AutoPass: Evidence-Guided LLM Agents for Compiler Performance Tuning

2026-03-18 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, extended

Summary

AutoPass is a multi-agent framework designed for LLVM compiler performance tuning, leveraging Large Language Models (LLMs) to guide optimization decisions. Unlike traditional black-box autotuning, AutoPass integrates LLMs directly with compiler-internal optimization states and intermediate representation analysis. Operating in an inference-only, training-free setting, it requires no offline training or task-specific fine-tuning, making it highly adaptable. Implemented on the LLVM compiler, AutoPass was evaluated on server-grade x86-64 and embedded ARM64 systems. It consistently outperformed expert-tuned heuristics and classical autotuning methods, achieving geometric-mean speedups of 1.043x on x86-64 and 1.117x on ARM64 over the LLVM -O3 baseline. The framework comprises a Score Agent for hotspot identification, an Analysis Agent for feature extraction, a Reasoning Agent for core optimization, and an Evaluation Agent for performance feedback, demonstrating architecture-aware optimization and robust iterative refinement.

Key takeaway

For Machine Learning Engineers or Compiler Engineers optimizing code performance, particularly on diverse architectures like x86-64 and ARM64, you should explore multi-agent LLM frameworks for compiler tuning. AutoPass demonstrates that integrating LLMs with compiler-internal feedback and runtime measurements can yield superior speedups over -O3, such as 1.043x on x86-64 and 1.117x on ARM64, without requiring offline training. Implement iterative refinement and a rollback policy to ensure robust, architecture-aware optimization and mitigate initial performance regressions.

Key insights

LLMs effectively guide compiler optimization by integrating compiler-internal signals and runtime feedback without offline training.

Principles

• Context-sensitive LLM reasoning improves compiler tuning.
• Compiler-internal signals enhance LLM optimization decisions.
• Iterative refinement with runtime feedback boosts robustness.

Method

A multi-agent LLM framework iteratively refines compiler pass sequences and parameters, guided by compiler remarks, IR analysis, and measured runtime performance feedback.

In practice

• Query compiler artifacts for transformation effects.
• Employ LLM agents for pass ordering and parameter selection.
• Implement a rollback mechanism for performance regressions.

Topics

• LLM Agents
• Compiler Optimization
• Performance Tuning
• LLVM
• Multi-Agent Systems
• Architecture-Aware Optimization

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Software Engineer

Related on AIssential

• AutoPass: Evidence-Guided LLM Agents for Compiler Performance Tuning — AutoPass guides LLM agents with compiler and runtime evidence for superior, training-free compiler performance tuning.
• When AI Learns to Tune Itself: How ML Is Rewriting the Rules of Compiler Optimization — Published on June 11th, 2026, by Ankit Sinha, a Tech Lead at Google specializing in ML/AI infrastructure, this article examines the profound impact of machine learning on compiler optimization.
• Prior Knowledge or Search? A Study of LLM Agents in Hardware-Aware Code Optimization — LLM agents in code optimization primarily leverage pretrained priors over iterative feedback or agentic exploration.
• Outrunning LLM Cutoffs: A Live Kernel Crash Resolution Benchmark for All — LLM-based kernel bug fixing requires dynamic, contamination-free evaluation and standardized execution environments to accurately assess agent performance.
• Fine-tuning a Large Language Model (LLM) usually feels like a battle against CUDA out-of-memory errors and broken environments. Unsloth AI Releases Studio: A Local No-Code Interface For High-Performance LLM Fine-Tuning With 70% Less VRAM Usage..... — Unsloth AI has released Studio, an open-source, local no-code Web UI designed to simplify and accelerate high-performance Large Language Model (LLM) fine-tuning.
• ChatNeuroSim: An LLM Agent Framework for Automated Compute-in-Memory Accelerator Deployment and Optimization — ChatNeuroSim automates CIM accelerator design and optimization using LLM agents and design space pruning for faster, more efficient exploration.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.