While model serving has unlocked unprecedented capabilities, the high cost of serving large-scale models continues to be a significant barrier to widespread accessibility and rapid innovation. Compiler optimizations have long driven substantial performance improvements, but existing compilers struggle with neural workloads due to the exponentially large and highly interdependent space of possible transformations. Although existing stochastic search techniques can be effective, they are often sample-inefficient and fail to leverage the structural context underlying compilation decisions. We set out to investigate the research question of whether reasoning with large language models (LLMs), without any retraining, can leverage the context-aware decision space of compiler optimization to significantly improve sample efficiency. To that end, we introduce a novel compilation framework (dubbed REASONING COMPILER) that formulates optimization as a sequential, context-aware decision process, guided by a large language model and structured Monte Carlo tree search (MCTS). The LLM acts as a proposal mechanism, suggesting hardware-aware transformations that reflect the current program state and accumulated performance feedback. Monte Carlo tree search (MCTS) incorporates the LLM-generated proposals to balance exploration and exploitation, facilitating structured, context-sensitive traversal of the expansive compiler optimization space. By achieving substantial speedups with markedly fewer samples than leading neural compilers, our approach demonstrates the potential of LLM-guided reasoning to transform the landscape of compiler optimization.
Industrial Control Systems (ICS) rely heavily on Programmable Logic Controllers (PLCs) to manage critical infrastructure, yet analyzing PLC executables remains challenging due to diverse proprietary compilers and limited access to source code. To bridge this gap, we introduce PLC-BEAD, a comprehensive dataset containing 2431 compiled binaries from 700+ PLC programs across four major industrial compilers (CoDeSys, GEB, OpenPLC-V2, OpenPLC-V3). This novel dataset uniquely pairs each binary with its original Structured Text source code and standardized functionality labels, enabling both binary-level and source-level analysis. We demonstrate the dataset's utility through PLCEmbed, a transformer-based framework for binary code analysis that achieves 93\% accuracy in compiler provenance identification and 42\% accuracy in fine-grained functionality classification across 22 industrial control categories. Through comprehensive ablation studies, we analyze how compiler optimization levels, code patterns, and class distributions influence model performance. We provide detailed documentation of the dataset creation process, labeling taxonomy, and benchmark protocols to ensure reproducibility. Both PLC-BEAD and PLCEmbed are released as open-source resources to foster research in PLC security, reverse engineering, and ICS forensics, establishing new baselines for data-driven approaches to industrial cybersecurity.
Large language models (LLMs) demonstrate remarkable capabilities but face deployment challenges due to their massive parameter counts. While existing compression techniques like pruning can reduce model size, it leads to significant accuracy degradation under high compression ratios. We present a novel perspective inspired by constant folding in compiler optimization. Our approach enables parameter reduction by treating activation functions in LLMs as linear functions. However, recent LLMs use complex non-linear activations like GELU that prevent direct application of this technique. We propose TARDIS, which enables optimization of LLMs with non-linear activations by partially approximating them with linear functions in frequently occurring input ranges. For outlier inputs, TARDIS employs an online predictor to dynamically fall back to original computations. Our experiments demonstrate that TARDIS achieves 80% parameter reduction in feed-forward networks, while significantly outperforming state-of-the-art pruning methods Wanda and RIA with up to 65% higher accuracy. In practical deployments for a 7B model, TARDIS achieves 1.6x end-to-end inference speedup when integrated with the vLLM serving system, and 1.4x speedup with the widely adopted HuggingFace implementation, while incurring only a 10.9% accuracy trade-off.
