The old caching was a mess of spaghetti code layered over pasta code.
The new code is well-commented, is clear about what it's trying to do,
and uses a blacklist of keyword arguments instead of a whitelist while
generating identifiers for dep caching which makes it much more robust
for future changes.
The only side-effect of forgetting about a new keyword argument would
be that the dependency would not be cached unless the values of that
keyword arguments were the same in the cached and new dependency.
There are also more tests which identify scenarios that were broken
earlier.
All our cached_dep magic was totally useless since we ended up using
the same identifier for native and cross deps. Just nuke all this
cached_dep code since it is very error-prone and improve the
identifier generation instead.
For instance, this is broken *right now* with the `type_name` kwarg.
Add a bunch of tests to ensure that all this actually works...
Closes https://github.com/mesonbuild/meson/issues/1736
Currently only strings can be passed to the link_depends argument of
executable and *library, which solves many cases, but not every one.
This patch allows generated sources and Files to be passed as well.
On the implementation side, it uses a helper method to keep the more
complex logic separated from the __init__ method. This also requires
that Targets set their link_depends paths as Files, and the backend is
responsible for converting to strings when it wants them.
This adds tests for the following cases:
- Using a file in a subdir
- Using a configure_file as an input
- Using a custom_target as an input
It does not support using a generator as an input, since currently that
would require calling the generator twice, once for the -Wl argument,
and once for the link_depends.
Also updates the docs.
In this test, we try to manually link against the generated library to
create an executable and then run it to verify that it works.
Also test for all possible library versioning in the versioning tests on
Windows. Even though they yield the same dll naming, we should still
test it.
Just checking that the version retrieved from the pkg-config file
matches is not enough. It's nearly tautological since it just checks
that we aren't returning garbage in dep.version(). Instead, check in the
test executable that the pkg-config version retrieved matches the
ZLIB_VERSION exported by zlib.
Also don't require zlib 1.2.8 since RHEL (EPEL) 7 still has 1.2.7 and we
don't really need 1.2.8 anyway.
Sometimes we want to restrict the acceptable versions to a list of
versions, or a smallest-version + largest-version, or both. For
instance, GStreamer's opencv plugin is only compatible with
3.1.0 >= opencv >= 2.3.0
In the case the main project set a subproject for a dependency another
subprojects uses, that other subproject should rather use the first
subproject rather that using native dependency.
For example in gst-all we set all GStreamer modules as subprojects
and, gst-plugins-base is set after gstreamer core, and
we want gst-plugins-base to always use GStreamer core from the subproject
and not the possibly avalaible native one.
And remove the InternalDependencyHolder class.
In some cases we need to know the type of dependency we are
dealing with. For example in GStreamer if the dependency
is not an internal one, then we need to get some env var
from pkg-config to know where to find some plugins necessary
to run some tests.
It is extremely common to need to know within a given dependency if
a given header, symbol, member, function, etc exists that cannot be
determined from the version number alone.
Without passing dependency objects to the various compiler/linker
checks and with many libraries headers/libraries being located in
their own subdirs of the standard prefix, the check for the library
would not find the header/function/symbol/etc.
This commit allows passing dependency objects to the compiler checks so
that the test program can be compiled/linked/run with the necessary
compilation and/or linking flags for that library.
Also add new tests for the platform-specific and compiler-specific
versioning scheme.
A rough summary is:
1. A bug in how run_tests.py:validate_install checked for files has been
fixed. Earlier it wasn't checking the install directory properly.
2. Shared libraries are no longer installed in common tests, and the
library name/path testing is now done in platform-specific tests.
3. Executables are now always called something?exe in the
installed_files.txt file, and the suffix automatically corrected
depending on the platform.
4. If a test installs a file called 'no-installed-files', the installed
files for that test are not validated. This is required to implement
compiler-specific tests for library names/paths such as MSVC vs MinGW
5. The platform-specific file renaming in run_tests.py has been mostly
removed since it is broken for shared libraries and isn't needed for
static libraries.
6. run_tests.py now reports all missing and extra files. The logic for
finding these has been reworked.
Nirbheek Chauhan
Dylan Baker
Jussi Pakkanen
Jon Turney
Thibault Saunier
Matthew Waters