[SYCL][AMD] Fix kernel stub calling convention for AMD

We need to ensure an amdgpu_kernel calling convention on the entry
point, otherwise the kernel is not properly registered and will be
removed from the module.

Author: frasercrmck

clang/include/clang/AST/GlobalDecl.h

@@ -164,7 +164,8 @@ class GlobalDecl {
   }
 
   static KernelReferenceKind getDefaultKernelReference(const FunctionDecl *D) {
-    return (D->hasAttr<OpenCLKernelAttr>() || D->getLangOpts().CUDAIsDevice)
+    return (D->hasAttr<OpenCLKernelAttr>() || D->getLangOpts().CUDAIsDevice ||
+            D->hasAttr<SYCLKernelAttr>())
                ? KernelReferenceKind::Kernel
                : KernelReferenceKind::Stub;
   }

clang/lib/AST/Decl.cpp

@@ -5515,7 +5515,8 @@ FunctionDecl *FunctionDecl::CreateDeserialized(ASTContext &C, GlobalDeclID ID) {
 }
 
 bool FunctionDecl::isReferenceableKernel() const {
-  return hasAttr<CUDAGlobalAttr>() || hasAttr<OpenCLKernelAttr>();
+  return hasAttr<CUDAGlobalAttr>() || hasAttr<OpenCLKernelAttr>() ||
+         hasAttr<SYCLKernelAttr>();
 }
 
 BlockDecl *BlockDecl::Create(ASTContext &C, DeclContext *DC, SourceLocation L) {

clang/lib/CodeGen/Targets/AMDGPU.cpp

@@ -304,6 +304,8 @@ class AMDGPUTargetCodeGenInfo : public TargetCodeGenInfo {
   void setTargetAttributes(const Decl *D, llvm::GlobalValue *GV,
                            CodeGen::CodeGenModule &M) const override;
   unsigned getOpenCLKernelCallingConv() const override;
+  void
+  setOCLKernelStubCallingConvention(const FunctionType *&FT) const override;
 
   llvm::Constant *getNullPointer(const CodeGen::CodeGenModule &CGM,
       llvm::PointerType *T, QualType QT) const override;
@@ -434,6 +436,12 @@ unsigned AMDGPUTargetCodeGenInfo::getOpenCLKernelCallingConv() const {
   return llvm::CallingConv::AMDGPU_KERNEL;
 }
 
+void AMDGPUTargetCodeGenInfo::setOCLKernelStubCallingConvention(
+    const FunctionType *&FT) const {
+  FT = getABIInfo().getContext().adjustFunctionType(
+      FT, FT->getExtInfo().withCallingConv(CC_AMDGPUKernelCall));
+}
+
 // Currently LLVM assumes null pointers always have value 0,
 // which results in incorrectly transformed IR. Therefore, instead of
 // emitting null pointers in private and local address spaces, a null