Pith. sign in

REVIEW 3 major objections 5 minor 34 references

BMOA: Baseline-Mechanism-Outcome Attribution for Compiler-Induced Numerical Deviations

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read Compiler-induced floating-point differences cannot be judged by one pass/fail flag: each difference needs a separate attribution of baseline meaning, supported mechanism, and reference-relative accuracy outcome.

desk verdict The B/M/O separation is a genuine contribution; the headline empirical claim is reference-dependent, but the paper is honest about it and deserves a serious referee. read the letter →

arxiv 2607.27270 v1 pith:NXS7MBFY submitted 2026-07-29 cs.PL

classification cs.PL
keywords floatingpointcompilertestingnumericalaccuracyattributioncompiler-induceddeviationBMOAreproducibilityreference-relativeoutcomes
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that when a compiler produces a different floating-point result, "does it match?" and "is it more accurate?" are different questions that must be answered separately. It introduces BMOA, a diagnostic framework that labels every observed difference with a baseline (which comparison gives it meaning), a mechanism (which compiler behavior the evidence supports), and an outcome (improvement, degradation, or no change relative to a named higher-precision reference). Across 1,276 baseline-specific records and 162 controlled compiler configurations on six scientific kernels, BMOA finds cases where a result deviates from strict floating-point yet is closer to the reference, and other cases where the same mechanism improves one input and degrades another. The point of the framework is to make the evidence behind each label explicit and auditable, so that a pass/fail mismatch stops being mistaken for an accuracy verdict.

What carries the argument

The central object is the ⟨B,M,O⟩ attribution record. B is one of five named baselines—strict floating point, numerical reference, mechanism-local, reproducibility, or across-compiler—that fixes what a difference means. M is a mechanism category such as reassociation, fused multiply-add (FMA) contraction, or compound fast-math; it is marked supported only when a controlled pair D_m(x) exceeds the threshold and code-generation evidence is consistent, otherwise mixed, ambiguous, or unknown. O is the outcome—improvement, degradation, neutral, or unknown—obtained by ordering the errors of candidate and strict configurations against the same higher-precision reference. The mechanism-local compari

What would settle it

Find a kernel-input family where, across all reasonable higher-precision references (different precisions, rounding modes, interval bounds), the sign of the strict-FP deviation always predicts the sign of the reference-error difference (e.g., y_c > y_s always means E_c > E_s). Such a class would collapse baseline and outcome, refuting the paper's claim that none of the three record components follows from the other two; the paper's own 20/162 reference disagreements predict no such class exists.

Watch

Extended reading notes

Core claim

BMOA's central claim is that an observed numerical difference between two compiler configurations is not a single fact but a conjunction of three independent facts. The comparison that gives the difference its meaning (the baseline), the compiler behavior supported by controlled evidence (the mechanism), and the numerical consequence measured against a named reference (the outcome) must be recorded separately, because none of the three components follows from the other two. The framework emits a ⟨B,M,O⟩ record per evaluated comparison, supports a mechanism category only when a controlled paired ablation changes the output and assembly evidence is consistent, and preserves mixed, ambiguous, a

Load-bearing premise

Accuracy outcome labels depend on the chosen higher-precision reference (Decimal with 90 digits over the exact represented float32 inputs) correctly representing the intended real arithmetic of the source expression; if a different reference is semantically right for a kernel, every improvement/degradation label can flip.

