[CUDA] Make sure device-side __global__ functions are always visible.

__global__ functions are a special case in CUDA.

Even when the symbol would normally not be externally
visible according to C++ rules, they still must be visible
in CUDA GPU object so host-side stub can launch them.

Differential Revision: http://reviews.llvm.org/D19748

llvm-svn: 268299

diff --git a/clang/lib/AST/ASTContext.cpp b/clang/lib/AST/ASTContext.cpp
index 4f7f8ee..7a09b0b 100644 (file)
--- a/clang/lib/AST/ASTContext.cpp
+++ b/clang/lib/AST/ASTContext.cpp
@@ -8418,22 +8418,29 @@ static GVALinkage basicGVALinkageForFunction(const ASTContext &Context,
   return GVA_DiscardableODR;
 }
 
-static GVALinkage adjustGVALinkageForAttributes(GVALinkage L, const Decl *D) {
+static GVALinkage adjustGVALinkageForAttributes(const ASTContext &Context,
+                                                GVALinkage L, const Decl *D) {
   // See http://msdn.microsoft.com/en-us/library/xa0d9ste.aspx
   // dllexport/dllimport on inline functions.
   if (D->hasAttr<DLLImportAttr>()) {
     if (L == GVA_DiscardableODR || L == GVA_StrongODR)
       return GVA_AvailableExternally;
-  } else if (D->hasAttr<DLLExportAttr>() || D->hasAttr<CUDAGlobalAttr>()) {
+  } else if (D->hasAttr<DLLExportAttr>()) {
     if (L == GVA_DiscardableODR)
       return GVA_StrongODR;
+  } else if (Context.getLangOpts().CUDA && Context.getLangOpts().CUDAIsDevice &&
+             D->hasAttr<CUDAGlobalAttr>()) {
+    // Device-side functions with __global__ attribute must always be
+    // visible externally so they can be launched from host.
+    if (L == GVA_DiscardableODR || L == GVA_Internal)
+      return GVA_StrongODR;
   }
   return L;
 }
 
 GVALinkage ASTContext::GetGVALinkageForFunction(const FunctionDecl *FD) const {
-  return adjustGVALinkageForAttributes(basicGVALinkageForFunction(*this, FD),
-                                       FD);
+  return adjustGVALinkageForAttributes(
+      *this, basicGVALinkageForFunction(*this, FD), FD);
 }
 
 static GVALinkage basicGVALinkageForVariable(const ASTContext &Context,
@@ -8490,8 +8497,8 @@ static GVALinkage basicGVALinkageForVariable(const ASTContext &Context,
 }
 
 GVALinkage ASTContext::GetGVALinkageForVariable(const VarDecl *VD) {
-  return adjustGVALinkageForAttributes(basicGVALinkageForVariable(*this, VD),
-                                       VD);
+  return adjustGVALinkageForAttributes(
+      *this, basicGVALinkageForVariable(*this, VD), VD);
 }
 
 bool ASTContext::DeclMustBeEmitted(const Decl *D) {

diff --git a/clang/test/CodeGenCUDA/ptx-kernels.cu b/clang/test/CodeGenCUDA/ptx-kernels.cu
index 6280e60..034cef6 100644 (file)
--- a/clang/test/CodeGenCUDA/ptx-kernels.cu
+++ b/clang/test/CodeGenCUDA/ptx-kernels.cu
@@ -19,8 +19,17 @@ __global__ void global_function() {
 
 // Make sure host-instantiated kernels are preserved on device side.
 template <typename T> __global__ void templated_kernel(T param) {}
-// CHECK-LABEL: define weak_odr void @_Z16templated_kernelIiEvT_
-void host_function() { templated_kernel<<<0,0>>>(0); }
+// CHECK-DAG: define weak_odr void @_Z16templated_kernelIiEvT_(
+
+namespace {
+__global__ void anonymous_ns_kernel() {}
+// CHECK-DAG: define weak_odr void @_ZN12_GLOBAL__N_119anonymous_ns_kernelEv(
+}
+
+void host_function() {
+  templated_kernel<<<0, 0>>>(0);
+  anonymous_ns_kernel<<<0,0>>>();
+}
 
 // CHECK: !{{[0-9]+}} = !{void ()* @global_function, !"kernel", i32 1}
 // CHECK: !{{[0-9]+}} = !{void (i32)* @_Z16templated_kernelIiEvT_, !"kernel", i32 1}