return %2 : tensor<1x3xf32>
    }
  }
)mlir";

constexpr absl::string_view kModuleXlaCallModule = R"mlir(
  module {
    func.func @main(%arg0: tensor<?x2xf32> {tf_saved_model.index_path = ["input_tensor"]}) -> (tensor<?x2xf32>) {
      %0 = stablehlo.constant dense<[-0.211145893, -0.708605706]> : tensor<2xf32>

      // Matches the `tf_saved_model.index_path = ["__tf_file_prefix"]`.
      const auto index_path_attr_itr =
          node_def.attr().find(kTfSavedModelIndexPathAttr.str());
      if (index_path_attr_itr != node_def.attr().end()) {
        const auto& index_paths = index_path_attr_itr->second.list().s();
        if (absl::c_find(index_paths, kTfFilePrefix.str()) !=
            index_paths.end()) {

  // The pass shouldn't do anything in this case.

  // CHECK: func @f(%arg0: tensor<!tf_type.resource<tensor<f32>>>  {tf_saved_model.index_path = [0]})
  func.func @f(%arg0: tensor<!tf_type.resource<tensor<f32>>>  {tf_saved_model.index_path = [0]})
  attributes {tf_saved_model.exported_names = ["f"]} {
    %val = "tf.ReadVariableOp"(%arg0) : (tensor<!tf_type.resource<tensor<f32>>>) -> tensor<f32>

    and use the standardized MLIR `sym_name` attribute to hold it.

    This dialect annotates func ops with two argument attributes
    `tf_saved_model.index_path` and `tf_saved_model.bound_input` in order to
    model the calling API of a function with SavedModel semantics.
    In particular, func's in SavedModel are called with a structured value, and

# CHECK-SAME: tf.versions
# CHECK-SAME: bad_consumers
# CHECK-SAME: min_consumer
# CHECK-SAME: producer

# CHECK:      func {{@[a-zA-Z_0-9]+}}(
# CHECK-SAME:   [[ARG:%.*]]: tensor<*xf32> {tf_saved_model.index_path = ["x"]}

# CHECK: [[shape:%.*]] = "tf.Shape"([[ARG]])
# CHECK-NEXT: [[batch_size:%.*]] = "tf.StridedSlice"([[shape]],
# CHECK-NEXT: [[result:%.*]] = "tf.Pack"([[batch_size]],
# CHECK-NEXT: return [[result]] : tensor<2xi32>

// function to represent the path which one would use to index into a structured
// value to reach a given tensor.
inline constexpr StringRef kTfSavedModelIndexPathAttr =
    "tf_saved_model.index_path";

// Name of the attribute that inidicates the type of initializer. It should be
// on a function and the function should exist in the initializers attribute of
// the SessionInitializerOp.

  // Test case: Sanity check handling of non-bound inputs.
  // The pass shouldn't do anything in this case.

  // CHECK: func @f(%arg0
  func.func @f(%arg0: tensor<!tf_type.string> {tf_saved_model.index_path = [0]})
  attributes {tf_saved_model.exported_names = ["f"]} {

//   2) The `tf_saved_model` dialect requires that the arguments with
//   `tf_saved_model.index_path` attributes should precede those with
//   `tf_saved_model.bound_input` attributes. The init function of type
//   `kTfSavedModelInitializerRestoreType` usually has an argument with
//   `tf_saved_model.index_path`, whereas the init function of type
//   `kTfSavedModelInitializerInitType` may have arguments with

// RUN: tf-quant-opt %s -split-input-file -quant-add-dump-tensor-op='debugger_type=float_per_layer' | FileCheck --check-prefix=FloatPerLayer %s

module {
  func.func @matmul2(%arg0: tensor<?x2xf32> {tf_saved_model.index_path = ["input_tensor"]}) -> (tensor<?x2xf32>) {
    %0 = stablehlo.constant dense<[-0.211145893, -0.708605706]> : tensor<2xf32>
    %1 = stablehlo.constant dense<[[-0.630731344, 0.54962182], [0.180364341, -0.764542698]]> : tensor<2x2xf32>

// FloatPerLayer-DAG: return %[[output1_unquantized]]
}

// -----

module {
  func.func @matmul2(%arg0: tensor<2x2xf32> {tf_saved_model.index_path = ["input_tensor"]}) -> (tensor<2x2xf32>) {
    %cst = "tf.Const"() {device = "", value = dense<[[-0.630731344, 0.54962182], [0.180364341, -0.764542698]]> : tensor<2x2xf32>} : () -> tensor<2x2xf32>