在C11中具有非原子的栅栏(Fences with non-atomics in C11)

 有没有办法使用栅栏来推断C11中非原子操作的行为？ 具体来说，我想在某些字段需要为int以与旧接口兼容的情况下使代码安全，这些旧接口可能会读取和写入数据结构到文件或将它们作为系统调用参数传递。 因为没有要求atomic_int甚至与int大小相同，所以我不能使用atomic_int 。 
 
 这是一个最小的工作示例，遗憾的是根据5.1.2.4第25节产生了未定义的行为，因为数据竞争ready ： 
 
#include <stdatomic.h>
#include <stdio.h>
#include <threads.h>

int ready;  /* purposely NOT _Atomic */
int value;

void
p1()
{
  value = 1;
  atomic_thread_fence(memory_order_release);
  ready = 1;
}

void
p2(void *_ignored)
{
  while (!ready)
    ;
  atomic_thread_fence(memory_order_acquire);
  printf("%d\n", value);
}

int
main()
{
  thrd_t t;
  thrd_create(&t, p2, NULL);
  p1();
  thrd_join(&t, NULL);
}
 
 我的具体问题是，是否可以修复上述代码以保证打印1而不用ready _Atomic 。 （我可以ready一个volatile ，但是在规范中没有看到任何有助于此的建议。） 
 
 一个相关的问题是，无论如何编写上面的代码是否安全，因为我的代码运行的任何机器都具有缓存一致性？ 我知道当C11程序包含所谓的良性竞赛时， 很多 事情都会出错 ，所以我真的在寻找一个合理的编译器和架构可以对上面的代码做什么的细节，而不是关于数据竞争和未定义的一般警告行为。 

Is there any way to use fences to reason about the behavior of non-atomic operations in C11? Specifically, I'd like to make code safe in situations where certain fields are required to be ints for compatibility with old interfaces that might, say, read and write data structures to files or pass them as system call arguments. As there's no requirement that an atomic_int even be the same size as an int, I can't use an atomic_int.
 
Here's a minimal working example that unfortunately produces undefined behavior according to section 5.1.2.4 paragraph 25, because of the data race on ready:
 
#include <stdatomic.h>
#include <stdio.h>
#include <threads.h>

int ready;  /* purposely NOT _Atomic */
int value;

void
p1()
{
  value = 1;
  atomic_thread_fence(memory_order_release);
  ready = 1;
}

void
p2(void *_ignored)
{
  while (!ready)
    ;
  atomic_thread_fence(memory_order_acquire);
  printf("%d\n", value);
}

int
main()
{
  thrd_t t;
  thrd_create(&t, p2, NULL);
  p1();
  thrd_join(&t, NULL);
}
 
My specific question is whether it's possible to fix the above code to guarantee printing 1 without changing ready to an _Atomic. (I could make ready a volatile, but don't see any suggestion in the spec that this would help.)
 
A related question is whether it's safe to write the above code anyway, because any machine my code will run on has cache coherence? I'm aware that many things go wrong when C11 programs contain so-called benign races, so I'm really looking for the specifics of what a plausible compiler and architecture could do to the above code rather than general warnings about data races and undefined behavior.

原文：https://stackoverflow.com/questions/42122610