An atomic operation can be optimized?

[wangrunji0408]

When I perf a program based on Tokio threaded scheduler on ARM64 Linux, I found this atomic operation consumes a lot of time:

tokio/tokio/src/runtime/thread_pool/idle.rs

Line 136 in 2e05399

let state = State(self.state.fetch_add(0, SeqCst));

It can be optimized to an atomic load, which prevents cache invalidation on other cores:

let state = State(self.state.load(SeqCst));

After that the overhead of Idle::worker_to_notify reduced from 10% to 4% on my workload.

And I'm not sure if it is safe to relax the memory consistency from SeqCst to Acquire and Release. If possible, it will improve the performance of ISAs with weak memory model, such as ARM.

Is that reasonable?

[sfackler]

That fetch_add is there intentionally: #1788