Forcing serialization on when writing out the build rule makes very
little sense. It was always "forced" on because we mandated a couple of
environment variables due to legacy reasons.
Add an attribute to RunTarget to say that a given target doesn't *need*
those environment variables, and let ninja optimize them away and run
the command directly if set.
That method had nothing specific to the backend, it's purely a Target
method. This allows to cache the OptionOverrideProxy object on the
Target instance instead of creating a new one for each option lookup.
In this case, the test fname might have an implicit extension and cannot
be found by `os.path.isfile()`.
We cannot use `shutil.which()` to handle platform differences, because
not all test fnames are executable -- for example Java jars.
The test representation does have an "is built" attribute which in
theory should work here, because all built targets definitely have their
full filename known to Meson, but it turns out to be misnamed. Rename it
correctly and add an actual "is built" attribute to check.
Tests which aren't built by Meson can be assumed to exist without
consulting their existence on the filesystem.
Fixes#10027
Using future annotations, type annotations become strings at runtime and
don't impact performance. This is not possible to do with T.cast though,
because it is a function argument instead of an annotation.
Quote the type argument everywhere in order to have the same effect as
future annotations. This also allows linters to better detect in some
cases that a given import is typing-only.
This was a nice idea in theory, but in practice it had various problems:
- On the only platform where ldconfig is expected to be run, it is
really slow, even when the user uses a non-default prefix and ldconfig
doesn't even have permission to run, nor can do anything useful due to
ld.so.conf state
- On FreeBSD, it bricked the system: #9592
- On cross builds, it should not be used and broke installing, because
ldconfig may not be runnable without binfmt + qemu: #9707
- it prints weird and confusing errors in the common "custom prefix"
layout: #9241
Some of these problems can be or have been fixed. But it's a constant
source of footguns and complaints and for something that was originally
supposed to be just "it's the right thing to do anyway, so just do it
automatically" it is entirely too risky.
Ultimately I do not think there is justification for keeping this
feature in since it doesn't actually make everyone happy. Better for
users to decide whether they need this themselves.
This is anyways the case for cmake and autotools and generally any other
build system, so it should not be too intimidating...
Fixes#9721
Store in TestSerialisation whether a particular test must always be logged
verbosely. This is particularly useful for long-running tests or when a
single Meson test() is wrapping an external test harness. In this case,
TAP can be used by the external harness and Meson will log each subtest as
it runs.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
For example:
```
meson builddir \
--native-file vs2019-paths.txt \
--native-file vs2019-win-x64.txt \
--cross-file vs2019-paths.txt \
--cross-file vs2019-win-arm64.txt
```
This was causing the error:
> ERROR: Multiple producers for Ninja target "/path/to/vs2019-paths.txt". Please rename your targets.
Fix it by using a set() when generating the list of regen files, and
add a test for it too.
In some cases, init variables that accept None as a sentinel and
immediately overwrite with [], are migrated to dataclass field
factories. \o/
Note: dataclasses by default cannot provide eq methods, as they then
become unhashable. In the future we may wish to opt into declaring them
frozen, instead/additionally.
Names used in init functions are sometimes pointlessly different from
the class instance attributes they are immediately assigned to. They
would make more sense if defined properly.
Since we scan all dependencies for build-only RPATHs now (which are
removed on install), we must take care not to add build-only RPATHs
pointing to directories that dependencies explicitly add -Wl,-rpath
link args for, otherwise the paths will get wiped on install.
Caught by LinuxlikeTests::test_usage_pkgconfig_prefixes
If a pkg-config dependency has multiple libraries in it, which is the
most common case when it has a Requires: directive, or when it has
multiple -l args in Libs: (rare), then we don't add -Wl,-rpath
directives to it when linking.
The existing test wasn't catching it because it was linking to
a pkgconfig file with a single library in it. Update the test to
demonstrate this.
This function was originally added for shared libraries in the source
directory, which explains the name:
https://github.com/mesonbuild/meson/pull/2397
However, since now it is also used for linking to *all* non-system
shared libraries that we link to with absolute paths:
https://github.com/mesonbuild/meson/pull/3092
But that PR is incomplete / wrong, because only adding RPATHs for
dependencies that specify a single library, which is simply
inconsistent. Things will work for some dependencies and not work for
others, with no logical reason for it.
We should add RPATHs for *all* libraries. There are no special length
limits for RPATHs that I can find.
For ELF, DT_RPATH or DT_RUNPATH are used, which are just stored in
a string table (DT_STRTAB). The maximum length is only a problem when
editing pre-existing tags.
For Mach-O, each RPATH is stored in a separate LC_RPATH entry so there
are no length issues there either.
Fixes https://github.com/mesonbuild/meson/issues/9543
Fixes https://github.com/mesonbuild/meson/issues/4372
Meson was passing only the first output and warning about it. To do this
easily, refactor construct_target_rel_path to return a list.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
For libraries installed to libdir, it's not expected to have rpath
hooked up. But for non-default libdirs, the path might not get searched
by default. `ldconfig -m <libdir>` is convenient here, as it will
programmatically add a new directory to search for shared libraries, so
the resulting installed programs work out of the box.
Include the dragonfly BSD platform name, which doesn't match the 'bsd'
catch-all pattern.
This previously worked because we were accidentally doing this via
mutation. However, doing this via mutation is not a good way to do it,
we should be explicit.
Fixes#9542
This replaces the absolute hack of using
```
install_subdir('nonexisting', install_dir: 'share')
```
which requires you to make sure you don't accidentally or deliberately
have a completely different directory with the same name in your source
tree that is full of files you don't want installed. It also avoids
splitting the name in two and listing them in the wrong order.
You can also set the install mode of each directory component by listing
them one at a time in order, and in fact create nested structures at
all.
Fixes#1604
Properly fixes#2904
This really is more of a struct than a dict, as the types are disjoint
and they are internally handled, (ie, not from user input). This cleans
some things up, in addition I spotted a bug in the ModuleState where the
dict with the version and license is passed to a field that expects just
the version string.
We have a lot of these. Some of them are harmless, if unidiomatic, such
as `if (condition)`, others are potentially dangerous `assert(...)`, as
`assert(condtion)` works as expected, but `assert(condition, message)`
will result in an assertion that never triggers, as what you're actually
asserting is `bool(tuple[2])`, which will always be true.
Eli Schwartz
Xavier Claessens
Paolo Bonzini
Nirbheek Chauhan
Dylan Baker
complexoctonion
Pablo Correa Gómez
Crend King
Jussi Pakkanen
Filipe Laíns