[echoix]

Summary

• Adds opt-in clang -fprofile-instr-generate -fcoverage-mapping -fprofile-update=atomic instrumentation to the macOS build (macos_install.sh) and to the Linux pytest build (build_ubuntu-24.04.sh), gated by GRASS_LLVM_COVERAGE=1. macOS keeps its native clang; Linux switches to clang when opted in.
• Sets LLVM_PROFILE_FILE=$PWD/llvm-cov-raw/grass-%p-%9m.profraw via GITHUB_ENV before the test steps, so every xdist worker and every C-tool subprocess writes its own .profraw file. This eliminates the write-collision that made re-runs non-deterministic on files the tests never touched.
• Adds utils/llvm_coverage_export.sh to merge .profraw → .profdata (llvm-profdata merge -sparse, deterministic for a given input set) and export lcov for Codecov, filtering system, conda, and build-tree paths. The -object list is filtered by ELF/Mach-O magic bytes so executable Python scripts don't break llvm-cov.
• Adds utils/coverage_disk_usage.sh to log .profraw count/size and df -h before combine, so a run approaching the runner storage budget is visible in the log.
• Fixes a latent ordering bug where utils/coverage_mapper.py ran before coverage combine and could not read a .coverage that did not yet exist in parallel mode. The mapper now runs after combine.
• Adds coverage, pytest-cov, llvm-tools to the macOS conda env; clang, llvm to the Linux apt list.
• Wires two new Codecov flags on macOS (macos-pytest-python, macos-llvm-c) and one on Linux (ubuntu-llvm-c). The macOS combine+upload runs after the gunittest step so .profraw from both pytest and gunittest are included.
• Gitignores *.profraw, *.profdata, coverage-llvm.lcov, llvm-cov-raw/, coverage_html_report/. The coverage-prep step also find . -name default.profraw -deletes stray files the instrumented build itself emits (Makefiles run each tool with --interface-description to dump XML).

Determinism investigation

The previous attempt saw coverage of untouched files change between runs. I reproduced it locally against an instrumented HEAD build, then narrowed and fixed:

1. Minimal repro (outside GRASS): 8 parallel processes writing to a single default.profraw produced 20 vs 12 counts on identical inputs. Adding %p-%9m to LLVM_PROFILE_FILE made two runs bit-identical (32/32).
2. Full GRASS repro: 6 parallel grass --tmp-project --exec sessions each calling g.*, r.mapcalc, r.info, r.univar. Two runs produced 244 differing lcov lines — every one inside r.univar_main.c process_raster.omp_outlined_debug__, i.e. #pragma omp parallel for. That's LLVM's non-atomic coverage counters racing between threads.
3. Fix: -fprofile-update=atomic in the coverage CFLAGS. Verified locally in isolation (a small OpenMP loop went from 284296/304669 counts across runs → byte-identical) and against GRASS: two runs of the same 6-way parallel workload are now bit-identical on line coverage (DA:). The only residual diff is workload-timing noise in parser.c (±1 line hit between runs of the same command set) and inherent floating-point reduction order in r.univar's lock-protected min/max reduction, both of which reproduce identically without any coverage flags in play.

Test plan

• macOS workflow builds with GRASS_LLVM_COVERAGE=1 and produces coverage-llvm.lcov non-empty
• macOS Codecov shows macos-llvm-c and macos-pytest-python flags with non-zero coverage
• pytest workflow (Ubuntu, clang) builds and produces coverage-llvm.lcov non-empty
• Linux Codecov shows ubuntu-llvm-c alongside the existing pytest-python-<ver>
• Disk-usage step reports .profraw count/size within runner budget (<1 GB expected)
• coverage combine runs before coverage_mapper.py and no ordering warning appears

Large use of AI assistance (Claude Code) helped draft the workflow diffs, the utility scripts, and the local determinism reproduction. Human: EC.

---

Generated by Claude Code