> 0; c != 0) {
+      return c;
+    }
+  }
+  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
 // want to exclude from coverage since it's only testing related.
 
 // GCOVR_EXCL_START
 
+#if SHOW_MEMORY
+inline int64_t mallocBytes = 0;
+inline int64_t peakMallocBytes = 0;
+#endif
+
+inline thread_local int64_t mallocBytesDelta = 0;
+
+#ifndef NDEBUG
+constexpr auto kMallocHeaderSize = 16;
+#endif
+
+// malloc that aborts on OOM and thus always returns a non-null pointer. Must be
+// paired with `safe_free`.
 __attribute__((always_inline)) inline void *safe_malloc(size_t s) {
-  if (void *p = malloc(s)) {
+  mallocBytesDelta += s;
-    return p;
+#if SHOW_MEMORY
+  mallocBytes += s;
+  if (mallocBytes > peakMallocBytes) {
+    peakMallocBytes = mallocBytes;
   }
-  abort();
+#endif
+  void *p = malloc(s
+#ifndef NDEBUG
+                   + kMallocHeaderSize
+#endif
+  );
+  if (p == nullptr) {
+    abort();
+  }
+#ifndef NDEBUG
+  memcpy(p, &s, sizeof(s));
+  (char *&)p += kMallocHeaderSize;
+#endif
+  return p;
+}
+
+// Must be paired with `safe_malloc`.
+//
+// There's nothing safer about this than free. Only called safe_free for
+// symmetry with safe_malloc.
+__attribute__((always_inline)) inline void safe_free(void *p, size_t s) {
+  mallocBytesDelta -= s;
+#if SHOW_MEMORY
+  mallocBytes -= s;
+#endif
+#ifndef NDEBUG
+  (char *&)p -= kMallocHeaderSize;
+  size_t expected;
+  memcpy(&expected, p, sizeof(expected));
+  assert(s == expected);
+#endif
+  free(p);
 }
 
 // ==================== BEGIN ARENA IMPL ====================
@@ -89,6 +191,7 @@ inline void *operator new[](size_t size, std::align_val_t align,
 
 /// align must be a power of two
 template <class T> T *align_up(T *t, size_t align) {
+  assert(std::popcount(align) == 1);
   auto unaligned = uintptr_t(t);
   auto aligned = (unaligned + align - 1) & ~(align - 1);
   return reinterpret_cast<T *>(reinterpret_cast<char *>(t) + aligned -
@@ -97,6 +200,7 @@ template <class T> T *align_up(T *t, size_t align) {
 
 /// align must be a power of two
 constexpr inline int align_up(uint32_t unaligned, uint32_t align) {
+  assert(std::popcount(align) == 1);
   return (unaligned + align - 1) & ~(align - 1);
 }
 
@@ -112,12 +216,10 @@ struct Arena::ArenaImpl {
   uint8_t *begin() { return reinterpret_cast<uint8_t *>(this + 1); }
 };
 
-static_assert(sizeof(Arena::ArenaImpl) == 16);
-static_assert(alignof(Arena::ArenaImpl) == 8);
-
 inline Arena::Arena(int initialSize) : impl(nullptr) {
   if (initialSize > 0) {
-    auto allocationSize = align_up(initialSize + sizeof(ArenaImpl), 16);
+    auto allocationSize =
+        align_up(initialSize + sizeof(ArenaImpl), alignof(ArenaImpl));
     impl = (Arena::ArenaImpl *)safe_malloc(allocationSize);
     impl->prev = nullptr;
     impl->capacity = allocationSize - sizeof(ArenaImpl);
@@ -128,7 +230,7 @@ inline Arena::Arena(int initialSize) : impl(nullptr) {
 inline void onDestroy(Arena::ArenaImpl *impl) {
   while (impl) {
     auto *prev = impl->prev;
-    free(impl);
+    safe_free(impl, sizeof(Arena::ArenaImpl) + impl->capacity);
     impl = prev;
   }
 }
@@ -153,7 +255,7 @@ inline void *operator new(size_t size, std::align_val_t align, Arena &arena) {
                           (arena.impl ? std::max<int>(sizeof(Arena::ArenaImpl),
                                                       arena.impl->capacity * 2)
                                       : 0)),
-        16);
+        alignof(Arena::ArenaImpl));
     auto *impl = (Arena::ArenaImpl *)safe_malloc(allocationSize);
     impl->prev = arena.impl;
     impl->capacity = allocationSize - sizeof(Arena::ArenaImpl);
@@ -192,10 +294,75 @@ template <class T> struct ArenaAlloc {
   void deallocate(T *, size_t) noexcept {}
 };
 
-template <class T> using Vector = std::vector<T, ArenaAlloc<T>>;
+template <class T> struct Vector {
-template <class T> auto vector(Arena &arena) {
+  static_assert(std::is_trivially_destructible_v<T>);
-  return Vector<T>(ArenaAlloc<T>(&arena));
+  static_assert(std::is_trivially_copyable_v<T>);
-}
+
+  explicit Vector(Arena *arena)
+      : arena(arena), t(nullptr), size_(0), capacity(0) {}
+
+  void append(std::span<const T> slice) {
+    if (size_ + int(slice.size()) > capacity) {
+      grow(std::max<int>(size_ + slice.size(), capacity * 2));
+    }
+    if (slice.size() > 0) {
+      memcpy(const_cast<std::remove_const_t<T> *>(t) + size_, slice.data(),
+             slice.size() * sizeof(T));
+    }
+    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; }
+  T *end() { return t + size_; }
+  T *data() { return t; }
+  T &back() {
+    assert(size_ > 0);
+    return t[size_ - 1];
+  }
+  T &operator[](int i) {
+    assert(i >= 0 && i < size_);
+    return t[i];
+  }
+
+  void pop_back() {
+    assert(size_ > 0);
+    --size_;
+  }
+
+  int size() const { return size_; }
+
+  operator std::span<const T>() const { return std::span(t, size_); }
+
+private:
+  void grow(int newCapacity) {
+    capacity = newCapacity;
+    auto old = std::span<const T>(*this);
+    t = (T *)new (std::align_val_t(alignof(T)), *arena)
+        uint8_t[capacity * sizeof(T)];
+    size_ = 0;
+    append(old);
+  }
+
+  Arena *arena;
+  T *t;
+  int size_;
+  int capacity;
+};
+
+template <class T> auto vector(Arena &arena) { return Vector<T>(&arena); }
+
 template <class T, class C> using Set = std::set<T, C, ArenaAlloc<T>>;
 template <class T, class C = std::less<T>> auto set(Arena &arena) {
   return Set<T, C>(ArenaAlloc<T>(&arena));
@@ -348,42 +515,16 @@ inline uint32_t Arbitrary::bounded(uint32_t s) {
 
