OptimizationMOI: const f in MOIOptimizationNLPEvaluator to avoid large field copies

[ChrisRackauckas-Claude]

Summary

MOIOptimizationNLPEvaluator is a mutable struct that stores the instantiated OptimizationFunction inline as f::F. At n=10 with AutoForwardDiff + Hessian, sizeof(ev.f) is ~39 KB (DI Hessian preps), so every evaluator.f load from the mutable container copied the whole field on Julia 1.12. Declaring const f::F (supported since Julia 1.8; OptimizationMOI compat is julia = "1.10") lets the compiler elide those copies. Grep confirms nothing reassigns evaluator.f.

Before / after (Julia 1.12.7, JULIA_NUM_THREADS=1, taskset -c 95, @benchmark minima)

Rosenbrock + 3 constraints, AutoForwardDiff, Ipopt NLP callbacks. Identical callback values and Ipopt solutions before/after (n=2: 25 iters → [0.9999785, 0.9999569]; n=10: 12 iters → same u / objective).

callback (n=10)	before (ns)	after (ns)
eval_objective_gradient	1268	156
eval_constraint_jacobian	1347	240
eval_hessian_lagrangian	6240	5647

n=2 (small f): gradient 65→38 ns, jacobian 115→95 ns, hessian 439→408 ns.

Julia 1.10.12: callback values and Ipopt solves match the 1.12 results byte-for-byte on u/obj/iters. Absolute MOI-callback times on 1.10 were ~µs-scale both before and after in this env (other overhead appears to dominate; const still compiles and is correct).

Test group tails

• OPTIMIZATION_TEST_GROUP=Core (Julia 1.12, lib/OptimizationMOI): Core | 40 Pass, 2 Broken, 42 Total 4m43.3s; Testing OptimizationMOI tests passed
• OPTIMIZATION_TEST_GROUP=QA (Julia 1.12): Quality Assurance | 20 Pass, 1 Broken, 21 Total 2m30.7s; Testing OptimizationMOI tests passed
• Runic / typos on lib/OptimizationMOI/src/nlp.jl: exit 0 (diff is the single const token; file was already Runic-clean on master)

Not verified

• Full monorepo GROUP=All / other sublibraries
• Downstream packages beyond OptimizationMOI Core + QA
• Whether making additional large fields (reinit_cache, matrix buffers) const would help further
• A production Ipopt wall-time win at large n (callback microbenchmarks only; solve time was already dominated by Ipopt itself)

Reviewer pushback

• Is const f the right layer, or should hot paths use getfield(evaluator, :f) and leave the field mutable for theoretical reassignment through MOIOptimizationNLPCache’s setproperty!?
• Should other large inline fields on the same mutable evaluator get the same treatment in a follow-up?
• Julia 1.10 did not show the same microbenchmark win here; confirm whether that matches expectations given the existing getproperty override.

Please ignore until reviewed by @ChrisRackauckas.

Behavior note: because MOIOptimizationNLPCache's setproperty! forwards to the evaluator, cache.f = g now throws (const field). Nothing in the repo assigns it; reinit! only touches reinit_cache.

Risk assessment

• Risk: low
• Blast radius: OptimizationMOI NLP evaluator field layout only; f was never reassigned
• Evidence: 1.12 microbenchmarks + identical Ipopt solutions on 1.12 and 1.10; Core/QA green
• Independent review: Claude Code (claude-opus-5-5) MERGE, risk low (confidence high): Julia compat 1.10 allows const fields; nothing in the repo reassigns f; n=10 eval_objective_gradient 1161 -> 226 ns, constraint_jacobian 1249 -> 292 ns, hessian_lagrangian 6214 -> 5522 ns; Ipopt n=2 and n=10 results identical; OptimizationMOI Core passes.
• Merge: needs review

🤖 Generated with Cursor Agent 2026.10.01-14929f9 (model: unknown (Cursor auto)), transcript /home/crackauc/sandbox/goals/performance/jobs/opt/opt-moi/log.txt on amdci2; orchestrated by Claude Code (claude-opus-5-5[1m]) https://claude.ai/code/session_01LPHREnnonfLg1VcE1EJovv

Made with Cursor

Independent review: Devin Fusion (fusion-claude-opus-5-5-high-sidekick-swe-2-medium) rated it low, verdict MERGE: #1404 (comment)

[ChrisRackauckas]

🤖 Automated comment from an AI agent running as @ChrisRackauckas — not written or reviewed by Chris.

Independent review (Devin CLI 3000.11.3, model fusion-claude-opus-5-5-high-sidekick-swe-2-medium): MERGE, risk low.

Full review

VERDICT: MERGE
RISK: low

PR: #1404. The diff is one token: lib/OptimizationMOI/src/nlp.jl:6 f::F → const f::F on the mutable MOIOptimizationNLPEvaluator. The PR body links no issue and no tracking issue.

Blocking findings

None.

