[clang] [CUDA] Add device-side kernel launch support (PR #165519)

[via cfe-commits]

================
@@ -22,6 +22,116 @@ using namespace CodeGen;
 
 CGCUDARuntime::~CGCUDARuntime() {}
 
+static llvm::Value *emitGetParamBuf(CodeGenFunction &CGF,
+                                    const CUDAKernelCallExpr *E) {
+  auto *GetParamBuf = CGF.getContext().getcudaGetParameterBufferDecl();
+  const FunctionProtoType *GetParamBufProto =
+      GetParamBuf->getType()->getAs<FunctionProtoType>();
+
+  DeclRefExpr *DRE = DeclRefExpr::Create(
+      CGF.getContext(), {}, {}, GetParamBuf,
+      /*RefersToEnclosingVariableOrCapture=*/false, GetParamBuf->getNameInfo(),
+      GetParamBuf->getType(), VK_PRValue);
+  auto *ImpCast = ImplicitCastExpr::Create(
+      CGF.getContext(), CGF.getContext().getPointerType(GetParamBuf->getType()),
+      CK_FunctionToPointerDecay, DRE, nullptr, VK_PRValue, FPOptionsOverride());
+
+  CGCallee Callee = CGF.EmitCallee(ImpCast);
+  CallArgList Args;
+  // Use 64B alignment.
+  Args.add(RValue::get(CGF.CGM.getSize(CharUnits::fromQuantity(64))),
+           CGF.getContext().getSizeType());
+  // Calculate parameter sizes.
+  const PointerType *PT = E->getCallee()->getType()->getAs<PointerType>();
+  const FunctionProtoType *FTP =
+      PT->getPointeeType()->getAs<FunctionProtoType>();
+  CharUnits Offset = CharUnits::Zero();
+  for (auto ArgTy : FTP->getParamTypes()) {
+    auto TInfo = CGF.CGM.getContext().getTypeInfoInChars(ArgTy);
+    Offset = Offset.alignTo(TInfo.Align);
+    Offset += TInfo.Width;
+  }
+  Args.add(RValue::get(CGF.CGM.getSize(Offset)),
----------------
darkbuck wrote:

Offset is first aligned to each param's alignment. Then, the offset is increased by that type's size (width). That was the original intention: to put the offset modification into 2 lines (for a different purpose).

https://github.com/llvm/llvm-project/pull/165519