CC '26: Proceedings of the 35th ACM SIGPLAN International Conference on Compiler Construction

Optimizing compilers often sacrifice binary size in pursuit of higher run-time performance. In the absence of method execution profiles, they uniformly apply performance-oriented optimizations, typically various forms of code duplication. Duplications in methods that are rarely or never executed only increase binary size without improving performance. Modern static profiler use ML to predict branch profiles, yet they do not identify which methods will be frequently executed at run time. Doing so would enable more selective optimizations, reducing binary size while preserving or only minimally affecting run-time performance.

We present GraalMHC, a machine–learning–based static profiler that predicts method hotness. GraalMHC uses the XGBoost ensemble to classify methods as cold and warm. For cold methods, GraalMHC enables code-size-reducing optimizations, and for warm methods, it enables performance-improving optimizations. In this way, GraalMHC enables binary-size reductions with no or minimal impact on run-time performance. In addition, GraalMHC allows users to choose between three different size-optimization levels: (S1) 9–13% binary-size reduction with 1–2% performance loss, (S2) 15–25% reduction with 3–5% performance loss, and (S3) 17–35% reduction with 5–7% performance loss. We integrate GraalMHC into the Oracle GraalVM Native Image compiler, delivering a complete end-to-end solution.

The rise of asynchronous execution and specialized concurrent thread groups has reshaped GPU programming, but it also introduces complex timing and coordination challenges. Developers must carefully manage data readiness, concurrency, and hardware-specific tensor units that current low-level primitives make error-prone and hardware-dependent.

We present Async Graphene, a set of high-level abstractions for asynchronous and concurrent GPU programming. It builds on three pillars: (1) Async Tensor Types, which encode temporal relationships in the type system to prevent synchronization errors; (2) Concurrency Primitives, which structure complex control flow through specialization and pipelining; and (3) Tensor-Unit Orchestration, which simplifies low-level tensor hardware management.

On critical workloads such as GEMM and FlashAttention, Async Graphene delivers performance competitive with hand-optimized frameworks while significantly reducing development complexity. By making temporal behavior explicit, it allows compilers to address asynchrony and concurrency challenges without obscuring performance-critical details.

Sparse tensor algebra plays an important role in many scientific and engineering applications, yet existing sparse libraries and compilers face challenges when the output tensor is sparse. Array-based storage formats, such as CSR, require costly memory reallocations and rely on intermediate tensors (workspaces) to handle sparse scattering into the output, which limits performance and scalability. We introduce a new approach that employs our proposed flexible map-based storage format to directly support sparse scattering into the output without requiring extra workspaces. Our system then applies code and storage-specific optimizations to maximize efficiency. Experimental results across a range of kernels and datasets demonstrate an average speedup of 8.06× over a state-of-the-art compiler and 5.28× over a sparse tensor library.

Compilers apply optimizations such as function specialization and constant propagation to eliminate redundant work at compile time. However, because compilers must prove that values are constant, many profitable optimization opportunities remain unrealized. In this paper, we propose run-time invariant function specialization (RIFS), a profile-guided compiler technique that specializes functions based on runtime invariant call-site-specific argument values. RIFS introduces a novel value-profiling LLVM pass to identify runtime invariant arguments, even though they cannot be proven constant statically. A subsequent LLVM transformation pass generates specialized function variants tailored to these value profiles. To efficiently select among potentially thousands of specialization candidates, we develop a predictive cost model that estimates the performance benefit of each candidate prior to code generation. We integrate our passes seamlessly into the existing PGO-enabled LLVM toolchain. We evaluate RIFS across 11 real-world applications, demonstrating substantial improvements over state-of-the-art optimization techniques. RIFS achieves an average speedup of 6.3% and an instruction reduction of 2.5% over the LLVM -O3+PGO baseline.

The ONNX Optimizer, part of the official ONNX repository and widely adopted for graph-level model optimizations, is used by default to optimize ONNX models. Despite its popularity, its ability to preserve model correctness has not been systematically evaluated. We present DiTOX, an automated framework for comprehensively assessing the correctness of the ONNX Optimizer using differential testing, fault localization, and evaluation techniques that generalize to other compiler optimizers. DiTOX applies optimization passes to a corpus of ONNX models, executes both original and optimized versions on user-defined inputs, and detects discrepancies in behavior or optimizer failures. When divergences are observed, DiTOX isolates the responsible optimization pass through iterative, fine-grained analysis. We evaluated DiTOX on 130 models from the ONNX Model Hub spanning vision and language tasks. We found that 9.2% of model instances crashed the optimizer or produced invalid models under default settings. Moreover, output discrepancies occurred in 30% of classification models and 16.6% of object detection and segmentation models, while text-based models were largely robust. Overall, DiTOX uncovered 15 issues—14 previously unknown—affecting 9 of the 47 optimization passes as well as the optimizer infrastructure. All issues were reported to the ONNX Optimizer developers. Our results demonstrate that DiTOX provides a simple and effective approach for validating AI model optimizers and is readily extensible beyond ONNX.

