CUDA: mark implicit intrinsics properly

[eliben]

Summary

r218624 implemented target inference for implicit special members. However, other entities can be implicit - for example intrinsics. These can not have inference running on them, so they should be marked host device as before. This is the safest and most flexible setting, since by construction these functions don't invoke anything, and we'd like them to be invokable from both host and device code. LLVM's intrinsics definitions (where these intrinsics come from in the case of CUDA/NVPTX) have no notion of target, so both host and device intrinsics can be supported this way.

[rnk]

lgtm

[eliben]

r218688.

lib/Sema/SemaCUDA.cpp

Sema::CUDAFunctionTarget Sema::IdentifyCUDATarget(const FunctionDecl *D) {
 
   if (D->hasAttr<CUDAGlobalAttr>())
     return CFT_Global;
 
   if (D->hasAttr<CUDADeviceAttr>()) {
     if (D->hasAttr<CUDAHostAttr>())
       return CFT_HostDevice;
     return CFT_Device;
+  } else if (D->hasAttr<CUDAHostAttr>()) {
+    return CFT_Host;
+  } else if (D->isImplicit()) {
+    // Some implicit declarations (like intrinsic functions) are not marked.
+    // Set the most lenient target on them for maximal flexibility.
+    return CFT_HostDevice;
   }
 
   return CFT_Host;
 }
 
 bool Sema::CheckCUDATarget(const FunctionDecl *Caller,
                            const FunctionDecl *Callee) {
   return CheckCUDATarget(IdentifyCUDATarget(Caller),

test/SemaCUDA/implicit-intrinsic.cu

+// RUN: %clang_cc1 -std=gnu++11 -triple nvptx64-unknown-unknown -fsyntax-only -verify %s
+
+#include "Inputs/cuda.h"
+
+// expected-no-diagnostics
+__device__ void __threadfence_system() {
+  // This shouldn't produce an error, since __nvvm_membar_sys is inferred to
+  // be __host__ __device__ and thus callable from device code.
+  __nvvm_membar_sys();
+}