#include "ConflictSet.h"

#include <algorithm>
#include <cassert>
#include <cstdint>
#include <cstring>
#include <limits>
#include <map>
#include <set>
#include <span>
#include <string>
#include <string_view>
#include <utility>
#include <vector>

#ifdef HAS_AVX
#include <immintrin.h>
#elif defined(HAS_ARM_NEON)
#include <arm_neon.h>
#endif

#define DEBUG_VERBOSE 0

// GCOVR_EXCL_START

__attribute__((always_inline)) inline void *safe_malloc(size_t s) {
  if (void *p = malloc(s)) {
    return p;
  }
  abort();
}

// ==================== BEGIN ARENA IMPL ====================

/// Group allocations with similar lifetimes to amortize the cost of malloc/free
struct Arena {
  explicit Arena(int initialSize = 0);
  /// O(log n) in the number of allocations
  ~Arena();
  struct ArenaImpl;
  Arena(const Arena &) = delete;
  Arena &operator=(const Arena &) = delete;
  Arena(Arena &&other) noexcept;
  Arena &operator=(Arena &&other) noexcept;

private:
  ArenaImpl *impl = nullptr;
  friend void *operator new(size_t size, std::align_val_t align, Arena &arena);
};

inline void operator delete(void *, std::align_val_t, Arena &) {}
void *operator new(size_t size, std::align_val_t align, Arena &arena);
void *operator new(size_t size, std::align_val_t align, Arena *arena) = delete;

inline void operator delete(void *, Arena &) {}
inline void *operator new(size_t size, Arena &arena) {
  return operator new(size, std::align_val_t(alignof(std::max_align_t)), arena);
}
inline void *operator new(size_t size, Arena *arena) = delete;

inline void operator delete[](void *, Arena &) {}
inline void *operator new[](size_t size, Arena &arena) {
  return operator new(size, arena);
}
inline void *operator new[](size_t size, Arena *arena) = delete;

inline void operator delete[](void *, std::align_val_t, Arena &) {}
inline void *operator new[](size_t size, std::align_val_t align, Arena &arena) {
  return operator new(size, align, arena);
}
inline void *operator new[](size_t size, std::align_val_t align,
                            Arena *arena) = delete;

/// align must be a power of two
template <class T> T *align_up(T *t, size_t align) {
  auto unaligned = uintptr_t(t);
  auto aligned = (unaligned + align - 1) & ~(align - 1);
  return reinterpret_cast<T *>(reinterpret_cast<char *>(t) + aligned -
                               unaligned);
}

/// align must be a power of two
constexpr inline int align_up(uint32_t unaligned, uint32_t align) {
  return (unaligned + align - 1) & ~(align - 1);
}

/// Returns the smallest power of two >= x
constexpr inline uint32_t nextPowerOfTwo(uint32_t x) {
  return x <= 1 ? 1 : 1 << (32 - __builtin_clz(x - 1));
}

struct Arena::ArenaImpl {
  Arena::ArenaImpl *prev;
  int capacity;
  int used;
  uint8_t *begin() { return reinterpret_cast<uint8_t *>(this + 1); }
};

static_assert(sizeof(Arena::ArenaImpl) == 16);
static_assert(alignof(Arena::ArenaImpl) == 8);

Arena::Arena(int initialSize) : impl(nullptr) {
  if (initialSize > 0) {
    auto allocationSize = align_up(initialSize + sizeof(ArenaImpl), 16);
    impl = (Arena::ArenaImpl *)safe_malloc(allocationSize);
    impl->prev = nullptr;
    impl->capacity = allocationSize - sizeof(ArenaImpl);
    impl->used = 0;
  }
}

void onDestroy(Arena::ArenaImpl *impl) {
  while (impl) {
    auto *prev = impl->prev;
    free(impl);
    impl = prev;
  }
}

Arena::Arena(Arena &&other) noexcept
    : impl(std::exchange(other.impl, nullptr)) {}
Arena &Arena::operator=(Arena &&other) noexcept {
  onDestroy(impl);
  impl = std::exchange(other.impl, nullptr);
  return *this;
}

Arena::~Arena() { onDestroy(impl); }

void *operator new(size_t size, std::align_val_t align, Arena &arena) {
  int64_t aligned_size = size + size_t(align) - 1;
  if (arena.impl == nullptr ||
      (arena.impl->capacity - arena.impl->used) < aligned_size) {
    auto allocationSize = align_up(
        sizeof(Arena::ArenaImpl) +
            std::max<int>(aligned_size,
                          (arena.impl ? std::max<int>(sizeof(Arena::ArenaImpl),
                                                      arena.impl->capacity * 2)
                                      : 0)),
        16);
    auto *impl = (Arena::ArenaImpl *)safe_malloc(allocationSize);
    impl->prev = arena.impl;
    impl->capacity = allocationSize - sizeof(Arena::ArenaImpl);
    impl->used = 0;
    arena.impl = impl;
  }
  auto *result =
      align_up(arena.impl->begin() + arena.impl->used, size_t(align));
  auto usedDelta = (result - arena.impl->begin()) + size - arena.impl->used;
  arena.impl->used += usedDelta;
  return result;
}

/// STL-friendly allocator using an arena
template <class T> struct ArenaAlloc {
  typedef T value_type;

  ArenaAlloc() = delete;
  explicit ArenaAlloc(Arena *arena) : arena(arena) {}

  Arena *arena;

  template <class U> constexpr ArenaAlloc(const ArenaAlloc<U> &other) noexcept {
    arena = other.arena;
  }

