Programming Language Design and Implementation for the Machine Learning Era: A Personal(…)
summary
Machine learning already augments and sometimes outperforms handcrafted compiler components: meta optimization (2003) evolves priority functions and yields ~25–50% speedups; imitation learning on SLP matches or beats its ILP teacher; a learned cost model trained on 1.4M basic blocks halves prediction error versus Intel models; an architecture‑manual approach produces vegan generated dot‑product code that outpaces LLVM for some kernels; a verified compiler prototype produces 40k lines, 34 days with proof artifacts. Three futures emerge: augment, generate, or replace traditional compilers, with large energy and cost stakes if performance improves by ~10%.
AI-written summary · about
chapters (11)
- 0:00Introduction and overview
- 2:59Future of compilers
- 6:33Lessons from the past
- 10:28Data-driven cost models
- 16:35Generating compilers with ML
- 20:01Credible compilation
- 23:09ML replacing compilers
- 26:08Redesigning programming
- 34:29Source of record and future
- 44:00A plea to the PLDI community
- 49:42Concluding thoughts and Q&A
related talks
34mOptimizing AI Inference with ML Compilers & Hardware
59mPrinciples of Programming Language Translators
48mCross-Domain Compilation: Exploiting Synergies Across the CS Community
22mA Vision of Miking: Interactive Programmatic Modeling, Language Composition, and Compilation
43mMachine Learning for Developer Productivity
27mImproving Vectorization Heuristics in a Dynamic Compiler with Machine Learning Models
01h 00mBetter Learning through Programming Languages
44mMachine Learning for Faster Optimization