[clang] [Clang][AMDGPU] Stop defaulting to `one-as` for all atomic scopes (PR #120095)

[Joseph Huber via cfe-commits]

================
@@ -537,7 +537,8 @@ AMDGPUTargetCodeGenInfo::getLLVMSyncScopeID(const LangOptions &LangOpts,
     break;
   }
 
-  if (Ordering != llvm::AtomicOrdering::SequentiallyConsistent) {
+  if (LangOpts.OpenCL &&
----------------
jhuber6 wrote:

I'm sure it's using it somewhere since it's the OpenCL spec I think? This currently miscompiles on C/C++ w/ OpenMP so I think it should be fixed. Possible performance regressions but I don't think it's a valid optimization on non-OpenCL targets since we don't explicitly expose the address spaces as a part of the language.

https://github.com/llvm/llvm-project/pull/120095