 // ==================== END ARBITRARY IMPL ====================
 
-// ==================== BEGIN UTILITIES IMPL ====================
-
-// Call Stepwise::step for each element of remaining until it returns true.
-// Applies a permutation to `remaining` as a side effect.
-template <class Stepwise> void runInterleaved(std::span<Stepwise> remaining) {
-  while (remaining.size() > 0) {
-    for (int i = 0; i < int(remaining.size());) {
-      bool done = remaining[i].step();
-      if (done) {
-        if (i != int(remaining.size()) - 1) {
-          using std::swap;
-          swap(remaining[i], remaining.back());
-        }
-        remaining = remaining.subspan(0, remaining.size() - 1);
-      } else {
-        ++i;
-      }
-    }
-  }
-};
-
-template <class Stepwise> void runSequential(std::span<Stepwise> remaining) {
-  for (auto &r : remaining) {
-    while (!r.step()) {
-    }
-  }
-}
-
 struct ReferenceImpl {
-  explicit ReferenceImpl(int64_t oldestVersion) : oldestVersion(oldestVersion) {
+  explicit ReferenceImpl(int64_t oldestVersion)
+      : oldestVersion(oldestVersion), newestVersion(oldestVersion) {
     writeVersionMap[""] = oldestVersion;
   }
   void check(const ConflictSet::ReadRange *reads, ConflictSet::Result *results,
              int count) const {
     for (int i = 0; i < count; ++i) {
-      if (reads[i].readVersion < oldestVersion) {
+      if (reads[i].readVersion < oldestVersion ||
+          reads[i].readVersion < newestVersion - 2e9) {
         results[i] = ConflictSet::TooOld;
         continue;
       }
@@ -403,7 +544,10 @@ struct ReferenceImpl {
                                                      : ConflictSet::Commit;
     }
   }
-  void addWrites(const ConflictSet::WriteRange *writes, int count) {
+  void addWrites(const ConflictSet::WriteRange *writes, int count,
+                 int64_t writeVersion) {
+    assert(writeVersion >= newestVersion);
+    newestVersion = writeVersion;
     for (int i = 0; i < count; ++i) {
       auto begin =
           std::string((const char *)writes[i].begin.p, writes[i].begin.len);
@@ -411,7 +555,6 @@ struct ReferenceImpl {
           writes[i].end.len == 0
               ? begin + std::string("\x00", 1)
               : std::string((const char *)writes[i].end.p, writes[i].end.len);
-      auto writeVersion = writes[i].writeVersion;
       auto prevVersion = (--writeVersionMap.upper_bound(end))->second;
       for (auto iter = writeVersionMap.lower_bound(begin),
                 endIter = writeVersionMap.lower_bound(end);
@@ -424,16 +567,21 @@ struct ReferenceImpl {
   }
 
   void setOldestVersion(int64_t oldestVersion) {
-    assert(oldestVersion >= oldestVersion);
+    assert(oldestVersion >= this->oldestVersion);
     this->oldestVersion = oldestVersion;
   }
 
   int64_t oldestVersion;
+  int64_t newestVersion;
   std::map<std::string, int64_t> writeVersionMap;
 };
 
 using Key = ConflictSet::Key;
 
+inline Key operator""_s(const char *str, size_t size) {
+  return {reinterpret_cast<const uint8_t *>(str), int(size)};
+}
+
 [[maybe_unused]] static Key toKey(Arena &arena, int n) {
   uint8_t *buf = new (arena) uint8_t[sizeof(n)];
   memcpy(buf, &n, sizeof(n));
@@ -461,59 +609,70 @@ 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()));
 }
 
+inline const char *resultToStr(ConflictSet::Result r) {
+  switch (r) {
+  case ConflictSet::Commit:
+    return "commit";
+  case ConflictSet::Conflict:
+    return "conflict";
+  case ConflictSet::TooOld:
+    return "too old";
+  }
+  abort();
+}
+
 namespace {
 
-template <class ConflictSetImpl> struct TestDriver {
+template <class ConflictSetImpl, bool kEnableAssertions = true>
-  Arbitrary arbitrary;
+struct TestDriver {
-  explicit TestDriver(const uint8_t *data, size_t size)
+  Arbitrary *arbitrary;
-      : arbitrary({data, size}) {}
+  explicit TestDriver(Arbitrary &a) : arbitrary(&a) {
+#if DEBUG_VERBOSE && !defined(NDEBUG)
+    fprintf(stderr, "%p Initial version: {%" PRId64 "}\n", this, writeVersion);
+#endif
+  }
 
-  int64_t writeVersion = 0;
+  int64_t oldestVersion = arbitrary->next();
-  int64_t oldestVersion = 0;
+  int64_t writeVersion = oldestVersion;
   ConflictSetImpl cs{oldestVersion};
   ReferenceImpl refImpl{oldestVersion};
 
-  constexpr static auto kMaxKeyLen = 32;
+  constexpr static auto kMaxKeySuffixLen = 8;
 
   bool ok = true;
 
-  static const char *resultToStr(ConflictSet::Result r) {
+  const int prefixLen = arbitrary->bounded(512);
-    switch (r) {
+  const int prefixByte = arbitrary->randT<uint8_t>();
-    case ConflictSet::Commit:
-      return "commit";
-    case ConflictSet::Conflict:
-      return "conflict";
-    case ConflictSet::TooOld:
-      return "too old";
-    }
-    abort();
-  }
 
   // Call until it returns true, for "done". Check internal invariants etc
   // between calls to next.
   bool next() {
-    if (!arbitrary.hasEntropy()) {
+    assert(cs.getBytes() >= 0);
+    if (!arbitrary->hasEntropy()) {
       return true;
     }
     Arena arena;
     {
-      int numPointWrites = arbitrary.bounded(100);
+      int numPointWrites = arbitrary->bounded(100);
-      int numRangeWrites = arbitrary.bounded(100);
+      int numRangeWrites = arbitrary->bounded(100);
-      int64_t v = ++writeVersion;
+      int64_t v =
+          (writeVersion +=
+           arbitrary->bounded(10) ? arbitrary->bounded(10) : arbitrary->next());
       auto *writes =
           new (arena) ConflictSet::WriteRange[numPointWrites + numRangeWrites];
       auto keys = set<std::string_view>(arena);
       while (int(keys.size()) < numPointWrites + numRangeWrites * 2) {
-        if (!arbitrary.hasEntropy()) {
+        if (!arbitrary->hasEntropy()) {
           return true;
         }
-        int keyLen = arbitrary.bounded(kMaxKeyLen);
+        int keyLen = prefixLen + arbitrary->bounded(kMaxKeySuffixLen);
         auto *begin = new (arena) uint8_t[keyLen];
-        arbitrary.randomBytes(begin, keyLen);
+        memset(begin, prefixByte, prefixLen);
+        arbitrary->randomBytes(begin + prefixLen, keyLen - prefixLen);
         keys.insert(std::string_view((const char *)begin, keyLen));
       }
 