Reinforcement learning (RL) has emerged as a powerful paradigm for enhancing large language models (LLMs) in code generation and optimization. This survey systematically reviews RL-driven techniques across the code development lifecycle, from compiler-level optimizations and resource allocation strategies to end-to-end code synthesis frameworks. We first examine classical and modern RL algorithms -- spanning policy gradients, actor-critic methods, human-feedback alignment, and preference-based optimization -- and their adaptations to the unique challenges of code generation, such as sparse and delayed rewards. Next, we analyze key benchmarks, datasets, and evaluation metrics that drive progress in RL-augmented Code LLMs. Finally, we identify open problems, including the need for richer feedback sources, support for low-level and domain-specific languages, and methods to reduce computational overhead. By consolidating current insights and outlining future directions, this work aims to guide researchers and practitioners in leveraging RL to produce more robust, efficient, and human-aligned code generation systems.
Term rewriting plays a crucial role in software verification and compiler optimization. With dozens of highly parameterizable techniques developed to prove various system properties, automatic term rewriting tools work in an extensive parameter space. This complexity exceeds human capacity for parameter selection, motivating an investigation into automated strategy invention. In this paper, we focus on confluence, an important property of term rewrite systems, and apply machine learning to develop the first learning-guided automatic confluence prover. Moreover, we randomly generate a large dataset to analyze confluence for term rewrite systems. Our results focus on improving the state-of-the-art automatic confluence prover CSI: When equipped with our invented strategies, it surpasses its human-designed strategies both on the augmented dataset and on the original human-created benchmark dataset Cops, proving/disproving the confluence of several term rewrite systems for which no automated proofs were known before.
The real-world effectiveness of deep neural networks often depends on their latency, thereby necessitating optimization techniques that can reduce a model's inference time while preserving its performance. One popular approach is to sequentially rewrite the input computation graph into an equivalent but faster one by replacing individual subgraphs. This approach gives rise to the so-called phase-ordering problem in which the application of one rewrite rule can eliminate the possibility to apply an even better one later on. Recent work has shown that equality saturation, a technique from compiler optimization, can mitigate this issue by first building an intermediate representation (IR) that efficiently stores multiple optimized versions of the input program before extracting the best solution in a second step. In practice, however, memory constraints prevent the IR from capturing all optimized versions and thus reintroduce the phase-ordering problem in the construction phase. In this paper, we present a tensor graph rewriting approach that uses Monte Carlo tree search to build superior IRs by identifying the most promising rewrite rules. We also introduce a novel extraction algorithm that can provide fast and accurate runtime estimates of tensor programs represented in an IR. Our approach improves the inference speedup of neural networks by up to 11% compared to existing methods.
Large Language Models (LLMs) have demonstrated remarkable capabilities across a variety of software engineering and coding tasks. However, their application in the domain of code and compiler optimization remains underexplored. Training LLMs is resource-intensive, requiring substantial GPU hours and extensive data collection, which can be prohibitive. To address this gap, we introduce Meta Large Language Model Compiler (LLM Compiler), a suite of robust, openly available, pre-trained models specifically designed for code optimization tasks. Built on the foundation of Code Llama, LLM Compiler enhances the understanding of compiler intermediate representations (IRs), assembly language, and optimization techniques. The model has been trained on a vast corpus of 546 billion tokens of LLVM-IR and assembly code and has undergone instruction fine-tuning to interpret compiler behavior. LLM Compiler is released under a bespoke commercial license to allow wide reuse and is available in two sizes: 7 billion and 13 billion parameters. We also present fine-tuned versions of the model, demonstrating its enhanced capabilities in optimizing code size and disassembling from x86_64 and ARM assembly back into LLVM-IR. These achieve 77% of the optimising potential of an autotuning search, and 45% disassembly round trip (14% exact match). This release aims to provide a scalable, cost-effective foundation for further research and development in compiler optimization by both academic researchers and industry practitioners.
