The current setup is configuring the "affected projects" as well
as their dependencies, and run `ninja all` followed by
`ninja check-all`.
This is quite a pessimization for leaf project which don't need to
build and run the tests for their dependencies.
For example a patch affecting only MLIR shouldn't need to build
and run LLVM and clang tests.
This patch changes this by running checks only for the affected
project. For example a patch touching `mlir` is affecting `mlir`
and `flang`. However `flang` depends on `clang`. So the list of
projects to configure is `mlir;flang;clang;llvm;`, but we want
to test only mlir and flang ; we'll run only `ninja check-mlir
check-flang`.
In practice in this example running `ninja all` builds 5658 targets
and `ninja check-all` after that adds 716 more targets. On the other
hands `ninja check-flang check-mlir` results in 3997 targets total.
Concretely the contract with premerge_checks.py is changed so that
the expected argument for the --projects flag is only the list of
affected project, dependencies are automatically added.
Now "report" step combines result in a uniform way and processes unit test
results XML output. It works for sub-builds only started from the 'premerge'
pipeline, i.e. non-recursive. One downside is that now one has to wait until
all jobs have finished.
- Add instructions to setup python environment
- added option to do full report cycle but not call Phabricator
- use "annotations" to show build status. That lifts the need to filter ninja
and other output (thus `ph_no_filter_output` param removed) and output
everything. That is nice as script failures no longer lead to loss of logs.
- improved annotate() usability
- misc fixes
fixes#192
additionally:
- install python libs after scripts checkout, so we don't need to
rebuild images and restart agents only to add a new python dependency
- updated lib versions
- similar scripts checkout in steps
Previous approach had drawbacks:
- every step had to implement exporting of results in fixed format
- if step failed then failure will not be detected
Now report step will fetch results directly from Buildkite.
Agents have to be updated to have BUILDKITE_API_TOKEN env.
Previously list of projects was resolved when job run at target OS.
Now project resolution is moved to the premerge pipeline and logic
should be there.
Now it's possible to allow sub-projects to define own checks and skip
"generic" ones.
To properly accomodate affected projects that might not have special
treatment we:
1. extend the set of affected projecs with dependent (e.g. add 'libc' if
'clang' was modified)
2. add custom steps for projects that define own workflow. At the moment
it's only libcxx and it has a custom trigger pipeline so it's noop.
3. add dependent projects and run generic check on them.
To illustrate: imagine that we have a dependency graph:
llvm -> clang -> openmp
and only clang was modified in a diff; also clang defines own checks.
Thus list of affected projects will be [clang, openmp].
After adding custom checks and removing their projecst: [openmp].
After adding dependencies: [llvm, clang, openmp]. Generic linux /
windows checks will be run on thouse 3 projects.
So as you can see in some scenarios projects with custom checks will
still go through generic checks.
Note that clang-format and clang-tidy checks are run only for "generic"
checks at the moment.
