Pith. sign in

REVIEW 2 major objections 1 minor 35 references

Fine-grained interprocedural code layout becomes feasible for warehouse-scale applications through evolutionary optimization guided by hardware performance measurements.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-06-29 06:02 UTC pith:SZXCQN5E

load-bearing objection The paper extends Propeller via evolutionary search and real-hardware rewards to reach interprocedural layouts at warehouse scale, but the reported gains rest on an abstract with no experimental controls. the 2 major comments →

arxiv 2606.00131 v1 pith:SZXCQN5E submitted 2026-05-28 cs.SE cs.AIcs.LGcs.PL

AI-PROPELLER: Warehouse-Scale Interprocedural Code Layout Optimization with AlphaEvolve

Chaitanya Mamatha Ananda , Rajiv Gupta , Mircea Trofin , Aiden Grossman , Sriraman Tallam , Xinliang David Li , Amir Yazdanbakhsh This is my paper
classification cs.SE cs.AIcs.LGcs.PL
keywords code layoutinterprocedural optimizationevolutionary searchhardware feedbackperformance optimizationpost-link optimizationwarehouse scale
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper establishes that an evolutionary approach can make fine-grained interprocedural code layout practical for large industrial applications. It generates layout variants and measures their performance directly on hardware to provide feedback for refining the optimization policy. This is important because previous post-link optimizers were limited to arrangements within individual functions, leaving interprocedural opportunities unexplored. The reported results show that this method can deliver additional performance improvements of 0.23% to 1.6% on top of existing state-of-the-art techniques for warehouse-scale binaries.

Core claim

The central claim is that an evolutionary workflow can transform a post-link optimizer into a fine-grained interprocedural code layout tool. Layout variants are executed on hardware to supply accurate performance data as the search reward. This has been shown to optimize large warehouse-scale applications for the first time, producing gains of 0.23% to 1.6% beyond current feedback-directed and post-link methods.

What carries the argument

The hardware-in-the-loop evolutionary search that refines interprocedural layout decisions by measuring real performance counters on variant executions.

Load-bearing premise

Generating and running multiple layout variants on hardware supplies a reward signal precise and unbiased enough for the evolutionary process to find better layouts overall.

What would settle it

An experiment that applies the resulting layouts to new inputs or different hardware and observes that the performance advantage is not maintained.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The paper presents AI-PROPELLER, which applies the Magellan agentic workflow (built on AlphaEvolve) to evolve Propeller's post-link optimization heuristics into a fine-grained interprocedural code layout optimizer. Multiple layout variants are generated and executed on real hardware to obtain performance-counter rewards, bypassing static cost models; the approach is evaluated on large warehouse-scale applications and reports 0.23–1.6 % gains on top of state-of-the-art FDO and PLO, claimed to be the first such demonstration of interprocedural layout at industrial scale.

Significance. If the reported gains prove robust, the work would be significant for demonstrating that evolutionary search with hardware feedback can practically address the combinatorially hard interprocedural layout problem at warehouse scale, where prior PLO systems were limited to intraprocedural techniques. The hardware-reward design is a methodological strength that could influence future profile-guided and post-link optimizers.

major comments (2)
  1. [Abstract] Abstract: the central performance claims (0.23 %–1.6 % gains) are stated without any reference to the number of independent runs per variant, run-to-run variance, statistical significance testing, or the exact Propeller baseline configuration, rendering the headline result impossible to assess from the provided text.
  2. [Evaluation] The description of the evolutionary loop does not address whether the hardware reward signal was collected on held-out workloads or whether measurement noise and benchmark-specific artifacts were controlled for, which is load-bearing for the claim that the method discovers globally superior interprocedural layouts rather than overfitting to the measured executions.
minor comments (1)
  1. [Introduction] The abstract and introduction repeatedly use the phrase 'first time ever' without a dedicated related-work subsection that systematically compares against prior interprocedural attempts (even if unsuccessful) in the literature.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and indicate where revisions to the manuscript are planned.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central performance claims (0.23 %–1.6 % gains) are stated without any reference to the number of independent runs per variant, run-to-run variance, statistical significance testing, or the exact Propeller baseline configuration, rendering the headline result impossible to assess from the provided text.

    Authors: We agree that the abstract should reference these methodological details to allow proper assessment of the claims. The evaluation section already specifies the number of independent runs, variance, and that the baseline is the standard Propeller configuration combined with FDO and PLO. We will revise the abstract to briefly incorporate these elements. revision: yes

  2. Referee: [Evaluation] The description of the evolutionary loop does not address whether the hardware reward signal was collected on held-out workloads or whether measurement noise and benchmark-specific artifacts were controlled for, which is load-bearing for the claim that the method discovers globally superior interprocedural layouts rather than overfitting to the measured executions.

