Contest

This is a small experiment to test the effects of unsynchronized, non-atomic writes to a single u64.

The Experiment

The following code is basically pseudocode in Rust syntax. It won't compile because it's not safe.

Let's start with a single 64-bit unsigned integer.

let n: u64 = 0;

Then, if we generate a random number r and xor it with n an even number of times, n should still be zero.

let r: u64 = rng.gen();
for _ in 0..1024 { // the upper bound is any even number
    n = n ^ r
}

Things get more interesting if we involve concurrency. The tricky part is:

n = n ^ r

If the load on the right and the assignment on the left aren't done atomically, it's possible that another thread will operate on n in the middle of our "fetch-xor" operation. This will break our invariant of xoring an even number of times.

Implementation

The code accomplishes unsynchronized data access with UnsafeCell:

#[derive(Clone)]
pub struct SharedUnsync(Arc<UnsafeCell<u64>>);

impl SharedUnsync {
    fn fetch_xor(&self, other: u64) {
        // SAFETY: very unsafe.
        unsafe { *self.0.get() ^= other }
    }
}

We also have something like a control group using AtomicU64:

#[derive(Clone)]
pub struct SharedAtomic(Arc<AtomicU64>);

fn fetch_xor(&self, other: u64) {
    self.0.fetch_xor(other, Ordering::Relaxed);
}

Rust Version

$ rustc --version
rustc 1.58.1 (db9d1b20b 2022-01-20)

Results

Both unsync and atomic

The two fetch_xor operations ("atomic" and "unsync") are executed sequentially in the source. When both are enabled, we see that the unsynchronized value is corrupted while the atomic stays intact.

unsync: 0011000101111101110010000001000111001011011101010000010010110111
atomic: 0000000000000000000000000000000000000000000000000000000000000000

Only unsync

When we comment out the atomics-synchronized fetch_add, we actually get zero back for unsync. This was surprising and I feel like there's some spooky UB optimization going on. It also completes much more quickly (300x as quickly) without the atomics. This is also surprising because a relaxed load should compile to a regular load, as far as I know.

unsync: 0000000000000000000000000000000000000000000000000000000000000000

Only atomic

Finally, if we only leave the atomic fetch_add operation in the source and comment out the unsynchronized writes (effectively commenting out the UB) we get zero back like we'd expect.

atomic: 0000000000000000000000000000000000000000000000000000000000000000

Note that we left both values initialized to zero in each of the above variations.

vlmutolo/contest