mlir::FunctionType func_type = main_func.getFunctionType();

  outputs->clear();
  outputs->reserve(func_type.getNumResults());
  resource_updates->clear();
  resource_updates->reserve(func_type.getNumResults());

  std::vector<xla::Shape> shapes;
  shapes.reserve(func_type.getNumResults());

  llvm::SmallDenseMap<unsigned, unsigned> output_to_input_alias;

  std::vector<int32_t> rhs_transpose;
  std::vector<int32_t> out_transpose;
  lhs_transpose.reserve(dnums.lhs_rhs_out.size() + dnums.lhs_out.size() +
                        dnums.lhs_rhs.size());
  rhs_transpose.reserve(dnums.lhs_rhs_out.size() + dnums.rhs_out.size() +
                        dnums.lhs_rhs.size());
  out_transpose.reserve(dnums.lhs_rhs_out.size() + dnums.lhs_out.size() +
                        dnums.rhs_out.size());

  llvm::SmallVector<Value, 8> new_packed_inputs;
  llvm::SmallVector<llvm::SmallVector<Value, 8>, 8> replicated_inputs;
  replicated_inputs.reserve(replicate.GetNumReplicatedBlockArguments());
  new_packed_inputs.reserve(replicate.GetNumPackedBlockArguments());
  for (const auto& arg : replicate.GetReplicatedBlockArguments()) {
    replicated_inputs.emplace_back();

    Block& block = parallel_execute.GetRegionBlockWithIndex(i);
    auto results = parallel_execute.GetRegionOutputs(i);
    auto& results_to_remove = results_to_remove_by_region[i];
    results_to_remove.reserve(results.size());
    Operation* terminator = block.getTerminator();
    for (auto result : llvm::enumerate(results)) {
      TF::AssignVariableOp assign_var =
          GetSingleUseResourceWrite(parallel_execute, result.value());

	CBZ	R4, nocgo
	MOVD	$_rt0_arm64_netbsd_lib_go(SB), R0
	MOVD	$0, R1
	SUB	$16, RSP		// reserve 16 bytes for sp-8 where fp may be saved.
	BL	(R4)
	ADD	$16, RSP
	B	restore

nocgo:
	MOVD	$0x800000, R0                     // stacksize = 8192KB
	MOVD	$_rt0_arm64_netbsd_lib_go(SB), R1
	MOVD	R0, 8(RSP)
	MOVD	R1, 16(RSP)
	MOVD	$runtime·newosproc0(SB),R4
	BL	(R4)

restore:

  computeInputsOutputs(ops, &inputs, &outputs);

  std::vector<Type> input_types;
  std::vector<Type> output_types;

  input_types.reserve(inputs.size());
  for (Value v : inputs) {
    input_types.push_back(v.getType());
  }
  output_types.reserve(outputs.size());
  for (Value v : outputs) {
    output_types.push_back(v.getType());
  }

  // Create the function.

type ReservePlugin interface {
	Plugin
	// Reserve is called by the scheduling framework when the scheduler cache is
	// updated. If this method returns a failed Status, the scheduler will call
	// the Unreserve method for all enabled ReservePlugins.
	Reserve(ctx context.Context, state *CycleState, p *v1.Pod, nodeName string) *Status
	// Unreserve is called by the scheduling framework when a reserved pod was

                        llvm::SmallVectorImpl<int64_t>* shape) {
  shape->reserve(input_shape.dims());
  for (const auto& d : input_shape) {
    shape->push_back(d.size == kTFDynamicSize ? ShapedType::kDynamic : d.size);
  }
}

Status ConvertToMlirShape(const TensorShapeProto& input_shape,
                          llvm::SmallVectorImpl<int64_t>* shape) {
  shape->reserve(input_shape.dim_size());
  auto& dims = input_shape.dim();

  std::vector<tensorflow::Flag> flag_list;
  tensorflow::AppendMarkForCompilationPassFlags(&flag_list);
  auto usage = tensorflow::Flags::Usage(real_argv[0], flag_list);

  std::vector<char*> args;

  args.reserve(real_argc + 1);
  for (int i = 0; i < real_argc; i++) {
    args.push_back(real_argv[i]);
  }

  struct FreeDeleter {
    void operator()(char* ptr) { free(ptr); }
  };

	CBZ	R4, nocgo
	MOVD	$_rt0_arm64_linux_lib_go(SB), R0
	MOVD	$0, R1
	SUB	$16, RSP		// reserve 16 bytes for sp-8 where fp may be saved.
	BL	(R4)
	ADD	$16, RSP
	B	restore

nocgo:
	MOVD	$0x800000, R0                     // stacksize = 8192KB
	MOVD	$_rt0_arm64_linux_lib_go(SB), R1
	MOVD	R0, 8(RSP)
	MOVD	R1, 16(RSP)
	MOVD	$runtime·newosproc0(SB),R4
	BL	(R4)

restore:
	// Restore callee-save registers.