RedDragon: LLM-assisted IR analysis of broken/incomplete code across languages (github.com)

RedDragon is a groundbreaking framework designed for reverse-engineering incomplete code, particularly from legacy migrations and decompiled binaries, across multiple programming languages. It incorporates three main approaches: deterministic language frontends with LLM-assisted repair for handling malformed syntax, full LLM frontends for languages lacking tree-sitter support, and a virtual machine (VM) that integrates LLMs to address issues with missing dependencies and unresolved imports. By maximizing deterministic coverage and utilizing a universal flattened three-address code intermediate representation (IR), RedDragon enables reliable analysis and execution of code even in the presence of ambiguities. This project is significant for the AI/ML community as it demonstrates a practical application of large language models (LLMs) in improving code interpretation and execution, thereby facilitating code analysis in settings where traditional methods might fail. RedDragon showcases a seamless pipeline from parsing to execution, allowing for type-aware execution and detailed data flow analysis without unnecessary LLM calls. The framework ultimately empowers developers to work with incomplete codebases more effectively, promising enhanced tools for software maintenance and migration in an increasingly diverse coding landscape.