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

[Alexander Kornienko via llvm-commits]

alexfh wrote:

We're seeing a ~40% regression in a macrobenchmark after this change. Is Skylake-X considered modern enough to benefit from `lock or` or would `mfence` still be better for this microarchitecture?

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