reversed-stride in WENO7 coefficients (amdflang 23.2.x false-nuw)

[sbryngelson]

What

Rewrites the five reversed-stride array sections in the WENO7 branch of s_compute_weno_coefficients (the only negative-stride sections in the codebase) as explicit element-wise assignments. Values are bit-identical — array subtraction is element-wise — so there are no golden-file changes.

Why

amdflang from AMD's AFAR drops 23.2.0/23.2.1 carries flang commit llvm/llvm-project#184573 ("Apply nusw nuw flags on array_coor gep's") without its correctness fix llvm/llvm-project#198014 (merged upstream 2026-05-20, fixes llvm/llvm-project#197393). In that window, flang stamps unsigned-no-wrap (nuw) claims on the address arithmetic of descriptor-array accesses. For negative-stride sections — like s_cb(i+4:i-3:-1) here — those claims are false, i.e. poison, and the optimizer legally miscompiles: the WENO7 coefficient tables come out slightly wrong, which the ill-conditioned WENO weights amplify to abs 2.1e-4 density errors at shock fronts (weno7 golden tests fail on gfx90a builds at -O2/-O3; correct on drop 23.1.0).

Reported to AMD with full root-cause evidence: ROCm/llvm-project#3471.

Element-wise assignment lowers to plain indexed accesses whose no-wrap claims are true, removing the miscompile at its source. Each site carries a comment so the slices don't get "cleaned up" back into reversed sections. Beyond the specific bug, this makes the one-time setup code robust against this whole class of section-lowering issues on any compiler, at zero cost (init-only code, value-identical).

Testing

• amdflang AFAR 23.2.0, gfx90a OpenMP offload, -O3: all four weno_order=7 golden tests (1D/2D/3D + wenoz/mapped) fail before this change (abs 2.1e-4) and pass after it.
• Values are bit-identical by construction; CI validates the four gated compilers.
• ./mfc.sh precheck clean.

[Copilot]

Pull request overview

This PR hardens the WENO7 coefficient setup in src/simulation/m_weno.fpp against an amdflang 23.2.x optimizer miscompile triggered by negative-stride array sections, by replacing the affected reversed-stride descriptor-array sections with equivalent explicit element-wise assignments.

Changes:

• Replaced the single reversed-stride w = s_cb(i+4:i-3:-1) - s_cb(i) assignment with explicit w(1:8) element-wise assignments to avoid negative-stride section lowering.
• Replaced four reversed-stride y = ... - ... assignments (used for left-side polynomial coefficient setup) with explicit y(1:4) element-wise assignments.
• Added in-code comments documenting why reversed-stride sections must not be reintroduced (compiler correctness rationale).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 59.49%. Comparing base (0857ace) to head (72a483d).

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1660      +/-   ##
==========================================
+ Coverage   59.46%   59.49%   +0.03%     
==========================================
  Files          83       83              
  Lines       21170    21189      +19     
  Branches     3135     3135              
==========================================
+ Hits        12588    12607      +19     
  Misses       6476     6476              
  Partials     2106     2106              

☔ 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.