  [[nodiscard]] T *allocate(size_t n) {
    if (n > 0xfffffffffffffffful / sizeof(T)) { // NOLINT
      __builtin_unreachable();
    }

    return static_cast<T *>((void *)new (std::align_val_t(alignof(T)), *arena)
                                uint8_t[n * sizeof(T)]); // NOLINT
  }

  void deallocate(T *, size_t) noexcept {}
};

template <class T> using Vector = std::vector<T, ArenaAlloc<T>>;
template <class T> auto vector(Arena &arena) {
  return Vector<T>(ArenaAlloc<T>(&arena));
}
template <class T> using Set = std::set<T, std::less<T>, ArenaAlloc<T>>;
template <class T> auto set(Arena &arena) {
  return Set<T>(ArenaAlloc<T>(&arena));
}

template <class T, class U>
bool operator==(const ArenaAlloc<T> &lhs, const ArenaAlloc<U> &rhs) {
  return lhs.arena == rhs.arena;
}

template <class T, class U>
bool operator!=(const ArenaAlloc<T> &lhs, const ArenaAlloc<U> &rhs) {
  return !(lhs == rhs);
}

// ==================== END ARENA IMPL ====================

// ==================== BEGIN ARBITRARY IMPL ====================

/// Think of `Arbitrary` as an attacker-controlled random number generator.
/// Usually you want your random number generator to be fair, so that you can
/// sensibly analyze probabilities. E.g. The analysis that shows that quicksort
/// is expected O(n log n) with a random pivot relies on the random pivot being
/// selected uniformly from a fair distribution.
///
/// Other times you want your randomness to be diabolically unfair, like when
/// looking for bugs and fuzzing. The random-number-like interface is still
/// convenient here, but you can potentially get much better coverage by
/// allowing the possibility of e.g. flipping heads 100 times in a row.
///
/// When it runs out of entropy, it always returns 0.
struct Arbitrary {
  Arbitrary() = default;

  explicit Arbitrary(std::span<const uint8_t> bytecode) : bytecode(bytecode) {}

  /// Draws an arbitrary uint32_t
  uint32_t next() { return consume<4>(); }

  /// Draws an arbitrary element from [0, s)
  uint32_t bounded(uint32_t s);

  /// Fill `bytes` with `size` arbitrary bytes
  void randomBytes(uint8_t *bytes, int size) {
    int toFill = std::min<int>(size, bytecode.size());
    if (toFill > 0) {
      memcpy(bytes, bytecode.data(), toFill);
    }
    bytecode = bytecode.subspan(toFill, bytecode.size() - toFill);
    memset(bytes + toFill, 0, size - toFill);
  }

  /// Fill `bytes` with `size` random hex bytes
  void randomHex(uint8_t *bytes, int size) {
    for (int i = 0; i < size;) {
      uint8_t arbitrary = consume<1>();
      bytes[i++] = "0123456789abcdef"[arbitrary & 0xf];
      arbitrary >>= 4;
      if (i < size) {
        bytes[i++] = "0123456789abcdef"[arbitrary & 0xf];
      }
    }
  }

  template <class T, class = std::enable_if_t<std::is_trivially_copyable_v<T>>>
  T randT() {
    T t;
    randomBytes((uint8_t *)&t, sizeof(T));
    return t;
  }

  bool hasEntropy() const { return bytecode.size() != 0; }

private:
  uint8_t consumeByte() {
    if (bytecode.size() == 0) {
      return 0;
    }
    auto result = bytecode[0];
    bytecode = bytecode.subspan(1, bytecode.size() - 1);
    return result;
  }

  template <int kBytes> uint32_t consume() {
    uint32_t result = 0;
    static_assert(kBytes <= 4);
    for (int i = 0; i < kBytes; ++i) {
      result <<= 8;
      result |= consumeByte();
    }
    return result;
  }

  std::span<const uint8_t> bytecode;
};

Arbitrary gArbitrary;

uint32_t Arbitrary::bounded(uint32_t s) {
  if (s == 1) {
    return 0;
  }
  switch (32 - __builtin_clz(s - 1)) {
  case 1:
  case 2:
  case 3:
  case 4:
  case 5:
  case 6:
  case 7:
  case 8:
    return consume<1>() % s;
  case 9:
  case 10:
  case 11:
  case 12:
  case 13:
  case 14:
  case 15:
  case 16:
    return consume<2>() % s;
  case 17:
  case 18:
  case 19:
  case 20:
  case 21:
  case 22:
  case 23:
  case 24:
    return consume<3>() % s;
  default:
    return consume<4>() % 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) {
    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) {
        results[i] = ConflictSet::TooOld;
        continue;
      }
      auto begin =
          std::string((const char *)reads[i].begin.p, reads[i].begin.len);
      auto end =
          reads[i].end.len == 0
              ? begin + std::string("\x00", 1)
              : std::string((const char *)reads[i].end.p, reads[i].end.len);
      int64_t maxVersion = oldestVersion;
      for (auto iter = --writeVersionMap.upper_bound(begin),
                endIter = writeVersionMap.lower_bound(end);
           iter != endIter; ++iter) {
        maxVersion = std::max(maxVersion, iter->second);
      }
      results[i] = maxVersion > reads[i].readVersion ? ConflictSet::Conflict
                                                     : ConflictSet::Commit;
    }
  }
  void addWrites(const ConflictSet::WriteRange *writes, int count) {
    for (int i = 0; i < count; ++i) {
      auto begin =
          std::string((const char *)writes[i].begin.p, writes[i].begin.len);
      auto end =
          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);
           iter != endIter;) {
        iter = writeVersionMap.erase(iter);
      }
      writeVersionMap[begin] = writeVersion;
      writeVersionMap[end] = prevVersion;
    }
  }

  void setOldestVersion(int64_t oldestVersion) {
    this->oldestVersion = oldestVersion;
  }

  void printLogical(std::string &result) {
    for (const auto &[k, v] : writeVersionMap) {
      std::string key;
      for (uint8_t c : k) {
        key += "x";
        key += "0123456789abcdef"[c / 16];
        key += "0123456789abcdef"[c % 16];
      }
      result += key + " -> " + std::to_string(v) + "\n";
    }
  }

  int64_t oldestVersion;
  std::map<std::string, int64_t> writeVersionMap;
};

using Key = ConflictSet::Key;

Key toKey(Arena &arena, int n) {
  constexpr int kMaxLength = 8;
  int i = kMaxLength;
  uint8_t *itoaBuf = new (arena) uint8_t[kMaxLength];
  memset(itoaBuf, '0', kMaxLength);
  do {
    itoaBuf[--i] = "0123456789abcdef"[n % 16];
    n /= 16;
  } while (n);
  return Key{itoaBuf, kMaxLength};
}