Transient execution attacks exploit speculative execution to leak confidential data through unauthorized transient memory accesses. We make the observation that transient attacks can be identified by one unusual memory access, the transient sensitive data access. To protect systems from such attacks while minimizing performance overhead, we propose leveraging compile-time information to identify memory operations that cannot extract sensitive data and can therefore be deemed safe. Safe memory operations are allowed to execute transiently, causing no extra performance cost. Unsafe memory operations delay accessing the memory system until they are no longer in a speculative state, preventing unauthorized transient accesses to sensitive data. To communicate this information to the microarchitecture, we introduce the set safe memory region (ssmr) instruction. Inserted automatically by the compiler, it establishes the memory regions that may be accessed transiently by a sequence of instructions. This defense incurs only a 7% performance overhead compared to the insecure baseline and mitigates at least two variants of transient execution attacks.

Fully Homomorphic Encryption (FHE) enables computations to be performed directly on encrypted data without requiring decryption, providing strong privacy guarantees. However, FHE remains computationally expensive, and writing efficient FHE programs is a complex, error-prone, and time-consuming task that demands significant cryptographic expertise. Programmers are often unaware of available optimizations, and applying them manually requires substantial effort. In this paper, we present CHEHAB, a compiler that automatically vectorizes scalar code, optimizes it, and generates highly efficient FHE programs. CHEHAB supports both structured and unstructured code and takes as input programs written in a domain-specific language embedded in C++. It relies on a Term Rewriting System (TRS) based on equality saturation to simplify and transform programs. CHEHAB targets two key challenges in FHE compilation: (1) automatic vectorization of scalar code, and (2) reduction of instruction execution latency and ciphertext noise growth. By leveraging equality saturation, CHEHAB explores a large optimization space to reduce instruction count and circuit depth while improving vector utilization. Experimental evaluation on a set of representative kernels shows that CHEHAB outperforms Coyote, a state-of-the-art vectorizing compiler for FHE. On average, CHEHAB generates code that is 7.38× faster at runtime, incurs 2.49× less accumulated noise, and achieves 251× faster compilation time. CHEHAB is released as an open-source compiler to support reproducibility and further research in FHE compilation.

Equality saturation enables compilers to explore many semantically equivalent program variants, deferring optimization decisions to a final extraction phase. However, existing frameworks exhibit sequential execution and hard-coded saturation loops. This limits scalability and requires significant engineering effort to customize saturation behavior.

This paper addresses these limitations using three novel techniques. First, it shows how saturation can be parallelized thanks to the use of thread-safe data structures and the notion of deferred e-graph updates. Second, it provides an extensible mechanism to express custom and composable saturation strategies. Third, it generalizes e-graph metadata to support custom e-graph annotations.

The implementation, written in Scala, is evaluated on four use-cases: classical program optimization, idiom recognition, scalability strategies and incremental equality saturation. The results show that it outperforms several existing equality saturation engines, including the highly optimized egglog library. When used to reimplement an existing idiom recognition technique, the new design finds higher-quality idioms, 16× faster. Additionally, the design is able to natively express state-of-the-art custom equality saturation behavior such as incremental equality saturation and multi-phase rewriting strategies without any modification to the core library.

Linear algebra libraries and tensor domain-specific languages are able to deliver high performance for modern scientific and machine learning workloads. While there has been recent work in automatically translating legacy software to use these libraries/DSLs using pattern matching and program lifting, this has been largely limited to dense linear algebra.

This paper tackles the more challenging problem of porting legacy sparse linear algebra code to libraries and DSLs. It exploits the power of large language models to predict a sketch of the solution and then uses type-based program synthesis to search the space of possible code to target parameter bindings. We implement this in a tool named SLEB (Sparse LiftEr with Binding) and evaluate it on a large set of benchmarks and real-world datasets, comparing against two state-of-the-art compiler techniques, LiLAC and SpEQ; and GPT 4.o. Overall, we lift 94% of programs compared to 11%, 17%, and 48% for LiLAC, SpEQ, and GPT 4.o respectively. This delivers a geomean speedup of 2.6x and 7.2x on a CPU and GPU platform, respectively.

