[llvm] [NVPTX] Improve kernel byval parameter lowering (PR #136008)

[Alexander Kornienko via llvm-commits]

alexfh wrote:

@AlexMaclean this seems to be causing clang crashes when compiling https://github.com/NVIDIA/cccl/blob/9019a6a28f2b6947d76c3b7502764a0a240f5111/thrust/testing/complex_transform.cu:
```
assertion failed at llvm/include/llvm/Support/Casting.h:662 in decltype(auto) llvm::dyn_cast(From *) [To = llvm::GEPOperator, From = const llvm::Value]:
detail::isPresent(Val) && "dyn_cast on a non-existent value"
    @     0x559920198f84  __assert_fail
    @     0x55991f87517d  llvm::getUnderlyingObject()
    @     0x55991f875900  llvm::getUnderlyingObjects()
    @     0x55991d635d76  canLowerToLDG()
    @     0x55991d62d28c  llvm::NVPTXDAGToDAGISel::tryLoad()
    @     0x55991d62cfd3  llvm::NVPTXDAGToDAGISel::Select()
    @     0x55991e6ba49a  llvm::SelectionDAGISel::DoInstructionSelection()
    @     0x55991e6b9379  llvm::SelectionDAGISel::CodeGenAndEmitDAG()
    @     0x55991e6b6e92  llvm::SelectionDAGISel::SelectAllBasicBlocks()
    @     0x55991e6b3fee  llvm::SelectionDAGISel::runOnMachineFunction()
    @     0x55991d62cea2  llvm::NVPTXDAGToDAGISel::runOnMachineFunction()
    @     0x55991e6b12c9  llvm::SelectionDAGISelLegacy::runOnMachineFunction()
    @     0x55991e2c56ba  llvm::MachineFunctionPass::runOnFunction()
    @     0x55991fbf9d4e  llvm::FPPassManager::runOnFunction()
    @     0x55991fc01e3d  llvm::FPPassManager::runOnModule()
    @     0x55991fbfa5ec  llvm::legacy::PassManagerImpl::run()
    @     0x55991a63fc4f  clang::emitBackendOutput()
    @     0x55991a31be4a  clang::BackendConsumer::HandleTranslationUnit()
    @     0x55991b208da8  clang::ParseAST()
    @     0x55991af53f2a  clang::FrontendAction::Execute()
    @     0x55991aecdded  clang::CompilerInstance::ExecuteAction()
    @     0x55991a31ac6b  clang::ExecuteCompilerInvocation()
    @     0x55991a30ed60  cc1_main()
```



https://github.com/llvm/llvm-project/pull/136008