@@ -523,7 +682,7 @@ template <class ConflictSetImpl> struct TestDriver {
                rangesRemaining = numRangeWrites;
            pointsRemaining > 0 || rangesRemaining > 0; ++i) {
         bool pointRead = pointsRemaining > 0 && rangesRemaining > 0
-                             ? bool(arbitrary.bounded(2))
+                             ? bool(arbitrary->bounded(2))
                              : pointsRemaining > 0;
         if (pointRead) {
           assert(pointsRemaining > 0);
@@ -542,48 +701,121 @@ template <class ConflictSetImpl> struct TestDriver {
           ++iter;
           --rangesRemaining;
         }
-        writes[i].writeVersion = v;
-#if DEBUG_VERBOSE && !defined(NDEBUG)
-        if (writes[i].end.len == 0) {
-          fprintf(stderr, "Write: {%s} -> %d\n",
-                  printable(writes[i].begin).c_str(),
-                  int(writes[i].writeVersion));
-        } else {
-          fprintf(stderr, "Write: [%s, %s) -> %d\n",
-                  printable(writes[i].begin).c_str(),
-                  printable(writes[i].end).c_str(),
-                  int(writes[i].writeVersion));
-        }
-#endif
       }
       assert(iter == keys.end());
       assert(i == numPointWrites + numRangeWrites);
 
+      // Test non-canonical writes
+      if (numPointWrites > 0) {
+        int overlaps = arbitrary->bounded(numPointWrites);
+        for (int i = 0; i < numPointWrites + numRangeWrites && overlaps > 0;
+             ++i) {
+          if (writes[i].end.len == 0) {
+            int keyLen = prefixLen + arbitrary->bounded(kMaxKeySuffixLen);
+            auto *begin = new (arena) uint8_t[keyLen];
+            memset(begin, prefixByte, prefixLen);
+            arbitrary->randomBytes(begin + prefixLen, keyLen - prefixLen);
+            writes[i].end.len = keyLen;
+            writes[i].end.p = begin;
+            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);
+            } else if (c == 0) {
+              // It's a point write after all, I guess
+              writes[i].end.len = 0;
+            }
+            --overlaps;
+          }
+        }
+      }
+      if (arbitrary->bounded(2)) {
+        // Shuffle writes
+        for (int i = numPointWrites + numRangeWrites - 1; i > 0; --i) {
+          int j = arbitrary->bounded(i + 1);
+          if (i != j) {
+            using std::swap;
+            swap(writes[i], writes[j]);
+          }
+        }
+      }
+
+      oldestVersion +=
+          arbitrary->bounded(10) ? arbitrary->bounded(10) : arbitrary->next();
+      oldestVersion = std::min(oldestVersion, writeVersion);
+
+#ifdef THREAD_TEST
+      std::latch ready{1};
+      std::thread thread2{[&]() {
+        ready.count_down();
+        ConflictSet::MetricsV1 *m;
+        int count;
+        cs.getMetricsV1(&m, &count);
+        for (int i = 0; i < count; ++i) {
+          m[i].getValue();
+        }
+      }};
+      ready.wait();
+#endif
+
+#if DEBUG_VERBOSE && !defined(NDEBUG)
+      for (int i = 0; i < numPointWrites + numRangeWrites; ++i) {
+        if (writes[i].end.len == 0) {
+          fprintf(stderr, "%p Write: {%s}\n", this,
+                  printable(writes[i].begin).c_str());
+        } else {
+          fprintf(stderr, "%p Write: [%s, %s)\n", this,
+                  printable(writes[i].begin).c_str(),
+                  printable(writes[i].end).c_str());
+        }
+      }
+      fprintf(stderr, "%p Write @ %" PRId64 "\n", this, v);
+#endif
+
       CALLGRIND_START_INSTRUMENTATION;
-      cs.addWrites(writes, numPointWrites + numRangeWrites);
+      cs.addWrites(writes, numPointWrites + numRangeWrites, v);
       CALLGRIND_STOP_INSTRUMENTATION;
 
-      refImpl.addWrites(writes, numPointWrites + numRangeWrites);
+      if constexpr (kEnableAssertions) {
+        refImpl.addWrites(writes, numPointWrites + numRangeWrites, v);
+      }
 
-      oldestVersion = std::max<int64_t>(writeVersion - arbitrary.bounded(10),
+#if DEBUG_VERBOSE && !defined(NDEBUG)
-                                        oldestVersion);
+      fprintf(stderr, "%p Set oldest version: %" PRId64 "\n", this,
+              oldestVersion);
+#endif
+      CALLGRIND_START_INSTRUMENTATION;
       cs.setOldestVersion(oldestVersion);
-      refImpl.setOldestVersion(oldestVersion);
+      CALLGRIND_STOP_INSTRUMENTATION;
+
+      if constexpr (kEnableAssertions) {
+        refImpl.setOldestVersion(oldestVersion);
+      }
+
+#ifdef THREAD_TEST
+      thread2.join();
+#endif
     }
     {
-      int numPointReads = arbitrary.bounded(100);
+      int numPointReads = arbitrary->bounded(100);
-      int numRangeReads = arbitrary.bounded(100);
+      int numRangeReads = arbitrary->bounded(100);
-      int64_t v = std::max<int64_t>(writeVersion - arbitrary.bounded(10), 0);
+
+      int64_t v = std::max<int64_t>(writeVersion - (arbitrary->bounded(10)
+                                                        ? arbitrary->bounded(10)
+                                                        : arbitrary->next()),
+                                    0);
       auto *reads =
           new (arena) ConflictSet::ReadRange[numPointReads + numRangeReads];
       auto keys = set<std::string_view>(arena);
       while (int(keys.size()) < numPointReads + numRangeReads * 2) {
-        if (!arbitrary.hasEntropy()) {
+        if (!arbitrary->hasEntropy()) {
           return true;
         }
-        int keyLen = arbitrary.bounded(kMaxKeyLen);
+        int keyLen = prefixLen + arbitrary->bounded(kMaxKeySuffixLen);
         auto *begin = new (arena) uint8_t[keyLen];
-        arbitrary.randomBytes(begin, keyLen);
+        memset(begin, prefixByte, prefixLen);
+        arbitrary->randomBytes(begin + prefixLen, keyLen - prefixLen);
         keys.insert(std::string_view((const char *)begin, keyLen));
       }
 
