Recoleta
GitHub
Source note

AutoPass: Evidence-Guided LLM Agents for Compiler Performance Tuning

Authors
Zepeng Li, Jie Ren, Zhanyong Tang, Jie Zheng, Zheng Wang
Source
arXiv
Published
Collection
Software Intelligence
Compiler OptimizationLLM AgentsLlvmCode IntelligencePerformance Tuning
Open arXivMarkdown source

Summary

AutoPass is an inference-only LLM agent system for LLVM pass tuning. It uses compiler evidence and measured runtime feedback to choose, repair, test, and revise optimization pass pipelines.

Problem

  • Runtime compiler tuning matters because fixed LLVM -O3 pipelines do not perform best on every program and target.
  • The search space is large: LLVM 17 exposes over 100 transformation passes, and the paper evaluates 74 passes with sequences up to 107 passes.
  • Runtime speed is noisy and hardware-dependent, so PGO, black-box autotuning, and learned policies can miss profitable pass orders or need more data and runs than a deployment budget allows.

Approach

  • AutoPass starts from the LLVM -O3 pipeline and uses four agents: Score, Analysis, Reasoning, and Evaluation.
  • The Score Agent ranks functions using call graphs and IR features such as block, loop, call, and conditional-branch counts.
  • The Analysis Agent reads LLVM IR plus -Rpass, -Rpass-missed, and -Rpass-analysis remarks, then emits structured JSON about semantic hints and missed optimizations.
  • The Reasoning Agent edits pass order and pass parameters, such as unroll factors or inlining thresholds, using compiler evidence and prior runtime results.
  • The Evaluation Agent compiles, verifies, runs candidates, measures runtime and hardware counters, keeps only faster valid pipelines, and rolls back to -O3 when no candidate improves.

Results

  • Across 5 benchmark suites with 64 workloads, AutoPass R3 reports the best result in 9 of 10 platform-suite settings against Instrumented PGO, CSSPGO, AutoFDO, and OpenTuner under a 3-iteration budget.
  • The abstract reports geometric-mean speedups over LLVM -O3 of 1.043× on x86-64 and 1.117× on ARM64.
  • On x86-64, AutoPass R3 reports 1.059× on cBench, 1.009× on PolyBench, 1.137× on CoreMark, 1.008× on MiniFE, and 1.102× on LULESH.
  • On ARM64, AutoPass R3 reports 1.111× on cBench, 1.149× on PolyBench, 1.091× on CoreMark, 1.068× on MiniFE, and 1.046× on LULESH.
  • The motivating Qsort and BitCount experiment reports an average 1.259× speedup over -O3 across x86-64 and ARM64.
  • One-shot AutoPass R1 is weaker than R3 in the table, for example ARM64 PolyBench improves from 1.129× at R1 to 1.149× at R3, and ARM64 CoreMark improves from 1.006× to 1.091×.

Link