Repository navigation
Conversation
The absolute RUNPATH tied bin/kepler-formal to the directory it was built in, so a moved or re-pointed checkout loaded another build's libnaja_*.so. Use $ORIGIN/../lib, and @loader_path/../lib on macOS. Fixes The-OpenROAD-Project#4717. Signed-off-by: Skot <48910715+saikatchatrg@users.noreply.github.com>
There was a problem hiding this comment.
Code Review
This pull request replaces the manual CMake compilation of the kepler-formal tool and its dependencies with a Bazel-based installation using bazelisk. It removes the tools/kepler-formal submodule, cleans up related dependencies from etc/DependencyInstaller.sh, and introduces Bazel configuration files under etc/kepler-formal. Feedback on these changes includes ignoring Bazel build outputs in .dockerignore to keep the Docker build context small, correcting the ignored symlink name in .gitignore to match the workspace name orfs-kepler-formal, and avoiding the use of a personal fork as a Bazel registry in .bazelrc due to security and reliability risks.
A small Bazel module that only pins kepler-formal; ORFS itself does not
become a Bazel project. From this folder,
bazelisk run //:install -- <prefix>
installs bin/kepler-formal with everything it loads next to it.
Bumping kepler-formal is the version line in MODULE.bazel. README.md
says how this maps to DependencyInstaller.sh, and which lines go away
as kepler-formal's registry entry and install target merge upstream.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
Instead of building kepler-formal with CMake, run its install target from etc/kepler-formal, as Build.sh already builds OpenROAD with Bazel. This drops the per-architecture -march choice, the EL9 TBB define, the LTO flags, the build and install RPATHs (The-OpenROAD-Project#4717) and the gcc-toolset-13 setup that only kepler-formal needed. The binary still lands in tools/install/kepler-formal/bin/kepler-formal, and it loads its libraries from next to itself, wherever the install is moved. The Docker builder copies etc/kepler-formal into the image, and the tools image tag covers it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
build_openroad.sh installs kepler-formal from the version pinned in etc/kepler-formal, so the submodule, and the five submodules inside it, are no longer checked out. Its build directory leaves .dockerignore and its commit leaves the tools image tag. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
kepler-formal no longer builds against host packages: capnproto and TBB on Ubuntu, EL and Homebrew, the oneTBB source build on EL8 and gcc-toolset-13 on EL8 go. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
d078245 to
2aa8f0a
Compare
|
@vvbandeira ready for review. The GitHub now serves a different archive for that slang commit than the one recorded in the Bazel Central Registry entry The step this PR changes passed in the same Jenkins runs. On both Ubuntu 26.04 and 24.04, the tools image ran |
|
@maliberty @povik @vvbandeira Regarding unrelated pr-merge/head failure with github sv lang .tar.gz updated hash. Once we're switched to Bazel only in OpenROAD, we can get all yosys+sv-lang, etc. into bazel bcr instead of git submodules. Then we're no longer vulnerable to hash upates on github and it is bazel idiomatic. That's a bit of an ordeal to untangle, I only have a vague sense of all the details, but worth it as it makes life for everyone simpler going forward. |
🔍 QoR checkMetrics reflect the PR merge build — i.e. what will land on the target branch. Commit 62 design(s) checked — 0 with regression(s), 0 without a comparable baseline. |
|
Checked #4732 at 2aa8f0a for the #4717 case. #4727 fixes the same RUNPATH with a one-line change to the CMake build this PR removes; I'll close it once this lands. |
|
Thanks for checking the #4717 case on a moved install, @saikatchatrg. The Agreed on closing #4727 once this lands: its RUNPATH fix is one line in a CMake build that this PR deletes. Why this replaces the CMake build rather than patching it:
|
|
Thanks for the explanation and for taking this on. Building kepler-formal through its own |
|
@nanocoh any concerns? |
| # TODO(keplertech/kepler-formal#292): delete this override once a | ||
| # kepler-formal version with the //:install target is on the registry, | ||
| # and move the version above to it. | ||
| git_override( | ||
| module_name = "kepler-formal", | ||
| commit = "99d66cac88978312853a10b82657781f8e99d843", | ||
| remote = "https://github.com/oharboe/kepler-formal.git", | ||
| ) |
There was a problem hiding this comment.
