const std::optional<DictionaryAttr>& devices) {
  auto result = region.walk([&](Operation* op) -> WalkResult {
    if (RequiresReplicaIDAttribute(op)) {
      op->setAttr(kReplicaIdAttr, builder.getI64IntegerAttr(replica_id));
      return WalkResult::advance();
    }

    if (auto placeholder = dyn_cast<TF::_TPUDeviceOrdinalPlaceholderOp>(op)) {
      int64_t device_ordinal;

  builder.insert(metadata_op.clone());
  for (Operation& op : func.getOps()) {
    if (!IsTPUOp(&op)) continue;
    op.setAttr(TF::kReplicationInfoAttr, new_group);
  }
  TF::TPUCompilationResultOp new_result = compilation_op.clone();
  new_result->setAttr(kTpuCompilationStatus, new_group);
  builder.insert(new_result);
}

template <typename OpType>
LogicalResult FindAndExcludeOp(func::FuncOp func,

      op_to_update->setAttr(kDeviceAttr, new_device_attr);
      PopulateDeviceForOpResults(*op_to_update, new_device_attr.getValue(),
                                 value_to_device);

      if (auto sink =
              llvm::dyn_cast<tf_executor::NextIterationSinkOp>(op_to_update)) {
        auto source = sink.GetSource();
        source->setAttr(kDeviceAttr, new_device_attr);

      for (const auto& attr_name : attr_names) {
        if (!global_iter_id_op->hasAttr(attr_name)) continue;
        offset->setAttr(attr_name, global_iter_id_op->getAttr(attr_name));
        new_global_iter_id->setAttr(attr_name,
                                    global_iter_id_op->getAttr(attr_name));
      }
      // Make the next function to get inlined use a different offset.
      ++offset_value;
    }

    } else {
      auto handle = resource_tensor.scalar<tensorflow::ResourceHandle>()();
      is_variable_initialized =
          IsVariableInitialized(var_op, handle.device(), mgr, session);
    }
    var_op->setAttr("_is_initialized",
                    BoolAttr::get(context, is_variable_initialized));
  }
  return success();
}

LogicalResult MarkInitializedVariablesInFunction(ModuleOp module,

      auto float_attr =
          mlir::FloatAttr::get(float_type, static_cast<float>(cost));
      device_costs.push_back({device_identifier, float_attr});
    }

    op->setAttr("per_device_costs",
                mlir::DictionaryAttr::get(module.getContext(), device_costs));
  });
}

}  // namespace

//////////// Importer ////////////

namespace tensorflow {

void AddControlInput(TF_Graph* graph, TF_Operation* op, TF_Operation* input);

// Changes an attr value in the node_def Protocol Buffer and sets a status upon
// completion.
void SetAttr(TF_Graph* graph, TF_Operation* op, const char* attr_name,
             TF_Buffer* attr_value_proto, TF_Status* status);

// Clears the attr in the node_def Protocol Buffer and sets a status upon
// completion.

          // Set interface_name & target to the call_op as well.
          new_call->setAttr(kInterfaceNameAttr,
                            builder->getStringAttr(interface_name));
          new_call->setAttr(
              kDevice,
              builder->getStringAttr(preferred_inference_device_type.hardware));
          new_call->setAttr(
              kInferenceType,
              builder->getStringAttr(GetInferenceString(

  Builder builder(&getContext());

  // Adds the `tf._noinline = true` attribute to the function if the name
  // matches.
  if (noinline_functions.contains(func_op.getSymName())) {
    func_op->setAttr(kTfNoinlineAttr, builder.getBoolAttr(true));
    LLVM_DEBUG(llvm::dbgs()
               << "Marked tf._noinline = true: " << func_op.getSymName());
  }
}

static PassRegistration<MarkFunctionsNoinlinePass> pass{};

              getElementTypeOrSelf(operand.getType()))) {
        legalize_to_tfl = false;
        return WalkResult::interrupt();
      }
    }
    return WalkResult::advance();
  });
  module->setAttr(kLegalizeTflVariables,
                  BoolAttr::get(context, legalize_to_tfl));
}

}  // namespace

std::unique_ptr<OperationPass<ModuleOp>> CreateAnalyzeVariablesPass() {