Source linked

XCheck Finds 2,034 Compiler Bugs by Catching Platform Interactions That Fuzzing Misses

arxiv.org@systems_wire3 hours ago·Artificial Intelligence·0 comments

A new tool based on cross-layer constraints uncovered memory overflows, integer overflows, and silent miscompilations in TVM and ONNX-MLIR that traditional input generation approaches overlooked.

xchecktvmonnx mlirdeep learning compilerscompiler testingsoftware verification

2,034 bug-revealing cases across three widely-used DL compilers. That's what XCheck found, and it didn't get there by fuzzing random models.

Why Existing Testing Misses the Real Bugs

Most DL compiler testing tools constrain input generation with type rules, then watch for crashes or coverage gains. They overlook the bugs that bite you when you actually deploy: compiler-platform interaction bugs that arise from interleaved effects across compilation passes and execution environments. TVM and ONNX-MLIR both suffer from this blind spot.

Full-Stack Constraints and Assertions

XCheck's key insight: these bugs stem from violated assumptions that span compilation passes and hardware platforms. Instead of constraining just input generation, XCheck extracts full-stack constraints that jointly guide model generation and characterize compilation behaviors. It prioritizes constraints that expose interaction-sensitive behaviors, so generated models exercise deep compilation logic. Then it automatically inserts assertions to monitor symptoms that coverage or pass/fail signals never catch.

The Tally: Memory Overflows, Integer Overflows, Silent Failures

XCheck found 2,034 bug-revealing cases across three compilers. That includes memory overflows, integer overflows, and silent unexpected compilations - cases where the compiler produced wrong output without crashing. Every bug was rooted in compiler-platform interactions, not type mismatches or simple codegen errors.

I've spent years annoyed by testing tools that claim to find deep bugs but only surface crash dumps. XCheck's results suggest the real problem wasn't missing test coverage - it was missing the cross-layer constraints that define correct behavior.

If you're shipping models through TVM or ONNX-MLIR, XCheck's approach gives you a concrete reason to start hunting for similar latent bugs in your own pipeline.

Source: Finding Compiler-Platform Interaction Bugs in Deep Learning Pipelines via Cross-Layer Constraints
Domain: arxiv.org

Read original source ->

External source stays available while the OJO article and comment thread stay local.

More in Artificial Intelligence

view topic
Budget-Aware Adaptive Patches Expose Query-Visibility Tradeoffs in Black-Box Object Detection

New attack method simultaneously optimizes patch location, texture, and size while adapting to limited query budgets, achieving strong suppression on YOLOv5 and Faster R-CNN with minimal visual footprint.

RegimeVGGT Cuts Cross-Frame Attention Cost 6.7x With Regime-Aware Compression

A training-free acceleration method for VGGT identifies three distinct attention regimes and applies U-shaped compression to achieve 6.7x speedup without quality loss.

CaVe-VLM-CoT: Agentic RAG Pipeline Hits 87% on ScienceQA by Routing Verification Failures

CaVe-VLM-CoT detects ungrounded claims and triggers re-retrieval, achieving 87.1% accuracy on ScienceQA while introducing CaVeScore for measuring citation faithfulness.

PROPEL Doubles Useful Training Tasks by Predicting Solver Pass Rate in One Forward Pass

Training a single software-engineering task candidate can take tens of minutes; PROPEL replaces costly solver rollouts with a lightweight probe, boosting learnable-frontier tasks from 10.1% to 20.0% for a 3B coding...

Comments load interactively on the live page.