Bridging the differences across CPU instruction set architectures—LoongArch architecture software porting technique based on large language models

The evolution of China's processor chip industry has given rise to the domestic instruction set architecture LoongArch, which necessitates the support of a corresponding software ecosystem. However, due to the differences between instruction set architectures, developers often need to dedicate a substantial amount of time and professional expertise to identify architecture-related components in software packages and formulate corresponding porting strategies during the software adaptation process. The prevailing technologies utilized in the domain of instruction set architecture porting primarily employ binary translation techniques. While these techniques have substantially expedited the process of software adaptation, they concomitantly introduce performance overhead and complicate the task of debugging when errors arise. Furthermore, these principles do not universally apply to all software systems. Consequently, there is an urgent need for effective source code porting assistance technology to fundamentally solve the instruction set architecture challenges of software and its subsequent evolved versions. To address these challenges, this paper proposes a porting technology framework called ArchPorting to facilitate the migration and adaptation of software for the domestic instruction set architecture LoongArch. ArchPorting constructs a multi-dimensional instruction set architecture feature knowledge base and can identify architecture-related components in software packages written in various programming languages. For the code snippets that have been identified as requiring adaptation, the tool utilizes the code generation capabilities of large models and the knowledge base to customize prompts. These prompts contain migration background knowledge based on basic architecture knowledge, architecture mapping relationships, and code migration examples. Subsequently, it generates code porting strategies that are semantically coherent with the context, thereby assisting developers in completing software adaptation. Results indicate that ArchPorting attains an accuracy rate of 90.5% and a recall rate of 99.1% in identifying architecture-related components. It has been determined that 80.29% of the code snippets can reduce the cost for developers to adapt software. ArchPorting has been demonstrated to enhance the efficiency and precision ofsoftware migration, thereby providing substantial support for the development of LoongArch CPUS.