This paper introduces Schedgehammer, a general-purpose auto-scheduling framework that optimizes program execution across diverse compiler infrastructures. Unlike existing auto-schedulers that are tightly coupled to specific intermediate representations or rely on template-based search, Schedgehammer provides a generic, reusable framework for optimization schedules by modeling them as graph-structured objects. This approach captures dependencies among transformations and parameters across compilers, enabling systematic mutation and validation.

We demonstrate Schedgehammer’s flexibility on TVM and TACO, showing that it effectively optimizes dense and sparse computations. Across benchmarks, it achieves performance comparable to specialized auto-schedulers such as Ansor, highlighting that a unified, extensible abstraction can generalize scheduling beyond individual compiler ecosystems.

Tiny machine learning (TinyML) enables low-power microcontrollers to leverage the power of artificial intelligence without relying on remote computing resources. Typically, developing a TinyML application relies primarily on existing TinyML frameworks, which provide runtime APIs for loading and executing machine learning models. However, such framework-based TinyML development has limitations in terms of portability, programmability, and resource efficiency, motivating the need for a new approach to the TinyML development and deployment process.

To address these challenges, this work introduces TinyGen, a new code generation framework for TinyML. TinyGen generates portable high-level code directly from a target model without depending on external runtime APIs. It statically analyzes the tensor and operator usage of the target model to enable compact code generation. This work demonstrates that TinyGen reduces binary code size by 31.2% on average compared to an existing TinyML framework while requiring fewer lines of code to write TinyML applications.

Software testing is a critical stage in the software development lifecycle, ensuring that programs behave correctly and reliably across diverse environments. Unlike general software testing, compiler testing requires well-structured input programs that systematically exercise the compiler’s various optimization passes, since incomplete coverage can easily mask critical issues. Currently, a range of techniques exists for compiler testing, including differential testing, fuzzing, slicing, sanitizers and formal methods based validation tools, While these methods are effective at detecting compiler crashes, undefined behavior, or general correctness bugs, they are largely inadequate for systematically evaluating compiler optimizations. Tools like Csmith have pioneered random program generation for compiler testing, but they lack precise control over program structure, which is critical when the goal is to stress specific optimization passes or identify performance regressions. To address this limitation, we developed CPerf- Smith , an extension of Csmith that provides fine-grained control over code constructs, enabling targeted generation of programs that exercise particular compiler optimizations. Through our experiments, CPerfSmith was able to generate over 100 programs exhibiting significant runtime differences across industry-standard C compilers (GCC, Clang, and AOCC). Furthermore, we successfully created more than 30 test cases that revealed significant performance regressions between two subsequent release versions of Clang, demonstrating the tool’s ability to detect missing or improperly applied optimizations.

Recent efforts in open-source GPU research are opening new avenues in a domain that has long been tightly coupled with a few commercial vendors. Emerging open GPU architectures define SIMT functionality through their own ISAs, but executing existing GPU programs and optimizing performance on these ISAs relies on a compiler framework that is technically complex and often undercounted in open-hardware development costs.

To address this challenge, the Vortex-Optimized Lightweight Toolchain (VOLT) has been proposed. This paper presents its design principles, overall structure, and the key compiler transformations required to support SIMT execution on Vortex. VOLT enables SIMT code generation and optimization across multiple levels of abstraction through a hierarchical design that accommodates diverse front-end languages and open GPU hardware. To ensure extensibility as GPU architectures evolve, VOLT centralizes fundamental SIMT-related analyses and optimizations in the middle-end, allowing them to be reused across front-ends and easily adapted to emerging open-GPU variants. Through two case studies on ISA extensions and host-runtime API, this paper also demonstrates how VOLT can support extensions.

The proliferation of Python DSLs for developing kernels has democratized GPU programming. While kernel development is now Python-native, performance analysis and optimization still rely on external tools and fragmented workflows.

We introduce Nsight Python, a Python profiling toolkit that bridges this gap by bringing performance analysis for GPU kernels into the Python ecosystem. Nsight Python is framework-agnostic and works seamlessly with any Python framework through simple decorators and context managers that mark code regions for analysis. The tool automatically handles performance measurement, data collection, and result aggregation across multiple configurations.

Key innovations include: (1) Python-native interface enabling direct GPU kernel analysis within existing Python workflows, (2) configuration-driven execution model that correlates measured metrics with input parameters, (3) access to thousands of core metrics and enabling custom derived and relative metric computations through parameter correlation, (4) built-in thermal management to prevent GPU throttling during batch analysis, and (5) structured customizable visualization with automatic plot generation based on parameter roles.

