Phase-Aware Counterfactual Optimization: Learning When Less Information Helps

You know how Theodoridis and Su's 2024 paper revealed this counterintuitive phenomenon where giving compilers more information sometimes makes things worse? They found 59 cases where additional information degraded optimization in GCC and LLVM. This is fascinating because it suggests we're fundamentally misunderstanding something about how compiler phases interact. What if we could learn to predict these cases? My idea is to build a system that uses counterfactual machine learning - essentially asking "what would happen if I didn't provide this information?" - to identify when additional program analysis will backfire. We'd train models on thousands of program variants, comparing optimized outputs with and without extra information. The key innovation is using graph neural networks to model the complex phase ordering interactions that cause these degradations. This goes beyond just detecting problems (like Theodoridis and Su did) to actively preventing them. Imagine a compiler that can say "Actually, don't tell me about this dead code - it'll mess up my register allocation later." That's a paradigm shift from "more information is always better" to "the right information at the right time."

References:

1. Refined Input, Degraded Output: The Counterintuitive World of Compiler Behavior. Theodoros Theodoridis, Zhendong Su (2024). Proc. ACM Program. Lang..