Keep calc_forces!' per-panel loads out of line so Julia 1.12 gets linearize's ForwardDiff Jacobian right on AVX-512 Zen

[1-Bort-1]

TL;DR

On Julia 1.12 with --check-bounds=yes (as Pkg.test runs), LLVM 18's SLP vectorizer miscompiles calc_forces!'s per-panel force assembly for 10-partial Duals on AMD Zen 4/5 CPUs with AVX-512 (JuliaLang/julia#62368). linearize's default ForwardDiff Jacobian came out about 5% off there, and whether CI drew such a runner was the whole #360 flake. The loads and body moment now come from a @noinline panel_body_loads, which keeps the Jacobian exact there, and a new check makes sure the Jacobian does not depend on the ForwardDiff chunk size.

This replaces #378, cut down as asked on #360: only the workaround and the chunk-size test. The POLAR_MATRICES check stays on main's ram-air wing; the affine tables are gone.

What was wrong

Diagnostic CI ran repro379.jl (in #379) against main on 16 Julia 1.12.7 runners per round. It computes linearize's Jacobian with chunk 10 and with chunk 5, which must agree to rounding:

runner	--check-bounds=yes	default bounds checks	--check-bounds=yes, SLP off
EPYC 9V74 (znver4) / 9V45 (znver5), +avx512f	rel. 0.042 or 0.050: 7 of 7 in the rounds that logged CPU features	0	3e-19
EPYC 9V74 with AVX-512 masked	3e-19	0	3e-19
EPYC 7763, Intel Sapphire/Granite Rapids, Ice Lake	3e-19	0	3e-19

• JULIA_LLVM_ARGS=-vectorize-slp=false makes it exact, which places it in Julia 1.12 and LLVM 18 vectorization defect JuliaLang/julia#62368. That issue has the LLVM fix identified (llvm/llvm-project@5d7cf504), still waiting on a backport to Julia 1.12's LLVM.
• Julia 1.13 (LLVM 20) was correct on the same CPUs. Ordinary runs, without --check-bounds=yes, are not affected. The finite-difference Jacobian was right everywhere.
• Comparing one Dual{10} solve with two stitched Dual{5} solves put the first wrong value at sol.f_body_3D in calc_forces!. Recompiling copies of it with single statements removed: leaving out either the moment_dist[i] line or the m_body_3D store fixes it. The same loop in a function of its own, or in a standalone script, compiles correctly, so there is no package-free reproducer.

What changed

• panel_body_loads(axes, dirs, q_dyn, cl, cd, cm, c_span, arm) (src/panel_aerodynamics.jl, @noinline) returns panel_loads plus body_moment, the moment about the reference point. calc_forces! and calculate_results both call it; calculate_results carried the same moment line, so the formula has one source now. Results do not move: norm_fwd was bit-identical with and without the change.
• test/solver/test_forwarddiff.jl: a nested testset on the POLAR_MATRICES case, AutoForwardDiff(chunksize=5) matches the default chunk to 1e-12. That is the class this bug belongs to, checked with no finite-difference noise.
• Changelog: a Fixed entry under Unreleased.

Where I would push back

• @noinline works around the bug and does not fix it. Other code with the same shape could hit the same SLP bug, and the chunk-size testset catches that only on an AVX-512 Zen runner, and only for this case. Julia 1.12's LLVM SLP vectorizer miscompiles calc_forces! for Dual{10} on AVX-512 Zen 4/5 under --check-bounds=yes (JuliaLang/julia#62368) #379 tracks removing it once Julia 1.12 ships the LLVM backport.
• The comment above the POLAR_MATRICES case in the test still blames Flaky POLAR_MATRICES forwarddiff test on Windows / Julia 1.12 (knot-proximity in piecewise-linear polar interpolation) #360 on grid knots. The miscompile explains the failures instead, but I left it, since you asked for only these two changes.
• I did not file a second upstream issue, since Julia 1.12 and LLVM 18 vectorization defect JuliaLang/julia#62368 already covers this. Posting our reproducer there is a public comment on a third-party repo and waits for your yes.

Verification

• Reproduced first (on Keep calc_forces!' per-panel loads out of line so Julia 1.12 gets linearize's ForwardDiff Jacobian right on AVX-512 Zen #378's history): PR run 35666027777, ubuntu 1.12: 0.04894348185682458 < 0.0001. Control runs on main (35730023185, 35736962099): repro379.jl rel. 0.042 / 0.050 on every AVX-512 Zen runner.
• Workaround proven on the failing hardware: run 35740722868, 16/16 green, three AVX-512 Zen runners (2× 9V74, 1× 9V45) give repro379.jl rel. 0 under --check-bounds=yes. That run carried the affine test; the src/ change is the same one here.
• After merging main (e62fc15): the 11 solver and body-aerodynamics test files, 5243/5243 pass locally (juliaserver, Julia 1.13.0). There is no local red: this box's i5-13500 has no AVX-512.
• Local CI mirror (agent ci-local, Julia 1.13): started · GitHub CI: running
• No REUSE in this repo · up to date with main
• Risk: another force loop that LLVM 18 SLP-vectorizes the same way, which only an AVX-512 Zen runner under --check-bounds=yes would show.

Scope

+34 / −10 across 6 files. src/ is +21 / −10: the new 13-line function, and its two callers trading their inline moment line for the call. The test gains the 8-line chunk-size testset; the docs page and the changelog take one entry each.

Closes #360

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[1-Bart-1]

Does this affect performance?