The git_override is gone: kepler-formal@0.5.0-20261026-a64adfaf comes from the BCR (bazelbuild/bazel-central-registry#11201 merged), and it has //:install. 65fab0f
Looks good as long as the regress is clean and we can occasionally use git_overide when bcr is not up to date. |
Yes, you still can. This can be convenient when making improvements that ORFS cares about and then using git_override() to pick them up until next release is available. You could also carry a patch in ORFS on top of an official release, which I sometimes prefer over git_override() because it is is naturally scrubbed out in a bazel bcr bump. git_override() should come with a comment as to why the release isn't being used. Which approach is better while waiting for a bazel bcr bump of a module all depends on the specifics of each circumstance; normally I'd only carry tiny patches and use git_override() if there are lots of updates. |
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
kepler-formal 0.5.0-20261026-a64adfaf is on the Bazel Central Registry and has the //:install target, so the git_override on a fork goes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
kepler-formal keeps its clang toolchain a dev_dependency on purpose, so a module that depends on it registers the toolchain itself; the README no longer says these lines go away. llvm moves to 0.8.26, the version kepler-formal builds with. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
kepler-formal 1.0.0-20261105-6870bd1a is on the Bazel Central Registry. Its .github/consumer-test registers the downloaded clang for x86-64 Linux only, so this module does the same and other platforms use the host compiler, as the README says. The dynamic_mode comment gives kepler-formal's reason for the flag. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
|
@nanocoh Good to go? |
|
If passing all tests so looks good to me |
vvbandeira
left a comment
There was a problem hiding this comment.
Two items from the earlier review are still open:
-
EL8 needs gcc-toolset-13 for Yosys. Yosys is still built with CMake before the Kepler step, and its pinned source requires C++20. The EL8 package list installs GCC 8, and the OpenROAD installer now pinned on master is Bazel-only and installs no host compiler. Kepler's downloaded clang does not apply to the Yosys build. Please keep the gcc-toolset-13 install in DependencyInstaller.sh and source it in build_openroad.sh before Yosys (and keep it active for the bazelisk run), and update the gcc-toolset-13 row in etc/kepler-formal/README.md.
-
Dependency notices are no longer in the image. kepler-formal's package installs only its own LICENSE.rst and LICENSES/, while the binary links yaml-cpp, spdlog, kissat and naja. The Docker license collector only searches tools/, so it cannot find notices in the Bazel cache. Please package the dependency notices upstream or collect them here.
Smaller items:
- .dockerignore: please exclude tools/kepler-formal/. An existing checkout keeps the old submodule tree, which then enters the Docker context and the license collector.
- etc/kepler-formal/.bazelrc: the comment says clang "must win resolution on Linux", but the line below only sets dynamic_mode=off. Clang is chosen by register_toolchains in MODULE.bazel, on x86-64 Linux only.
The earlier -cmake/Bazelisk comment no longer applies: the OpenROAD installer pinned on master has dropped -cmake.
Yosys is still built with CMake before kepler-formal, and the pinned Yosys sets CMAKE_CXX_STANDARD 20, which the EL8 default GCC 8 cannot compile. The OpenROAD installer pinned here is Bazel-only and installs no host compiler, and the clang that kepler-formal downloads does not apply to the Yosys build. Keep installing gcc-toolset-13 on EL8 and keep sourcing it in build_openroad.sh, where it stays active for the kepler-formal install: on aarch64 no downloaded clang is registered and the host compiler is used. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
A checkout made before the submodule was removed keeps the old tools/kepler-formal tree. It would enter the Docker build context and the license collector, which searches tools/. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
The comment said the downloaded clang must win resolution on Linux, but the line below only links statically. Clang is chosen by register_toolchains in MODULE.bazel, on x86-64 Linux only. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
kepler-formal's //:install ships only its own LICENSE.rst and LICENSES/, while the binary links naja, yaml-cpp, spdlog, kissat and more, statically links libc++, and carries an embedded Python in its runfiles. The Docker license collector searches tools/ and cannot find notices in the Bazel cache. etc/kepler-formal/install_licenses.sh, run by build_openroad.sh after the install, takes the repositories cquery finds in the binary's target configuration and copies their license notices to share/kepler-formal/licenses/<repository>/ under the same prefix. Build tools that run in the exec configuration are left out. It fails if it finds none. On x86-64 Linux it installs the notices of 40 modules, 1.1 MB. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
|
Answering #4732 (review) point by point. master is merged in (e965a18, no conflicts).
