Dataset Card for IR-OptSet

IR-OptSet is a publicly available dataset designed to advance the use of large language models (LLMs) in compiler optimization. It leverages LLVM—one of the most widely adopted modern compilers—and focuses on its intermediate representation (IR) as the foundation of the dataset. IR-OptSet contains 170,000 IR samples curated from 1,704 GitHub repositories across diverse domains. The dataset provides a comprehensive resource for training and evaluating models in the field of IR-level compiler optimizations.

---

Dataset Details

Description

• Languages: LLVM Intermediate Representation (LLVM IR)
• Size: ~170,000 IR samples
• Effective Optimizations: >4.3 million annotations

Source

• Repositories: 1,704 open-source GitHub repositories
• Link: IR-OptSet on Hugging Face

---

Intended Uses

IR-OptSet is suitable for the following use cases:

1. IR Understanding: Train models to extract structural and semantic information from LLVM IR code.
2. IR Analysis: Evaluate model ability to understand and analyze basic properties of LLVM IR, such as dominator trees, loop structures, and memory dependencies.
3. Optimized Code Generation: Use LLMs to generate optimized IR from unoptimized input.

Out-of-Scope Uses

• Non-Compiler Tasks: The dataset is specifically tailored to IR-level optimization tasks and may not generalize well to unrelated domains.

---

Dataset Structure

Each record in IR-OptSet includes:

• original_ir: Unoptimized LLVM IR (via clang/clang++ -Xclang -disable-llvm-passes)
• preprocessed_ir: Cleaned version of the original IR
• o3_ir: Optimized IR generated via opt -passes='default<O3>'
• o3_active_passes: Effective passes during O3 compiling
• structural_hash: Unique structurehash of the IR, via opt -passes='print<structural-hash><detailed>
• repo_name: Source GitHub repository
• repo_file_path: File path within the repository
• function_name: IR function name
• repo_license: Associated license

Each .parquet file contains ~8,500 examples.

---

Dataset Creation

Collection and Preprocessing

• Selection: Repositories were manually filtered for compilation relevance and diversity across domains.
• Preprocessing:
  • Local variable renaming
  • Basic block renaming
  • Struct name normalization
  • Whitespace and comment removal
  • Insertion of missing basic block headers where needed

All transformations preserve the semantic and structural fidelity of the IR.

• Domains:
  • High-Performance Computing (HPC)
  • Machine Learning
  • Multimedia
  • Embedded Systems
  • System Software
  • Security
  • Reusable Libraries
  • Algorithms

Domain	Description	#Repos	#LLVM IR	#Total Eff. Opt.	Avg. Eff. Opt. Steps
High-Performance Computing (HPC)	Loop-intensive, memory-bound workloads; key targets for vectorization, parallelism, and memory locality optimizations.	275	17,145	399,110	23.28
Machine Learning	Compute-bound code; used for parallel execution.	95	9,366	249,467	26.64
Multimedia	SIMD-intensive, throughput-sensitive workloads; targets loop optimizations.	174	15,019	338,699	22.55
Embedded Systems	Resource-constrained (size, energy, real-time); compilers optimize for minimal footprint and energy efficiency.	108	5,942	129,449	21.79
System Software	Includes OS kernels, runtime systems, allocators; requires low-level control, safety, and instruction-level optimization.	93	6,581	136,688	20.77
Security	Demands high performance and strict constant-time behavior; requires careful register use, vectorization, and avoidance of side-channel vulnerabilities during optimization.	94	9,252	184,329	19.92
Reusable Libraries	Common code reused across domains; compilers apply inlining, and target-specific tuning.	106	7,664	157,162	20.51
Algorithms	Classical algorithms (e.g., sorting); compiler optimizations target computation and memory.	759	99,595	2,754,505	27.66
Total	-	1,704	170,564	4,349,409	Avg: 22.89

Citation

BibTeX

@misc{iropti2025,
  title={IR-Optset: An Optimization-Sensitive Dataset for Advancing LLM-Based IR Optimizer},
  author={Zi Yang, Lei Qiu, Fang Lyu, Ming Zhong, Zhilei Chai, Haojie Zhou, Huimin Cui, Xiaobing Feng},
  year={2025},
  url={https://huggingface.co/datasets/YangziResearch/IR-OptSet}
}

We also provide an open-source toolchain for building similar datasets. If you're interested in generating your own optimization corpus, feel free to use our tools. (https://github.com/yilingqinghan/IR-OptSet)