HCC: Always produce one-as syncscopes until optimization in the backe…

…nd is implemented

Change-Id: I68d0b58824c75b5780a38f014a8246c8cea6ddda

lib/CodeGen/TargetInfo.cpp

@@ -8003,7 +8003,17 @@ AMDGPUTargetCodeGenInfo::getLLVMSyncScopeID(const LangOptions &LangOpts,
     Name = "wavefront";
   }
 
-  if (Ordering != llvm::AtomicOrdering::SequentiallyConsistent) {
+  // FIXME: It is incorrect for HIP and HCC to use one-as since
+  // it appears their memory model has a single happens-before
+  // relation across all address spaces. However, the recent fix
+  // in the backend to produce correct code has caused a
+  // performance regression in HIP/HCC. This can be mitigated by
+  // a backend optimization that is planned. Until that is
+  // implemented, this change makes HIP/HCC continue to generate
+  // the incorrect code. This will only break code that relies
+  // on a single happens-before between address spaces.
+  if (!LangOpts.OpenCL ||
+      Ordering != llvm::AtomicOrdering::SequentiallyConsistent) {
     if (!Name.empty())
       Name = Twine(Twine(Name) + Twine("-")).str();