Key toKeyAfter(Arena &arena, int n) {
  constexpr int kMaxLength = 8;
  int i = kMaxLength;
  uint8_t *itoaBuf = new (arena) uint8_t[kMaxLength + 1];
  memset(itoaBuf, '0', kMaxLength);
  itoaBuf[kMaxLength] = 0;
  do {
    itoaBuf[--i] = "0123456789abcdef"[n % 16];
    n /= 16;
  } while (n);
  return Key{itoaBuf, kMaxLength + 1};
}

// ==================== END UTILITIES IMPL ====================

// ==================== BEGIN IMPLEMENTATION ====================

// GCOVR_EXCL_STOP

struct Entry {
  int64_t pointVersion;
  int64_t rangeVersion;
};

enum class Type : int8_t {
  Node4,
  Node16,
  Node48,
  Node256,
  Invalid,
};
struct Node {
  /* begin section that's copied to the next node */
  Node *parent = nullptr;
  int64_t maxVersion;
  Entry entry;
  int16_t numChildren = 0;
  bool entryPresent = false;
  uint8_t parentsIndex = 0;
  /* end section that's copied to the next node */

  Type type = Type::Invalid;
};
Node *getChild(Node *self, uint8_t index);
int getChildGeq(Node *self, int child);
Node *&getOrCreateChild(Node *&self, uint8_t index);
Node *newNode();
void eraseChild(Node *self, uint8_t index);

struct Node4 : Node {
  // Sorted
  uint8_t index[4] = {};
  Node *children[4] = {};
  Node4() { this->type = Type::Node4; }
};

Node *newNode() { return new (safe_malloc(sizeof(Node4))) Node4; }

struct Node16 : Node {
  // Sorted
  uint8_t index[16] = {};
  Node *children[16] = {};
  Node16() { this->type = Type::Node16; }
};

struct Node48 : Node {
  int8_t nextFree = 0;
  int8_t index[256];
  Node *children[48] = {};
  Node48() {
    this->type = Type::Node48;
    memset(index, -1, 256);
  }
};

struct Node256 : Node {
  Node *children[256] = {};
  Node256() { this->type = Type::Node256; }
};

int getNodeIndex(Node4 *self, uint8_t index) {
  for (int i = 0; i < self->numChildren; ++i) {
    if (self->index[i] == index) {
      return i;
    }
  }
  return -1;
}

int getNodeIndex(Node16 *self, uint8_t index) {
#ifdef HAS_AVX
  // Based on https://www.the-paper-trail.org/post/art-paper-notes/

  // key_vec is 16 repeated copies of the searched-for byte, one for every
  // possible position in child_keys that needs to be searched.
  __m128i key_vec = _mm_set1_epi8(index);

  // Compare all child_keys to 'index' in parallel. Don't worry if some of the
  // keys aren't valid, we'll mask the results to only consider the valid ones
  // below.
  __m128i indices;
  memcpy(&indices, self->index, sizeof(self->index));
  __m128i results = _mm_cmpeq_epi8(key_vec, indices);

  // Build a mask to select only the first node->num_children values from the
  // comparison (because the other values are meaningless)
  int mask = (1 << self->numChildren) - 1;

  // Change the results of the comparison into a bitfield, masking off any
  // invalid comparisons.
  int bitfield = _mm_movemask_epi8(results) & mask;

  // No match if there are no '1's in the bitfield.
  if (bitfield == 0)
    return -1;

  // Find the index of the first '1' in the bitfield by counting the leading
  // zeros.
  return __builtin_ctz(bitfield);
#elif defined(HAS_ARM_NEON)
  // Based on
  // https://community.arm.com/arm-community-blogs/b/infrastructure-solutions-blog/posts/porting-x86-vector-bitmask-optimizations-to-arm-neon

  uint8x16_t indices;
  memcpy(&indices, self->index, sizeof(self->index));
  // 0xff for each match
  uint16x8_t results =
      vreinterpretq_u16_u8(vceqq_u8(vdupq_n_u8(index), indices));
  uint64_t mask = self->numChildren == 16
                      ? uint64_t(-1)
                      : (uint64_t(1) << (self->numChildren * 4)) - 1;
  // 0xf for each match in valid range
  uint64_t bitfield =
      vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(results, 4)), 0) & mask;
  if (bitfield == 0)
    return -1;
  return __builtin_ctzll(bitfield) / 4;
#else
  for (int i = 0; i < self->numChildren; ++i) {
    if (self->index[i] == index) {
      return i;
    }
  }
  return -1;
#endif
}

#ifdef HAS_AVX
int firstNonNeg1(const int8_t x[16]) {
  __m128i key_vec = _mm_set1_epi8(-1);
  __m128i indices;
  memcpy(&indices, x, 16);
  __m128i results = _mm_cmpeq_epi8(key_vec, indices);
  uint32_t bitfield = _mm_movemask_epi8(results) ^ 0xffff;
  if (bitfield == 0)
    return -1;
  return __builtin_ctz(bitfield);
}

int lastNonNeg1(const int8_t x[16]) {
  __m128i key_vec = _mm_set1_epi8(-1);
  __m128i indices;
  memcpy(&indices, x, 16);
  __m128i results = _mm_cmpeq_epi8(key_vec, indices);
  uint32_t bitfield = _mm_movemask_epi8(results) ^ 0xffff;
  if (bitfield == 0)
    return -1;
  return 31 - __builtin_clz(bitfield);
}
#endif

#ifdef HAS_ARM_NEON
int firstNonNeg1(const int8_t x[16]) {
  uint8x16_t indices;
  memcpy(&indices, x, 16);
  uint16x8_t results = vreinterpretq_u16_u8(vceqq_u8(vdupq_n_u8(-1), indices));
  uint64_t bitfield =
      ~vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(results, 4)), 0);
  if (bitfield == 0)
    return -1;
  return __builtin_ctzll(bitfield) / 4;
}

int lastNonNeg1(const int8_t x[16]) {
  uint8x16_t indices;
  memcpy(&indices, x, 16);
  uint16x8_t results = vreinterpretq_u16_u8(vceqq_u8(vdupq_n_u8(-1), indices));
  uint64_t bitfield =
      ~vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(results, 4)), 0);
  if (bitfield == 0)
    return -1;
  return 15 - __builtin_clzll(bitfield) / 4;
}
#endif