    Authors: The manuscript describes hardware-based rewards from the target warehouse-scale applications but does not explicitly discuss held-out sets or noise controls. We will revise the evaluation section to clarify that repeated executions were used to average out measurement noise and to discuss the implications of using the same production workloads for both evolution and final measurement, including any cross-application validation performed. revision: partial

Circularity Check

0 steps flagged

No significant circularity; reward from independent hardware measurements

full rationale

The paper describes an evolutionary process (AlphaEvolve/Magellan) that generates layout variants and obtains the reward signal directly from executing those variants on actual hardware to measure performance counters. No equations, parameter fits, self-definitional relations, or load-bearing self-citations are present in the abstract or described method. The central result is framed as an empirical outcome of hardware execution rather than a derivation that reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract provides no explicit free parameters, axioms, or invented entities; the central claim rests on the unstated premise that the evolutionary search plus hardware feedback is effective and that the reported gains generalize beyond the evaluated benchmarks.

pith-pipeline@v0.9.1-grok · 5804 in / 1229 out tokens · 22020 ms · 2026-06-29T06:02:22.683874+00:00 · methodology

0 comments
read the original abstract

Post-link optimizers (PLOs) such as Propeller and BOLT have demonstrated that precise, profile-guided code layout can extract significant performance gains from heavily optimized binaries. However, these systems are currently restricted to intraprocedural techniques, leaving the global potential of interprocedural layout largely untapped. Interprocedural code layout is historically difficult due to a combinatorially intractable search space and complex call-return semantics that are challenging to model. Consequently, the performance potential of fine-grained interprocedural layout remains unproven in practice. AI-PROPELLER uses Magellan, an agentic workflow that evolves the compiler heuristic in Propeller into a fine-grained interprocedural optimizer and fine-tunes the resulting policy hyperparameters. To ensure high-fidelity, we move away from approximate static cost models and the agentic workflow generates multiple layout variants that are executed on actual hardware to measure real performance counters, providing a precise reward signal for the evolutionary loop. AI-PROPELLER has been evaluated on several benchmarks including large warehouse-scale applications and experiments show performance improvements of 0.23% to 1.6% optimized with state-of-the-art FDO and PLO which is significant for real-world binaries. This is the first time ever that large warehouse-scale applications in industrial settings have been optimized with fine-grained interprocedural code layout.

Figures

Figures reproduced from arXiv: 2606.00131 by Aiden Grossman, Amir Yazdanbakhsh, Chaitanya Mamatha Ananda, Mircea Trofin, Rajiv Gupta, Sriraman Tallam, Xinliang David Li.

Figure 1
Figure 1. Figure 1: Motivating Example from a real-world application, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: AI-PROPELLER Overview: Application of Magellan to Propeller formance improvements ranging from 0.23% to 1.6%. We want to stress that, while naively these improvements may seem small, they are both significant from a operational cost / business perspective, and very hard to achieve in industrial settings already employing the state of the art optimization techniques (profile-guided optimizations, link-time … view at source ↗
Figure 3
Figure 3. Figure 3: Discovered Policy from AI￾PROPELLER for LLVM Clang [18] 0 10 20 30 40 50 60 Number of Partitions 10 0 10 1 10 2 10 3 10 4 Number of Hot Functions (Log Scale) Distribution of Function Partitions 2 Partitions > 2 Partitions [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Execution Times of baseline and rela￾tive changes in execution times achieved by CD￾Sort, Ext-TSP and AI-PROPELLER. clang leveldb redis 0 10 20 30 40 50 60 % CPU Cycles Stalled (Frontend Bound) Baseline CDSort Ext-TSP AI-Propeller [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: % of CPU pipeline slots spent in retiring instructions. Higher is better as it indicates more [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    AsmDB: understanding and mitigating front-end stalls in warehouse-scale computers,

    G. Ayers, N. P. Nagendra, D. I. August, H. K. Cho, S. Kanev, C. Kozyrakis, T. Krishnamurthy, H. Litz, T. Moseley, and P. Ranganathan, “AsmDB: understanding and mitigating front-end stalls in warehouse-scale computers,” inISCA, 2019

    2019

  2. [2]

    Interprocedural conditional branch elimination,

    R. Bod ´ık, R. Gupta, and M. L. Soffa, “Interprocedural conditional branch elimination,” in PLDI, 1997

    1997

  3. [3]

