// Add new control input if its defining op is not already a defining
      // op for some other operand. Update defining_ops.
      if (defining_ops.insert(new_control_input.getDefiningOp()).second) {
        operands.push_back(new_control_input);
      }
    }
    state.addOperands(operands);
    Operation* new_op = builder.create(state);
    item.replaceAllUsesWith(new_op);
    new_op->setAttrs(item.getAttrDictionary());

  }
  auto iter = backprops_.find(src);
  if (iter != backprops_.end()) {
    auto* grads = &iter->second;
    grads->push_back(dst_grad);
    if (--pending_[src.node()->id()] == 0) {
      ready_.push_back(src.node());
    }
  }
  return absl::OkStatus();
}

std::vector<bool> SymbolicGradientBuilder::GetReachableNodes() {

    } else {
      reduced_axes.push_back(*axis);
    }
  }
  // Get the corresponding dimensions for each reduced axis.
  std::vector<Output> reduced_dims_inputs;
  reduced_dims_inputs.reserve(reduced_axes.size());
  for (const int i : reduced_axes) {
    if (i < 0) {
      reduced_dims_inputs.push_back(
          Gather(scope, input_shape, Add(scope, Size(scope, input_shape), i)));

  // contain zero scale for zero values.
  llvm::SmallVector<double> scales;
  for (float scale : quant_params.scale) {
    if (scale == 0) {
      scales.push_back(std::numeric_limits<float>::min());
      continue;
    }
    scales.push_back(scale);
  }

  // Scale size can't be zero as it is checked before.
  if (quant_params.scale.size() != 1) {
    return mlir::quant::UniformQuantizedPerAxisType::get(

    Type result_type;

    // Include info about the activation function if applicable.
    if (fuse_activation) {
      locations.push_back(activation->getLoc());
      fused_ops.push_back(
          StringAttr::get(context, activation->getName().stripDialect()));
      result_type = activation->getResultTypes().front();
    } else {

        if (!mlir::isa<BlockArgument>(value))
          island_control_operands.push_back(value);
      } else {
        inner_op_operands.push_back(value);
      }
    }

    auto island = rewriter.create<tf_executor::IslandOp>(
        loc, new_types, island_control_operands);
    island.getBody().push_back(new mlir::Block);

    rewriter.setInsertionPointToEnd(&island.getBody().front());

  for (auto& device_spec : specified_device_specs) {
    auto device = GetCanonicalHardwareName(device_spec);

    if (device == "CPU") cpu_include = true;
    device_specs->push_back(device);
  }
  if (!cpu_include) {
    device_specs->push_back("CPU");
  }

  // Make sure all the devices are registered.
  for (const std::string& device : *device_specs) {
    if (GetTargetHardware(device) == nullptr) {

        // GradientTape which uses ones.
        Gradient* zero = input_tensors[target_index].ZerosLike();
        new_zeros.push_back(zero);
        in_grads.push_back(zero);
      } else {
        in_grads.push_back(nullptr);
      }
    } else {
      in_grads.push_back(current_grad->second);
    }
  }

  // Avoid infinite recursion. Whichever forward function we run, it'll end up

      for (auto list_input : list_op.getOperands()) {
        auto cast_op = dyn_cast_or_null<CastOp>(list_input.getDefiningOp());
        if (!cast_op) return failure();
        list_inputs.push_back(cast_op.getArg());
        list_input_types.push_back(cast_op.getInputElementType());
      }
      CastValuesToSameType(rewriter, call_op.getLoc(), list_input_types,
                           list_inputs);

      auto dequantize_op = llvm::cast<DequantizeOp>(returned_op);
      Value dequantized_result = dequantize_op.getInput();
      output_types.push_back(dequantized_result.getType());
      terminator->setOperand(i, dequantized_result);
      returned_op->erase();
    } else {
      output_types.push_back(returned_value.getType());
    }
  }
  auto new_func_type = builder.getFunctionType(input_types, output_types);