Node *getChild(Node *self, uint8_t index) {
  if (self->type == Type::Node4) {
    auto *self4 = static_cast<Node4 *>(self);
    int i = getNodeIndex(self4, index);
    if (i >= 0) {
      return self4->children[i];
    }
    return nullptr;
  } else if (self->type == Type::Node16) {
    auto *self16 = static_cast<Node16 *>(self);
    int i = getNodeIndex(self16, index);
    if (i >= 0) {
      return self16->children[i];
    }
    return nullptr;
  } else if (self->type == Type::Node48) {
    auto *self48 = static_cast<Node48 *>(self);
    int secondIndex = self48->index[index];
    if (secondIndex >= 0) {
      return self48->children[secondIndex];
    }
    return nullptr;
  } else {
    auto *self256 = static_cast<Node256 *>(self);
    return self256->children[index];
  }
}

// Precondition - an entry for index must exist in the node
Node *&getChildExists(Node *self, uint8_t index) {
  if (self->type == Type::Node4) {
    auto *self4 = static_cast<Node4 *>(self);
    return self4->children[getNodeIndex(self4, index)];
  } else if (self->type == Type::Node16) {
    auto *self16 = static_cast<Node16 *>(self);
    return self16->children[getNodeIndex(self16, index)];
  } else if (self->type == Type::Node48) {
    auto *self48 = static_cast<Node48 *>(self);
    int secondIndex = self48->index[index];
    if (secondIndex >= 0) {
      return self48->children[secondIndex];
    }
  } else {
    auto *self256 = static_cast<Node256 *>(self);
    return self256->children[index];
  }
  __builtin_unreachable(); // GCOVR_EXCL_LINE
}

int getChildGeq(Node *self, int child) {
  if (self->type == Type::Node4) {
    auto *self4 = static_cast<Node4 *>(self);
    for (int i = 0; i < self->numChildren; ++i) {
      if (i > 0) {
        assert(self4->index[i - 1] < self4->index[i]);
      }
      if (self4->index[i] >= child) {
        return self4->index[i];
      }
    }
  } else if (self->type == Type::Node16) {
    auto *self16 = static_cast<Node16 *>(self);
    for (int i = 0; i < self->numChildren; ++i) {
      if (i > 0) {
        assert(self16->index[i - 1] < self16->index[i]);
      }
      if (self16->index[i] >= child) {
        return self16->index[i];
      }
    }
  } else if (self->type == Type::Node48) {
    auto *self48 = static_cast<Node48 *>(self);
#if defined(HAS_AVX) || defined(HAS_ARM_NEON)
    int i = child;
    for (; (i & 0xf) != 0; ++i) {
      if (self48->index[i] >= 0) {
        assert(self48->children[self48->index[i]] != nullptr);
        return i;
      }
    }
    for (; i < 256; i += 16) {
      auto result = firstNonNeg1(self48->index + i);
      if (result != -1) {
        return i + result;
      }
    }
#else
    for (int i = child; i < 256; ++i) {
      if (self48->index[i] >= 0) {
        assert(self48->children[self48->index[i]] != nullptr);
        return i;
      }
    }
#endif
  } else {
    auto *self256 = static_cast<Node256 *>(self);
    // For some reason gcc can't auto vectorize this, and the plain loop is
    // faster.
#if defined(__clang__)
    int i = child;
    constexpr int kUnrollCount = 8; // Must be a power of two and <= 8
    for (; (i & (kUnrollCount - 1)) != 0; ++i) {
      if (self256->children[i]) {
        return i;
      }
    }
    for (; i < 256; i += kUnrollCount) {
      uint8_t nonNull[kUnrollCount];
      for (int j = 0; j < kUnrollCount; ++j) {
        nonNull[j] = self256->children[i + j] != nullptr ? 0xff : 0;
      }
      uint64_t word;
      memcpy(&word, nonNull, kUnrollCount);
      if (word) {
        return i + __builtin_ctzll(word) / 8;
      }
    }
#else
    for (int i = child; i < 256; ++i) {
      if (self256->children[i]) {
        return i;
      }
    }
#endif
  }
  return -1;
}

int getChildLeq(Node *self, int child) {
  if (self->type == Type::Node4) {
    auto *self4 = static_cast<Node4 *>(self);
    for (int i = self->numChildren - 1; i >= 0; --i) {
      if (i > 0) {
        assert(self4->index[i - 1] < self4->index[i]);
      }
      if (self4->index[i] <= child) {
        return self4->index[i];
      }
    }
  } else if (self->type == Type::Node16) {
    auto *self16 = static_cast<Node16 *>(self);
    for (int i = self->numChildren - 1; i >= 0; --i) {
      if (i > 0) {
        assert(self16->index[i - 1] < self16->index[i]);
      }
      if (self16->index[i] <= child) {
        return self16->index[i];
      }
    }
  } else if (self->type == Type::Node48) {
    auto *self48 = static_cast<Node48 *>(self);
    // TODO the plain loop is faster?
#if defined(HAS_AVX) || defined(HAS_ARM_NEON)
    int i = child;
    if (i < 0) {
      return -1;
    }
    for (; (i & 0xf) != 0; --i) {
      if (self48->index[i] >= 0) {
        assert(self48->children[self48->index[i]] != nullptr);
        return i;
      }
    }
    if (self48->index[i] >= 0) {
      assert(self48->children[self48->index[i]] != nullptr);
      return i;
    }
    i -= 16;
    for (; i >= 0; i -= 16) {
      auto result = lastNonNeg1(self48->index + i);
      if (result != -1) {
        return i + result;
      }
    }
#else
    for (int i = child; i >= 0; --i) {
      if (self48->index[i] >= 0) {
        assert(self48->children[self48->index[i]] != nullptr);
        return i;
      }
    }
#endif
  } else {
    auto *self256 = static_cast<Node256 *>(self);
    // TODO: The plain loop is faster?
#if defined(__clang__)
    int i = child;
    constexpr int kUnrollCount = 8; // Must be a power of two and <= 8
    for (; (i & (kUnrollCount - 1)) != 0; --i) {
      if (self256->children[i]) {
        return i;
      }
    }
    if (self256->children[i]) {
      return i;
    }
    i -= kUnrollCount;
    for (; i >= 0; i -= kUnrollCount) {
      uint8_t nonNull[kUnrollCount];
      for (int j = 0; j < kUnrollCount; ++j) {
        nonNull[j] = self256->children[i + j] != nullptr ? 0xff : 0;
      }
      uint64_t word;
      memcpy(&word, nonNull, kUnrollCount);
      if (word) {
        return i + 7 - __builtin_clzll(word) / 8;
      }
    }
#else
    for (int i = child; i >= 0; --i) {
      if (self256->children[i]) {
        return i;
      }
    }
#endif
  }
  return -1;
}

