Software vulnerabilities, caused by unintentional flaws in source codes, are the main root cause of cyberattacks. Source code static analysis has been used extensively to detect the unintentional defects, i.e. vulnerabilities, introduced into the source codes by software developers. In this paper, we propose a deep learning approach to detect vulnerabilities from their LLVM IR representations based on the techniques that have been used in natural language processing. The proposed approach uses a hierarchical process to first identify source codes with vulnerabilities, and then it identifies the lines of codes that contribute to the vulnerability within the detected source codes. This proposed two-step approach reduces the false alarm of detecting vulnerable lines. Our extensive experiment on real-world and synthetic codes collected in NVD and SARD shows high accuracy (about 98\%) in detecting source code vulnerabilities.