Proof sketch that using Relaxed is okay? about consume HOT 1 OPEN

A possible workaround would be to insert compiler_fence(Ordering::Acquire) after Consume loads on platforms where Consume is mapped to Relaxed. This would be heavier than necessary, as it would prevent all accesses from being reordered, not just dependent ones, and also has the disadvantage of needing to be manually done unless custom atomic types are created, but it's better than nothing.

For example, given the following:

use consume::Consume;
use std::sync::atomic::{compiler_fence, Ordering};

pub fn consume_fence() {
    if (Consume == Ordering::Relaxed) {
        compiler_fence(Ordering::Acquire);
    }
}

the example above can be made to work by inserting a call to consume_fence:

let at = AtomicUsize::new(0);
let mut arr = [UnsafeCell::new(4.2); 2];  // Note that this array is very small

/* ... stuff happens, including in other threads ... */

let i = at.load(Consume);
consume_fence();
let x = unsafe { *arr[i].get() };

from consume.