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

Summary:
The documentation at
https://llvm.org/docs/AMDGPUUsage.html#memory-scopes states that these
'one-as' modifiers are more specific versions of the scopes that only
apply to a specific address space. This doesn't make sense for fences
which have no associated address space to use, and it's a more
restrictive version the normal scope. This should not tbe the default
behavior, but it is currently emitted in all cases except for
sequentially consistent.

Author: jhuber6

clang/lib/CodeGen/Targets/AMDGPU.cpp

@@ -530,7 +530,11 @@ AMDGPUTargetCodeGenInfo::getLLVMSyncScopeID(const LangOptions &LangOpts,
     break;
   }
 
-  if (Ordering != llvm::AtomicOrdering::SequentiallyConsistent) {
+  // OpenCL assumes by default that atomic scopes are per-address space for
+  // non-sequentially consistent operations.
+  if (Scope >= SyncScope::OpenCLWorkGroup &&
+      Scope <= SyncScope::OpenCLSubGroup &&
+      Ordering != llvm::AtomicOrdering::SequentiallyConsistent) {
     if (!Name.empty())
       Name = Twine(Twine(Name) + Twine("-")).str();