Modular Storage Mode Analysis
summary
Claims a modular storage mode analysis that works across compilation units and is now adopted in MLKit, replacing the old conservative local analysis. Implementation ties region inference and a region flow graph to Reml’s gradual annotations, supports region resets, and yields concrete memory wins on a TCP/IP-layer case study: region-only runs hit ~34 MB, a tag-based GC plus regions drops to ~234 KB (2:1 live ratio), and the precision GC experiment reaches ~184 KB (2:1) or ~533 KB at 10:1. Progress/preservation proofs are underway.
AI-written summary · about
chapters (8)
- 0:00Modular storage mode analysis
- 2:51Region-based memory management
- 12:54Reverse function storage
- 17:54Old storage limits
- 20:06New modular approach
- 26:17Type system soundness
- 27:26Practical GC applications
- 30:45Future work and conclusion
related talks
37mContractions Are Abstractions - Towards Effective Memory Management in ReML
01h 03mMMTk and The Case for Modular VM Development
30mCompositional Deep Argument Flattening
14mSound and Modular Activity Analysis for Automatic Differentiation in MLIR
18mParallelism in a Region Inference Context
15mA Modular Cost Analysis for Probabilistic Programs
21mFlow-Analysis-Based Closure Optimization
23mVerified Compilation of C Programs with a Nominal Memory Model