Dynamic Dispatch of Context-Sensitive Optimizations

on YouTube ↗09m

" summary

Implements a runtime DFA-based context tracking that avoids cloning whole call chains, implemented as an LLVM link-time pass and published with code on Github. Claim is that moving context bookkeeping into a state machine yields much smaller binaries and low overhead: on SPEC CPU2017 and LLVM tests the DFA approach produces binaries about 3x smaller than cloning (cloning adds ~8.5x bytes) while compile-time remains under a minute and actual runs show >3% speed-up in 11 programs.

AI-written summary · about

" chapters (11)

1. 0:00Intro
2. 0:15Interprocedural, context-sensitive analyses
3. 0:46Interprocedural, context-sensitive optimizations
4. 1:28Inlining
5. 2:44Full Cloning
6. 4:24Our approach (DFA)
7. 4:59DFA - Overview
8. 6:35Compile time & binary sizes
9. 7:16Running time
10. 7:39Scalability
11. 8:31Summary

" related talks

• 21mContext Sensitivity without Contexts: A Cut-Shortcut Approach to Fast and Precise(…)Yue Li, Tian Tan, Chang Xu, Xiaoxing Ma, Wenjie Ma, Shengyuan Yang · PLDI 2023
• 22mRevisiting Dynamic Dispatch for Modern ArchitecturesDave Mason · VMIL 2023
• 16mReconciling Optimization With Secure CompilationAlbert Cohen, Arnaud de Grandmaison, Karine Heydemann, Son Tuan Vu, Christophe Guillon · OOPSLA 2021
• 28mReducing Feedback PollutionFilip Krikava, Jan Vitek, Sebastián Krynski, Michal Štěpánek, Filip Říha · VMIL 2024
• 01h 07mC++20 Concepts: A Day in the LifeSaar Raz · CppCon 2019
• 01h 04mJust-in-Time Compilation: The Next Big Thing?Kris Jusiak, Ben Deane · CppCon 2020
• 24mTowards Understanding the Costs of Avoiding Out-of-Thin-Air ResultsPeizhao Ou · OOPSLA 2018
• 15mDuplo: A Framework for OCaml Post-Link OptimisationNandor Licker, Timothy M. Jones · ICFP 2020