Index: include/clang/Driver/Action.h
===================================================================
--- include/clang/Driver/Action.h
+++ include/clang/Driver/Action.h
@@ -160,14 +160,16 @@
 class CudaHostAction : public Action {
   virtual void anchor();
   ActionList DeviceActions;
+  const char *DeviceTriple;
 
 public:
-  CudaHostAction(std::unique_ptr<Action> Input,
-                 const ActionList &DeviceActions);
+  CudaHostAction(std::unique_ptr<Action> Input, const ActionList &DeviceActions,
+                 const char *DeviceTriple);
   ~CudaHostAction() override;
 
   ActionList &getDeviceActions() { return DeviceActions; }
   const ActionList &getDeviceActions() const { return DeviceActions; }
+  const char *getDeviceTriple() const { return DeviceTriple; }
 
   static bool classof(const Action *A) { return A->getKind() == CudaHostClass; }
 };
Index: lib/Driver/Action.cpp
===================================================================
--- lib/Driver/Action.cpp
+++ lib/Driver/Action.cpp
@@ -66,8 +66,10 @@
 void CudaHostAction::anchor() {}
 
 CudaHostAction::CudaHostAction(std::unique_ptr<Action> Input,
-                               const ActionList &DeviceActions)
-    : Action(CudaHostClass, std::move(Input)), DeviceActions(DeviceActions) {}
+                               const ActionList &DeviceActions,
+                               const char *DeviceTriple)
+    : Action(CudaHostClass, std::move(Input)), DeviceActions(DeviceActions),
+      DeviceTriple(DeviceTriple) {}
 
 CudaHostAction::~CudaHostAction() {
   for (auto &DA : DeviceActions)
Index: lib/Driver/Driver.cpp
===================================================================
--- lib/Driver/Driver.cpp
+++ lib/Driver/Driver.cpp
@@ -1247,13 +1247,18 @@
 buildCudaActions(const Driver &D, const ToolChain &TC, DerivedArgList &Args,
                  const Arg *InputArg, std::unique_ptr<Action> HostAction,
                  ActionList &Actions) {
+  // Figure out which NVPTX triple to use for device-side compilation based on
+  // whether host is 64-bit.
+  const char *DeviceTriple = TC.getTriple().isArch64Bit()
+                                 ? "nvptx64-nvidia-cuda"
+                                 : "nvptx-nvidia-cuda";
   Arg *PartialCompilationArg = Args.getLastArg(options::OPT_cuda_host_only,
                                                options::OPT_cuda_device_only);
   // Host-only compilation case.
   if (PartialCompilationArg &&
       PartialCompilationArg->getOption().matches(options::OPT_cuda_host_only))
     return std::unique_ptr<Action>(
-        new CudaHostAction(std::move(HostAction), {}));
+        new CudaHostAction(std::move(HostAction), {}, DeviceTriple));
 
   // Collect all cuda_gpu_arch parameters, removing duplicates.
   SmallVector<const char *, 4> GpuArchList;
@@ -1291,12 +1296,6 @@
     }
   }
 
-  // Figure out which NVPTX triple to use for device-side compilation based on
-  // whether host is 64-bit.
-  const char *DeviceTriple = TC.getTriple().isArch64Bit()
-                                 ? "nvptx64-nvidia-cuda"
-                                 : "nvptx-nvidia-cuda";
-
   // Figure out what to do with device actions -- pass them as inputs to the
   // host action or run each of them independently.
   bool DeviceOnlyCompilation = PartialCompilationArg != nullptr;
@@ -1332,7 +1331,7 @@
   // Return a new host action that incorporates original host action and all
   // device actions.
   return std::unique_ptr<Action>(
-      new CudaHostAction(std::move(HostAction), DeviceActions));
+      new CudaHostAction(std::move(HostAction), DeviceActions, DeviceTriple));
 }
 
 void Driver::BuildActions(const ToolChain &TC, DerivedArgList &Args,