void setChildrenParents(Node *node) {
  for (int i = getChildGeq(node, 0); i >= 0; i = getChildGeq(node, i + 1)) {
    getChildExists(node, i)->parent = node;
  }
}

Node *&getOrCreateChild(Node *&self, uint8_t index) {
  if (self->type == Type::Node4) {
    auto *self4 = static_cast<Node4 *>(self);
    {
      int i = getNodeIndex(self4, index);
      if (i >= 0) {
        return self4->children[i];
      }
    }
    if (self->numChildren == 4) {
      auto *newSelf = new (safe_malloc(sizeof(Node16))) Node16;
      memcpy((void *)newSelf, self, offsetof(Node, type));
      memcpy(newSelf->index, self4->index, 4);
      memcpy(newSelf->children, self4->children, 4 * sizeof(void *));
      free(std::exchange(self, newSelf));
      setChildrenParents(self);
      goto insert16;
    } else {
      ++self->numChildren;
      for (int i = 0; i < int(self->numChildren) - 1; ++i) {
        if (int(self4->index[i]) > int(index)) {
          memmove(self4->index + i + 1, self4->index + i,
                  self->numChildren - (i + 1));
          memmove(self4->children + i + 1, self4->children + i,
                  (self->numChildren - (i + 1)) * sizeof(void *));
          self4->index[i] = index;
          self4->children[i] = nullptr;
          return self4->children[i];
        }
      }
      self4->index[self->numChildren - 1] = index;
      self4->children[self->numChildren - 1] = nullptr;
      return self4->children[self->numChildren - 1];
    }
  } else if (self->type == Type::Node16) {
  insert16:
    auto *self16 = static_cast<Node16 *>(self);
    {
      int i = getNodeIndex(self16, index);
      if (i >= 0) {
        return self16->children[i];
      }
    }
    if (self->numChildren == 16) {
      auto *newSelf = new (safe_malloc(sizeof(Node48))) Node48;
      memcpy((void *)newSelf, self, offsetof(Node, type));
      newSelf->nextFree = 16;
      int i = 0;
      for (auto x : self16->index) {
        newSelf->children[i] = self16->children[i];
        newSelf->index[x] = i;
        ++i;
      }
      assert(i == 16);
      free(std::exchange(self, newSelf));
      setChildrenParents(self);
      goto insert48;
    } else {
      ++self->numChildren;
      for (int i = 0; i < int(self->numChildren) - 1; ++i) {
        if (int(self16->index[i]) > int(index)) {
          memmove(self16->index + i + 1, self16->index + i,
                  self->numChildren - (i + 1));
          memmove(self16->children + i + 1, self16->children + i,
                  (self->numChildren - (i + 1)) * sizeof(void *));
          self16->index[i] = index;
          self16->children[i] = nullptr;
          return self16->children[i];
        }
      }
      self16->index[self->numChildren - 1] = index;
      self16->children[self->numChildren - 1] = nullptr;
      return self16->children[self->numChildren - 1];
    }
  } else if (self->type == Type::Node48) {
  insert48:
    auto *self48 = static_cast<Node48 *>(self);
    int secondIndex = self48->index[index];
    if (secondIndex >= 0) {
      return self48->children[secondIndex];
    }
    if (self->numChildren == 48) {
      auto *newSelf = new (safe_malloc(sizeof(Node256))) Node256;
      memcpy((void *)newSelf, self, offsetof(Node, type));
      for (int i = 0; i < 256; ++i) {
        if (self48->index[i] >= 0) {
          newSelf->children[i] = self48->children[self48->index[i]];
        }
      }
      free(std::exchange(self, newSelf));
      self = newSelf;
      setChildrenParents(self);
      goto insert256;
    } else {
      ++self->numChildren;
      assert(self48->nextFree < 48);
      self48->index[index] = self48->nextFree;
      self48->children[self48->nextFree] = nullptr;
      return self48->children[self48->nextFree++];
    }
  } else {
  insert256:
    auto *self256 = static_cast<Node256 *>(self);
    if (!self256->children[index]) {
      ++self->numChildren;
    }
    return self256->children[index];
  }
}

