Potential performance increases found

[JackThomson2]

Hi love this library and using it a lot!

I was having a poke through the code and was looking at the adaptive backoff spinning code and putting it into godbolt to see what it was doing. I noticed that LLVM wasn't actually unrolling the loops for the backoff spinning, so I had try at forcing it to do so.

What I found was actually fairly significant performance increases over current master. You can view the code I changed here:
https://rust.godbolt.org/z/f3Psh5end

Here is an example of the performance increases I was getting:

Compared to master here:

Just wondering what your thoughts where? I ran this benchmark multiple times to be sure but the results seemed to be consistently better. If you like the sound of it I can create a PR to make the changes.

Many Thanks

Jack

[Amanieu]

This is slightly counter-intuitive, but throughput increases when the spin takes longer. The reason for this is that a spinning core isn't fighting with another core for exclusive ownership of a cacheline, which allows a single core to quickly lock/unlock the mutex in succession without interference. While this is good for throughput, it is bad for overall system performance because you end up with a lot of core spinning instead of doing useful work (e.g. background processes).

[JackThomson2]

I see, but as far as I can tell this isn't doing any less or more spinning, it just seems to call the pause instructions more efficiently.

I could be completely wrong however.