#include "tensorflow/cc/ops/sparse_ops.h"
#include "tensorflow/cc/ops/state_ops.h"
#include "tensorflow/cc/ops/string_ops.h"
#include "tensorflow/cc/ops/training_ops.h"
#include "tensorflow/cc/ops/user_ops.h"

  toco_flags.set_reduce_type_precision(reduce_type_precision);
  // Read list of user select ops.
  llvm::SmallVector<llvm::StringRef, 2> user_ops;
  (llvm::StringRef(select_user_tf_ops))
      .split(user_ops, ',', /*MaxSplit=*/-1,
             /*KeepEmpty=*/false);
  llvm::for_each(user_ops, [&toco_flags](llvm::StringRef op_name) {
    *(toco_flags.add_select_user_tf_ops()) = op_name.str();
  });

  std::string result;

        "manip_ops",
        "nn_ops",
        "no_op",
        "parsing_ops",
        "random_ops",
        "sparse_ops",
        "state_ops",
        "string_ops",
        "training_ops",
        "user_ops",
    ],
    other_hdrs = [
        "ops/array_ops.h",
        "ops/const_op.h",
        "ops/math_ops.h",
        "ops/dataset_ops.h",
        "ops/experimental_dataset_ops.h",

    // Skip the original dequant op and connect the op before dequantize to the
    // user op.
    user_op->setOperand(user_idx, op_before_dequantize);

    // Wire input/output nodes.
    new_dequantize_op->setOperand(0, user_op->getResult(0));
    new_dequantize_op->getResult(0).setType(user_op->getResult(0).getType());
    user_op->getResult(0).replaceAllUsesExcept(new_dequantize_op->getResult(0),

    Operation* user_op, int argument_index, ElementsAttr value,
    llvm::SmallVector<std::pair<Region*, int>, 4>* work_list,
    llvm::MapVector<Operation*, llvm::SmallVector<unsigned int, 4>>*
        arguments_to_erase) {
  if (auto read_variable_op = dyn_cast<TF::ReadVariableOp>(user_op)) {
    (*arguments_to_erase)[read_variable_op];
    PropagateUsage(read_variable_op, value);
  } else if (auto call = dyn_cast<CallOpInterface>(user_op)) {

    bool all_uses_replaced = true;

    for (mlir::OpOperand &use :
         llvm::make_early_inc_range(op.getResult().getUses())) {
      mlir::Operation *user_op = use.getOwner();
      StringAttr device_attr = user_op->getAttrOfType<StringAttr>(kDeviceAttr);
      if (!device_attr) {
        all_uses_replaced = false;
        continue;
      }

  // CHECK: %[[VAL_0:.*]] = "tf._FusedConv2D"(%arg2, %arg1, %arg0) <{data_format = "NHWC", dilations = [1, 1, 1, 1], epsilon = 0.000000e+00 : f32, explicit_paddings = [], fused_ops = ["BiasAdd"], num_args = 1 : i64, operandSegmentSizes = array<i32: 1, 1, 1, 0>, padding = "SAME", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true}> {TArgs = [f32]} : (tensor<8x32x32x3xf32>, tensor<1x1x3x128xf32>, tensor<128xf32>) -> tensor<*xf32>...

    // together; epsilon (only with FusedBatchNorm).
    std::vector<NamedAttribute> attrs = contraction->getAttrs();
    ArrayAttr fused_ops_attr = ArrayAttr::get(context, fused_ops);
    attrs.push_back(
        NamedAttribute(StringAttr::get(context, "fused_ops"), fused_ops_attr));
    // Epsilon is used only in fusions with the FusedBatchNorm op, so we zero it
    // here.
    Attribute epsilon = rewriter.getF32FloatAttr(0);

  }

  Predicate* new_inner_op = MakeAndOrImpl(factored_ops, is_and);
  std::vector<Predicate*> outer_ops;
  outer_ops.push_back(new_inner_op);
  outer_ops.insert(outer_ops.end(), common_inner_operands.begin(),
                   common_inner_operands.end());
  return MakeAndOrImpl(outer_ops, !is_and);
}

class DeadnessAnalysisImpl : public DeadnessAnalysis {
 public:

  %conv2d = "tf._FusedConv2D"(%input, %filter, %bias, %act, %input_scale, %side_input_scale) {
    data_format = "NHWC", dilations = [1, 1, 1, 1], epsilon = 9.99999974E-5 : f32, explicit_paddings = [], filter_format = "HWIO", fused_ops = ["BiasAdd", "Relu"], leakyrelu_alpha = 2.000000e-01 : f32, num_args = 2 : i64, operandSegmentSizes = array<i32: 1, 1, 2, 2>, padding = "SAME", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true