@@ -592,7 +824,7 @@ template <class ConflictSetImpl> struct TestDriver {
       for (int pointsRemaining = numPointReads, rangesRemaining = numRangeReads;
            pointsRemaining > 0 || rangesRemaining > 0; ++i) {
         bool pointRead = pointsRemaining > 0 && rangesRemaining > 0
-                             ? bool(arbitrary.bounded(2))
+                             ? bool(arbitrary->bounded(2))
                              : pointsRemaining > 0;
         if (pointRead) {
           assert(pointsRemaining > 0);
@@ -614,12 +846,12 @@ template <class ConflictSetImpl> struct TestDriver {
         reads[i].readVersion = v;
 #if DEBUG_VERBOSE && !defined(NDEBUG)
         if (reads[i].end.len == 0) {
-          fprintf(stderr, "Read: {%s} @ %d\n",
+          fprintf(stderr, "%p Read: {%s} @ %" PRId64 "\n", this,
-                  printable(reads[i].begin).c_str(), int(reads[i].readVersion));
+                  printable(reads[i].begin).c_str(), reads[i].readVersion);
         } else {
-          fprintf(stderr, "Read: [%s, %s) @ %d\n",
+          fprintf(stderr, "%p Read: [%s, %s) @ %" PRId64 "\n", this,
                   printable(reads[i].begin).c_str(),
-                  printable(reads[i].end).c_str(), int(reads[i].readVersion));
+                  printable(reads[i].end).c_str(), reads[i].readVersion);
         }
 #endif
       }
@@ -630,32 +862,85 @@ template <class ConflictSetImpl> struct TestDriver {
       auto *results2 =
           new (arena) ConflictSet::Result[numPointReads + numRangeReads];
 
+#ifdef THREAD_TEST
+      auto *results3 =
+          new (arena) ConflictSet::Result[numPointReads + numRangeReads];
+      std::latch ready{1};
+      std::thread thread2{[&]() {
+        ready.count_down();
+        // Call all const methods
+        cs.check(reads, results3, numPointReads + numRangeReads);
+        cs.getBytes();
+        ConflictSet::MetricsV1 *m;
+        int count;
+        cs.getMetricsV1(&m, &count);
+        for (int i = 0; i < count; ++i) {
+          m[i].getValue();
+        }
+      }};
+      ready.wait();
+#endif
+
       CALLGRIND_START_INSTRUMENTATION;
       cs.check(reads, results1, numPointReads + numRangeReads);
       CALLGRIND_STOP_INSTRUMENTATION;
 
-      refImpl.check(reads, results2, numPointReads + numRangeReads);
+      if constexpr (kEnableAssertions) {
-      for (int i = 0; i < numPointReads + numRangeReads; ++i) {
+        // Call remaining const methods
-        if (results1[i] != results2[i]) {
+        cs.getBytes();
-          if (reads[i].end.len == 0) {
+        ConflictSet::MetricsV1 *m;
-            fprintf(stderr,
+        int count;
-                    "Expected %s, got %s for read of {%s} at version %" PRId64
+        cs.getMetricsV1(&m, &count);
-                    "\n",
+        for (int i = 0; i < count; ++i) {
-                    resultToStr(results2[i]), resultToStr(results1[i]),
+          m[i].getValue();
-                    printable(reads[i].begin).c_str(), reads[i].readVersion);
+        }
-          } else {
+
-            fprintf(
+        refImpl.check(reads, results2, numPointReads + numRangeReads);
-                stderr,
+      }
-                "Expected %s, got %s for read of [%s, %s) at version %" PRId64
+
-                "\n",
+      auto compareResults = [reads, this](ConflictSet::Result *results1,
-                resultToStr(results2[i]), resultToStr(results1[i]),
+                                          ConflictSet::Result *results2,
-                printable(reads[i].begin).c_str(),
+                                          int count) {
-                printable(reads[i].end).c_str(), reads[i].readVersion);
+        for (int i = 0; i < count; ++i) {
+          if (results1[i] != results2[i]) {
+            if (reads[i].end.len == 0) {
+              fprintf(
+                  stderr,
+                  "%p Expected %s, got %s for read of {%s} at version %" PRId64
+                  "\n",
+                  (void *)this, resultToStr(results2[i]),
+                  resultToStr(results1[i]), printable(reads[i].begin).c_str(),
+                  reads[i].readVersion);
+            } else {
+              fprintf(stderr,
+                      "%p Expected %s, got %s for read of [%s, %s) at version "
+                      "%" PRId64 "\n",
+                      (void *)this, resultToStr(results2[i]),
+                      resultToStr(results1[i]),
+                      printable(reads[i].begin).c_str(),
+                      printable(reads[i].end).c_str(), reads[i].readVersion);
+            }
+            return false;
           }
+        }
+        return true;
+      };
+
+      if constexpr (kEnableAssertions) {
+        if (!compareResults(results1, results2,
+                            numPointReads + numRangeReads)) {
           ok = false;
           return true;
         }
       }
+
+#ifdef THREAD_TEST
+      thread2.join();
+      if (!compareResults(results3, results2, numPointReads + numRangeReads)) {
+        ok = false;
+        return true;
+      }
+#endif
     }
     return false;
   }

