false_share.h

template <int align> struct alignas(align) A {
  volatile uint64_t count;
};

constexpr size_t iterations = 100'000'000;

void counter(volatile uint64_t *counterpt) {
  for (size_t i = 0; i < iterations; i++) {
    *counterpt = 3 * *counterpt * *counterpt + 1;
  }
}

template <int align> uint64_t threaded_counter() {
  A<align> a[4];
  static_assert(sizeof(a) == 4 * align);
  a[0].count = 1;
  a[1].count = 2;
  a[2].count = 3;
  a[3].count = 4;

  std::thread t1(counter, &a[0].count);
  std::thread t2(counter, &a[1].count);
  std::thread t3(counter, &a[2].count);
  std::thread t4(counter, &a[3].count);

  t1.join();
  t2.join();
  t3.join();
  t4.join();
  return a[0].count + a[1].count + a[2].count + a[3].count;
}