Building X-Ray for enterprise-scale software
summary
Clearblue lifts production binaries into LMIR and builds a whole‑program PDG stored on disk to answer value‑flow queries; it analyzes the Linux kernel (≈381,000 functions, ~30M LOC) with six levels of inlining and full path sensitivity, but deliberately gives up soundness and limits context depth (typically 6) and path matches (≈10) to finish. Precomputed CFL reachability summaries on disk yield up to 200–300× perceived speedups; hybrid use of LLMs (DAINFO, ~80% accuracy) recovers container aliasing and reduces false positives when combined with static patterns.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 3:36Static analysis needs X-ray
- 12:52Clear Blue vision
- 16:34Clear Blue architecture
- 19:52Value flow analysis innovations
- 25:12Sensitivity and context flow
- 31:57Optimizing reachability queries
- 37:00LLMs with static analysis
- 42:30Container data flow case study
- 45:06Deadlocks with interrupt handlers
- 48:40Memory wrapper detection
- 54:06Takeaways and future
- 56:22Q&A
related talks
30mEnhancing CTI Processes with Code Search Technology
28mImproving Build Time SBoMs
26mProfiling Mastery: Decode Your Data and Elevate Observability
46mThe Hidden Secrets of Debugging
59m“GDB - A Lot More Than You Knew"
14mWhen Sampling Lies: Trustworthy Performance Profiling for Flat Workloads with Blink
20mLDB: An Efficient Latency Profiling Tool for Multithreaded Applications
01h 03mApplied Best Practices