pm.addPass(mlir::createSymbolDCEPass());
  pm.addPass(mlir::createCanonicalizerPass());
  pm.addPass(mlir::TF::CreateTFShapeInferencePass());
  pm.addPass(mlir::createCanonicalizerPass());
  pm.addPass(mlir::TF::CreateTensorListOpsDecompositionPass());
  pm.addNestedPass<func::FuncOp>(
      mlir::TFDevice::CreateDecomposeResourceOpsPass());

  // FreezeVariables only freezes variables for TF v1 types. Separately handle

  // pass undo some CSE because shape_inference is not correctly able to
  // identify the shapes of TensorList initialization ops.
  // This pass requires CanonicalizerPass before
  // CreateTensorListOpsDecompositionPass for clean-ups.
  if (lower_to_xla_hlo) {
    pm.addNestedPass<mlir::func::FuncOp>(
        mlir::TF::CreateReplicateTensorListInitOpsPass());
  }

// Converts tensor list operations into operations on buffers and sizes. Needs
// static shapes and known max element count.
std::unique_ptr<OperationPass<ModuleOp>> CreateTensorListOpsDecompositionPass();

// Converts tensor array ops into operations on local variables, which can later
// be removed by resource lifting. Requires known sizes and known element shapes

  if (failed(DecomposeTensorListOps(&main.front(), module))) {
    signalPassFailure();
  }
}

}  // namespace

namespace TF {
std::unique_ptr<OperationPass<ModuleOp>>
CreateTensorListOpsDecompositionPass() {
  return std::make_unique<TensorListOpsDecompositionPass>();
}
}  // namespace TF

  %next_index_in_list = "tf.AddV2"(%index_in_list, %one) : (tensor<1xi32>, tensor<1xi32>) -> tensor<1xi32>
  return
}
```
  }];

  let constructor = "TF::CreateTensorListOpsDecompositionPass()";
}

def TPUParallelExecuteSinkResourceWritePass : Pass<"tf-tpu-parallel-execute-sink-resource-write", "mlir::func::FuncOp"> {