Towards Understanding What Code Language Models Learned