SLaDe-AGBCC: Neural Matching Decompilation

SLaDe-AGBCC is a research project investigating neural matching decompilation for the Game Boy Advance. Unlike traditional decompilers, it does not prioritize readable structure or name recovery; instead, it optimizes for producing C code that recompiles to byte-identical assembly when processed by the agbcc compiler. The project draws inspiration from AlphaFold 2, treating C recovery as a folding problem and utilizing a compiler-in-the-loop recycling signal to refine predictions. The author reports that the core hypothesis—that transformers can defeat the agbcc optimizer to achieve byte-exactness—is currently being tested rather than assumed. Development involves an automated pipeline where disassembled objects are converted to C and iteratively refined. The project relies on a curated corpus of GBA decompilation projects and decomp.me data. Recent experiments include v4 recycling, which demonstrated progress in loop-based matching, though the author notes that recycling must be explicitly trained. The project documentation credits Claude Fable 5 for co-authorship on various implementation and diagnostic tasks.

Sources: Source 1