[clang] [HIP] use offload wrapper for non-device-only non-rdc (PR #132869)

[Yaxun Liu via cfe-commits]

================
@@ -4945,12 +4959,15 @@ Action *Driver::BuildOffloadingActions(Compilation &C,
       }
     }
 
-    // Compiling HIP in non-RDC mode requires linking each action individually.
+    // Compiling HIP in device-only non-RDC mode requires linking each action
+    // individually.
     for (Action *&A : DeviceActions) {
       if ((A->getType() != types::TY_Object &&
            A->getType() != types::TY_LTO_BC) ||
           Kind != Action::OFK_HIP ||
-          Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, false))
+          Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc,
+                       false) ||
+          !offloadDeviceOnly())
----------------
yxsamliu wrote:

For HIP no-rdc device-only compilation, the existing HIP apps expect the output to be bundled or unbundled code objects so that they can be loaded directly without JIT at run time. We could use linker wrapper to generate them without wrapping them in a host object, but that would be a separate patch.

https://github.com/llvm/llvm-project/pull/132869