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}/'
 }
 
@@ -36,6 +36,61 @@ pipeline {
                 sh 'pre-commit run --all-files --show-diff-on-failure'
             }
         }
+        stage('64 bit versions') {
+            agent {
+                dockerfile {
+                    args '-v /home/jenkins/ccache:/ccache'
+                    reuseNode true
+                }
+            }
+            steps {
+                CleanBuildAndTest("-DCMAKE_CXX_FLAGS=-DUSE_64_BIT=1")
+            }
+        }
+        stage('Debug') {
+            agent {
+                dockerfile {
+                    args '-v /home/jenkins/ccache:/ccache'
+                    reuseNode true
+                }
+            }
+            steps {
+                CleanBuildAndTest("-DCMAKE_BUILD_TYPE=Debug")
+            }
+        }
+        stage('SIMD fallback') {
+            agent {
+                dockerfile {
+                    args '-v /home/jenkins/ccache:/ccache'
+                    reuseNode true
+                }
+            }
+            steps {
+                CleanBuildAndTest("-DUSE_SIMD_FALLBACK=ON")
+            }
+        }
+        stage('Release [clang]') {
+            agent {
+                dockerfile {
+                    args '-v /home/jenkins/ccache:/ccache'
+                    reuseNode true
+                }
+            }
+            steps {
+                CleanBuildAndTest("-DCMAKE_CXX_FLAGS=-DNVALGRIND")
+                recordIssues(tools: [clang()])
+                sh '''
+                cd build
+                cpack -G DEB
+                cpack -G RPM
+                '''
+                sh '''
+                cd paper
+                make
+                '''
+                minio bucket: 'jenkins', credentialsId: 'jenkins-minio', excludes: '', host: 'minio.weaselab.dev', includes: 'build/*.deb,build/*.rpm,paper/*.pdf', targetFolder: '${JOB_NAME}/${BUILD_NUMBER}/${STAGE_NAME}/'
+            }
+        }
         stage('Release [gcc]') {
             agent {
                 dockerfile {
@@ -44,18 +99,25 @@ pipeline {
                 }
             }
             steps {
-                CleanBuildAndTest("-DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ -DCMAKE_BUILD_TYPE=Release")
+                CleanBuildAndTest("-DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ -DCMAKE_CXX_FLAGS=-DNVALGRIND")
                 recordIssues(tools: [gcc()])
+            }
+        }
+        stage('Release [clang,aarch64]') {
+            agent {
+                dockerfile {
+                    args '-v /home/jenkins/ccache:/ccache'
+                    reuseNode true
+                }
+            }
+            steps {
+                CleanBuildAndTest("-DCMAKE_TOOLCHAIN_FILE=../aarch64-toolchain.cmake -DCMAKE_CXX_FLAGS=-DNVALGRIND")
                 sh '''
                 cd build
                 cpack -G DEB
+                cpack -G RPM
                 '''
-                minio bucket: 'jenkins', credentialsId: 'jenkins-minio', excludes: '', host: 'minio.weaselab.dev', includes: 'build/*.deb', targetFolder: '${JOB_NAME}/${BUILD_NUMBER}/${STAGE_NAME}/'
+                minio bucket: 'jenkins', credentialsId: 'jenkins-minio', excludes: '', host: 'minio.weaselab.dev', includes: 'build/*.deb,build/*.rpm', targetFolder: '${JOB_NAME}/${BUILD_NUMBER}/${STAGE_NAME}'
-                sh '''
-                cd paper
-                make
-                '''
-                minio bucket: 'jenkins', credentialsId: 'jenkins-minio', excludes: '', host: 'minio.weaselab.dev', includes: 'paper/*.pdf', targetFolder: '${JOB_NAME}/${BUILD_NUMBER}/${STAGE_NAME}/'
             }
         }
         stage('Coverage') {
@@ -66,11 +128,18 @@ pipeline {
                 }
             }
             steps {
-                CleanBuildAndTest("-DCMAKE_C_FLAGS=--coverage -DCMAKE_CXX_FLAGS=--coverage -DCMAKE_BUILD_TYPE=Debug")
+                script {
-                sh '''
+                    gcov_args = "-f ConflictSet.cpp -f LongestCommonPrefix.h -f Metrics.h --gcov-executable 'llvm-cov gcov' --exclude-noncode-lines"
-                gcovr --gcov-executable "llvm-cov-15 gcov" --exclude '.*third_party.*' --cobertura > build/coverage.xml
+                }
-                '''
+                CleanBuildAndTest("-DCMAKE_C_FLAGS=--coverage -DCMAKE_CXX_FLAGS=--coverage -DCMAKE_BUILD_TYPE=Debug -DDISABLE_TSAN=ON")
-                cobertura autoUpdateHealth: false, autoUpdateStability: false, coberturaReportFile: 'build/coverage.xml', conditionalCoverageTargets: '70, 0, 0', failUnhealthy: false, failUnstable: false, lineCoverageTargets: '80, 0, 0', maxNumberOfBuilds: 0, methodCoverageTargets: '80, 0, 0', onlyStable: false, sourceEncoding: 'ASCII', zoomCoverageChart: false
+                sh """
+                gcovr ${gcov_args} --cobertura > build/coverage.xml
+                """
+                recordCoverage qualityGates: [[criticality: 'NOTE', metric: 'MODULE']], tools: [[parser: 'COBERTURA', pattern: 'build/coverage.xml']]
+                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

@@ -1,70 +1,61 @@
 A data structure for optimistic concurrency control on ranges of bitwise-lexicographically-ordered keys.
 
-Intended to replace FoundationDB's skip list.
+Intended as an alternative to FoundationDB's skip list.
 
-# FoundationDB's benchmark
+Hardware for all benchmarks is an AMD Ryzen 9 7900 with (2x32GB) 5600MT/s CL28-34-34-89 1.35V RAM.
+
+Compiler is `Ubuntu clang version 20.0.0 (++20241029082144+7544d3af0e28-1~exp1~20241029082307.506)`.
+
+# Microbenchmark
 
 ## Skip list
 
-```
+|               ns/op |                op/s |    err% |          ins/op |          cyc/op |    IPC |         bra/op |   miss% |     total | benchmark
-New conflict set: 4.189 sec
+|--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-                  0.298 Mtransactions/sec
+|              159.65 |        6,263,576.52 |    1.6% |        2,972.36 |          820.37 |  3.623 |         504.59 |    0.0% |      0.01 | `point reads`
-                  1.194 Mkeys/sec
+|              156.32 |        6,397,320.65 |    0.7% |        2,913.62 |          806.87 |  3.611 |         490.19 |    0.0% |      0.01 | `prefix reads`
-Detect only:      3.990 sec
+|              229.18 |        4,363,293.65 |    1.2% |        3,541.05 |        1,219.75 |  2.903 |         629.33 |    0.0% |      0.01 | `range reads`
-                  0.313 Mtransactions/sec
+|              363.37 |        2,752,026.30 |    0.3% |        5,273.63 |        1,951.54 |  2.702 |         851.66 |    1.7% |      0.01 | `point writes`
-                  1.253 Mkeys/sec
+|              364.99 |        2,739,787.02 |    0.3% |        5,250.92 |        1,958.54 |  2.681 |         839.24 |    1.7% |      0.01 | `prefix writes`
-Skiplist only:    2.849 sec
+|              242.26 |        4,127,796.58 |    2.9% |        3,117.33 |        1,304.41 |  2.390 |         541.07 |    2.8% |      0.02 | `range writes`
-                  0.439 Mtransactions/sec
+|              562.48 |        1,777,855.27 |    0.8% |        7,305.21 |        3,034.34 |  2.408 |       1,329.30 |    1.3% |      0.01 | `monotonic increasing point writes`
-                  1.755 Mkeys/sec
+|          122,688.57 |            8,150.72 |    0.7% |      798,766.00 |      666,842.00 |  1.198 |     144,584.50 |    0.1% |      0.01 | `worst case for radix tree`
-Performance counters:
+|               41.71 |       23,976,459.34 |    1.7% |          885.00 |          219.17 |  4.038 |         132.00 |    0.0% |      0.01 | `create and destroy`
-               Build: 0.0913
-                 Add: 0.0998
-              Detect: 3.99
-              D.Sort: 0.808
-           D.Combine: 0.0309
-         D.CheckRead: 1.67
-   D.CheckIntraBatch: 0.0305
-        D.MergeWrite: 1.18
-      D.RemoveBefore: 0.265
-```
 
