#include <iostream>
#include <thread>
#include <mutex>

#define CACHE_LINE_SIZE 64

// struct alignas(64) data {
struct data {
   char data[2*CACHE_LINE_SIZE];
   void store(int value) {
      int* ip = (int*)(data + CACHE_LINE_SIZE - sizeof(int) / 2);
      *ip = value;
   }
   int load() {
      int* ip = (int*)(data + CACHE_LINE_SIZE - sizeof(int) / 2);
      return *ip;
   }
};
__attribute__((aligned(CACHE_LINE_SIZE))) struct data data;

std::mutex cout_mutex;

void dosth() {
   for (int i = 0; i < 10000; ++i) {
      int value = data.load();
      int val1 = value & 0xffff;
      int val2 = (value >> 16) & 0xffff;
      if (val1 != val2) {
	 std::lock_guard<std::mutex> lock(cout_mutex);
	 std::cout << "val1 = " << val1 << ", val2 = " << val2 << std::endl;
      }
      value = ((val1 + 1) << 16) | (val1 + 1);
      data.store(value);
   }
}

int main() {
   std::thread threads[10];
   for (int i = 0; i < 10; ++i) {
      threads[i] = std::thread(dosth);
   }
   for (auto& t: threads) t.join();
}