final delegate = Mock(DeferredExecutionAwareStep)
    final work = Stub(UnitOfWork)

    def setup() {
        _ * work.displayName >> displayName
        _ * work.identify(_, _) >> identity
    }

    protected void assertNoOperation() {
        assert buildOperationRunner.log.records.empty
    }

  Value control_fetch = *cloned_fetch_op.getFetches().begin();
  cloned_fetch_op.erase();

  return control_fetch;
}

// Creates a new `IdentityOp` wrapped by an `IslandOp`. The identity op returns
// the `main_file_prefix_arg` and has control dependencies to `control_inputs`.
IslandOp CreateFilePrefixIdentityOp(const BlockArgument main_file_prefix_arg,

        // capturing the normalized path and the snapshot of the raw contents, so we are using these to determine the identity.
        // We do this because external artifact transforms typically need to identify themselves redundantly many times during a build.

// CHECK: return %[[CONST_0]] : tensor<*xf32>
}

func.func @identity(%arg0: tensor<2xf32>) -> tensor<*xf32> {
  %cst_1 = "tf.Const"() {device = "", value = dense<1.000000e-03> : tensor<f32>} : () -> tensor<f32>
  %0 = "tf.Identity"(%arg0) {device = ""} : (tensor<2xf32>) -> tensor<*xf32>
  %1 = "tf.AddV2"(%0, %cst_1) {device = ""} : (tensor<*xf32>, tensor<f32>) -> tensor<*xf32>

			// To migrate away from encryption, user can specify an identity transformer higher up
			// (in the config file) than the encryption transformer. In that scenario, the identity transformer needs to
			// identify (during reads from disk) whether the data being read is encrypted or not. If the data is encrypted,

        Map<String, OutputPropertySpec> outputs = outputFileSpecs + outputDirSpecs

        return new MutableUnitOfWork() {
            boolean executed

            @Override
            UnitOfWork.Identity identify(Map<String, ValueSnapshot> identityInputs, Map<String, CurrentFileCollectionFingerprint> identityFileInputs) {
                new SimpleIdentity("myId")
            }

            @Override

        %cast = "tf.Cast"(%output) : (tensor<1024x24x24x3xi32>) -> tensor<1024x24x24x3xf32>
        %identity = "tf.Identity"(%scale) : (tensor<f32>) -> tensor<f32>
        %mul = "tf.Mul"(%cast, %identity) : (tensor<1024x24x24x3xf32>, tensor<f32>) -> tensor<1024x24x24x3xf32>
        func.return %mul : tensor<1024x24x24x3xf32>
      }
    }
  )mlir";

    for (auto resource_user : resource_users_to_visit) {
      add_resource_op_to_list(resource_user);

      // Account for pass-through identity ops.
      if (auto pass_through_identity =
              llvm::dyn_cast<TF::IdentityOp>(resource_user)) {
        for (auto identity_user :
             pass_through_identity.getOutput().getUsers()) {
          new_resource_users.emplace_back(identity_user);

  }
};

// Replaces the Identity op with its input in either of the following scenarios
// : 1) The Identity op's input and output have same types/shapes. 2) The result
// of Identity op is only used by TF ops.
struct RemoveIdentity : public OpRewritePattern<TF::IdentityOp> {
  using OpRewritePattern<TF::IdentityOp>::OpRewritePattern;

  LogicalResult matchAndRewrite(TF::IdentityOp identity,

  func.func private @internal_conv3d_fn(
                         %input : tensor<*xf32>, %filter : tensor<*xi8>) -> tensor<*xf32> {

   // Use identity op to avoid the filter being constant-folded.
    %identity = "tf.Identity"(%filter) : (tensor<*xi8>) -> tensor<*xi8>
    %2 = "tf.Cast"(%identity) {Truncate = false} : (tensor<*xi8>) -> tensor<*xf32>
    %3 = "tf.Conv3D"(%input, %2) {
      padding = "VALID", strides = [1, 1, 1, 1, 1],