 ## Radix tree (this implementation)
 
+|               ns/op |                op/s |    err% |          ins/op |          cyc/op |    IPC |         bra/op |   miss% |     total | benchmark
+|--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
+|               12.63 |       79,186,868.18 |    1.4% |          241.61 |           64.76 |  3.731 |          31.64 |    0.8% |      0.01 | `point reads`
+|               14.48 |       69,078,073.40 |    0.3% |          292.42 |           74.69 |  3.915 |          41.49 |    0.5% |      0.01 | `prefix reads`
+|               34.37 |       29,094,694.11 |    0.2% |          759.53 |          179.77 |  4.225 |         100.38 |    0.2% |      0.01 | `range reads`
+|               19.34 |       51,713,896.36 |    0.7% |          369.70 |          101.81 |  3.631 |          47.88 |    0.6% |      0.01 | `point writes`
+|               39.16 |       25,538,968.61 |    0.2% |          653.16 |          206.77 |  3.159 |          89.62 |    0.8% |      0.01 | `prefix writes`
+|               40.58 |       24,642,681.12 |    4.7% |          718.44 |          216.44 |  3.319 |          99.28 |    0.6% |      0.01 | `range writes`
+|               78.77 |       12,694,520.69 |    3.8% |        1,395.55 |          421.73 |  3.309 |         249.81 |    0.1% |      0.01 | `monotonic increasing point writes`
+|          287,760.50 |            3,475.11 |    0.5% |    3,929,266.50 |    1,550,225.50 |  2.535 |     639,064.00 |    0.0% |      0.01 | `worst case for radix tree`
+|              104.76 |        9,545,250.65 |    3.1% |        2,000.00 |          552.82 |  3.618 |         342.00 |    0.0% |      0.01 | `create and destroy`
+
+# "Real data" test
+
+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.
+
+## skip list
+
 ```
-New conflict set: 2.965 sec
+Check: 4.39702 seconds, 370.83 MB/s, Add: 4.50025 seconds, 124.583 MB/s, Gc ratio: 29.1333%, Peak idle memory: 5.51852e+06
-                  0.422 Mtransactions/sec
-                  1.686 Mkeys/sec
-Detect only:      2.761 sec
-                  0.453 Mtransactions/sec
-                  1.811 Mkeys/sec
-Skiplist only:    1.580 sec
-                  0.791 Mtransactions/sec
-                  3.165 Mkeys/sec
-Performance counters:
-               Build: 0.0902
-                 Add: 0.107
-              Detect: 2.76
-              D.Sort: 0.809
-           D.Combine: 0.0309
-         D.CheckRead: 0.658
-   D.CheckIntraBatch: 0.0294
-        D.MergeWrite: 0.921
-      D.RemoveBefore: 0.305
 ```
 
-# Our benchmark
+## radix tree
 
-|               ns/op |                op/s |    err% |     total | benchmark
+```
-|--------------------:|--------------------:|--------:|----------:|:----------
+Check: 0.987757 seconds, 1650.76 MB/s, Add: 1.24815 seconds, 449.186 MB/s, Gc ratio: 41.4675%, Peak idle memory: 2.02872e+06
-|              263.10 |        3,800,889.33 |    1.6% |      0.64 | `skip list (point reads)`
+```
-|              256.36 |        3,900,800.96 |    1.7% |      0.61 | `skip list (prefix reads)`
+
-|              526.95 |        1,897,714.85 |    1.6% |      1.23 | `skip list (range reads)`
+## hash table
-|              546.54 |        1,829,676.97 |    0.5% |      1.31 | `skip list (point writes)`
+
-|              539.23 |        1,854,511.73 |    0.5% |      1.29 | `skip list (prefix writes)`
+(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)
-|              266.68 |        3,749,799.99 |    0.6% |      0.65 | `skip list (range writes)`
+
-|               18.31 |       54,612,706.60 |    0.4% |      0.04 | `radix tree (point reads)`
+```
-|               48.78 |       20,498,870.77 |    0.2% |      0.12 | `radix tree (prefix reads)`
+Check: 0.84256 seconds, 1935.23 MB/s, Add: 0.697204 seconds, 804.146 MB/s, Gc ratio: 35.4091%
-|              347.16 |        2,880,474.86 |    0.9% |      0.83 | `radix tree (range reads)`
+```
-|               34.80 |       28,734,706.67 |    3.6% |      0.09 | `radix tree (point writes)`
-|               70.34 |       14,216,970.16 |    1.4% |      0.17 | `radix tree (prefix writes)`
-|               82.41 |       12,134,555.86 |    1.0% |      0.20 | `radix tree (range writes)`

RealDataBench.cpp

@@ -0,0 +1,147 @@
+#include <ConflictSet.h>
+
+#include <cerrno>
+#include <chrono>
+#include <cstdio>
+#include <cstring>
+#include <fcntl.h>
+#include <string_view>
+#include <sys/mman.h>
+#include <sys/stat.h>
+#include <unistd.h>
+#include <vector>
+
+using namespace weaselab;
+
+double now() {
+  return std::chrono::duration_cast<std::chrono::nanoseconds>(
+             std::chrono::steady_clock::now().time_since_epoch())
+             .count() *
+         1e-9;
+}
+
+inline size_t getPageSize() {
+  static size_t kPageSize = sysconf(_SC_PAGESIZE);
+  return kPageSize;
+}
+
+/// Helper for rounding up to page size (or some other alignment)
+constexpr inline size_t rightAlign(size_t offset, size_t alignment) {
+  return offset % alignment == 0 ? offset
+                                 : ((offset / alignment) + 1) * alignment;
+}
+
+int main(int argc, const char **argv) {
+  // Use with this dataset https://snap.stanford.edu/data/memetracker9.html
+  // Preprocess the files with `sed -i'' '/^Q/d'`
+
+  double checkTime = 0;
+  double addTime = 0;
+  double gcTime = 0;
+  double checkBytes = 0;
+  double addBytes = 0;
+
+  ConflictSet cs{0};
+  int64_t version = 0;
+  double timer = 0;
+
+  int64_t peakMemory = 0;
+
+  for (int i = 1; i < argc; ++i) {
+    int fd = open(argv[i], O_RDONLY);
+    struct stat st;
+    if (fstat(fd, &st) == -1) {
+      int err = errno;
+      fprintf(stderr, "stat error %s - %s\n", argv[i], strerror(err));
+      fflush(stderr);
+      abort();
+    }
+
+    int64_t size = rightAlign(st.st_size, getPageSize());
+    const uint8_t *begin =
+        (uint8_t *)mmap(0, size, PROT_READ, MAP_PRIVATE, fd, 0);
+    madvise((void *)begin, size, MADV_SEQUENTIAL);
+    auto *const mapOriginal = begin;
+    const auto sizeOriginal = size;
+
+    using StringView = std::basic_string_view<uint8_t>;
+
+    StringView write;
+    std::vector<StringView> reads;
+    std::vector<ConflictSet::ReadRange> readRanges;
+    std::vector<ConflictSet::Result> results;
+
+    for (uint8_t *end = (uint8_t *)memchr(begin, '\n', size); end != nullptr;) {
+      StringView line{begin, static_cast<size_t>(end - begin)};
+      size -= end - begin + 1;
+      begin = end + 1;
+      end = (uint8_t *)memchr(begin, '\n', size);
+
+      if (line.size() > 0 && line[0] == 'P') {
+        write = line.substr(2, line.size());
+      } else if (line.size() > 0 && line[0] == 'L') {
+        reads.push_back(line.substr(2, line.size()));
+      } else if (line.empty()) {
+        {
+          readRanges.resize(reads.size());
+          auto iter = readRanges.begin();
+          for (const auto &read : reads) {
+            iter->begin.p = (const uint8_t *)read.data();
+            iter->begin.len = read.size();
+            checkBytes += read.size();
+            iter->end.len = 0;
+            iter->readVersion = version - 100;
+            ++iter;
+          }
+        }
+        results.resize(readRanges.size());
+
+        timer = now();
+        cs.check(readRanges.data(), results.data(), readRanges.size());
+        checkTime += now() - timer;
+
+        // Add unconditionally so that the load doesn't actually depend on the
+        // conflict rate
+        ConflictSet::WriteRange w;
+        w.begin.p = (const uint8_t *)write.data();
+        w.begin.len = write.size();
+        w.end.len = 0;
+
+        addBytes += write.size();
+
+        timer = now();
+        cs.addWrites(&w, 1, ++version);
+        addTime += now() - timer;
+
+        write = {};
+        reads.clear();
+
+        if (cs.getBytes() > peakMemory) {
+          peakMemory = cs.getBytes();
+        }
+
+        timer = now();
+        cs.setOldestVersion(version - 10000);
+        gcTime += now() - timer;
+      }
+    }
+    munmap((void *)mapOriginal, sizeOriginal);
+    close(fd);
+  }
+
+  ConflictSet::MetricsV1 *metrics;
+  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());
+  }
+
+  printf("Check: %g seconds, %g MB/s, Add: %g seconds, %g MB/s, Gc ratio: "
+         "%g%%, Peak idle memory: %g\n",
+         checkTime, checkBytes / checkTime * 1e-6, addTime,
+         addBytes / addTime * 1e-6, gcTime / (gcTime + addTime) * 1e2,
+         double(peakMemory));
+}

