FLUX: Finding Bugs with LLVM IR Based Unit Test Crossovers
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Optimizing compilers are as ubiquitous as they are crucial to software development. However, bugs in compilers are not uncommon. Among the most serious are bugs in compiler optimizations, which can cause unexpected behavior in compiled binaries. Existing approaches for detecting such bugs have focused on end-to-end compiler fuzzing, which limits their ability for targeted exploration of a compiler's optimizations. This paper proposes FLUX (Finding bugs with LLVM IR based Unit test cross(X)overs), a fuzzer that is designed to generate test cases that stress compiler optimizations. Previous compiler fuzzers are overly constrained by having to construct well-formed inputs. FLUX sidesteps this constraint by using human-written unit test suites as a starting point, and then selecting random combinations of them to generate new tests. We hypothesize that tests generated this way will be able to explore new execution paths through compiler optimizations and find new bugs. Our evaluation of FLUX on LLVM indicates that it is able to increase path coverage over the baseline LLVM unit test suite and explores more edge coverage than previous work. Further, we demonstrate FLUX's ability to generate miscompiled and crash-producing IR on LLVM's optimizations. After a month of fuzzing, FLUX found 28 unique bugs in LLVM's active development branch. We have reported 11 of these bugs which led to 6 of them being patched by LLVM developers. 22 of these are crashes that are triggered by well-formed input programs, and 6 of these are miscompilation bugs that silently produced incorrect code.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.001 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it