Bayesian optimization is a powerful method for automating tuning of compilers. The complex landscape of autotuning provides a myriad of rarely considered structural challenges for black-box optimizers, and the lack of standardized benchmarks has limited the study of Bayesian optimization within the domain. To address this, we present CATBench, a comprehensive benchmarking suite that captures the complexities of compiler autotuning, ranging from discrete, conditional, and permutation parameter types to known and unknown binary constraints, as well as both multi-fidelity and multi-objective evaluations. The benchmarks in CATBench span a range of machine learning-oriented computations, from tensor algebra to image processing and clustering, and uses state-of-the-art compilers, such as TACO and RISE/ELEVATE. CATBench offers a unified interface for evaluating Bayesian optimization algorithms, promoting reproducibility and innovation through an easy-to-use, fully containerized setup of both surrogate and real-world compiler optimization tasks. We validate CATBench on several state-of-the-art algorithms, revealing their strengths and weaknesses and demonstrating the suite's potential for advancing both Bayesian optimization and compiler autotuning research.
This article optimizes the inference performance of the Qwen-1.8B model by performing Int8 quantization, vectorizing some operators in llama.cpp, and modifying the compilation script to improve the compiler optimization level. On the Yitian 710 experimental platform, the prefill performance is increased by 1.6 times, the decoding performance is increased by 24 times, the memory usage is reduced to 1/5 of the original, and the accuracy loss is almost negligible.
Generative pre-trained transformers (GPT's) are a type of large language machine learning model that are unusually adept at producing novel, and coherent, natural language. In this study the ability of GPT models to generate novel and correct versions, and notably very insecure versions, of implementations of the cryptographic hash function SHA-1 is examined. The GPT models Llama-2-70b-chat-h, Mistral-7B-Instruct-v0.1, and zephyr-7b-alpha are used. The GPT models are prompted to re-write each function using a modified version of the localGPT framework and langchain to provide word embedding context of the full source code and header files to the model, resulting in over 150,000 function re-write GPT output text blocks, approximately 50,000 of which were able to be parsed as C code and subsequently compiled. The generated code is analyzed for being compilable, correctness of the algorithm, memory leaks, compiler optimization stability, and character distance to the reference implementation. Remarkably, several generated function variants have a high implementation security risk of being correct for some test vectors, but incorrect for other test vectors. Additionally, many function implementations were not correct to the reference algorithm of SHA-1, but produced hashes that have some of the basic characteristics of hash functions. Many of the function re-writes contained serious flaws such as memory leaks, integer overflows, out of bounds accesses, use of uninitialised values, and compiler optimization instability. Compiler optimization settings and SHA-256 hash checksums of the compiled binaries are used to cluster implementations that are equivalent but may not have identical syntax - using this clustering over 100,000 novel and correct versions of the SHA-1 codebase were generated where each component C function of the reference implementation is different from the original code.
This work studies a Reinforcement Learning (RL) problem in which we are given a set of trajectories collected with K baseline policies. Each of these policies can be quite suboptimal in isolation, and have strong performance in complementary parts of the state space. The goal is to learn a policy which performs as well as the best combination of baselines on the entire state space. We propose a simple imitation learning based algorithm, show a sample complexity bound on its accuracy and prove that the the algorithm is minimax optimal by showing a matching lower bound. Further, we apply the algorithm in the setting of machine learning guided compiler optimization to learn policies for inlining programs with the objective of creating a small binary. We demonstrate that we can learn a policy that outperforms an initial policy learned via standard RL through a few iterations of our approach.
We introduce a novel paradigm in compiler optimization powered by Large Language Models with compiler feedback to optimize the code size of LLVM assembly. The model takes unoptimized LLVM IR as input and produces optimized IR, the best optimization passes, and instruction counts of both unoptimized and optimized IRs. Then we compile the input with generated optimization passes and evaluate if the predicted instruction count is correct, generated IR is compilable, and corresponds to compiled code. We provide this feedback back to LLM and give it another chance to optimize code. This approach adds an extra 0.53% improvement over -Oz to the original model. Even though, adding more information with feedback seems intuitive, simple sampling techniques achieve much higher performance given 10 or more samples.
