Can std::atomic cancel out increments with decrements?

Question

Using relaxed memory order, e.g. for a reference counting pointer, would the compiler be allowed to optimize away a subsequent increment and decrement?

std::atomic_int32_t ai;
for (size_t i = 0; i < 10000; i++)
{
    ai.fetch_add(1, std::memory_order_relaxed);
    ai.fetch_sub(1, std::memory_order_relaxed);
}

Looking at disassembly it doesn't look like. But since reordering is allowed and the atomic is behaving like a counter, just thread-safe, one could argue that he could optimize as if it would be a plain int.

Answer 1

I believe it can be optimized, unless declared volatile. The reason is that for any schedule that interleaves some thread in between, there exist valid schedule that does not. I believe it to be the case for drf-sc memory model too.

It would not be the case if this thread reads something in between.