    Using profile information to assist classic code optimizations,

    P. P. Chang, S. A. Mahlke, and W.-m. W. Hwu, “Using profile information to assist classic code optimizations,”Softw. Pract. Exper., 1991

    1991

  4. [4]

    Magellan: Autonomous discovery of novel compiler optimization heuristics with alphaevolve,

    H. Chen, A. Novikov, N. V ˜u, H. Alam, Z. Zhang, A. Grossman, M. Trofin, and A. Yazdan- bakhsh, “Magellan: Autonomous discovery of novel compiler optimization heuristics with alphaevolve,” inarXiv, 2026

    2026

  5. [5]

    Evan Martin, “Ninja,” https://ninja-build.org/, version X.Y .Z, accessed May 8, 2026

    2026

  6. [6]

    Trace scheduling: A technique for global microcode compaction,

    Fisher, “Trace scheduling: A technique for global microcode compaction,”IEEE Transactions on Computers, 1981

    1981

  7. [7]

    Ghemawat and J

    S. Ghemawat and J. Dean,LevelDB, 2011. [Online]. Available: https://github.com/google/ leveldb

    2011

  8. [8]

    Procedure placement using temporal ordering information,

    N. Gloy, T. Blackwell, M. Smith, and B. Calder, “Procedure placement using temporal ordering information,” inProceedings of 30th Annual International Symposium on Microarchitecture, 1997

    1997

  9. [9]

    Golovin, B

    D. Golovin, B. Solnik, S. Moitra, G. Kochanski, J. E. Karro, and D. Sculley, Eds., Google Vizier: A Service for Black-Box Optimization, 2017. [Online]. Available: http: //www.kdd.org/kdd2017/papers/view/google-vizier-a-service-for-black-box-optimization

    2017

  10. [10]

    Beyond malloc efficiency to fleet efficiency: a hugepage-aware memory allocator,

    A. H. Hunter, C. Kennelly, D. Gove, P. Ranganathan, P. J. Turner, and T. J. Moseley, “Beyond malloc efficiency to fleet efficiency: a hugepage-aware memory allocator,” inOSDI, 2021

    2021

  11. [11]

    Achieving high instruction cache performance with an optimizing compiler,

    W. W. Hwu and P. P. Chang, “Achieving high instruction cache performance with an optimizing compiler,” inISCA, 1989

    1989

  12. [12]

    ThinLTO: scalable and incremental LTO,

    T. Johnson, M. Amini, and X. D. Li, “ThinLTO: scalable and incremental LTO,” inCGO, 2017

    2017

  13. [13]

    Profiling a warehouse-scale computer,

    S. Kanev, J. P. Darago, K. M. Hazelwood, P. Ranganathan, T. Moseley, G. Wei, and D. M. Brooks, “Profiling a warehouse-scale computer,” inISCA, 2015

    2015

  14. [14]

    Kitware, Inc., “Cmake,” https://cmake.org/, version X.Y .Z, accessed May 8, 2026

    2026

  15. [15]

    [RFC] Machine Function Splitter - Split out cold blocks from machine func- tions using profile data,

    S. Kumar, “[RFC] Machine Function Splitter - Split out cold blocks from machine func- tions using profile data,” https://discourse.llvm.org/t/rfc-machine-function-splitter-split-out- cold-blocks-from-machine-functions-using-profile-data/56093, August 2020, lLVM Discus- sion Forums

    2020

  16. [16]

    Codestitcher: inter-procedural basic block layout opti- mization,

    R. Lavaee, J. Criswell, and C. Ding, “Codestitcher: inter-procedural basic block layout opti- mization,” inCC, 2019

    2019

  17. [17]

    ACE: An AI-Driven Code Efficiency Optimizer for Warehouse Scale Computers,

    H. Lin, M. Maas, M. Roquemore, A. Hasanzadeh, F. Lewis, Y . Simonson, A. Shringi, H. Dai, P. Musau, T.-W. Yang, A. Y . Bakhsh, D. Altinb ¨uken, F. Papa, M. N. Edmonds, A. Patil, D. Schwarz, S. Chandra, C. Kennelly, M. Hashemi, and P. Ranganathan, “ACE: An AI-Driven Code Efficiency Optimizer for Warehouse Scale Computers,” inOSDI, 2026

    2026

  18. [18]

    [Online]

    LLVM Developer Group,Clang: A C language family frontend for LLVM, 2007. [Online]. Available: https://clang.llvm.org