Non-blocking findings

1. Behavior change: cache.f = g now throws (confirmed by running). MOIOptimizationNLPCache's setproperty! (nlp.jl:45-50) forwards to setfield!(cache.evaluator, :f, x). On master, cache2.f = cache2.f succeeds. With the PR it throws ErrorException: setfield!: const field .f of type MOIOptimizationNLPEvaluator cannot be changed, on both 1.12.7 and 1.10.12. The PR body discloses this. On master the setter was already limited to a value of the exact same concrete type F, because setfield! does not convert. I searched lib/ (including OptimizationBase) and ~/.julia/packages/SciMLBase for assignments to cache.f/.f = and found none, so no in-repo or SciMLBase caller depends on the setter. The setter is undocumented internal behaviour, so I don't consider this a SemVer break. Maintainers may still want a patch bump of lib/OptimizationMOI/Project.toml (1.4.2) when releasing; the PR doesn't include one.
2. No regression test (from reading). This is a pure performance/layout change and the Core suite does exercise every callback. A one-line @test isconst(OptimizationMOI.MOIOptimizationNLPEvaluator, :f) would stop someone from silently reverting it. That would be nice to have, not required.
3. The author's suggested follow-up (making other large inline fields such as reinit_cache, J, H const) is out of scope here and correctly left out. evaluator.iteration += 1 (nlp.jl:258) mutates iteration, so that field must stay mutable; I also checked that cache.iteration = 3 still works with the PR. reinit! mutates the ReInitCache object, not the evaluator field, so whether that field could be const is a separate question.

What I ran

All runs were on this Mac (M2 Max) with -t1. I used scratch envs outside the checkout: env_pr develops the PR's lib/OptimizationBase + lib/OptimizationMOI, and env_master develops the same libs taken from git archive origin/master. The probe is scratch/probe.jl. Rosenbrock (n=2, n=10) + 3 constraints, AutoForwardDiff, Ipopt. It runs MOI.initialize([:Grad,:Jac,:Hess]), evaluates objective/gradient/Jacobian/Hessian-of-Lagrangian at a fixed x, uses @benchmark minima, does solve!, does reinit! + solve!, and tries the cache.f setter.

• Correctness, confirmed by running: on Julia 1.12.7, every printed value is byte-identical between PR and master: obj, G, J, H, sol.u, objective, retcode, iteration count, and the reinit solve (diff of the outputs after filtering timing lines: only precompile log lines differ). I hand-checked the n=2 Hessian against the analytic Rosenbrock + constraint Hessian at x=(0.1, 0.9), σ=1, μ=(0.3, 0.7, 1.1): [-345.4, -39.3, 200.6] is correct. The n=2 and n=10 solves converge to the all-ones minimizer with retcode Success. The 1.10.12 PR run gives the same values as 1.12.

• Performance claim, reproduced on 1.12.7 (minimum ns, master → PR). sizeof(ev.f) = 920 B at n=2 and 38 936 B at n=10, which matches the body's "~39 KB".

  • n=10: gradient 773 → 109, Jacobian 791 → 172, Hessian 5486 → 4924
  • n=2: gradient 37 → 16, Jacobian 86 → 82, Hessian 333 → 299

  The absolute numbers differ from the body's (it ran on different hardware), but the direction and magnitude match.

• OPTIMIZATION_TEST_GROUP=Core julia +1.12 --project=. -e 'using Pkg; Pkg.test()' on a copy of the PR's lib/ ran to Core | 40 Pass, 2 Broken, 42 Total 2m29.2s, Testing OptimizationMOI tests passed. That matches the body's 40/2/42. The 2 Broken are pre-existing, since the PR touches no test file.

• julia +1.12 --project=@runic -m Runic --check lib/OptimizationMOI/src/nlp.jl exited 0, and typos over the diff exited 0.

• Commit trailers carry Agent-Harness/Model/Session, and the PR body carries the attribution. The body ends with "Please ignore until reviewed by @ChrisRackauckas" instead of opening with the required first-line banner. That is a formatting nit and doesn't affect the code.

What I did not verify

• The QA group (Aqua/JET/ExplicitImports). I did not rerun it. A const field annotation cannot plausibly affect it, and the body reports 20 Pass, 1 Broken.
• Master's timings on Julia 1.10. I only ran the PR on 1.10 (it allocates and runs at µs scale there). So I can't confirm the body's claim that 1.10 showed no win before or after.
• Core on Julia 1.10 or the pre channel. I only checked that the const field compiles and gives correct results on 1.10.12.
• Downstream packages and other sublibraries. A grep found no assignment of .f on this evaluator or cache in lib/ or SciMLBase.

---

🤖 Posted by an AI agent — harness: Devin CLI 3000.11.3 (review), Claude Code 2.1.285 (posting) · model: fusion-claude-opus-5-5-high-sidekick-swe-2-medium
Conversation: local Claude session 3c311569-dda6-4687-bc9f-65febb212c33 (fleet master)