XCheck tool uncovers 2,034 DL compiler bugs via cross-layer constraints

A team of researchers from UCLA and other institutions has introduced XCheck, an automated framework designed to uncover compiler-platform interaction bugs in deep learning (DL) compilers. Unlike existing testing approaches that primarily focus on type constraints and monitor compilation crashes or coverage gains, XCheck derives full-stack constraints that jointly guide model generation and characterize compilation behaviors. The key insight: bugs often stem from violated assumptions due to interactions across multiple compilation passes and hardware platforms. XCheck prioritizes constraints that expose interaction-sensitive behaviors, ensuring generated models exercise deep compilation logic. It also automatically inserts assertions to monitor distinct compilation symptoms that coverage or pass/fail signals miss.

Evaluated on TVM and ONNX-MLIR—two widely-used DL compilers—XCheck found 2,034 bug-revealing cases, including memory overflows, integer overflows, and silent unexpected compilations. These bugs are rooted in compiler-platform interactions and are often missed by traditional testing. The work addresses a critical gap: as AI deployment grows, ensuring compiler correctness across diverse hardware is essential. XCheck's approach is scalable and automated, moving beyond input-model generation to consider the entire compilation stack. The researchers believe this methodology could be extended to other compiler testing scenarios, improving reliability in production AI systems.