Binary code analysis is the foundation of crucial tasks in the security domain; thus building effective binary analysis techniques is more important than ever. Large language models (LLMs) although have brought impressive improvement to source code tasks, do not directly generalize to assembly code due to the unique challenges of assembly: (1) the low information density of assembly and (2) the diverse optimizations in assembly code. To overcome these challenges, this work proposes a hierarchical attention mechanism that builds attention summaries to capture the semantics more effectively and designs contrastive learning objectives to train LLMs to learn assembly optimization. Equipped with these techniques, this work develops Nova, a generative LLM for assembly code. Nova outperforms existing techniques on binary code decompilation by up to 14.84 -- 21.58% (absolute percentage point improvement) higher Pass@1 and Pass@10, and outperforms the latest binary code similarity detection techniques by up to 6.17% Recall@1, showing promising abilities on both assembly generation and understanding tasks.
Compiler correctness is crucial, as miscompilation can falsify program behaviors, leading to serious consequences. Fuzzing has been studied to uncover compiler defects. However, compiler fuzzing remains challenging: Existing arts focus on black- and grey-box fuzzing, which generates tests without sufficient understanding of internal compiler behaviors. Meanwhile, traditional white-box techniques, like symbolic execution, are computationally inapplicable to the giant codebase of compilers. Recent advances demonstrate that Large Language Models (LLMs) excel in code generation/understanding tasks. Nonetheless, guiding LLMs with compiler source-code information remains a missing piece of research in compiler testing. To this end, we propose WhiteFox, the first white-box compiler fuzzer using LLMs with source-code information to test compiler optimization, with a spotlight on detecting deep logic bugs in the deep learning (DL) compilers. WhiteFox adopts a multi-agent framework: an LLM-based analysis agent examines the low-level optimization source code and produces requirements on the high-level test programs that can trigger the optimization; an LLM-based generation agent produces test programs based on the summarized requirements. Additionally, optimization-triggering tests are used as feedback to enhance the generation on the fly. Our evaluation on the three most popular DL compilers (i.e., PyTorch Inductor, TensorFlow-XLA, and TensorFlow Lite) shows WhiteFox can generate high-quality test programs to exercise deep optimizations, practicing up to 8X more than state-of-the-art fuzzers. WhiteFox has found 101 bugs for the DL compilers, with 92 confirmed as previously unknown and 70 fixed. WhiteFox has been acknowledged by the PyTorch team and is being incorporated into its development workflow. Beyond DL compilers, WhiteFox can also be adapted for compilers in different domains.
We explore the novel application of Large Language Models to code optimization. We present a 7B-parameter transformer model trained from scratch to optimize LLVM assembly for code size. The model takes as input unoptimized assembly and outputs a list of compiler options to best optimize the program. Crucially, during training, we ask the model to predict the instruction counts before and after optimization, and the optimized code itself. These auxiliary learning tasks significantly improve the optimization performance of the model and improve the model's depth of understanding. We evaluate on a large suite of test programs. Our approach achieves a 3.0% improvement in reducing instruction counts over the compiler, outperforming two state-of-the-art baselines that require thousands of compilations. Furthermore, the model shows surprisingly strong code reasoning abilities, generating compilable code 91% of the time and perfectly emulating the output of the compiler 70% of the time.
An important challenge in Machine Learning compilers like XLA is multi-pass optimization and analysis. There has been recent interest chiefly in XLA target-dependent optimization on the graph-level, subgraph-level, and kernel-level phases. We specifically focus on target-independent optimization XLA HLO pass ordering: our approach aims at finding the optimal sequence of compiler optimization passes, which is decoupled from target-dependent optimization. However, there is little domain specific study in pass ordering for XLA HLO. To this end, we propose introducing deep Reinforcement Learning (RL) based search for optimal XLA HLO pass ordering. We also propose enhancements to the deep RL algorithms to further improve optimal search performance and open the research direction for domain-specific guidance for RL. We create an XLA Gym experimentation framework as a tool to enable RL algorithms to interact with the compiler for passing optimizations and thereby train agents. Overall, in our experimentation we observe an average of $13.3\%$ improvement in operation count reduction on a benchmark of GPT-2 training graphs and $10.4\%$ improvement on a diverse benchmark including GPT-2, BERT, and ResNet graphs using the proposed approach over the compiler's default phase ordering.
