Index: clang/lib/Sema/SemaCUDA.cpp
===================================================================
--- clang/lib/Sema/SemaCUDA.cpp
+++ clang/lib/Sema/SemaCUDA.cpp
@@ -790,9 +790,12 @@
   // If the caller is known-emitted, mark the callee as known-emitted.
   // Otherwise, mark the call in our call graph so we can traverse it later.
   bool CallerKnownEmitted = IsKnownEmitted(*this, Caller);
-  if (CallerKnownEmitted)
-    MarkKnownEmitted(*this, Caller, Callee, Loc);
-  else {
+  if (CallerKnownEmitted) {
+    // Host-side references to a __global__ function refer to the stub, so the
+    // function itself is never emitted and therefore should not be marked.
+    if (getLangOpts().CUDAIsDevice || IdentifyCUDATarget(Callee) != CFT_Global)
+      MarkKnownEmitted(*this, Caller, Callee, Loc);
+  } else {
     // If we have
     //   host fn calls kernel fn calls host+device,
     // the HD function does not get instantiated on the host.  We model this by
Index: clang/test/SemaCUDA/call-device-fn-from-host.cu
===================================================================
--- clang/test/SemaCUDA/call-device-fn-from-host.cu
+++ clang/test/SemaCUDA/call-device-fn-from-host.cu
@@ -83,3 +83,10 @@
 __host__ __device__ void fn_ptr_template() {
   auto* ptr = &device_fn;  // Not an error because the template isn't instantiated.
 }
+
+// Launching a kernel from a host function does not result in code generation
+// for it, so calling HD function which calls a D function should not trigger
+// errors.
+static __host__ __device__ void hd_func() { device_fn(); }
+__global__ void kernel() { hd_func(); }
+void host_func(void) { kernel<<<1, 1>>>(); }