On the Hidden Mystery of OCR in Large Multimodal Models