ServerBench.cpp

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

TestDriver.cpp

@@ -3,18 +3,68 @@
 #include <fstream>
 #include <sstream>
 
+#ifndef PERF_TEST
+#define PERF_TEST 0
+#endif
+
 int main(int argc, char **argv) {
   for (int i = 1; i < argc; ++i) {
-    printf("Running: %s\n", argv[i]);
     std::ifstream t(argv[i], std::ios::binary);
     std::stringstream buffer;
     buffer << t.rdbuf();
     auto str = buffer.str();
-    TestDriver<ConflictSet> driver{(const uint8_t *)str.data(), str.size()};
+    Arbitrary arbitrary({(const uint8_t *)str.data(), str.size()});
-    while (!driver.next())
+    TestDriver<ConflictSet, !PERF_TEST> driver1{arbitrary};
-      ;
+    TestDriver<ConflictSet, !PERF_TEST> driver2{arbitrary};
-    if (!driver.ok) {
+    bool done1 = false;
-      abort();
+    bool done2 = false;
+    for (;;) {
+      if (!done1) {
+        done1 = driver1.next();
+        if (!driver1.ok) {
+          abort();
+        }
+      }
+      if (!done2) {
+        done2 = driver2.next();
+        if (!driver2.ok) {
+          abort();
+        }
+      }
+      if (done1 && done2) {
+        break;
+      }
+    }
+
+    {
+      ConflictSet::MetricsV1 *metrics;
+      int metricsCount;
+      driver1.cs.getMetricsV1(&metrics, &metricsCount);
+      printf("#################### METRICS for ConflictSet 1 for %s "
+             "####################\n",
+             argv[i]);
+      for (int i = 0; i < metricsCount; ++i) {
+        printf("# HELP %s %s\n", metrics[i].name, metrics[i].help);
+        printf("# TYPE %s %s\n", metrics[i].name,
+               metrics[i].type == metrics[i].Counter ? "counter" : "gauge");
+        printf("%s %g\n", metrics[i].name, metrics[i].getValue());
+      }
+      puts("");
+    }
+    {
+      ConflictSet::MetricsV1 *metrics;
+      int metricsCount;
+      driver2.cs.getMetricsV1(&metrics, &metricsCount);
+      printf("#################### METRICS for ConflictSet 2 for %s "
+             "####################\n",
+             argv[i]);
+      for (int i = 0; i < metricsCount; ++i) {
+        printf("# HELP %s %s\n", metrics[i].name, metrics[i].help);
+        printf("# TYPE %s %s\n", metrics[i].name,
+               metrics[i].type == metrics[i].Counter ? "counter" : "gauge");
+        printf("%s %g\n", metrics[i].name, metrics[i].getValue());
+      }
+      puts("");
     }
   }
 }

aarch64-symbol-imports.txt

@@ -0,0 +1,11 @@
+__aarch64_ldadd8_relax
+__getauxval@GLIBC_2.17
+__stack_chk_fail@GLIBC_2.17
+__stack_chk_guard@GLIBC_2.17
+abort@GLIBC_2.17
+free@GLIBC_2.17
+malloc@GLIBC_2.17
+memcmp@GLIBC_2.17
+memcpy@GLIBC_2.17
+memmove@GLIBC_2.17
+memset@GLIBC_2.17

aarch64-toolchain.cmake

@@ -0,0 +1,8 @@
+set(CMAKE_SYSTEM_NAME Linux)
+set(CMAKE_SYSTEM_PROCESSOR aarch64)
+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-exports.txt

@@ -0,0 +1,19 @@
+_ConflictSet_addWrites
+_ConflictSet_check
+_ConflictSet_create
+_ConflictSet_destroy
+_ConflictSet_getBytes
+_ConflictSet_setOldestVersion
+__ZN8weaselab11ConflictSet16setOldestVersionEx
+__ZN8weaselab11ConflictSet9addWritesEPKNS0_10WriteRangeEix
+__ZN8weaselab11ConflictSetC1EOS0_
+__ZN8weaselab11ConflictSetC1Ex
+__ZN8weaselab11ConflictSetC2EOS0_
+__ZN8weaselab11ConflictSetC2Ex
+__ZN8weaselab11ConflictSetD1Ev
+__ZN8weaselab11ConflictSetD2Ev
+__ZN8weaselab11ConflictSetaSEOS0_
+__ZNK8weaselab11ConflictSet12getMetricsV1EPPNS0_9MetricsV1EPi
+__ZNK8weaselab11ConflictSet5checkEPKNS0_9ReadRangeEPNS0_6ResultEi
+__ZNK8weaselab11ConflictSet8getBytesEv
+__ZNK8weaselab11ConflictSet9MetricsV18getValueEv

