[llvm] [X86] Prefer `lock or` over mfence (PR #106555)

[Phoebe Wang via llvm-commits]

phoebewang wrote:

> Instead of a tuning flag I'm very tempted to say that we create a general avoidMFence() method that always returns true for 64-bit targets - similar to what we do for hasCLFLUSH() etc.
> 
> @phoebewang any thoughts?

Sounds good. I don't like the tedious tunings for all modern targets.

https://github.com/llvm/llvm-project/pull/106555