Editorial extensions

If this is right

  • A pass/fail mismatch is not a diagnosis: the same candidate result can carry a strict-FP deviation, a supported mechanism, and an improvement outcome simultaneously.
  • Compiler-induced deviation does not imply accuracy loss; the controlled matrix found 16 improvements and 4 degradations, and all three FMA-contraction deviations improved reference accuracy.
  • Mechanism attribution must be evidence-bounded: a vector or FMA instruction in the assembly does not establish that it caused the observed difference; the controlled matrix found vector instructions in 102/108 assembly files but no vectorization-enabled output deviations.
  • Outcome labels are reference-relative: Decimal and a stable float32 algorithm disagreed on 20/162 instances, always in the same direction, so any accuracy claim must name its reference.
  • By retaining inputs, bit patterns, metrics, and provenance, BMOA records give later formal specifications concrete comparison relations and proof obligations to target.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension is to apply the same B,M,O schema to GPU kernels and parallel reductions, where the reproducibility baseline would separate run-to-run variance from compiler causality—a distinction the current single-CPU, single-seed study cannot probe.
  • The 20/162 reference disagreements imply that reporting any single accuracy number without its reference and tolerance is under-specified; a practical takeaway is that accuracy claims should become "improvement relative to X under tolerance T" rather than bare percentages.
  • If BMOA records were emitted automatically by testing harnesses, compiler bug reports could carry replayable evidence bundles, letting maintainers distinguish "semantic violation" from "different rounding order but better result" before deciding whether a change is a defect.
  • The framework treats strict FP as one comparator among several, not as the truth; a testable consequence is that replacing strict FP with another operational configuration (e.g., a different rounding mode) would change B but should leave the outcome ordering unchanged if the reference is fixed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. BMOA formalizes post-detection diagnosis for compiler-induced floating-point deviations as <B,M,O> records: baseline defines the comparison relation and system boundary, mechanism identifies the compiler behavior category supported by controlled paired ablations, and outcome reports the reference-qualified accuracy result. The evaluation uses strict-FP, mechanism-local, reproducibility, cross-compiler, and higher-precision comparisons over six small kernels, nine configurations, and two Clang distributions on ARM64. The empirical components are a 1,276-record attribution corpus and a 162-instance controlled mechanism matrix. The main empirical findings are that baseline choice changes diagnosis; among 20 distinct nonzero deviations, 16 improve and 4 degrade relative to Decimal; cancellation and large dynamic range dominate the effects; and repeated runs and Apple/Homebrew comparisons are bit-identical. BMOA is explicitly not a proof system; it produces auditable, evidence-bounded records intended as a basis for future formal specifications.

Significance. The paper contributes a genuinely useful separation of concerns: pass/fail mismatch is inadequate as a diagnostic, and distinguishing comparison meaning, evidence-supported mechanism, and reference-relative outcome is valuable. The evaluation is careful in several respects: exact-bit comparisons, serialized inputs reused across configurations, repeated runs, Decimal–NumPy agreement on all 162 instances, identical Apple/Homebrew outcomes, explicit negative controls, and honest statements of scope. The central empirical claim that strict-FP deviation does not imply accuracy loss is, however, reference-conditional. The paper's own §5.4 reports 20 Decimal–stable-reference disagreements, with every disagreement being a Decimal improvement that the stable float32 reference labels as a degradation. Thus the headline 16-improvement/4-degradation split can shift substantially under a different semantically appropriate reference. This is a scoping and strength-of-evidence issue rather than a logical flaw in BMOA itself, and it is fixable with additional analysis and more careful claim qualification.

major comments (3)
  1. [§5.2, Table 5; §5.4, Table 7] The headline '16 improve, 4 degrade' is computed solely against Decimal. §5.4 reports that Decimal and the stable float32 reference disagree on 20 raw labels, corresponding to ten distinct combinations, and in every disagreement Decimal labels the candidate an improvement while the stable reference labels it a degradation. Since the ten disputed distinct cases are among the 16 Decimal improvements, adopting the stable reference as the intended semantics for those kernels changes the split to 6 improvements and 14 degradations. The paper is transparent about reference sensitivity, but the abstract and §5.2 present the Decimal-relative split as the main evidence that 'strict-FP deviation does not imply accuracy loss.' Please re-run the headline counts under the stable reference and report both, or explicitly restrict the claim to Decimal-relative outcomes and revise the abstract accordingl
  2. [§4, References, metrics, and scale; §5.4] The 'stable float32 algorithmic reference' is never defined. The paper only says that, where applicable, it changes the finite-precision algorithm; no concrete algorithm is given for the reduction, dot-product, polynomial, reciprocal, threshold, or FMA kernels. Because the outcome label O is reference-qualified and the paper itself demonstrates that the reference choice can flip labels, the omitted definition makes the reported outcome labels and the alternative count proposed above impossible to audit or recompute. Please provide exact algorithms and implementations per kernel, store them as artifacts, and use the same definitions in the re-run requested in the previous comment.
  3. [§3.3, Eq. (3)] The condition 'E_m(c,x) is consistent with m' in Eq. (3) is not operationalized. The text says assembly evidence corroborates but cannot establish causality, yet it never specifies what pattern counts or code-generation observations are sufficient to mark a mechanism as 'consistent.' Without a precise rule, third parties cannot reproduce S(c,x), and the distinction between 'supported,' 'mixed,' and 'unknown' is not fully auditable. Please define the consistency predicate for each mechanism category used in the evaluation (FMA, vectorization, reassociation, reciprocal math, etc.), or provide a code-level procedure implementing it.
minor comments (5)
  1. [§3.3, Eq. (3)] The notation E_m(c,x) appears in Eq. (3) but is not defined; define the evidence bundle component and its relation to the evidence table described in §3.5.
  2. [Table 5] The column headers 'Improve' and 'Degrade' should state explicitly that these are Decimal-relative outcomes, so the reference dependence of the counts is visible at the point of use.
  3. [Figure 2] State in the caption that all counts are deterministic and based on a single seed and one representative instance per kernel–input family; the text says this, but the figure alone could be misread as a prevalence estimate.
  4. [§5.4] The absence of positive reproducibility and cross-compiler cases is reported, but it would be clearer to state in the RQ4 conclusion that those categories are validated only as negative controls in this study, not as demonstrated diagnostic categories.
  5. [§7] Minor formatting issue: the author affiliation markers in the ACM reference format line contain a stray '1,'; fix the metadata formatting.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: BMOA's labels are fixed-rule classifications against named references, not fitted predictions, and the headline deviations are externally demonstrated with transparent reference sensitivity.

