arxiv.org/abs/2606.31238 A Multi-Dimensional, Per-Pass Empirical Study of the LLVM Optimization Pipeline
A study of the LLVM -O3 pipeline split it into 113 cumulative per-pass prefixes and took 84,750 measurements on 30 PolyBench/C kernels. It found 36% of significant runtime-changing transitions were regressions, gains were back-loaded and concentrated in a few passes, and runtime-targeted passes also saved 30-60% energy.
— via arXiv Programming Languages, Federico Bruzzone, Walter Cazzola
