Index: clang/lib/Basic/Targets.cpp
===================================================================
--- clang/lib/Basic/Targets.cpp
+++ clang/lib/Basic/Targets.cpp
@@ -1962,6 +1962,16 @@
     Opts.support("cl_khr_local_int32_base_atomics");
     Opts.support("cl_khr_local_int32_extended_atomics");
   }
+
+  CallingConvCheckResult checkCallingConvention(CallingConv CC) const override {
+    // CUDA compilations support all of the host's calling conventions.
+    //
+    // TODO: We should warn if you apply a non-default CC to anything other than
+    // a host function.
+    if (HostTarget)
+      return HostTarget->checkCallingConvention(CC);
+    return CCCR_Warning;
+  }
 };
 
 const Builtin::Info NVPTXTargetInfo::BuiltinInfo[] = {
Index: clang/test/SemaCUDA/cuda-inherits-calling-conv.cu
===================================================================
--- /dev/null
+++ clang/test/SemaCUDA/cuda-inherits-calling-conv.cu
@@ -0,0 +1,30 @@
+// RUN: %clang_cc1 -std=c++11 -triple i386-windows-msvc \
+// RUN:   -aux-triple nvptx-nvidia-cuda -fsyntax-only -verify %s
+
+// RUN: %clang_cc1 -std=c++11 -triple nvptx-nvidia-cuda \
+// RUN:   -aux-triple i386-windows-msvc -fsyntax-only \
+// RUN:   -fcuda-is-device -verify %s
+
+// RUN: %clang_cc1 -std=c++11 -triple nvptx-nvidia-cuda \
+// RUN:   -aux-triple x86_64-linux-gnu -fsyntax-only \
+// RUN:   -fcuda-is-device -verify -verify-ignore-unexpected=note \
+// RUN:   -DEXPECT_ERR %s
+
+// CUDA device code should inherit the host's calling conventions.
+
+template <class T>
+struct Foo;
+
+template <class T>
+struct Foo<T()> {};
+
+// On x86_64-linux-gnu, this is a redefinition of the template, because the
+// __fastcall calling convention doesn't exist (and is therefore ignored).
+#ifndef EXPECT_ERR
+// expected-no-diagnostics
+#else
+// expected-error@+4 {{redefinition of 'Foo}}
+// expected-warning@+3 {{calling convention '__fastcall' ignored}}
+#endif
+template <class T>
+struct Foo<T __fastcall()> {};