PrETi: Predicting Execution Time in Early Stage with LLVM and Machine Learning

We propose preti, a novel framework for predicting software execution time in the early stages of development. preti combines static analysis and simulation to extract timing-relevant features from LLVM intermediate representation (IR), which are then used to train machine learning models. Central to our approach is a control-flow graph (CFG) alignment algorithm that establishes a structural correspondence between IR basic blocks and their associated assembly instructions. To enhance the fidelity of timing predictions, preti simulates microarchitectural behaviors—including cache accesses, branch prediction, and register spills—thereby generating execution traces that approximate real-world behavior. This design enables accurate execution time prediction without requiring native execution on embedded hardware, making preti particularly suited for early-stage performance analysis.