Long Context vs. RAG for LLMs: An Evaluation and Revisits