Hugging Face's logo

Datasets:
sjkim1127
/
decbench-cpp-qualification

Tasks:
Other
Tags:
code
cpp
reverse-engineering
decompilation
cross-platform
elf
License:
Dataset card Files
xet
 Community
1
decbench-cpp-qualification / qualification-v2 /README.md
sjkim1127's picture
sjkim1127
Make the unchanged 45-image qualification pilot available on main
fa19039 verified
preview code
|
Raw History Blame Contribute Delete
6.91 kB

C++ qualification v2 staging

This is an unofficial, versioned curation pipeline. Raw corpus files remain unchanged. A staged binary is not automatically a qualified target, a scored function, or an official ranking entry.

Initial matrix

The counting unit is project × platform. There are 15 planned cells, 10 unique projects and 45 planned linked images.

Platform Project and selected target
Linux arm64 The Powder Toy powder; Snappy libsnappy.so.1.2.2; LevelDB libleveldb.so.1.23.0; SSDB ssdb-server; mergerfs mergerfs; OpenCC opencc
Windows x86_64 Detours withdll.exe; WinSparkle WinSparkle.dll; Ninja ninja.exe; Snappy snappy.dll; OpenCC opencc.exe; Tracy tracy-profiler.exe
macOS arm64 Ninja ninja; Snappy libsnappy.dylib; OpenCC opencc

All cells use O0, O2 and O2-noinline. Platform/compiler/architecture are separate fields. These modes are controlled compiler profiles, not promises of identical transformations across compilers. The initial matrix contains userspace systems software, a server, a GUI application, CLI programs and libraries. It does not contain a kernel or driver target. Future expansion should fill that gap.

Stage and qualification contract

  1. A candidate plan records exact repository revisions and linked target names.
  2. Packaging checks verify files, hashes, debug identities, unchanged executable code and isolated debug-free inputs.
  3. Oracle checks retain original sources, preprocessed TUs, build commands and concrete project function identities.
  4. Complete qualification additionally requires missing provenance/ownership gates to pass.
  5. Scoring requires platform-native identity/type adapters and an agreed C++ type policy. No score is assigned merely because a gate cannot run.

scoring_ready=false and qualified=false remain explicit in the preview. Inherited raw artifacts may lack full per-TU compiler/link command evidence. The DWARF evidence exporter reuses the pinned DecBench reader at 8eeceb82088e97ace88d15a592db1ee76ab37384. Its legacy variable summaries are not a complete source-level C++ type model. Original source and full debug information remain available for future policies. Source CU ownership in inherited raw evidence is weaker than full per-declaration ownership; do not count its header-defined dependency functions as a proven project-only population.

Observed packaging defects

  • Hub qualification revision d211137b0471b6534ccebef08677dcd947e0a6a4 has 24 files. Its legacy manifest has 1,086 paths, including PDB/object files in the binaries array; 1,065 referenced paths are absent. This legacy manifest cannot establish corpus completeness.
  • Ninja upstream enables Release LTO after a global IPO setting. An initially empty dSYM passed dwarfdump --verify; a native source function count exposed the problem. The macOS recipe uses a custom Qualification build configuration, preserving the upstream source tree.
  • Inherited SSDB O0 DWARF described optimized LevelDB dependency code, with no confirmed SSDB core functions. The SSDB recipe controls its actual CFLAGS and builds only the selected server and required libraries.
  • Windows PE/PDB imports match CodeView GUID and Age, and preserve executable section bytes in isolated inputs. Native PDB coordinates retain section/offset evidence; they are not borrowed from DWARF names.
  • Tracy also overrides a global IPO setting in its own configuration. The Windows recipe passes its explicit NO_LTO=ON option. The qualifier rejects project compilands marked LTCG.

Reproduce

Use Python 3.12 with huggingface_hub, pyelftools and lief; CMake; LLVM tools; Xcode for native macOS builds. Set PYTHONPATH to a checkout of the pinned DecBench evaluator for the DWARF exporter.

python scripts/prepare_qualification_sources.py --root /path/to/work --plan targets/qualification-v2/pilot.json
python scripts/build_qualification_macos.py --root /path/to/work --plan targets/qualification-v2/pilot.json
python scripts/stage_qualification_v2.py --root /path/to/work --plan targets/qualification-v2/pilot.json --raw-manifest /path/to/frozen/raw/manifest.json --local-raw /path/to/raw/cpp-compiled
python scripts/import_qualification_msvc.py --root /path/to/work --plan targets/qualification-v2/pilot.json
python scripts/verify_qualification_inputs.py --root /path/to/work
python scripts/inspect_qualification_dwarf.py /path/to/status.json --stage /path/to/work/staged/qualification-v2 --sources /path/to/work/sources
python scripts/seal_qualification_v2.py --root /path/to/work --plan targets/qualification-v2/pilot.json

The existing Windows evidence importer defaults to the Detours/WinSparkle layout of GitHub Actions run 38208011147. New builds use the same PROJECT/MODE/ evidence layout and explicit run provenance:

gh workflow run cpp-qualification-v2-windows.yml --repo sjkim1127/Decbench_SampleSet --ref codex/cpp-qualification-v2 \
  -f 'projects=["ninja","snappy","OpenCC","tracy"]' -f 'modes=["O0","O2","O2-noinline"]'
# Download each successful job's artifact into evidence/PROJECT/MODE before importing.
gh run download RUN_ID --repo sjkim1127/Decbench_SampleSet --name qualification-v2-ninja-O0 --dir evidence/ninja/O0
python scripts/import_qualification_msvc.py --root /path/to/work --plan targets/qualification-v2/pilot.json \
  --projects ninja snappy OpenCC tracy --evidence-root /path/to/evidence --github-run RUN_ID

The dedicated branch push trigger starts the full matrix; dispatch inputs support focused rebuilds. Jobs do not use continue-on-error. A failed build, missing PDB, failed preprocessing replay or failed PDB gate cannot become a passing artifact. build_qualification_windows.ps1 and preprocess_qualification_msvc.py require native Windows/MSVC. Their execution evidence is retained separately from preparation on macOS. New evidence includes pinned tracked sources, effective compile commands, /P replay outputs and the hosted runner image version. PDB source ownership is compiland-based; template/header definitions still require per-procedure ownership before complete qualification. Tracy's fetched dependency TUs retain preprocessed expansions, but dependency commit resolution, original archives and license files remain a separate provenance gate. A tracked main-tree source archive does not contain CPM's fetched source trees. Docker recipes build_qualification_tpt.sh and build_qualification_ssdb.sh require the pinned revisions through environment variables and the workspace mounted at /work.

Independent claims

  • Successful builds do not establish runtime parity.
  • LLVM debug verification does not establish project ownership or full debug completeness.
  • File/hash checks do not establish metric readiness.
  • No new Fission production changes or decompiler scores are part of this curation task.