Is it ever useful for only one thread to use memory_order_seq_cst?

user_7509065 0
Viewed 75

In a multithreaded program, do you ever get any guarantees from memory_order_seq_cst that you wouldn't also get from a weaker ordering, if you only have one thread that uses it? Example:

#include <stdatomic.h>

extern atomic_int x, y;

void doSomething(void);

void thread1(void) {
    atomic_store_explicit(&x, 1, memory_order_seq_cst);
    atomic_thread_fence(memory_order_seq_cst);
    if(atomic_load_explicit(&y, memory_order_seq_cst)) {
        doSomething();
    }
}

void thread2(void) {
    atomic_store_explicit(&y, 1, memory_order_release);
    atomic_thread_fence(memory_order_acq_rel);
    if(atomic_load_explicit(&x, memory_order_acquire)) {
        doSomething();
    }
}

If thread1 and thread2 are running in parallel, is it guaranteed that exactly one will always call doSomething?

stdatomic memory-barriers language-lawyer atomic c
1 Answers

I believe no, it doesn't give any extra guarantees if only one thread uses it.

The only difference between seq-cst and acqure/release is the existence of the global seq-cst operation order.

Other threads can't observe nor influence this order without performing seq-cst operations. This order thus copies the "sequenced-before" order of the single thread that uses seq-cst operations, since nothing else constrains it.

user_27520750
Related