// Precondition - an entry for index must exist in the node
void eraseChild(Node *self, uint8_t index) {
  if (self->type == Type::Node4) {
    auto *self4 = static_cast<Node4 *>(self);
    int nodeIndex = getNodeIndex(self4, index);
    memmove(self4->index + nodeIndex, self4->index + nodeIndex + 1,
            sizeof(self4->index[0]) * (self->numChildren - (nodeIndex + 1)));
    memmove(self4->children + nodeIndex, self4->children + nodeIndex + 1,
            sizeof(self4->children[0]) * // NOLINT
                (self->numChildren - (nodeIndex + 1)));
  } else if (self->type == Type::Node16) {
    auto *self16 = static_cast<Node16 *>(self);
    int nodeIndex = getNodeIndex(self16, index);
    memmove(self16->index + nodeIndex, self16->index + nodeIndex + 1,
            sizeof(self16->index[0]) * (self->numChildren - (nodeIndex + 1)));
    memmove(self16->children + nodeIndex, self16->children + nodeIndex + 1,
            sizeof(self16->children[0]) * // NOLINT
                (self->numChildren - (nodeIndex + 1)));
  } else if (self->type == Type::Node48) {
    auto *self48 = static_cast<Node48 *>(self);
    int8_t toRemoveChildrenIndex = std::exchange(self48->index[index], -1);
    int8_t lastChildrenIndex = --self48->nextFree;
    assert(toRemoveChildrenIndex >= 0);
    assert(lastChildrenIndex >= 0);
    if (toRemoveChildrenIndex != lastChildrenIndex) {
      self48->children[toRemoveChildrenIndex] =
          std::exchange(self48->children[lastChildrenIndex], nullptr);
      self48->index[self48->children[toRemoveChildrenIndex]->parentsIndex] =
          toRemoveChildrenIndex;
    }
  } else {
    auto *self256 = static_cast<Node256 *>(self);
    self256->children[index] = nullptr;
  }
  --self->numChildren;
  if (self->numChildren == 0 && !self->entryPresent &&
      self->parent != nullptr) {
    eraseChild(self->parent, self->parentsIndex);
  }
}

Node *nextPhysical(Node *node) {
  int index = -1;
  for (;;) {
    auto nextChild = getChildGeq(node, index + 1);
    if (nextChild >= 0) {
      return getChildExists(node, nextChild);
    }
    index = node->parentsIndex;
    node = node->parent;
    if (node == nullptr) {
      return nullptr;
    }
  }
}

Node *nextLogical(Node *node) {
  for (node = nextPhysical(node); node != nullptr && !node->entryPresent;
       node = nextPhysical(node))
    ;
  return node;
}

std::string printable(std::string_view key) {
  std::string result;
  for (uint8_t c : key) {
    result += "x";
    result += "0123456789abcdef"[c / 16];
    result += "0123456789abcdef"[c % 16];
  }
  return result;
}

std::string printable(const Key &key) {
  return printable(std::string_view((const char *)key.p, key.len));
}

std::string_view getSearchPath(Arena &arena, Node *n) {
  if (n->parent == nullptr) {
    return {};
  }
  auto result = vector<char>(arena);
  for (; n->parent != nullptr; n = n->parent) {
    result.push_back(n->parentsIndex);
  }
  std::reverse(result.begin(), result.end());
  return std::string_view((const char *)&result[0], result.size()); // NOLINT
}

void printLogical(std::string &result, Node *node) {
  Arena arena;
  for (Node *iter = node; iter != nullptr;) {
    auto *next = nextLogical(iter);
    std::string key;
    for (uint8_t c : getSearchPath(arena, iter)) {
      key += "x";
      key += "0123456789abcdef"[c / 16];
      key += "0123456789abcdef"[c % 16];
    }
    if (iter->entry.pointVersion == iter->entry.rangeVersion) {
      result += key + " -> " + std::to_string(iter->entry.pointVersion) + "\n";
    } else {
      result += key + " -> " + std::to_string(iter->entry.pointVersion) + "\n";
      if (next == nullptr || (getSearchPath(arena, next) !=
                              (std::string(getSearchPath(arena, iter)) +
                               std::string("\x00", 1)))) {
        result +=
            key + "x00 -> " + std::to_string(iter->entry.rangeVersion) + "\n";
      }
    }
    iter = next;
  }
}

Node *prevPhysical(Node *node) {
  assert(node->parent != nullptr);
  auto prevChild = getChildLeq(node->parent, node->parentsIndex - 1);
  assert(prevChild < node->parentsIndex);
  if (prevChild >= 0) {
    node = getChildExists(node->parent, prevChild);
    // Move down the right spine
    for (;;) {
      auto rightMostChild = getChildLeq(node, 255);
      if (rightMostChild >= 0) {
        node = getChildExists(node, rightMostChild);
      } else {
        return node;
      }
    }
  } else {
    return node->parent;
  }
}

struct Iterator {
  Node *n;
  int cmp;
};

std::string_view getSearchPath(Arena &arena, Node *n);

Iterator lastLeq(Node *n, const std::span<const uint8_t> key) {
  auto remaining = key;
  for (;;) {
    Arena arena;
    assert((std::string(getSearchPath(arena, n)) +
            std::string((const char *)remaining.data(), remaining.size()))
               .ends_with(std::string((const char *)key.data(), key.size())));
    if (remaining.size() == 0) {
      // We've found the physical node corresponding to search path `key`
      if (n->entryPresent) {
        return {n, 0};
      } else {
        break;
      }
    } else {
      int c = getChildLeq(n, remaining[0]);
      if (c == remaining[0]) {
        n = getChildExists(n, c);
        remaining = remaining.subspan(1, remaining.size() - 1);
      } else {
        // The physical node corresponding to search path `key` does not exist.
        // Let's find the physical node corresponding to the highest search key
        // (not necessarily present) less than key
        // Move down the right spine
        for (;;) {
          if (c >= 0) {
            n = getChildExists(n, c);
          } else {
            break;
          }
          c = getChildLeq(n, 255);
        }
        break;
      }
    }
  }
  // Iterate backwards along existing physical nodes until we find a present
  // entry
  for (; !n->entryPresent; n = prevPhysical(n)) {
  }
  return {n, -1};
}

void insert(Node **self_, std::span<const uint8_t> key, int64_t writeVersion) {
  for (;;) {
    auto &self = *self_;
    self->maxVersion = writeVersion;
    if (key.size() == 0) {
      auto l = lastLeq(self, key);
      self->entryPresent = true;
      self->entry.pointVersion = writeVersion;
      assert(l.n != nullptr);
      assert(l.n->entryPresent);
      self->entry.rangeVersion = l.n->entry.rangeVersion;
      return;
    }
    auto &child = getOrCreateChild(self, key.front());
    if (!child) {
      child = newNode();
      child->parent = self;
      child->parentsIndex = key.front();
    }
    self_ = &child;
    key = key.subspan(1, key.size() - 1);
  }
}

