inputs.push_back(std::move(tensor));
  }

  std::vector<tensorflow::TensorValue> input_values;
  for (tensorflow::Tensor& tensor : inputs) {
    input_values.emplace_back();
    input_values.back().tensor = &tensor;
  }

  tensorflow::OpKernelContext::Params params;
  params.inputs = input_values;
  params.device = runner->device();
  params.op_kernel = runner->op_kernel();

  void rewrite(AddOp op, PatternRewriter& rewriter) const override {
    DeferRhsTransposeForBinaryOp(op, rewriter);
  }
};

// Rewrites the `reduce_window(transpose(%activation), %init_value)` patterns to
// `transpose(reduce_window(%activation), %init_value)`, deferring the transpose
// to the result. The reduce function should be equivalent to
// `stablehlo.maximum`, representing max pooling.

  }

  template <typename T>
  void Run(const std::vector<Input::Initializer>& input_values,
           const std::vector<T>& expected_output_values) {
    ClientSession session(scope_);

    DCHECK_EQ(input_values.size(), inputs_.size());
    ClientSession::FeedType feeds;
    for (int i = 0; i < inputs_.size(); ++i) {
      feeds.emplace(inputs_[i], input_values[i]);
    }

    std::vector<Tensor> out_tensors;

  }

  void Run(std::initializer_list<int> input_values) {
    Run(outputs_, input_values);
  }

  void Run(const std::vector<TF_Output>& run_outputs,
           std::initializer_list<int> input_values) {
    DCHECK_EQ(inputs_.size(), input_values.size());
    std::vector<std::pair<TF_Operation*, TF_Tensor*>> inputs(inputs_.size());
    int i = 0;
    for (int v : input_values) {
      inputs[i] = {inputs_[i].oper, Int32Tensor(v)};

 public:
  using ConvertReduceOpToTfOp::ConvertReduceOpToTfOp;

  LogicalResult MatchInitValue(Value init_value) const override {
    auto type = mlir::cast<ShapedType>(init_value.getType()).getElementType();
    if (mlir::isa<FloatType>(type)) {
      float const_value;
      if (failed(GetConstantSplatValue<float>(init_value, const_value)) ||
          const_value != 1.0)
        return failure();

    } else {
      llvm::SmallVector<Value> input_values(node_output_tensors.size(),
                                            scalar_one);
      for (const auto& [output_index, tensor_index, tensor_value] :
           node_output_tensors) {
        input_values[tensor_index] = tensor_value;
      }
      identity_op = builder.create<TF::IdentityNOp>(
          new_loc, TypeRange(ValueRange(input_values)), input_values);
    }

constexpr char kVarOpName[] = "var";

GraphDef BuildSimpleGraph() {
  Scope root = Scope::NewRootScope();
  auto init_value = Const(root, 0);
  auto var = Variable(root.WithOpName(kVarOpName), TensorShape({}),
                      DataType::DT_INT32);
  auto assign = Assign(root.WithOpName(kAssignOpName), var, init_value);
  auto count = CountUpTo(root.WithOpName(kCountUpToOpName), var, 10);
  Square(root.WithOpName(kSquareOpName), var);  // NOLINT

    return "Unfreeze large constants.";
  }

  void runOnOperation() override;

 private:
  Option<int64_t> CreateSizeThresholdInBytesOption(const int64_t init_value) {
    return Option<int64_t>(
        *this, "size_threshold_in_bytes", llvm::cl::init(init_value),
        llvm::cl::desc(
            "Lower threshold of the constant size for unfreezing. Constants "

    }

    def "test Groovy eager collection types assignment for #description"() {
        def initValue = inputType.contains("Map<") ? "[:]" : "[]"
        def inputDeclaration = "$inputType input = $initValue"
        groovyBuildFile(inputDeclaration, inputValue, operation)

        expect:
        runAndAssert("myTask", expectedResult)

        where:

TEST(ResourceOperationSafetyAnalysisTest, Loop) {
  Scope root = Scope::NewRootScope().ExitOnError();

  Output init_value = ops::Placeholder(root.WithOpName("init"), DT_FLOAT);
  Output loop_cond = ops::Placeholder(root.WithOpName("init"), DT_BOOL);
  Output enter_value =
      ops::internal::Enter(root.WithOpName("enter"), init_value, "fr");
  ops::Merge iv(root.WithOpName("iv"), {enter_value, enter_value});