A Multi-Dimensional, Per-Pass Empirical Study of the LLVM Optimization Pipeline
Researchers conducted an empirical study analyzing the LLVM optimization pipeline on a per-pass basis across multiple dimensions. The work provides detailed insights into how individual optimization passes affect code performance and compilation metrics, offering a granular understanding of compiler optimization behavior.
Background
- The paper studies the **LLVM optimization pipeline** — the sequence of transformations (passes) a compiler runs to make code faster/smaller. Each pass does something like eliminating dead code or unrolling loops, and the order matters a lot.
- LLVM is the compiler infrastructure behind Clang (C/C++/Rust), Swift, Julia, and many languages. It's the dominant open-source compiler project.
- Traditionally, pass order was hand-tuned by experts and fixed behind flags like "-O2". This work is part of a newer wave trying to treat pass ordering as a search or machine-learning problem.
- The authors exhaustively tested all pass orderings on small programs — usually infeasible due to combinatorial explosion. Key finding: a pass's effect depends heavily on which passes ran before it, and many passes do nothing on average but are critical in specific positions. That challenges fixed pipelines and supports adaptive ordering.