[SYCL][CodeGen] Directly disable OCL kernel stub for SYCL (NFC) (#19213)

The way 1ce3803 ([SYCL] Disable kernel
stub for now (#28629)) and 3b170e3
([SYCL][AMD] Fix kernel stub calling convention for AMD) disabled the
OpenCL kernel stub for SYCL is weird, returning from
`TargetCodeGenInfo::setOCLKernelStubCallingConvention()` in SYCL mode.

Basically we were actually emitting code for the stub would but we
customized it to not change its calling convention.

Instead, add `getLangOpts().OpenCL` checks to
`CodeGenFunction::GenerateCode`
(to emit the kernel contents directly into the kernel) and to
`CodeGenModule::EmitGlobal` (to disable the deferred emission of the
stub).
With this change, we should never come across
`KernelReferenceKind::Stub`
in SYCL mode.

Author: Maetveis

clang/lib/CodeGen/CodeGenFunction.cpp

@@ -1932,7 +1932,7 @@ void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
     // Implicit copy-assignment gets the same special treatment as implicit
     // copy-constructors.
     emitImplicitAssignmentOperatorBody(Args);
-  } else if (FD->hasAttr<OpenCLKernelAttr>() &&
+  } else if (getLangOpts().OpenCL && FD->hasAttr<OpenCLKernelAttr>() &&
              GD.getKernelReferenceKind() == KernelReferenceKind::Kernel) {
     CallArgList CallArgs;
     for (unsigned i = 0; i < Args.size(); ++i) {

clang/lib/CodeGen/CodeGenModule.cpp

@@ -4341,7 +4341,8 @@ void CodeGenModule::EmitGlobal(GlobalDecl GD) {
 
   // Ignore declarations, they will be emitted on their first use.
   if (const auto *FD = dyn_cast<FunctionDecl>(Global)) {
-    if (FD->hasAttr<OpenCLKernelAttr>() && FD->doesThisDeclarationHaveABody())
+    if (LangOpts.OpenCL && FD->hasAttr<OpenCLKernelAttr>() &&
+        FD->doesThisDeclarationHaveABody())
       addDeferredDeclToEmit(GlobalDecl(FD, KernelReferenceKind::Stub));
 
     // Update deferred annotations with the latest declaration if the function

clang/lib/CodeGen/TargetInfo.cpp

@@ -119,10 +119,6 @@ unsigned TargetCodeGenInfo::getOpenCLKernelCallingConv() const {
 
 void TargetCodeGenInfo::setOCLKernelStubCallingConvention(
     const FunctionType *&FT) const {
-
-  if (getABIInfo().getContext().getLangOpts().SYCLIsNativeCPU)
-     return;
-
   FT = getABIInfo().getContext().adjustFunctionType(
       FT, FT->getExtInfo().withCallingConv(CC_C));
 }

clang/lib/CodeGen/Targets/AMDGPU.cpp

@@ -305,8 +305,6 @@ 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;
@@ -437,14 +435,6 @@ unsigned AMDGPUTargetCodeGenInfo::getOpenCLKernelCallingConv() const {
   return llvm::CallingConv::AMDGPU_KERNEL;
 }
 
-void AMDGPUTargetCodeGenInfo::setOCLKernelStubCallingConvention(
-    const FunctionType *&FT) const {
-  bool IsSYCL = getABIInfo().getContext().getLangOpts().isSYCL();
-  FT = getABIInfo().getContext().adjustFunctionType(
-      FT,
-      FT->getExtInfo().withCallingConv(!IsSYCL ? CC_C : 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

clang/lib/CodeGen/Targets/SPIR.cpp

@@ -394,9 +394,6 @@ void SPIRVTargetCodeGenInfo::setCUDAKernelCallingConvention(
 
 void CommonSPIRTargetCodeGenInfo::setOCLKernelStubCallingConvention(
     const FunctionType *&FT) const {
-  // Disable kernel stub for sycl
-  if (getABIInfo().getContext().getLangOpts().isSYCL())
-     return;
   FT = getABIInfo().getContext().adjustFunctionType(
       FT, FT->getExtInfo().withCallingConv(CC_SpirFunction));
 }