With Nsight Python, the entire kernel development process—from implementation to iterative optimization and analysis—is now truly Python-native, providing detailed insights without requiring external profilers or disrupting workflows.

Alias analysis is a technique to identify whether a memory location can be accessed in more than one way. An ideal alias analysis implementation should be both precise and scalable. However, in practice, implementations of alias analysis have to make a trade-off between precision and scalability. The alias analysis implementations that perform inter-procedural analysis are more precise (answer a higher number of alias queries with certainty), but expensive, making them infeasible for practical use. Most compiler developers opt for intra-procedural analysis over inter-procedural, thereby compromising the precision of alias analysis implementations to achieve scalability. For compilers, this compromise leads to a loss in optimization opportunities, limiting the performance achievable by the compiled program.

In this work, we present HORIZON, a tool that estimates the upper bound on alias analysis precision improvement and its impact on program performance, enabling an understanding of how compromised precision affects execution. HORIZON implements a profiling-based approach to gather precise alias information (must- and no-alias) missed by existing alias implementations and provides this information to the compiler to estimate potential performance gains. Integrated with LLVM, we applied HORIZON to evaluate the default LLVM alias analysis on standard benchmarks, demonstrating that precision improvement could range from 0% to 33%, with corresponding performance impact between 0% and 53%, indicating the effectiveness of our approach and tool.

With version 17, LLVM finalized the transition to opaque pointer types, eliminating explicit pointee‑type information from the Intermediate Representation (IR). Thus, starting from LLVM 17, each pointer type is represented in IR by the unique type ptr. Despite eliminating redundant pointer bitcasts and consequently reducing IR size and compile time, this change disrupts analyses that have reason to rely on pointee-type information, forcing existing compiler projects to depend on outdated LLVM versions. This information can in fact be insightful in fields like approximate computing, where the compiler can apply non-conservative optimizations, or in passes that require it to make analyses and transformations that do not impact the correctness of the program. To address this problem, we present a new Type Deduction Analysis pass that reconstructs transparent pointer types directly from opaque‑pointer IR. Moreover, we illustrate two different case-studies on existing LLVM projects, namely TAFFO and ASPIS, that demonstrate the need for pointee-type information in LLVM compilers.

Constraint-based points-to analysis using Andersen-style inclusion constraints is widely used for its convenience, generality, and precision in modeling complex program behaviors. Typically, such analyses generate constraints and resolve them by computing the transitive closure of a constraint graph. However, traditional constraint modeling often introduces redundancy during constraint generation, solving, or both. In the context of points-to analysis for object-oriented languages like Java, this redundancy primarily stems from the modeling of virtual calls and heap operations. As a result, the analysis produces redundant constraints and inflated constraint graphs, thereby increasing analysis time.

To address these limitations, we propose a novel constraint representation and solving system PInter, which extends the traditional inclusion constraint model. It presents a novel technique to represent the constraints in compact and expressive way. Further, it defers generation of constraints until relevant objects are discovered, enabling demand-driven and interleaved constraint generation and solving. We present a proof of correctness of PInter. We used PInter to implement two different Java points-to analyses, within the Soot framework, and evaluated it on 12 applications drawn from the DaCapo benchmark suite. As is standard, we used Tamiflex to handle dynamic features of Java benchmarks and performed a soundy evaluation. Our results show that, compared to traditional methods, PInter reduced the constraint count by 96% (geomean) and the analysis time by 78% (geomean). We also evaluated PInter against Soot’s Spark and the flow-, context- insensitive analysis in Doop, and found that PInter achieved significantly faster performance.

May happen in Parallel (MHP) analysis is one of the most foundational analysis in the context of programs written in parallel languages like Java. The currently known techniques for doing MHP analysis of Java applications suffer from two main challenges: (i) scalability to real-world large applications, (ii) precision in the presence of complex programs. In this manuscript, we address these two issues with a goal of making MHP analysis for Java applications a practical option. We propose a new MHP analysis scheme called GRIP-MHP. It includes techniques to reduce time taken to perform MHP analysis significantly; it does so by using novel schemes to reduce the size of the input graph (representing the original application). GRIP-MHP also addresses many drawbacks in existing techniques, thereby improving the applicability and precision of MHP analysis for real-world Java applications. We implemented GRIP-MHP in the Soot compiler framework. Our experiments show that GRIP-MHP runs successfully in reasonable time on all the tested benchmarks in DaCapo and Renaissance (geomean 20.176×improvement, over prior work). We find that GRIP-MHP performs more precise joining of threads significant leading to improvements in precision of the analysis results: geomean, 1.85×reduction in MHP pairs, over prior work.