apple-symbol-imports.txt

@@ -0,0 +1,10 @@
+___stack_chk_fail
+___stack_chk_guard
+__tlv_bootstrap
+_abort
+_bzero
+_free
+_malloc
+_memcpy
+_memmove
+dyld_stub_binder

conflict_set.py

@@ -0,0 +1,138 @@
+import ctypes
+import enum
+import os
+
+from typing import Optional
+
+
+class _Key(ctypes.Structure):
+    _fields_ = [("p", ctypes.POINTER(ctypes.c_ubyte)), ("len", ctypes.c_int)]
+
+
+class ReadRange(ctypes.Structure):
+    _fields_ = [
+        ("begin", _Key),
+        ("end", _Key),
+        ("readVersion", ctypes.c_int64),
+    ]
+
+
+class WriteRange(ctypes.Structure):
+    _fields_ = [("begin", _Key), ("end", _Key)]
+
+
+class Result(enum.Enum):
+    COMMIT = 0
+    CONFLICT = 1
+    TOO_OLD = 2
+
+
+def write(begin: bytes, end: Optional[bytes] = None) -> WriteRange:
+    b = (ctypes.c_ubyte * len(begin)).from_buffer(bytearray(begin))
+
+    if end is None:
+        e = (ctypes.c_ubyte * 0)()
+    else:
+        e = (ctypes.c_ubyte * len(end)).from_buffer(bytearray(end))
+    return WriteRange(_Key(b, len(b)), _Key(e, len(e)))
+
+
+def read(version: int, begin: bytes, end: Optional[bytes] = None) -> ReadRange:
+    b = (ctypes.c_ubyte * len(begin)).from_buffer(bytearray(begin))
+    if end is None:
+        e = (ctypes.c_ubyte * 0)()
+    else:
+        e = (ctypes.c_ubyte * len(end)).from_buffer(bytearray(end))
+    return ReadRange(_Key(b, len(b)), _Key(e, len(e)), version)
+
+
+class ConflictSet:
+    def __init__(
+        self,
+        version: int = 0,
+        build_dir: Optional[str] = None,
+        implementation: Optional[str] = None,
+    ) -> None:
+        self._lib = None
+        if build_dir is None:
+            build_dir = os.path.dirname(__file__) + "/build"
+        if implementation is None:
+            implementation = "radix_tree"
+        for f in (
+            build_dir + "/" + implementation + "/libconflict-set.so.0",
+            os.path.dirname(__file__)
+            + "/build/"
+            + implementation
+            + "/libconflict-set.0.dylib",
+        ):
+            try:
+                self._lib = ctypes.cdll.LoadLibrary(f)
+            except:
+                pass
+
+        if self._lib is None:
+            import sys
+
+            print(
+                "Could not find libconflict-set implementation " + implementation,
+                file=sys.stderr,
+            )
+            sys.exit(1)
+
+        self._lib.ConflictSet_create.argtypes = (ctypes.c_int64,)
+        self._lib.ConflictSet_create.restype = ctypes.c_void_p
+
+        self._lib.ConflictSet_check.argtypes = (
+            ctypes.c_void_p,
+            ctypes.POINTER(ReadRange),
+            ctypes.POINTER(ctypes.c_int),
+            ctypes.c_int,
+        )
+
+        self._lib.ConflictSet_addWrites.argtypes = (
+            ctypes.c_void_p,
+            ctypes.POINTER(WriteRange),
+            ctypes.c_int,
+            ctypes.c_int64,
+        )
+
+        self._lib.ConflictSet_setOldestVersion.argtypes = (
+            ctypes.c_void_p,
+            ctypes.c_int64,
+        )
+
+        self._lib.ConflictSet_destroy.argtypes = (ctypes.c_void_p,)
+
+        self._lib.ConflictSet_getBytes.argtypes = (ctypes.c_void_p,)
+        self._lib.ConflictSet_getBytes.restype = ctypes.c_int64
+
+        self.p = self._lib.ConflictSet_create(version)
+
+    def addWrites(self, version: int, *writes: WriteRange):
+        self._lib.ConflictSet_addWrites(
+            self.p, (WriteRange * len(writes))(*writes), len(writes), version
+        )
+
+    def check(self, *reads: ReadRange) -> list[Result]:
+        r = (ctypes.c_int * len(reads))()
+        self._lib.ConflictSet_check(
+            self.p, (ReadRange * len(reads))(*reads), r, len(reads)
+        )
+        return [Result(x) for x in r]
+
+    def setOldestVersion(self, version: int) -> None:
+        self._lib.ConflictSet_setOldestVersion(self.p, version)
+
+    def getBytes(self) -> int:
+        return self._lib.ConflictSet_getBytes(self.p)
+
+    def __enter__(self):
+        return self
+
+    def close(self) -> None:
+        if self.p is not None:
+            self._lib.ConflictSet_destroy(self.p)
+            self.p = None
+
+    def __exit__(self, exception_type, exception_value, exception_traceback):
+        self.close()

conflict_set_c_api_test.c

@@ -7,17 +7,19 @@ int main(void) {
   ConflictSet_WriteRange w;
   ConflictSet_Result result;
   ConflictSet_ReadRange r;
+  int64_t bytes;
   w.begin.p = (const uint8_t *)"0000";
   w.begin.len = 4;
   w.end.len = 0;
-  w.writeVersion = 1;
+  ConflictSet_addWrites(cs, &w, 1, 1);
-  ConflictSet_addWrites(cs, &w, 1);
   r.begin.p = (const uint8_t *)"0000";
   r.begin.len = 4;
   r.end.len = 0;
   r.readVersion = 0;
   ConflictSet_check(cs, &r, &result, 1);
   assert(result == ConflictSet_Conflict);
+  bytes = ConflictSet_getBytes(cs);
+  assert(bytes > 0);
   ConflictSet_destroy(cs);
   return 0;
 }

conflict_set_cxx_api_test.cpp

@@ -1,6 +1,9 @@
 #include "ConflictSet.h"
 
 #include <cassert>
+#include <cstdio>
+
+using namespace weaselab;
 
 int main(void) {
   ConflictSet cs(0);
@@ -8,8 +11,7 @@ int main(void) {
   w.begin.p = (const uint8_t *)"0000";
   w.begin.len = 4;
   w.end.len = 0;
-  w.writeVersion = 1;
+  cs.addWrites(&w, 1, 1);
-  cs.addWrites(&w, 1);
   ConflictSet::Result result;
   ConflictSet::ReadRange r;
   r.begin.p = (const uint8_t *)"0000";
@@ -18,4 +20,16 @@ int main(void) {
   r.readVersion = 0;
   cs.check(&r, &result, 1);
   assert(result == ConflictSet::Conflict);
+  int64_t bytes = cs.getBytes();
+  assert(bytes > 0);
+
+  ConflictSet::MetricsV1 *metrics;
+  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());
+  }
 }