cache miss on bounded crossbeam_channel::Receiver try_recv

[boranby]

Hi, I don't know whether this is an expected behavior of Atomics but when I run perf on my code I see that there are cache misses on try_recv call. I have a bounded queue, which has 8 senders and 1 receiver. In the sender sides, I couldn't find any notable miss, because of that I raise that issue for the receiver side. My call:

loop {
    while let Ok(mut messages) = self.parsed_feed_receiver.try_recv() {
        ...
    }
        ...
}

There are other small misses but they are negligible compared to these 3 lines. If you can help to figure this out, I'll appreciate that. Thanks in advance.

[taiki-e]

It seems that the access to the field is causing loading from memory, so this patch may improve the situation: https://github.com/crossbeam-rs/crossbeam/compare/master...taiki-e/local-mark-bit

[boranby]

@taiki-e thanks for your quick response! I tried with the new branch, unfortunately it's still same.

[taiki-e]

Hmm, since mfence does the serialization, the previous cache should not be available for the load from r15+0x80 (i.e., self.tail), but I thought the load from r15+0x190 was the access to self.mark_bit, so I thought that could be omitted. However, normal offset of self.mark_bit is actually 0x108...