| Age | Commit message (Collapse) | Author |
|---|
|
|
|
|
|
|
|
|
|
... as otherwise we will use the same string in two ways: literal and as a value
to the expanded, making proper quoting impossible. Moreover, it is not
necessary to expand TOOLCHAIN: pointers into the toolchain can be passed
through the "bin dirs" part of the toolchain.
|
|
dependencies
|
|
|
|
dependencies
|
|
|
|
|
|
When a shared library is built that depends on other shared libraries, it
instructs its consumers via the "run-link-args" to link only against this
library and not also against its dependencies.
|
|
When a shared library has static-library dependencies, it is linked against
these static libraries including all their link arguments. Thus, the
"link-deps" as well as the "link-args" provider of the shared library should be
empty.
|
|
|
|
... as using the built-in "[]" directly is cleaner and equally
readable.
|
|
|
|
... provided the configuration variable "LINT" is set.
|
|
|
|
... and honor those in ["lint", "targets"].
|
|
|
|
|
|
|
|
... collecting test results similar to the way "install" rules
are typically used; however in such a way that the provider "lint"
is forwarded.
|
|
|
|
While there, also properly transition "srcs" and "private-hdrs"
to the host version.
|
|
In case many "targets" are given, the field "lint" will contain
all the concatenation of the provider "lint" of the given targets.
There is, however, not need to lint the same file in the same
context twice, so deduplicate the targets first. While this does
not change the amount of lint actions carried out (as equal actions
are handled only once anyway), it keeps the summary clean by not
having dulicate entries.
|
|
|
|
|
|
If the configuration variable "LINT" is set, also provide information
on compile actions and header files (with preprocessing as described
command, in particular also providing the correct flags) in correct
dependency context. In this way, lint rules can request the needed
information for performing their checks.
|
|
... for certain fields, in particular, the "components".
|
|
... however not their runfiles, as those become the runfiles
of the resulting library.
|
|
|
|
... instead of hard-coding ["cqs"].
|
|
|
|
The existing rules allow a shell test script to explicitly 'keep'
created files in the resulting output directory, to be accessible
to any dependent target, thus allowing a chaining of data across
test scripts. This field however excludes directories, whose
content would need to be added one file at a time to the 'keep'
field.
Add instead also a 'keep-dirs' field to ["shell/test", "script"]
rule which allows the explicit staging of given directory paths the
test is expected to generate.
|
|
|
|
... to avoid staging conflicts with the toolchain
|
|
In this action we support a user provided toolchain, hence all the
components of the library need to go into a subdirectory to avoid
staiging conflicts.
|
|
... i.e., dependencies that are to be included into the library
itself. In this way, a large library (convenient for a user to have
single library to care about) can be defined as a collection of
smaller libraries.
Technically, components are like public dependencies on libraries
transitioned to object libraries with the following differences
- the header files (i.e., runfiles) of the components become header
files of the resulting libary, and
- the objects (i.e., artifacts) of the components become objects
of the library rather than link dependencies.
To achive the transfer of the object to the requesting library,
an object library can be instructed to drop the objects from the
link arguments; in order to continue to support tranditional object
libraries in the style of, e.g., cmake, this is done by a different
configuration variable that is transitioned as well. In particular,
the object-library test case (using a configure target) can be
left unchanged.
|
|
... instead of a property of the library itself. An object library is
not a meaningful concept in itself; it only exists, because a consumer
wants to link the library in its entirety. But consumer-specified
properties should be propagated through configuration transitions
and the definition of the library should not care about how it is
consumed; this is also the approach we follow with respect to building
a library position independent. As oposed to position-independent
building, however, the property of being included unconditionally
is not propagated transitively.
|
|
|
|
... of passed tests (as only for those, the time is meaningful).
Given that we read the timing information anyway, if available, we
can as well report more useful information.
|
|
... and add a rule allowing the summarizer to specify what it needs.
|
|
... as test meta data. Tests are executed in an unspecified directory,
assuming pass or fail is independent of the location where the test
is run. While this generally is true, test logs often contain the
working directory. So, in order to more easily compare different
execution orders of a potential race condition, it can be desirable
to compare logs "up to the execution directory". This, however,
requires that this directory is recored in the first place. Do so.
For consistency of the output format, also have a (fixed) artifact
pwd in the summary report.
|
|
|
|
|
|
|
|
|
|
If a test is run several times (as set by RUNS_PER_TEST), a summary of
the individual test runs is computed using a summarizer as configured
in the target layer of the rules. As the inputs for computing that
test summary are all the individual test runs, that action has a
large number of files as input, including a large number of identical
files, e.g., the ones indicating the outcome of an individual run.
Therefore, allow setting additional remote-execution properties
allowing to dispatch that action to a suitable end point.
|
|
... by running a shell prolog with the provided shell to set
PATH to the absolut paths of the shell-provided bin dirs.
|
|
|
