Skip to content

Research-2075: High-signal non-CUDA Cppcheck cleanup

Scope and diagnosis

At integration head e0d31c5bcb43a7aab7fb845a592488635b061566, a fresh GCC CPU build with DNN and float extractors enabled reproduced twelve owned exhaustive Cppcheck findings across eight native files: one unassignedVariable, three unreadVariable, two variableScope, two knownConditionTrueFalse, two clarifyCondition, one usleepCalled, and one redundantInitialization. Other whole-tree findings are outside this bounded batch and remain visible; no suppression or baseline changes are used.

Those findings came from configuration-wide variable lifetimes, assignments overwritten before their first read, a high-bit-depth shift whose zero case returned earlier, Xiph loop initializers parsed as assignment-plus-comparison expressions, and an obsolete POSIX delay call in a contention test. The SpEED ak locals are aliases of ap1; storing ap1 directly preserves the same float value and write order in both mirrored CPU helper TUs. Backend and pooled-score error locals move to the smallest scopes that consume them. The benchmark GPU initializer remains compiled and checked only when CUDA or SYCL support exists.

The first ordinary commit attempt then exposed 32 baselined HISS findings in the touched files. Project policy now requires touched-file zero, so the batch expanded rather than using a debt exception. The findings were four classes:

  • file-wide C++ linkage blocks and a C++ extern "C" wrapper that the scanner treated as functions;
  • long declarative SpEED option tables;
  • oversized CLI, benchmark, validation and picture-pool test functions; and
  • cleanup goto ladders in vmaf and vmaf_bench.

The DNN template contains only macros and static inline definitions, so its outer C-linkage wrapper had no linkage effect and was removed. vmaf.cpp now uses short, reopened anonymous-namespace blocks plus CliRunState / CliRunGuard; the blocks preserve internal linkage while remaining below the scanner's 60-line limit. Teardown still runs in the established order (context, GPU state, input readers, files, CLI settings, model arrays). Pixel-copy branches retain their original loop and sample order. GPU initialisation remains SYCL, CUDA, HIP, then Metal. vmaf_bench uses one cleanup helper for benchmark/validation resources and a separate SYCL profiling owner. The compact option rows retain their original registration order, names, aliases, defaults and bounds.

No alternatives: these are direct expression, lifetime and ownership cleanups under existing contracts. No new architecture or policy is added, so no ADR is needed. The updated package invariant records the structured cleanup ownership, and the bench guide records failure-status behavior.

Preservation controls

The scalar Xiph DCT, SpEED arithmetic expressions, pooling ranges, backend selection order and picture-pool delay durations remain unchanged. Existing DISTS high-bit-depth normalisation, SpEED, motion-minimum, picture-preallocation and AVX2 PSNR-HVS tests exercise the touched paths. The benchmark now propagates SYCL frame/flush errors and validation flush errors instead of overwriting them with later success. The before/after proof uses the same generated Meson compile database and Cppcheck 2.22.0 command:

cppcheck --enable=all --check-level=exhaustive --inline-suppr \
  --library=posix --library=scripts/ci/cppcheck-public-entrypoints.cfg \
  --suppressions-list=.cppcheck-suppressions.txt \
  --project=build/compile_commands.json --error-exitcode=1

The expected whole-tree Cppcheck exit remains non-zero for unrelated findings. Success for this batch means all twelve owned Cppcheck records disappear and standardsctl audit reports every touched file clean. The final governance scan reports 1304 active findings within the 1411-entry baseline, down from 1336 while exposing no new touched-file debt. This does not claim that excluded CUDA, framesync, CAMBI, model-tooling, libsvm, cJSON, Python or release findings are closed.