schets / consume Goto Github PK

So, I see that this crate defines Consume to be Relaxed on hardware architectures that are known to honor data dependencies in their cache coherence model (which can roughly be summarized at "anything but DEC Alpha") and Acquire elsewhere.

I can see the correctness argument at the hardware level, but can you give me a quick idea of why it should also work at the compiler optimizer level? For example, given this kind of code...

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

...why can I be sure that no Sufficiently Smart Compiler™ will ever load the full contents of arr in registers at some point before reading from at, and then give me a stale copy of either arr[0] or arr[1] depending on the value of i that was read from at?