Improving Assembly Code Performance with LLMss via Reinforcement Learning