    2007

  19. [19]

    Ispike: A post-link optimizer for the intel®itanium®architecture,

    C.-K. Luk, R. Muth, H. Patil, R. Cohn, and G. Lowney, “Ispike: A post-link optimizer for the intel®itanium®architecture,” inCGO, 2004

    2004

  20. [20]

    Program optimization for instruction caches,

    S. McFarling, “Program optimization for instruction caches,” inASPLOS, 1989. 9

    1989

  21. [21]

    Procedure merging with instruction caches,

    S. McFarling, “Procedure merging with instruction caches,” inPLDI, 1991

    1991

  22. [22]

    Avoiding conditional branches by code replication,

    F. Mueller and D. B. Whalley, “Avoiding conditional branches by code replication,” 1995

    1995

  23. [23]

    Improved basic block reordering,

    A. Newell and S. Pupyrev, “Improved basic block reordering,”IEEE Transactions on Comput- ers, 2020

    2020

  24. [24]

    AlphaEvolve: A coding agent for scientific and algorithmic discovery,

    A. Novikov, N. V ˜u, M. Eisenberger, E. Dupont, P.-S. Huang, A. Z. Wagner, S. Shirobokov, B. Kozlovskii, F. J. R. Ruiz, A. Mehrabian, M. P. Kumar, A. See, S. Chaudhuri, G. Holland, A. Davies, S. Nowozin, P. Kohli, and M. Balog, “AlphaEvolve: A coding agent for scientific and algorithmic discovery,” inarXiv, 2025

    2025

  25. [25]

    Optimizing function placement for large-scale data-center applica- tions,

    G. Ottoni and B. Maher, “Optimizing function placement for large-scale data-center applica- tions,” inCGO, 2017

    2017

  26. [26]

    BOLT: a practical binary optimizer for data centers and beyond,

    M. Panchenko, R. Auler, B. Nell, and G. Ottoni, “BOLT: a practical binary optimizer for data centers and beyond,” inCGO, 2019

    2019

  27. [27]

    Profile guided code positioning,

    K. Pettis and R. C. Hansen, “Profile guided code positioning,”SIGPLAN Not., 1990

    1990

  28. [28]

    llvm::codelayout::cdsortconfig struct reference,

    T. L. Project, “llvm::codelayout::cdsortconfig struct reference,” Doxygen API Documentation, 2025, https://llvm.org/doxygen/structllvm 1 1codelayout 1 1CDSortConfig.html

    2025

  29. [29]

    Sanfilippo,Redis, 2009

    S. Sanfilippo,Redis, 2009. [Online]. Available: https://redis.io

    2009

  30. [30]

    Propeller: A Profile Guided, Relinking Optimizer for Warehouse-Scale Applications,

    H. Shen, K. Pszeniczny, R. Lavaee, S. Kumar, S. Tallam, and X. D. Li, “Propeller: A Profile Guided, Relinking Optimizer for Warehouse-Scale Applications,” inASPLOS, 2023

    2023

  31. [31]

    Learning performance-improving code edits,

    A. Shypula, A. Madaan, Y . Zeng, U. Alon, J. Gardner, M. Hashemi, G. Neubig, P. Ran- ganathan, O. Bastani, and A. Yazdanbakhsh, “Learning performance-improving code edits,” inICLR (Spotlight), 2023

    2023

  32. [32]

    Automated high-level code optimization for warehouse performance,

    A. Shypula, A. Madaan, Y . Zeng, U. Alon, J. Gardner, M. Hashemi, G. Neubig, P. Ran- ganathan, O. Bastani, and A. Yazdanbakhsh, “Automated high-level code optimization for warehouse performance,”IEEE Micro, 2025

    2025

  33. [33]

    Mlgo: a machine learning guided compiler optimizations framework

    M. Trofin, Y . Qian, E. Brevdo, Z. Lin, K. Choromanski, and D. Li, “Mlgo: a machine learning guided compiler optimizations framework,” 2021. [Online]. Available: https://arxiv.org/abs/2101.04808

    work page arXiv 2021

  34. [34]

    Hardware Counted Profile-Guided Optimization

    B. Wicht, R. A. Vitillo, D. Chen, and D. Levinthal, “Hardware counted profile-guided optimization,” 2014. [Online]. Available: https://arxiv.org/abs/1411.6361

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  35. [35]

    A Top-Down method for performance analysis and counters architecture,

    A. Yasin, “A Top-Down method for performance analysis and counters architecture,” inIS- PASS, 2014. 10

    2014