AI Agents Assist in Decompiling Nintendo 64 Game Code

AI in Code Decompilation

AI agents were employed in the early stages of decompiling a Nintendo 64 game, specifically for identifying and matching standard-library code. This task is a common initial step in N64 decompilation projects, as many games utilize shared libraries like Nintendo's libultra and the libmus audio library, which have available source code.

Leveraging Existing Libraries

A key aspect of the process involved guiding AI agents to prioritize existing library source code over attempting to decompile these functions from scratch. Agents were instructed to use identified library functions as a starting point, exploring various SDK versions, compiler options, and conditional compilation paths before generating their own implementations. Tools like N64Sym aided in identifying probable library functions within the game's ROM.

Automated Matching and Efficiency

To improve efficiency, an agent developed a script that ran m2c against unmatched functions and automatically integrated exact matches. While this script only matched 17 out of 1,830 functions (a 0.93% success rate), it reduced the computational cost associated with individual agent attempts for these specific matches.

Learning Compiler Behavior

The project utilized AI's ability to learn and generalize compiler quirks. Agents were prompted to document observed IDO compiler behaviors in a shared DECOMPILATION_LEARNINGS.md file. This created a feedback loop where agents contributed to a knowledge base, which in turn helped subsequent agents more effectively match IDO code by providing insights into recurring compiler patterns.

Utilizing Specialized Tools

N64 Decomp Workbench proved to be a valuable resource for debugging late-stage MIPS decompilation mismatches. This tool can classify mismatches, account for relocations, and replay individual compiler passes, offering deeper insights than raw assembly differences. While requiring specific compiler binaries and project setup, it provides critical information for distinguishing structural problems from register-allocation issues.