full rationale

BMOA's derivation chain is a diagnostic workflow rather than a fitted model. Outcome labels are assigned by comparing E_c and E_s to a named reference (Section 2.1, Eq. 1; Section 3.3), with fixed thresholds (tau_m = 0, zero-tolerance ordering) rather than parameters tuned to observed outcomes. The headline result—strict-FP deviation does not necessarily imply accuracy loss—is supported by concrete records (e.g., Section 5.3 mixed-magnitude reduction: strict output 61, candidate 992, Decimal reference 1952, so E_c = 960 < E_s = 1891). The paper explicitly qualifies O as reference-relative and reports in Section 5.4 that Decimal and the stable float32 algorithm disagree on 20/162 labels, always in the same direction; this is a transparent scope limitation, not a circular argument. Mechanism labels are based on controlled paired ablations and assembly corroboration, and the component-independence of B, M, and O is an explicit design invariant enforced by the attribution order, not a consequence imported from self-citations or an imported uniqueness theorem. No parameter is fitted to the data to force the improvement/degradation counts; the conclusions could, as the paper acknowledges, differ for other references, kernels, seeds, or platforms without invalidating the framework. No load-bearing self-citations were found. Thus there is no circular step meeting the bar of Eq.-equals-input or fitted-parameter-renamed-as-prediction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The framework itself adds no fitted parameters to its central inference: it applies fixed attribution rules to externally defined references. The hand-chosen thresholds (tau_m=0, zero tolerance) and the 90-digit Decimal reference are design choices that affect labels but are not tuned to obtain the headline results. The main unstated background commitments are the semantic adequacy of the Decimal-over-float32 reference as an accuracy oracle and the assumption that paired flag differences are controlled interventions.

free parameters (3)
  • tau_m threshold = 0
    Constant in Eq. 3; set to 0 (exact bit inequality) for this evaluation. Hand-chosen design threshold that determines whether a mechanism pair counts as supported.
  • outcome tolerance = 0
    Finite reference errors are ordered with zero tolerance (§3.3). Hand-chosen decision threshold; determines improvement/degradation/neutral labels.
  • Decimal precision = 90 digits
    Higher-precision reference precision chosen for the controlled mechanism matrix (§4). It is not fitted, but it is a hand-chosen setting that affects reference values.
assumptions (4)
  • domain assumption IEEE 754 float arithmetic is rounded, order-sensitive, and compiler transformations can alter evaluation order, contraction, or fast-math behavior.
    Invoked throughout §1–2 as the standard model for floating-point behavior under compiler testing.
  • domain assumption A strict-FP configuration (-fno-fast-math -ffp-contract=off) provides an operational baseline encoding the default evaluation path.
    Used in Eq. 1 and every strict-FP record. The paper acknowledges formal semantic correctness would require tools such as LifeJacket, so this is an assumption rather than a proven equivalence.
  • domain assumption Decimal-90 over exact represented float32 inputs, and NumPy float64, evaluate the intended real-arithmetic expression.
    Outcome labels in §5.2 and §5.3 depend on this reference. §4 calls it 'not an exact oracle,' and §5.4 documents 20/162 disagreements with the stable float32 reference.
  • domain assumption Paired configurations differing by one flag isolate the named mechanism category, with compound interventions permitted to be labeled ambiguous/mixed.
    Eq. 2 computes D_m from paired configurations. The paper acknowledges compound flags and phase ordering, but the elementary comparison still assumes the flag difference is the controlled intervention.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BMOA: Baseline-Mechanism-Outcome Attribution for Compiler-Induced Numerical Deviations." pith.science (2026). https://pith.science/paper/NXS7MBFY

