AnICA: Analyzing Inconsistencies in Microarchitectural Code Analyzers
summary
AnICA applies differential testing plus an abstract basic block generalization algorithm to compare llvm-mca, uiCA, Ithemal and similar predictors on 10,000 random 4-instruction blocks and finds large disagreements (e.g., ~31% between uiCA and llvm-mca, many pairs ~20–50%, some ~50%). The tool reduces thousands of inconsistent cases to a few dozen abstract descriptions (150 per campaign, top 10 cover up to 82%), discovers an llvm-mca fxrstor crash later fixed, and reports AMD Zen model inaccuracies (five issues, 72 instruction schemes).
AI-written summary · about
chapters (6)
- 0:00Intro
- 0:13The Setting: Basic Block Throughput Prediction
- 8:21The Algorithm - Example
- 15:48The Abstract Interpretation Connection
- 19:23Evaluation - Effectiveness of Generalization
- 22:08Evaluation - Selected Case Studies
related talks
29mCross-Level Debugging for Static Analysers
20mMica: Automated Differential Testing for OCaml Modules
14mUnveiling Compiler Faults via Attribute-Guided Compilation Space Exploration
17mDemo: RISC-V AIA Expanding Interrupts: Applications, Implementation and Verification
38mStatic Code Analysis - A Behind-the-scenes Look
36mCode Analysis: Empowering developers to write performant, reliable, and safe C++
47mA Root Canal Static Analysis Based Audit of Zephyr
18mPreview: Verification (Correctness & Debugging)