avoid atomic contention on AsyncValue's refcount.
  }];

  let arguments = (ins
    TFTensorType:$operand
  );

  let results = (outs
    Variadic<TFTensorType>:$results
  );

  let assemblyFormat = "$operand attr-dict `:` `(` type($operand) `)` `->` `(` type($results) `)`";
}

def CreateOp: FallbackAsync_Op<"createop", [CoreRT_TypedAttributeTrait]> {

	cgofunc:   "cgo function",
}

// An operand represents an intermediate value during type checking.
// Operands have an (addressing) mode, the expression evaluating to
// the operand, the operand's type, a value for constants, and an id
// for built-in functions.
// The zero value of operand is a ready to use invalid operand.
type operand struct {
	mode operandMode
	expr syntax.Expr
	typ  Type

  return absl::OkStatus();
}

Status MlirAbstractOp::Create(ArrayRef<Value> operands,
                              OperationState** state) {
  state_->operands = llvm::to_vector<4>(operands);
  Builder builder(context_);

  if (current_ods_input_ != op_def_->input_arg_size())
    return InvalidArgument(absl::StrCat("Mismatch in operands number: got ",
                                        current_ods_input_, " expected ",

    return rewriter.notifyMatchFailure(
        concat, "Concatenate op should have at least two operands");

  auto first = concat.getVal()[0].getDefiningOp<mhlo::SliceOp>();
  auto second = concat.getVal()[1].getDefiningOp<mhlo::SliceOp>();
  if (!first || !second)
    return rewriter.notifyMatchFailure(concat, "operands are not slice ops");
  if (first.getOperand() != second.getOperand())

  metadata.ParseFromString(std::string(metadata_str.getValue()));
  int64_t num_replicas = replicate.getN();
  // Find the formattable operands of `execute`, which must be mirrored
  // variables (arguments of `replicate`), and must be pass-throughs from while
  // operands.
  for (const auto& mirrored_index : mirrored_variable_indices_attr) {
    int64_t replicate_arg = mlir::cast<IntegerAttr>(mirrored_index).getInt();

    // NOLINTNEXTLINE
    llvm::SmallVectorImpl<mlir::NamedAttribute> &attributes);

// While the last several tensors could be optional tensors for an tfl op, the
// number of input operands could vary. This function gets the min/max number of
// operands from tflite op name.
llvm::MinMax OperandNumbersMinMax(llvm::StringRef op_name);

// Populates the `custom_code` and `custom_options` to attributes.

    std::unique_ptr<OpQuantSpec> spec = GetTFOpQuantSpec(op);
    const absl::flat_hash_set<int> operands = spec->quantizable_operands;

    if (operands.size() != 1) return failure();
    int weight_operand_idx = *operands.begin();

    Operation* weight_op = op.getOperand(weight_operand_idx).getDefiningOp();
    DenseFPElementsAttr attr;

  SmallVector<Value, 4> else_returns;
  for (int i = 1; i < state_size + 1; i++) {
    else_returns.push_back(body_args[i]);
  }
  builder.create<TF::YieldOp>(loc,
                              /*operands=*/else_returns);

  // Then branch gets the data and calls the reduce_function.
  auto& then_branch = dataset_if.getThenBranch();
  then_branch.push_back(new Block);

    // The map contains all the quantization parameters which are required to
    // satisfy the same operands and results constraint. The keys of this map
    // are the values from `operand_states_` and `result_state_`.
    std::unordered_map<int, RequantizeState> rescale_states_;

    // Maps of indexes to the propagation state vector from the ops operands,
    // results and arguments.
    llvm::DenseMap<OpValue, int> operand_states_;

  for (Value operand : op.getOperands()) {
    if (mlir::isa<tf_executor::TokenType>(operand.getType())) continue;
    if (mlir::isa<tf_executor::ControlType>(operand.getType())) break;

    if (is_switch &&
        llvm::isa_and_nonnull<tf_executor::LoopCondOp>(operand.getDefiningOp()))
      continue;

    auto it = value_to_device.find(operand);