Many useful tasks in data science and machine learning applications can be written as simple variations of matrix multiplication. However, users have difficulty performing such tasks as existing matrix/vector libraries support only a limited class of computations hand-tuned for each unique hardware platform. Users can alternatively write the task as a simple nested loop but current compilers are not sophisticated enough to generate fast code for the task written in this way. To address these issues, we extend an open-source compiler to recognize and optimize these matrix multiplication-like tasks. Our framework, called Amulet, uses both database-style and compiler optimization techniques to generate fast code tailored to its execution environment. We show through experiments that Amulet achieves speedups on a variety of matrix multiplication-like tasks compared to existing compilers. For large matrices Amulet typically performs within 15% of hand-tuned matrix multiplication libraries, while handling a much broader class of computations.
Although code generation for Convolution Neural Network (CNN) models has been extensively studied, performing efficient data slicing and parallelization for highly-constrai\-ned Multicore Neural Processor Units (NPUs) is still a challenging problem. Given the size of convolutions' input/output tensors and the small footprint of NPU on-chip memories, minimizing memory transactions while maximizing parallelism and MAC utilization are central to any effective solution. This paper proposes a TensorFlow XLA/LLVM compiler optimization pass for Multicore NPUs, called Tensor Slicing Optimization (TSO), which: (a) maximizes convolution parallelism and memory usage across NPU cores; and (b) reduces data transfers between host and NPU on-chip memories by using DRAM memory burst time estimates to guide tensor slicing. To evaluate the proposed approach, a set of experiments was performed using the NeuroMorphic Processor (NMP), a multicore NPU containing 32 RISC-V cores extended with novel CNN instructions. Experimental results show that TSO is capable of identifying the best tensor slicing that minimizes execution time for a set of CNN models. Speed-ups of up to 21.7\% result when comparing the TSO burst-based technique to a no-burst data slicing approach. To validate the generality of the TSO approach, the algorithm was also ported to the Glow Machine Learning framework. The performance of the models were measured on both Glow and TensorFlow XLA/LLVM compilers, revealing similar results.
Finding the best way to schedule operations in a computation graph is a classical NP-hard problem which is central to compiler optimization. However, evaluating the goodness of a schedule on the target hardware can be very time-consuming. Traditional approaches as well as previous machine learning ones typically optimize proxy metrics, which are fast to evaluate but can lead to bad schedules when tested on the target hardware. In this work, we propose a new approach to scheduling by sampling proportionally to the proxy metric using a novel GFlowNet method. We introduce a technique to control the trade-off between diversity and goodness of the proposed schedules at inference time and demonstrate empirically that the pure optimization baselines can lead to subpar performance with respect to our approach when tested on a target model. Furthermore, we show that conditioning the GFlowNet on the computation graph enables generalization to unseen scheduling problems for both synthetic and real-world compiler datasets.
Finding the optimal pass sequence of compilation can lead to a significant reduction in program size and/or improvement in program efficiency. Prior works on compilation pass ordering have two major drawbacks. They either require an excessive budget (in terms of compilation steps) at compile time or fail to generalize to unseen programs. In this paper, for code-size reduction tasks, we propose a novel pipeline to find program-dependent pass sequences within 45 compilation calls. It first identifies a coreset of 50 pass sequences via greedy optimization of a submodular function, and then learns a policy with Graph Neural Network (GNN) to pick the optimal sequence by predicting the normalized values of the pass sequences in the coreset. Despite its simplicity, our pipeline outperforms the default -Oz flag by an average of 4.7% over a large collection (4683) of unseen code repositories from diverse domains across 14 datasets. In comparison, previous approaches like reinforcement learning on the raw pass sequence space may take days to train due to sparse reward, and may not generalize well in held-out ones from different domains. Our results demonstrate that existing human-designed compiler flags can be improved with a simple yet effective technique that transforms the raw action space into a small one with denser rewards.