Profiling-driven fixes (2026-07-25)

[gaelforget]

part 1

Fix A — Polygons.jl SparseArrays wrong dispatch (done)

to_Polygons and to_LineStrings2D used [a b c d e] syntax for 1×5 matrices.
With contaminated upstream types this dispatched through SparseArrays (~1,093 samples, ~81% of demo_grid runtime).
Fix: Float64[a b c d e] — forces typed_hcat, plain Matrix{Float64}.

Fix B — CatView deep nesting in define_regions (done)

Two sub-fixes:

src/types/gcmvector.jl — view(gcmarray, gcmvector): was building CatView(CatView(...)) iteratively — O(nFaces) nesting, O(depth) cost per element write (~6,254 samples).
Fix: views = [view(A[a], B[a]) for a in eachindex(A)]; CatView(views...) — single-level flat CatView.

src/Integration.jl — define_regions hot call sites: mask[findall(...)].=l desugared to dotview → view → CatView → broadcast.
Fix: changed .=l to =l — routes to setindex!(gcmarray, Number, gcmvector) which does a direct face-by-face loop.

Fix C — face-by-face broadcast copyto! (done)

Root cause: A .* A, A .+ B, scalar broadcasts etc. all went through an element-by-element gcmarray_getindex_evalf loop, allocating a new matrix per face per broadcast. Inconsistent with the fast face-by-face *, +, -, / methods in main.jl.

Fix in src/types/gcmarray.jl and src/types/gcmvector.jl: rewrote copyto! to loop for I in CartesianIndices(dest.f) and call broadcast!(bc.f, dest.f[I], face_args...). Helper _bc_face rewrites a Broadcasted tree by replacing each AbstractMeshArray arg with its face-I inner matrix.

Four bugs encountered and fixed during this rewrite:

1. gcmvector.similar left inner Array{T,1} as undef references — fixed to allocate each f[I] explicitly.
2. eachindex(dest.f) returns linear integers for Julia Arrays; I[1] on Int is a no-op. Fixed to CartesianIndices(dest.f) so I[1] always gives the face dimension.
3. Mixed 1D/2D gcmarray broadcast (e.g. GM_PsiX[:,k] .* DYG): _bc_face must use I[1] for 1D sources and full I for 2D. Fix: ndims(a.f) == 1 ? a.f[I[1]] : a.f[I].
4. exch_UV replaces face matrices with larger halo-padded ones but does NOT update fSize. So similar allocates dest faces at wrong size → DimensionMismatch. Fix: copyto! recomputes result shape from actual face args via Base.Broadcast.combine_axes, reallocates dest.f[I] in-place if size mismatches. filter(a -> isa(a, AbstractArray), face_args) guards against scalar args which have no axes.

part 2

• Integration.func: replace collect comprehension with explicit loop,
  hoist xymsk(b) mask out of inner k-loop — zero Vector allocations per call
• Operations.land_mask: use setindex! (=) not broadcast (.=) on gcmvector
  results to bypass CatView materialize! and its ~40-level recursion
• transport test: replace side-effect comprehension with for loop

part 3

• pre-compute basin mask in define_sums; replace func/func_h with face-level loops to eliminate per-call gcmarray allocations
• replace gcmarray broadcast+findall pattern in define_regions with face-level loop, eliminating CatView nesting

[codecov]

Codecov Report

❌ Patch coverage is 96.92308% with 6 lines in your changes missing coverage. Please review.
✅ Project coverage is 87.50%. Comparing base (b039c36) to head (124cc5a).

Files with missing lines	Patch %	Lines
src/types/gcmarray.jl	86.36%	3 Missing ⚠️
src/Polygons.jl	50.00%	2 Missing ⚠️
src/types/gcmvector.jl	93.75%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #226      +/-   ##
==========================================
+ Coverage   87.02%   87.50%   +0.47%     
==========================================
  Files          37       37              
  Lines        3446     3538      +92     
==========================================
+ Hits         2999     3096      +97     
+ Misses        447      442       -5     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.