The acyclic e-graph: Cranelift's mid-end optimizer

Article

• Cranelift (WebAssembly compiler) adopts acyclic e-graphs (aegraphs) for mid-end optimization
• Aegraphs represent multiple equivalent program forms simultaneously, enabling better optimization
• Solves the pass-ordering problem: no need to sequence GVN, DCE, and other passes manually
• Surprising finding: multi-value representation alone may not be the key benefit

Discussion

• Compiler writer argues pass-ordering problem is overstated; GVN+load elimination is the main real case
• E-graphs also applicable to type checkers and trait solvers, not just codegen (Rust’s trait solver cited)
• egg library (egraphs-good.github.io) referenced as the earlier popularizer of the approach
• Curiosity about compile-time overhead tradeoffs of the e-graph approach

Discuss on HN