main.cpp

// complie with:
// clang++ -O3 -std=c++11 -pthread -fsanitize=thread main.cpp

#include <atomic>
#include <thread>
#include <iostream>

static uint8_t data;
static std::atomic<bool> flag = {false};

static int produce(uint8_t d) {
    if (flag.load(std::memory_order_relaxed)) { // memory_order_consume ??
        return -1;
    }
    data = d;
    flag.store(true, std::memory_order_release);
    return 0;
}

static int consume() {
    if (!flag.load(std::memory_order_acquire)) {
        return -1;
    }
    uint8_t d = data;
    flag.store(false, std::memory_order_release);
    return d;
}

int sum1 = 0;
int sum2 = 0;

void foo() {
    for (int i = 0; i < 1000; i++) {
        int v = (uint8_t)i;
        while (produce(v) < 0) {}
        sum1 += v;
    }
}

void bar() {
    for (int i = 0; i < 1000; i++) {
        int v;
        while (v = consume(), v < 0) {}
        sum2 += v;
    }
}

int main() {
    std::thread t1(foo);
    std::thread t2(bar);
    t2.join();
    t1.join();
    std::cout << "t1: " << sum1 << "\n";
    std::cout << "t2: " << sum2 << "\n";
    return 0;
}