4 files changed, 19 insertions, 7 deletions
diff --git a/clang/lib/CodeGen/CGCall.cpp b/clang/lib/CodeGen/CGCall.cpp
index 2d8339ef697..064b5fcb102 100644
--- a/clang/lib/CodeGen/CGCall.cpp
+++ b/clang/lib/CodeGen/CGCall.cpp
@@ -255,6 +255,16 @@ CodeGenTypes::arrangeCXXMethodType(const CXXRecordDecl *RD,
       FTP->getCanonicalTypeUnqualified().getAs<FunctionProtoType>(), MD);
 }
 
+/// Set calling convention for CUDA/HIP kernel.
+static void setCUDAKernelCallingConvention(CanQualType &FTy, CodeGenModule &CGM,
+                                           const FunctionDecl *FD) {
+  if (FD->hasAttr<CUDAGlobalAttr>()) {
+    const FunctionType *FT = FTy->getAs<FunctionType>();
+    CGM.getTargetCodeGenInfo().setCUDAKernelCallingConvention(FT);
+    FTy = FT->getCanonicalTypeUnqualified();
+  }
+}
+
 /// Arrange the argument and result information for a declaration or
 /// definition of the given C++ non-static member function.  The
 /// member function must be an ordinary function, i.e. not a
@@ -264,7 +274,9 @@ CodeGenTypes::arrangeCXXMethodDeclaration(const CXXMethodDecl *MD) {
   assert(!isa<CXXConstructorDecl>(MD) && "wrong method for constructors!");
   assert(!isa<CXXDestructorDecl>(MD) && "wrong method for destructors!");
 
-  CanQual<FunctionProtoType> prototype = GetFormalType(MD);
+  CanQualType FT = GetFormalType(MD).getAs<Type>();
+  setCUDAKernelCallingConvention(FT, CGM, MD);
+  auto prototype = FT.getAs<FunctionProtoType>();
 
   if (MD->isInstance()) {
     // The abstract case is perfectly fine.
@@ -424,6 +436,7 @@ CodeGenTypes::arrangeFunctionDeclaration(const FunctionDecl *FD) {
   CanQualType FTy = FD->getType()->getCanonicalTypeUnqualified();
 
   assert(isa<FunctionType>(FTy));
+  setCUDAKernelCallingConvention(FTy, CGM, FD);
 
   // When declaring a function without a prototype, always use a
   // non-variadic type.
diff --git a/clang/lib/CodeGen/CodeGenModule.cpp b/clang/lib/CodeGen/CodeGenModule.cpp
index 814eda4381b..6595694ebd5 100644
--- a/clang/lib/CodeGen/CodeGenModule.cpp
+++ b/clang/lib/CodeGen/CodeGenModule.cpp
@@ -3671,8 +3671,6 @@ void CodeGenModule::EmitGlobalFunctionDefinition(GlobalDecl GD,
 
   MaybeHandleStaticInExternC(D, Fn);
 
-  if (D->hasAttr<CUDAGlobalAttr>())
-    getTargetCodeGenInfo().setCUDAKernelCallingConvention(Fn);
 
   maybeSetTrivialComdat(*D, *Fn);
 
diff --git a/clang/lib/CodeGen/TargetInfo.cpp b/clang/lib/CodeGen/TargetInfo.cpp
index 3ec00553607..b29bcce237c 100644
--- a/clang/lib/CodeGen/TargetInfo.cpp
+++ b/clang/lib/CodeGen/TargetInfo.cpp
@@ -7646,7 +7646,7 @@ public:
                             llvm::Function *BlockInvokeFunc,
                             llvm::Value *BlockLiteral) const override;
   bool shouldEmitStaticExternCAliases() const override;
-  void setCUDAKernelCallingConvention(llvm::Function *F) const override;
+  void setCUDAKernelCallingConvention(const FunctionType *&FT) const override;
 };
 }
 
@@ -7783,8 +7783,9 @@ bool AMDGPUTargetCodeGenInfo::shouldEmitStaticExternCAliases() const {
 }
 
 void AMDGPUTargetCodeGenInfo::setCUDAKernelCallingConvention(
-    llvm::Function *F) const {
-  F->setCallingConv(llvm::CallingConv::AMDGPU_KERNEL);
+    const FunctionType *&FT) const {
+  FT = getABIInfo().getContext().adjustFunctionType(
+      FT, FT->getExtInfo().withCallingConv(CC_OpenCLKernel));
 }
 
 //===----------------------------------------------------------------------===//
diff --git a/clang/lib/CodeGen/TargetInfo.h b/clang/lib/CodeGen/TargetInfo.h
index 5c19c714124..b530260ea48 100644
--- a/clang/lib/CodeGen/TargetInfo.h
+++ b/clang/lib/CodeGen/TargetInfo.h
@@ -302,7 +302,7 @@ public:
   /// as 'used', and having internal linkage.
   virtual bool shouldEmitStaticExternCAliases() const { return true; }
 
-  virtual void setCUDAKernelCallingConvention(llvm::Function *F) const {}
+  virtual void setCUDAKernelCallingConvention(const FunctionType *&FT) const {}
 };
 
 } // namespace CodeGen