if (tpu_replicated_input.getIsMirroredVariable()) {
      mirrored_variable_indices.push_back(pos_and_input.index());
    }
  }

  // Create replicate op.
  auto result_types = GetClusterResultTypes(cluster, partitioned_outputs);
  auto replicate_op = builder.create<mlir::tf_device::ReplicateOp>(
      cluster.getLoc(), num_replicas,
      llvm::SmallDenseMap<llvm::StringRef,

      // Create new ConstantOp-Dequantize-Operation sequences. At this moment,
      // old ConstantOp is guaranteed to have one F32->F16 cast regardless of
      // its number of users.
      rewriter.setInsertionPointAfter(op);
      auto new_const = rewriter.create<arith::ConstantOp>(
          op->getLoc(), new_result_type, new_value_attr);
      auto dq = rewriter.create<DQ>(op->getLoc(), old_result_type, new_const);

    func::FuncOp parent_func, ModuleOp module, const std::string& name) {
  // Moves all of the Operations in 'ops' into a newly created func.FuncOp
  // function named 'name' and replaces the original ops with a call to the
  // newly created function using a tf.StatefulPartitionedCall. Here,
  // 'parent_func' is the function that holds the original set of ops.

  }

  SmallVector<Value> return_values;
  for (Value result : results) {
    return_values.push_back(mapping.lookupOrNull(result));
  }
  builder.create<func::ReturnOp>(location, return_values);

  // Create a function call to the newly created function.
  StringAttr new_func_name =
      InsertToSymbolTable(*module, *wrap_func, func_name);
  builder.setInsertionPointAfter(result_op);
  ValueRange new_results =

  SliceInputs slice_inputs_int64 =
      MakeSliceIndexAndSizeInt64(host_scope, slice_inputs);

  // Create a list of all control dependencies to be copied when possibly
  // replacing nodes related to slice_size.
  Node* old_size;
  std::vector<const Edge*> old_size_ctrl_deps;
  TF_RETURN_IF_ERROR(slice->input_node(2, &old_size));

  Value tensor_i32 = builder.create<TF::CastOp>(
      loc, tensor_type.clone(builder.getIntegerType(32)), tensor);
  auto reduced =
      builder.create<TF::SumOp>(loc, tensor_i32, reduction_indices_value,
                                /*keep_dims=*/builder.getBoolAttr(true));
  auto mul_op = builder.create<TF::MulOp>(loc, zp, reduced);

      return failure();
    }

    // Compute multiplier = scale / sqrt(variance + epsilon)
    Value multiplier = rewriter.create<mhlo::AddOp>(
        bn_op.getLoc(), bn_op.getVariance(), epsilon);
    multiplier = rewriter.create<mhlo::RsqrtOp>(bn_op.getLoc(), multiplier);
    multiplier = rewriter.create<mhlo::MulOp>(bn_op.getLoc(), multiplier,
                                              bn_op.getScale());

// be created).
static void DumpModule(mlir::ModuleOp module, std::string file_prefix) {
  std::string prefix = GetDumpDirFromEnvVar();
  if (prefix.empty()) return;

  auto* env = tensorflow::Env::Default();
  auto status = env->RecursivelyCreateDir(prefix);
  if (!status.ok()) {
    LOG(WARNING) << "cannot create directory '" << prefix
                 << "': " << status.message();

        builder, block.getParent()->getLoc(), execute, device);

    builder->create<tf_device::ReturnOp>(block.getParent()->getLoc(),
                                               block_launch_op.getResults());
  }

  return success();
}

// Creates a `tf.TPUCompileSucceededAssert` operation that parses compilation
// status of `compile_op` to check whether compilation is successful.

  TF_Output output{dequeue_op, 0};
  TF_Tensor* ret;
  TF_SessionRun(session, /*run_options*/ nullptr,
                // input related parameters
                /*inputs*/ nullptr, /*input_values*/ nullptr, /*ninputs*/ 0,
                // output related parameters
                /*outputs*/ &output, /*output_values*/ &ret,
                /*noutputs*/ 1,
                /*targets*/ nullptr, /*ntargets*/ 0,