@misc{pith2026260727270,
  author       = {Pith},
  title        = {Pith review of: BMOA: Baseline-Mechanism-Outcome Attribution for Compiler-Induced Numerical Deviations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NXS7MBFY}},
  note         = {Machine review of arXiv:2607.27270}
}
read the original abstract

Formalizing compiler-aware numerical correctness requires distinguishing what an observed floating-point difference means, what compiler behavior the evidence supports, and what numerical consequence follows. Existing testing workflows often collapse these questions into a pass/fail mismatch. We introduce Baseline--Mechanism--Outcome Attribution (BMOA), a diagnostic framework that separates the comparison relation and system boundary, the evidence-supported compiler mechanism, and the reference-qualified accuracy outcome. BMOA combines operational strict floating-point, transformation-local, reproducibility, cross-compiler, and higher-precision comparisons, while preserving mixed, ambiguous, and unknown attributions when evidence is insufficient. Each record retains inputs, configurations, numerical metrics, and supporting artifacts for audit. We evaluate BMOA on six scientific-computing kernels, deterministic stress-input families, and controlled Clang configurations on ARM64. A 1,276-record attribution corpus and a 162-instance controlled mechanism matrix show that baseline choice changes diagnoses, compiler-induced deviation does not imply accuracy loss, and cancellation and large dynamic range expose the strongest effects within the targeted matrix. BMOA converts raw mismatches into explicit, auditable, evidence-bounded records. Although it is not itself a proof system, these records provide an empirical foundation for future formal specifications and proof obligations for compiler-aware numerical correctness.

Figures

Figures reproduced from arXiv: 2607.27270 by the authors.

Figure 1
Figure 1. BMOA workflow and running example. The mixed-magnitude reduction differs from strict FP, the paired ablation [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Controlled mechanism-matrix results after collapsing identical Apple clang 17 and Homebrew clang 22 outcomes. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 4 canonical work pages

  1. [1]

    Myreen, and Zachary Tatlock

    Heiko Becker, Eva Darulova, Magnus O. Myreen, and Zachary Tatlock. 2019. Icing: Supporting Fast-Math Style Optimizations in a Verified Compiler. InComputer Aided Verification (CA V). Springer, Cham, Switzerland, 155–173. doi:10.1007/978- 3-030-25543-5_10

  2. [2]

    Heiko Becker, Pavel Panchekha, Eva Darulova, and Zachary Tatlock. 2018. Com- bining Tools for Optimization and Analysis of Floating-Point Computations. InInternational Symposium on Formal Methods. Springer, Cham, Switzerland, 355–363. doi:10.1007/978-3-319-95582-7_21

  3. [3]

    Florian Benz, Andreas Hildebrandt, and Sebastian Hack. 2012. A Dynamic Pro- gram Analysis to Find Floating-Point Accuracy Problems. InACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI). Asso- ciation for Computing Machinery, New York, NY, USA, 453–462. doi:10.1145/ 2254064.2254118

  4. [4]

    Sangeeta Chowdhary and Santosh Nagarakatte. 2021. Parallel Shadow Execution to Accelerate the Debugging of Numerical Errors. InACM Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering (ESEC/FSE). Association for Computing Machinery, New York, NY, USA, 615–626. doi:10.1145/3468264.3468585

  5. [5]

    Sangeeta Chowdhary and Santosh Nagarakatte. 2022. Fast Shadow Execution for Debugging Numerical Errors Using Error Free Transformations.Proceedings of the ACM on Programming Languages6, OOPSLA2 (2022), 1845–1872. doi:10. 1145/3563353

  6. [6]

    Caroline Collange, David Defour, Stef Graillat, and Roman Iakymchuk. 2015. Numerical Reproducibility for the Parallel Reduction on Multi- and Many-Core Architectures.Parallel Comput.49 (2015), 83–97. doi:10.1016/j.parco.2015.09.001

  7. [7]