std::string printable(std::string_view key);
std::string printable(const Key &key);

struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
  void check(const ReadRange *reads, Result *result, int count) const {
    for (int i = 0; i < count; ++i) {
      const auto &r = reads[i];
      if (r.readVersion < oldestVersion) {
        result[i] = TooOld;
        continue;
      }
      // TODO support non-point reads
      assert(r.end.len == 0);
      auto [l, c] =
          lastLeq(root, std::span<const uint8_t>(r.begin.p, r.begin.len));
#if DEBUG_VERBOSE && !defined(NDEBUG)
      Arena arena;
      printf("LastLeq for `%s' got `%s'\n", printable(r.begin).c_str(),
             printable(getSearchPath(arena, l)).c_str());
#endif
      assert(l != nullptr);
      assert(l->entryPresent);
      result[i] = (c == 0 ? l->entry.pointVersion : l->entry.rangeVersion) >
                          r.readVersion
                      ? Conflict
                      : Commit;
    }
  }
  void addWrites(const WriteRange *writes, int count) {
    for (int i = 0; i < count; ++i) {
      const auto &w = writes[i];
      // TODO support non-point writes
      assert(w.end.len == 0);
      insert(&root, std::span<const uint8_t>(w.begin.p, w.begin.len),
             w.writeVersion);
    }
  }
  void setOldestVersion(int64_t oldestVersion) {
    this->oldestVersion = oldestVersion;
  }
  explicit Impl(int64_t oldestVersion) : oldestVersion(oldestVersion) {
    // Insert ""
    root = newNode();
    root->maxVersion = oldestVersion;
    root->entry.pointVersion = oldestVersion;
    root->entry.rangeVersion = oldestVersion;
    root->entryPresent = true;
  }
  ~Impl() {
    Arena arena;
    auto toFree = vector<Node *>(arena);
    toFree.push_back(root);
    while (toFree.size() > 0) {
      auto *n = toFree.back();
      toFree.pop_back();
      // Add all children to toFree
      for (int child = getChildGeq(n, 0); child >= 0;
           child = getChildGeq(n, child + 1)) {
        auto *c = getChildExists(n, child);
        assert(c != nullptr);
        toFree.push_back(c);
      }
      free(n);
    }
  }

  Node *root;
  int64_t oldestVersion;
};

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) {
  return impl->addWrites(writes, count);
}

void ConflictSet::setOldestVersion(int64_t oldestVersion) {
  return impl->setOldestVersion(oldestVersion);
}

ConflictSet::ConflictSet(int64_t oldestVersion, [[maybe_unused]] uint64_t seed)
    : impl(new(safe_malloc(sizeof(Impl))) Impl{oldestVersion}) {}

ConflictSet::~ConflictSet() {
  if (impl) {
    impl->~Impl();
    free(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) {
  ((ConflictSet::Impl *)cs)->addWrites(writes, count);
}
__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, uint64_t) {
  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();
  free(cs);
}
}

// ==================== END IMPLEMENTATION ====================

// GCOVR_EXCL_START

std::string printable(std::string_view key) {
  std::string result;
  for (uint8_t c : key) {
    result += "x";
    result += "0123456789abcdef"[c / 16];
    result += "0123456789abcdef"[c % 16];
  }
  return result;
}

std::string printable(const Key &key) {
  return printable(std::string_view((const char *)key.p, key.len));
}

std::string_view getSearchPath(Arena &arena, Node *n) {
  if (n->parent == nullptr) {
    return {};
  }
  auto result = vector<char>(arena);
  for (; n->parent != nullptr; n = n->parent) {
    result.push_back(n->parentsIndex);
  }
  std::reverse(result.begin(), result.end());
  return std::string_view((const char *)&result[0], result.size()); // NOLINT
}

void printLogical(std::string &result, Node *node) {
  Arena arena;
  for (Node *iter = node; iter != nullptr;) {
    auto *next = nextLogical(iter);
    std::string key;
    for (uint8_t c : getSearchPath(arena, iter)) {
      key += "x";
      key += "0123456789abcdef"[c / 16];
      key += "0123456789abcdef"[c % 16];
    }
    if (iter->entry.pointVersion == iter->entry.rangeVersion) {
      result += key + " -> " + std::to_string(iter->entry.pointVersion) + "\n";
    } else {
      result += key + " -> " + std::to_string(iter->entry.pointVersion) + "\n";
      if (next == nullptr || (getSearchPath(arena, next) !=
                              (std::string(getSearchPath(arena, iter)) +
                               std::string("\x00", 1)))) {
        result +=
            key + "x00 -> " + std::to_string(iter->entry.rangeVersion) + "\n";
      }
    }
    iter = next;
  }
}

void debugPrintDot(FILE *file, Node *node) {

  struct DebugDotPrinter {

    explicit DebugDotPrinter(FILE *file) : file(file) {}

    void print(Node *n) {
      assert(n != nullptr);
      if (n->entryPresent) {
        fprintf(file, " k_%p [label=\"m=%d p=%d r=%d\"];\n", (void *)n,
                int(n->maxVersion), int(n->entry.pointVersion),
                int(n->entry.rangeVersion));
      } else {
        fprintf(file, " k_%p [label=\"m=%d\"];\n", (void *)n,
                int(n->maxVersion));
      }
      for (int child = getChildGeq(n, 0); child >= 0;
           child = getChildGeq(n, child + 1)) {
        auto *c = getChildExists(n, child);
        fprintf(file, " k_%p -> k_%p [label=\"'%02x'\"];\n", (void *)n,
                (void *)c, child);
        print(c);
      }
    }
    FILE *file;
  };

  fprintf(file, "digraph ConflictSet {\n");
  assert(node != nullptr);
  DebugDotPrinter printer{file};
  printer.print(node);
  fprintf(file, "}\n");
}

void checkParentPointers(Node *node, bool &success) {
  for (int i = getChildGeq(node, 0); i >= 0; i = getChildGeq(node, i + 1)) {
    auto *child = getChildExists(node, i);
    if (child->parent != node) {
      Arena arena;
      fprintf(stderr, "%s child %d has parent pointer %p. Expected %p\n",
              printable(getSearchPath(arena, node)).c_str(), i,
              (void *)child->parent, (void *)node);
      success = false;
    }
    checkParentPointers(child, success);
  }
}

