Accurate static profile prediction is crucial for achieving optimal program performance in the absence of dynamic profiles. However, existing static profiling methods struggle to fully exploit the complex structure of the compiler’s intermediate representation (IR) and fail to effectively utilize the calling context information needed for accurate profile inference. To address these limitations, we introduce \textit{GraalNN}, a Graph Neural Network (GNN)-based static profiling framework that directly learns structural information from control-flow graphs (CFGs). This reduces the reliance on handcrafted features and minimizes the effort required for feature engineering while improving the model’s ability to predict profiles based on captured control-flow dependencies.

GraalNN follows a two-stage approach: it predicts context-insensitive profiles during the parsing phase and refines them with context-sensitive profiles during the inlining phase. This approach achieves a runtime speedup of 10.13% on a wide range of industry-standard benchmarks, surpassing existing static profiling techniques by over 2.5%. Additionally, our method limits the binary size increase to just 2.94%, maintaining compilation efficiency and outperforming all existing static profiling solutions by at least 0.9%.