[CUDA] Let NVPTX inherit the host's calling conventions.

[jlebar]

Summary

When compiling device code, we may still see host code with explicit
calling conventions. NVPTX needs to claim that it supports these CCs,
so that (a) we don't raise noisy warnings, and (b) we don't break
existing code which relies on the existence of these CCs when
specializing templates. (If a CC doesn't exist, clang ignores it, so
two template specializations which are different only insofar as one
specifies a CC are considered identical and therefore are an error if
that CC is not supported.)

clang/lib/Basic/Targets.cpp

void setSupportedOpenCLOpts() override {
 
     Opts.support("cl_khr_fp64");
     Opts.support("cl_khr_byte_addressable_store");
     Opts.support("cl_khr_global_int32_base_atomics");
     Opts.support("cl_khr_global_int32_extended_atomics");
     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[] = {
 #define BUILTIN(ID, TYPE, ATTRS)                                               \
   { #ID, TYPE, ATTRS, nullptr, ALL_LANGUAGES, nullptr },
 #define LIBBUILTIN(ID, TYPE, ATTRS, HEADER)                                    \
   { #ID, TYPE, ATTRS, HEADER, ALL_LANGUAGES, nullptr },
 #define TARGET_BUILTIN(ID, TYPE, ATTRS, FEATURE)                               \

clang/test/SemaCUDA/cuda-inherits-calling-conv.cu

+// 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()> {};