Tested: |
Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
| [[ -n "${files}" ]] || continue | ||
| mkdir -p "${dest}/${repo}" | ||
| while IFS= read -r f; do | ||
| (cd "${external}/${repo}" && cp -rL --parents "${f}" "${dest}/${repo}/") |
There was a problem hiding this comment.
- macOS breaks at the license step. install_licenses.sh uses cp -rL --parents, which only exists in GNU cp, not macOS. build_openroad.sh --local still supports Darwin, so a Mac build would now fail there. Separately, Darwin builds kepler-formal with the host Apple clang (the downloaded clang is registered on x86-64 Linux only), and nothing shows that was tested. Either make the copy portable (rsync -R, or mkdir -p "$(dirname …)" plus cp) or say clearly that Darwin is unsupported for kepler-formal.
There was a problem hiding this comment.
Fixed in 1eba9d5 as a best effort from the review: install_licenses.sh now makes each directory with mkdir -p and copies with cp -RL, so it no longer needs GNU cp --parents. The README now says that on macOS, kepler-formal is built with the host Apple clang and that this is untested. I have no Mac to test on. Testing and PRs welcome.
install_licenses.sh used cp -rL --parents, which macOS's cp does not have, so build_openroad.sh --local would fail there at the license step. Make each notice's directory with mkdir -p and copy it with cp -RL, which both cp's accept. The README says that on macOS kepler-formal is built with the host's Apple clang and that this has not been tested. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
Head branch was pushed to by a user without write access
Bumping kepler-formal becomes a one-line edit: the version in
etc/kepler-formal/MODULE.bazel. You don't move a submodule, and you don't chase new system packages on each platform.A build breakage like #4717 is fixed and tested in kepler-formal's own CI, not patched in ORFS's scripts.
What ORFS no longer maintains
build_openroad.sh:-marchchoice per architecture;gcc-toolset-13on EL8.tools/kepler-formal, along with the five submodules inside it.etc/DependencyInstaller.sh:gcc-toolset-13on EL8.etc/DockerTag.shand.dockerignore.What doesn't change for users
./build_openroad.sh --localis run the same way.tools/install/kepler-formal/bin/kepler-formal.env.sh,KEPLER_FORMAL_EXEandLEC_CHECKare unchanged.How
build_openroad.shruns kepler-formal's own install target,bazelisk run //:install, frometc/kepler-formal/, the way it already builds OpenROAD. That folder pins kepler-formal and nothing else. ORFS itself does not become a Bazel project. Nothing new is needed on the host. The folder'sREADME.mdexplains how to bump kepler-formal and how each step maps to whatDependencyInstaller.shdid.Why this is possible now
kepler-formal now installs this way after a number of PRs to kepler-formal and the Bazel Central Registry. Its install target, keplertech/kepler-formal#292, and its registry entries, bazelbuild/bazel-central-registry#11201 and bazelbuild/bazel-central-registry#11583, are merged, so
MODULE.bazeltakeskepler-formal@1.0.0-20261105-6870bd1afrom the registry with no override.So far the way to install kepler-formal has been CMake plus host packages. That costs kepler-formal the same per-platform work it costs ORFS, so I expect kepler-formal to streamline its own install too. Each step upstream only makes this folder shorter, and the interface doesn't change.
How it was tested
CI tests
build_openroad.shon every supported platform, so locally I only ran smoke tests:etc/kepler-formal:bazelisk run //:install -- <prefix>, from a cold start with the registry version, installs on x86-64 Linux with no host packages.kepler-formal --versionreports 1.0.0, including from a copy of the install in another directory.bash -n build_openroad.shandbash -n etc/DependencyInstaller.shpass.🤖 Generated with Claude Code