Index: cfe/trunk/lib/Sema/SemaDeclAttr.cpp
===================================================================
--- cfe/trunk/lib/Sema/SemaDeclAttr.cpp
+++ cfe/trunk/lib/Sema/SemaDeclAttr.cpp
@@ -3350,6 +3350,10 @@
   D->addAttr(::new (S.Context)
               CUDAGlobalAttr(Attr.getRange(), S.Context,
                              Attr.getAttributeSpellingListIndex()));
+
+  // Add implicit attribute((used)) so we don't eliminate kernels
+  // because there is nothing referencing them on device side.
+  D->addAttr(UsedAttr::CreateImplicit(S.Context));
 }
 
 static void handleGNUInlineAttr(Sema &S, Decl *D, const AttributeList &Attr) {
Index: cfe/trunk/test/CodeGenCUDA/ptx-kernels.cu
===================================================================
--- cfe/trunk/test/CodeGenCUDA/ptx-kernels.cu
+++ cfe/trunk/test/CodeGenCUDA/ptx-kernels.cu
@@ -1,7 +1,16 @@
+// Make sure that __global__ functions are emitted along with correct
+// annotations and are added to @llvm.used to prevent their elimination.
+// REQUIRES: nvptx-registered-target
+//
 // RUN: %clang_cc1 %s -triple nvptx-unknown-unknown -fcuda-is-device -emit-llvm -o - | FileCheck %s
 
 #include "Inputs/cuda.h"
 
+// Make sure that all __global__ functiona are added to @llvm.used
+// CHECK: @llvm.used = appending global
+// CHECK-SAME: @global_function
+// CHECK-SAME: @_Z16templated_kernelIiEvT_
+
 // CHECK-LABEL: define void @device_function
 extern "C"
 __device__ void device_function() {}
@@ -13,4 +22,10 @@
   device_function();
 }
 
+// Make sure host-instantiated kernels are preserved on device side.
+template <typename T> __global__ void templated_kernel(T param) {}
+// CHECK-LABEL: define linkonce_odr void @_Z16templated_kernelIiEvT_
+void host_function() { templated_kernel<<<0,0>>>(0); }
+
 // CHECK: !{{[0-9]+}} = !{void ()* @global_function, !"kernel", i32 1}
+// CHECK: !{{[0-9]+}} = !{void (i32)* @_Z16templated_kernelIiEvT_, !"kernel", i32 1}