You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
native-deep: first-run findings (coverage time, affinity tests under ASan, coverage-labels test) #769
The first run of the weekly report-only `native-deep.yml` workflow (from #762, run 43171495302) completed, with these findings to triage:
Coverage job time. The instrumented pytest took 2 h 25 min against the job's 240-minute timeout, which leaves little margin. Options: shard it, or cover the native tests plus a targeted Python subset.
Affinity tests under ASan and coverage. `test_the_process_pool_follows_the_affinity_mask` fails (`assert 1 == 2`) in both the extension-asan and the coverage jobs. It is probably the setarch/ASan environment narrowing CPU affinity. Either skip the test there with a stated reason, or widen the mask.
Coverage-labels test. `test_coverage_labels.py::test_label_collector_counts_concurrent_calls_and_escapes` raises StopIteration only in the coverage build.
Results to triage:
include-cleaner report (24 KB artifact);
coverage totals of about 76 % lines, 81 % functions and 76 % regions. Its dead-code list feeds the lean-plan native structure PR.
Trigger. The workflow also runs on a pull request that changes the workflow file itself. Keep that trigger, or drop it.
Addition: build the analysis presets at -O1 -g, not -O0.
The Python test suite and the benchmarks already run on Release builds (native/pyproject.toml: cmake.build-type = "Release"). Only the GoogleTest binary runs on a Debug build, in two native-quality CI jobs. Neither job needs -O0:
cpu-tidy (Debug plus clang-tidy) is static analysis, so the optimisation level makes no difference to what it finds. It only slows the build and the ctest run.
cpu-sanitize (Debug plus ASan, UBSan and Kokkos bounds checks) is the configuration where AddressSanitizer's documentation recommends -O1 for reasonable speed, with -fno-omit-frame-pointer for usable stacks.
Proposal:
Move both presets to CMAKE_BUILD_TYPE=RelWithDebInfo with -O1 -g -fno-omit-frame-pointer, or keep Debug and override CMAKE_CXX_FLAGS_DEBUG.
Keep NDEBUGundefined in both, so asserts still fire.
Use the same flags in native-deep.yml's ASan-extension and coverage jobs. Clang source coverage also works at -O1, and that would ease the coverage job's 2.5 h runtime noted above.
cpu-debug stays -O0 for interactive stepping. RelWithDebInfo (-O2 -g) is the documented choice for debugging real runs.
To check: the tidy and sanitize stages report the same findings before and after, and their CI wall time drops.
Activity
jtravs commentedon Nov 12, 2026
Addition: build the analysis presets at
-O1 -g, not-O0.The Python test suite and the benchmarks already run on Release builds (
native/pyproject.toml:cmake.build-type = "Release"). Only the GoogleTest binary runs on a Debug build, in twonative-qualityCI jobs. Neither job needs-O0:cpu-tidy(Debug plus clang-tidy) is static analysis, so the optimisation level makes no difference to what it finds. It only slows the build and the ctest run.cpu-sanitize(Debug plus ASan, UBSan and Kokkos bounds checks) is the configuration where AddressSanitizer's documentation recommends-O1for reasonable speed, with-fno-omit-frame-pointerfor usable stacks.Proposal:
CMAKE_BUILD_TYPE=RelWithDebInfowith-O1 -g -fno-omit-frame-pointer, or keepDebugand overrideCMAKE_CXX_FLAGS_DEBUG.NDEBUGundefined in both, soasserts still fire.native-deep.yml's ASan-extension and coverage jobs. Clang source coverage also works at-O1, and that would ease the coverage job's 2.5 h runtime noted above.cpu-debugstays-O0for interactive stepping.RelWithDebInfo(-O2 -g) is the documented choice for debugging real runs.To check: the tidy and sanitize stages report the same findings before and after, and their CI wall time drops.
🤖 Generated with Claude Code