    Clément Courbet. 2021. NSan: A Floating-Point Numerical Sanitizer. InACM SIGPLAN International Conference on Compiler Construction (CC). Association for Computing Machinery, New York, NY, USA, 83–93. doi:10.1145/3446804.3446848

  8. [8]

    Eva Darulova, Anastasiia Izycheva, Fariha Nasir, Fabian Ritter, Heiko Becker, and Robert Bastian. 2018. Daisy: Framework for Analysis and Optimization of Numerical Programs. InTools and Algorithms for the Construction and Analysis of Systems (TACAS). Springer, Cham, Switzerland, 270–287. doi:10.1007/978-3-319- 89960-2_15

Show all 34 references
  1. [9]

    James Demmel and Hong Diep Nguyen. 2013. Fast Reproducible Floating-Point Summation. InIEEE Symposium on Computer Arithmetic (ARITH). IEEE, Piscat- away, NJ, USA, 163–172. doi:10.1109/ARITH.2013.9

  2. [10]

    Christophe Denis, Pablo de Oliveira Castro, and Eric Petit. 2016. Verificarlo: Checking Floating Point Accuracy through Monte Carlo Arithmetic. InIEEE Symposium on Computer Arithmetic (ARITH). IEEE, Piscataway, NJ, USA, 55–62. doi:10.1109/ARITH.2016.31

  3. [11]

    Anthony Di Franco, Hui Guo, and Cindy Rubio-González. 2017. A Comprehensive Study of Real-World Numerical Bug Characteristics. InIEEE/ACM International Conference on Automated Software Engineering (ASE). IEEE, Piscataway, NJ, USA, 509–519. doi:10.1109/ASE.2017.8115662

  4. [12]

    Zhoulai Fu, Zhaojun Bai, and Zhendong Su. 2015. Automated Backward Error Analysis for Numerical Code. InACM SIGPLAN International Conference on Object-Oriented Programming, Systems, Languages, and Applications (OOPSLA). Association for Computing Machinery, New York, NY, USA, 6...

  5. [13]

    David Goldberg. 1991. What Every Computer Scientist Should Know About Floating-Point Arithmetic.Comput. Surveys23, 1 (1991), 5–48. doi:10.1145/ 103162.103163

  6. [14]

    Hui Guo, Ignacio Laguna, and Cindy Rubio-González. 2020. pLiner: Isolating Lines of Floating-Point Code for Compiler-Induced Variability. InInternational Conference for High Performance Computing, Networking, Storage and Analysis (SC). IEEE, Piscataway, NJ, USA, 1–14. doi:10.1...

  7. [15]

    IEEE. 2019. IEEE Standard for Floating-Point Arithmetic. 84 pages. doi:10.1109/ IEEESTD.2019.8766229

  8. [16]

    Ignacio Laguna. 2019. FPChecker: Detecting Floating-Point Exceptions in GPU Applications. InIEEE/ACM International Conference on Automated Software Engi- neering (ASE). IEEE, Piscataway, NJ, USA, 1126–1129. doi:10.1109/ASE.2019.00118

  9. [17]

    Ignacio Laguna. 2020. Varity: Quantifying Floating-Point Variations in HPC Systems Through Randomized Testing. InIEEE International Parallel and Dis- tributed Processing Symposium (IPDPS). IEEE, Piscataway, NJ, USA, 622–633. doi:10.1109/IPDPS47924.2020.00070

  10. [18]

    Vu Le, Mehrdad Afshari, and Zhendong Su. 2014. Compiler Validation via Equiv- alence Modulo Inputs. InACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI). Association for Computing Machinery, New York, NY, USA, 216–226. doi:10.1145/2594291.2594334

  11. [19]

    Dolores Miao, Ignacio Laguna, and Cindy Rubio-González. 2023. Expression Isolation of Compiler-Induced Numerical Inconsistencies in Heterogeneous Code. InInternational Conference on High Performance Computing. Springer, Cham, Switzerland, 381–401. doi:10.1007/978-3-031-32041-5_20

  12. [20]

    Dolores Miao, Ignacio Laguna, and Cindy Rubio-González. 2024. Input Range Generation for Compiler-Induced Numerical Inconsistencies. InACM Interna- tional Conference on Supercomputing (ICS). Association for Computing Machinery, New York, NY, USA, 201–212. doi:10.1145/3650200.3656618

  13. [21]

