LLM4EFFI: Leveraging Large Language Models to Enhance Code Efficiency and Correctness