int64_t checkMaxVersion(Node *node, bool &success) {
  int64_t expected =
      node->entryPresent
          ? std::max(node->entry.pointVersion, node->entry.rangeVersion)
          : std::numeric_limits<int64_t>::lowest();
  for (int i = getChildGeq(node, 0); i >= 0; i = getChildGeq(node, i + 1)) {
    auto *child = getChildExists(node, i);
    expected = std::max(expected, checkMaxVersion(child, success));
  }
  if (node->maxVersion != expected) {
    Arena arena;
    fprintf(stderr, "%s has max version %d. Expected %d\n",
            printable(getSearchPath(arena, node)).c_str(),
            int(node->maxVersion), int(expected));
    success = false;
  }
  return expected;
}

bool checkCorrectness(Node *node, ReferenceImpl &refImpl) {
  bool success = true;

  checkParentPointers(node, success);

  std::string logicalMap;
  std::string referenceLogicalMap;
  printLogical(logicalMap, node);
  refImpl.printLogical(referenceLogicalMap);
  if (logicalMap != referenceLogicalMap) {
    fprintf(stderr,
            "Logical map not equal to reference logical map.\n\nActual:\n"
            "%s\nExpected:\n%s\n",
            logicalMap.c_str(), referenceLogicalMap.c_str());
    success = false;
  }

  return success;
}
namespace std {
void __throw_length_error(const char *) { __builtin_unreachable(); }
} // namespace std

#ifdef ENABLE_MAIN
int main(void) {
  int64_t writeVersion = 0;
  ConflictSet::Impl cs{writeVersion};
  ReferenceImpl refImpl{writeVersion};
  Arena arena;
  constexpr int kNumKeys = 10;
  auto *write = new (arena) ConflictSet::WriteRange[kNumKeys];
  for (int i = 0; i < kNumKeys; ++i) {
    write[i].begin = toKey(arena, i);
    write[i].end.len = 0;
    write[i].writeVersion = ++writeVersion;
  }
  cs.addWrites(write, kNumKeys);
  refImpl.addWrites(write, kNumKeys);
  debugPrintDot(stdout, cs.root);
  return 0;
}
#endif

#ifdef ENABLE_FUZZ
extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
  // TODO call setOldestVersion, and check range writes/reads
  gArbitrary = Arbitrary{{data, size}};

  int64_t writeVersion = 0;
  ConflictSet::Impl cs{writeVersion};
  ReferenceImpl refImpl{writeVersion};

  while (gArbitrary.hasEntropy()) {
    Arena arena;
    {
      int numWrites = gArbitrary.bounded(10);
      int64_t v = ++writeVersion;
      auto *writes = new (arena) ConflictSet::WriteRange[numWrites];
      auto keys = set<std::string_view>(arena);
      while (int(keys.size()) < numWrites) {
        if (!gArbitrary.hasEntropy()) {
          // Tell the fuzzer it's not interesting
          return -1;
        }
        int keyLen = gArbitrary.bounded(8);
        auto *begin = new (arena) uint8_t[keyLen];
        gArbitrary.randomBytes(begin, keyLen);
        keys.insert(std::string_view((const char *)begin, keyLen));
      }
      auto iter = keys.begin();
      for (int i = 0; i < numWrites; ++i) {
        writes[i].begin.p = (const uint8_t *)iter->data();
        writes[i].begin.len = iter->size();
        ++iter;
        writes[i].end.len = 0;
        writes[i].writeVersion = v;
#if DEBUG_VERBOSE && !defined(NDEBUG)
        printf("Write: {%s} -> %d\n", printable(writes[i].begin).c_str(),
               int(writes[i].writeVersion));
#endif
      }
      assert(iter == keys.end());
      cs.addWrites(writes, numWrites);
      refImpl.addWrites(writes, numWrites);
    }
    bool success = checkCorrectness(cs.root, refImpl);
    if (!success) {
      abort();
    }
    {
      int numReads = gArbitrary.bounded(10);
      int64_t v = writeVersion - gArbitrary.bounded(10);
      auto *reads = new (arena) ConflictSet::ReadRange[numReads];
      auto keys = set<std::string_view>(arena);
      while (int(keys.size()) < numReads) {
        if (!gArbitrary.hasEntropy()) {
          // Tell the fuzzer it's not interesting
          return -1;
        }
        int keyLen = gArbitrary.bounded(8);
        auto *begin = new (arena) uint8_t[keyLen];
        gArbitrary.randomBytes(begin, keyLen);
        keys.insert(std::string_view((const char *)begin, keyLen));
      }
      auto iter = keys.begin();
      for (int i = 0; i < numReads; ++i) {
        reads[i].begin.p = (const uint8_t *)iter->data();
        reads[i].begin.len = iter->size();
        ++iter;
        reads[i].end.len = 0;
        reads[i].readVersion = v;
#if DEBUG_VERBOSE && !defined(NDEBUG)
        printf("Read: {%s} at %d\n", printable(reads[i].begin).c_str(),
               int(reads[i].readVersion));
#endif
      }
      assert(iter == keys.end());
      auto *results1 = new (arena) ConflictSet::Result[numReads];
      auto *results2 = new (arena) ConflictSet::Result[numReads];
      cs.check(reads, results1, numReads);
      refImpl.check(reads, results2, numReads);
      for (int i = 0; i < numReads; ++i) {
        if (results1[i] != results2[i]) {
          fprintf(stderr, "Expected %d, got %d for read of %s at version %d\n",
                  results2[i], results1[i], printable(reads[i].begin).c_str(),
                  int(reads[i].readVersion));
          std::string referenceLogicalMap;
          refImpl.printLogical(referenceLogicalMap);
          fprintf(stderr, "Logical map:\n\n%s\n", referenceLogicalMap.c_str());
          abort();
        }
      }
    }
  }
  return 0;
}
#endif

// GCOVR_EXCL_STOP