replace mfence with lock'ed inst on x86

[d-netto]

A dummy instruction with lock prefix should provide the same sequential consistency guarantees as an mfence on x86.

This had a large performance impact when benchmarking work-stealing queues for parallel marking and it would be interesting to see how/if it affects performance in general.

CC: @vchuravy

[gbaraldi]

This seems to be fine, linux changed to this a little while ago https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=450cbdd0125cfa5d7bbf9e2a6b6961cc48d29730
We might want to do the exact same instruction they do since apparently they did some effort on what's better

[Keno]

Though as the kernel people, I'm not sure why the compiler doesn't just do this and I agree with @gbaraldi that it makes sense to align with the Linux-kernel instructions.

[Keno]

Looks like LLVM has support to emit this since https://reviews.llvm.org/D58632 and even emits it in the fallback case when the processor has no mfence. I don't know why it still emits mfence for the llvm fence intrinsic other than that nobody bothered updating it.

[Keno]

@d-netto want to submit an upstream PR to remove the mfence case to see what they think?

[d-netto]

I believe @vchuravy already submitted a patch to LLVM to replace thefence intrinsic with the locked instruction.

[Keno]

Ah indeed: https://reviews.llvm.org/D129947

[vtjnash]

I may possibly say we should rather just wait for that to catch up with gcc and https://reviews.llvm.org/D129947, than worry about maintaining this if something changes again

[vchuravy]

It will take a long time for the compilers to catch up. It landed in GCC12 and it will be a second before we get it into Clang.