using DQ = quantfork::DequantizeCastOp;

template <typename LstmOp>
LogicalResult GetLstmProperty(LstmOp op,
                              operator_property::OpVariant* lstm_variant,
                              operator_property::OperatorProperty* op_property,
                              int activation_number_of_bits = 8) {
  if (llvm::isa<TFL::LSTMOp>(op.getOperation())) {
    lstm_variant->op_code = tflite::BuiltinOperator_LSTM;

  EXPECT_EQ(it->getName().getStringRef(),
            mlir::func::ReturnOp::getOperationName());
  it++;  // tensor_cast
  it++;  // lstm
  EXPECT_EQ(it->getName().getStringRef(),
            mlir::TFL::LSTMOp::getOperationName());
  EXPECT_EQ(it->getNumOperands(), 24);
  EXPECT_EQ(it->getNumResults(), 1);
  // cifg = false, so input2input is not None.
  EXPECT_FALSE(mlir::isa<NoneType>(it->getOperand(1).getType()));

  // TODO(b/176258587): Move to OpTrait if this should be generalized.
  // Add special handling for LSTM.
  OUT(2) << "if (auto lstm_op = llvm::dyn_cast<TFL::LSTMOp>(op)) {\n";
  OUT(4) << "return GetLstmOpQuantSpec<TFL::LSTMOp>(lstm_op);\n";
  OUT(2) << "} else if (auto lstm_op = "
            "llvm::dyn_cast<TFL::UnidirectionalSequenceLSTMOp>(op)) {\n";
  OUT(4) << "return "

      if (act_type == InputDataType::INT16) {
        if (is_toco) {
          // Conditions when using TOCO.
          if (absl::StrContains(op_name, "LSTMOp")) continue;
        } else {
          // Conditions when using MLIR.
          if (!(absl::StrContains(op_name, "LSTMOp") ||
                absl::StrContains(op_name, "SoftmaxOp") ||
                absl::StrContains(op_name, "LogisticOp") ||

  // consistent. Otherwise some FileCheck tests would fail.
  RewritePatternSet patterns_1(&getContext());
  if (quant_specs_.post_training_quantization) {
    patterns_1.add<PrepareLstmOutputScale<LSTMOp>>(ctx);
    patterns_1.add<PrepareLstmOutputScale<UnidirectionalSequenceLSTMOp>>(ctx);
  }
  if (is_qdq_conversion_ ||
      quant_specs_.qdq_conversion_mode != quant::QDQConversionMode::kQDQNone) {

  std::ostringstream oss;
  oss << "Create" << op_name.drop_front(4).str() << "erator";
  return oss.str();
}

static inline bool IsLstmOp(const StringRef op_name) {
  return op_name.take_back(6) == "LSTMOp";
}

static int HasOptions(const Record &def) {
  if (def.getValueAsBit("hasOptions")) {
    return 1;
  }
  if (def.getValueAsBit("hasOptions2")) {
    return 2;
  }
  return 0;
}

  // layer norm coefficients
  Value input_layer_norm_coefficients_;
  Value forget_layer_norm_coefficients_;
  Value cell_layer_norm_coefficients_;
  Value output_layer_norm_coefficients_;

  mlir::TFL::LSTMOp lstm_;

  Value none_;
  SmallVector<int64_t, 1> bias_slice_shape_;
  SmallVector<int64_t, 1> bias_size_values_;
  SmallVector<int64_t, 2> weight_slice_shape_;

//===----------------------------------------------------------------------===//
// LSTMOp
//===----------------------------------------------------------------------===//

mlir::LogicalResult LSTMOp::verify() {
  LSTMOp op = *this;
  auto operands = op.GetStatefulOperands();
  if (operands.size() != 2 || operands[0] != 18 || operands[1] != 19) {
    return op.emitOpError("LSTMOp expected to have two stateful operands");
  }

  auto func = getOperation();
  auto* ctx = func.getContext();
  TFL::populateWithGenerated(patterns);
  patterns.add<quant::FoldTrivalRequantizeOp<QuantizeOp>>(ctx);
  patterns.add<PruneUnusedOpsWithSideEffect<TFL::LSTMOp>>(ctx);
  patterns.add<PruneUnusedOpsWithSideEffect<TFL::UnidirectionalSequenceLSTMOp>>(
      ctx);
  patterns.add<PruneUnusedOpsWithSideEffect<TFL::SVDFOp>>(ctx);

  auto result_type = mlir::RankedTensorType::get(
      output_shape,
      mlir::cast<RankedTensorType>(input_.getType()).getElementType());
  lstm_ = builder_.create<mlir::TFL::LSTMOp>(
      fused_func_op_.getLoc(), result_type, input_, input2input_, input2forget_,
      input2cell_, input2output_, rec2input_, rec2forget_, rec2cell_,
      rec2output_, /*cell_to_input_weights*/ none_,