[llvm] [OpenMP] Replace AMDGPU fences with generic scoped fences (PR #119619)

[Joseph Huber via llvm-commits]

================
@@ -26,6 +26,14 @@ enum OrderingTy {
   seq_cst = __ATOMIC_SEQ_CST,
 };
 
+enum ScopeTy {
+  system = __MEMORY_SCOPE_SYSTEM,
+  device_ = __MEMORY_SCOPE_DEVICE,
----------------
jhuber6 wrote:

Yes, was too lazy to unify it right this moment since I don't want to delete the CUDA handling yet. Maybe someone like @AlexMaclean knows the status of lowering stuff like https://godbolt.org/z/cWs77s9Eq accurately for NVPTX without NVVM.

https://github.com/llvm/llvm-project/pull/119619