[llvm] BPF: Ensure __sync_fetch_and_add() always generate atomic_fetch_add insn (PR #101428)

[via llvm-commits]

yonghong-song wrote:

clang with default architecture and llc with bpf target works in some cases but not always, e.g. x86_64 target may generate (e.g. jump table) which bpf back will not support. In early days, when CORE is not available, we do use clang with default architecture and llc with bpf model, but that is complicated and need to access kernel internal headers. When CORE is available, the recommendation is to use 'clang --target=bpf ...' directly which is also we did in bpf selftests.

https://github.com/llvm/llvm-project/pull/101428