[llvm] [RFC][BPF] Do atomic_fetch_*() pattern matching with memory ordering (PR #107343)

[via llvm-commits]

eddyz87 wrote:

> More broadly speaking (non-BPF specific), if I know I'm ok with relaxed ordering, but some architecture only supports SEQ_CST, I'm fine with that as a writer of atomic code. It's much bigger PITA to work around all these version/arch specific issues if they become compile-time or runtime errors, IMO.

Fair enough.
@yonghong-song , what do you think?

https://github.com/llvm/llvm-project/pull/107343