    David Monniaux. 2008. The Pitfalls of Verifying Floating-Point Computations. ACM Transactions on Programming Languages and Systems30, 3 (2008), 12:1–12:41. doi:10.1145/1353445.1353446

  14. [22]

    Andres Nötzli and Fraser Brown. 2016. LifeJacket: Verifying Precise Floating- Point Optimizations in LLVM. InACM SIGPLAN International Workshop on State Of the Art in Program Analysis. Association for Computing Machinery, New York, NY, USA, 24–29. doi:10.1145/2931021.2931024

  15. [23]

    Wilcox, and Zachary Tatlock

    Pavel Panchekha, Alex Sanchez-Stern, James R. Wilcox, and Zachary Tatlock

  16. [24]

    Bailey, Costin Iancu, and David Hough

    Cindy Rubio-González, Cuong Nguyen, Hong Diep Nguyen, James Demmel, William Kahan, Koushik Sen, David H. Bailey, Costin Iancu, and David Hough

  17. [25]

    Alex Sanchez-Stern, Pavel Panchekha, Sorin Lerner, and Zachary Tatlock. 2018. Finding Root Causes of Floating Point Error. InACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI). Association for Com- puting Machinery, New York, NY, USA, 256–269. doi:...

  18. [26]

    Geof Sawaya, Michael Bentley, Ian Briggs, Ganesh Gopalakrishnan, and Dong H. Ahn. 2017. FLiT: Cross-Platform Floating-Point Result-Consistency Tester and Workload. InIEEE International Symposium on Workload Characterization (IISWC). IEEE, Piscataway, NJ, USA, 229–238. doi:10.1...

  19. [27]

    Baranowski, Ian Briggs, Charles Jacobsen, Zvonimir Rakamarić, and Ganesh Gopalakrishnan

    Alexey Solovyev, Marek S. Baranowski, Ian Briggs, Charles Jacobsen, Zvonimir Rakamarić, and Ganesh Gopalakrishnan. 2018. Rigorous Estimation of Floating- Point Round-Off Errors with Symbolic Taylor Expansions.ACM Transactions on Programming Languages and Systems41, 1 (2018), 2...

  20. [28]

    Jackson Vanover, Xuan Deng, and Cindy Rubio-González. 2020. Discovering Discrepancies in Numerical Libraries. InACM SIGSOFT International Symposium on Software Testing and Analysis (ISSTA). Association for Computing Machinery, New York, NY, USA, 488–501. doi:10.1145/3395363.3397380

  21. [29]

    Xuejun Yang, Yang Chen, Eric Eide, and John Regehr. 2011. Finding and Under- standing Bugs in C Compilers. InACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI). Association for Computing Machin- ery, New York, NY, USA, 283–294. doi:10.1145/1993498.1993532

  22. [30]

    Xin Yi, Hengbiao Yu, Liqian Chen, Xiaoguang Mao, and Ji Wang. 2024. FPCC: Detecting Floating-Point Errors via Chain Conditions.Proceedings of the ACM on Programming Languages8, OOPSLA2 (2024), 1504–1531. doi:10.1145/3689764

  23. [31]

    Hengbiao Yu, Xin Yi, Banghu Yin, Fa Li, Zhenbang Chen, and Chun Huang. 2023. Efficient Generation of Floating-Point Inputs for Compiler-Induced Variability. In IEEE International Conference on Software Analysis, Evolution and Reengineering (SANER). IEEE, Piscataway, NJ, USA, 2...

  24. [32]

    Daming Zou, Yuchen Gu, Yuanfeng Shi, Mingzhe Wang, Yingfei Xiong, and Zhendong Su. 2022. Oracle-Free Repair Synthesis for Floating-Point Programs. Proceedings of the ACM on Programming Languages6, OOPSLA2 (2022), 957–985. doi:10.1145/3563322

  25. [2013]

    InInter- national Conference for High Performance Computing, Networking, Storage and Analysis (SC)

    Precimonious: Tuning Assistant for Floating-Point Precision. InInter- national Conference for High Performance Computing, Networking, Storage and Analysis (SC). Association for Computing Machinery, New York, NY, USA, 1–12. BMOA: Baseline–Mechanism–Outcome Attribution for Compi...

  26. [2015]

    In ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI)

    Automatically Improving Accuracy for Floating Point Expressions. In ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI). Association for Computing Machinery, New York, NY, USA, 1–11. doi:10. 1145/2737924.2737959

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.