python compiled extensions should never need to expose any symbol other
than PyInit_* which is declared with default visibility via
PyMODINIT_FUNC on supported compilers.
Thus, a reasonably sane default is to mark any other symbols as hidden,
while still respecting any manually specified visibility.
Gate this on the version of python itself, as not all versions decorate
PyMODINIT_FUNC properly.
On Python 3.8.x and later, if the imported module requires non-system DLLs that
are not installed nor bundled with the module package, os.add_dll_directory()
must be called on every path that contains the required DLLs, so that the module
can be imported successfully by Python.
Make things easier for people by calling os.add_dll_directory() on the
valid directories in %PATH%, so that such module checks can be carried out
successfully with much less manual intervention.
It was originally added because proper detection was not working on
Debian, but that has been fixed since. It was causing annoying warning
by default when prefix is /usr/local that can only be avoided by setting
options.
This reverts commit 79c6075b56.
# Conflicts:
# docs/markdown/snippets/devenv.md
# mesonbuild/modules/python.py
# test cases/unit/91 devenv/test-devenv.py
PYTHONPATH cannot be reliably determined. The standard use case for
installing python modules with Meson is mixed pure sources (at least
`__init__.py`) and compiled extension_modules or configured files.
Unfortunately that doesn't actually work because python will not load
the same package hierarchy from two different directories, one a source
directory and one a (mandatory) out of tree build directory.
(It kind of can, but you need to do what this test case accidentally
stumbled upon, which is namespace packages. Namespace packages are a
very specific use case and you are NOT SUPPOSED to use them outside that
use case, so people are not going to use them just to circumvent Meson
devenv stuff as that would have negative install-time effects.)
Adding PYTHONPATH anyway will just lead to documentation commitments
which we cannot actually uphold, and confusing issues at time of use
because some imports *will* work... and some will *not*. The end result
will be a half-created tree of modules which just doesn't work together
at all, but because it partially works, users attempting to debug it
will spend time wondering why parts of it do import.
For any case where the automatic devenv would work correctly, it will
also work correctly to use `meson.add_devenv()` a single time, which is
very easy to manually get correct and doesn't provide any significant
value to automate.
In the long run, an uninstalled python package environment will require
"editable installs" support.
There are two cases where we can assume we found the python dependency
with its requisite libraries using sysconfig:
- we found the library with find_library and are prepared to link to it
- the library is not actually part of the dependency, so its presence or
absence is irrelevant
In the latter case, we should consider it found if link_libpython is
False. Originally we did this, but the logic was inverted in commit
5b422fce87 in an unrelated change and
without explanation, likely by accident.
Normally this doesn't much matter, since a python invariably comes with
a predictably located libpython and the first condition evaluates true.
But that is not true for pypy, and in fact that is the reason the
link_libpython check was originally added in commit
1bd14b52b2.
Restore that original logic.
Fixes#8570
A bunch of files have several T.TYPE_CHECKING blocks that each do some
things which could just as well be done once, with a single `if`
statement. Make them do so.
The default behavior of installing relative to prefix may be unexpected,
and is definitely wrong in many cases.
Give users control in order to specify that yes, they actually want to
install to a venv.
This is particularly useful for projects that use meson as a build
system for a python module, where *all* files shall be installed into
the python site-packages.
If the found python returns None from sysconfig.get_config_var('LIBPC')
then we cannot (and don't) set PKG_CONFIG_LIBDIR from it. In fact, we
can virtually guarantee we won't find a PkgConfigDependency either,
because any python that doesn't have a LIBPC is presumably not installed
to the system pkg-config directory (maybe it's an isolated relocatable
install, maybe it just doesn't have pkg-config support for who knows
what reason).
Trying to find one anyway using pkg-config's builtin search paths can
unexpectedly succeed, though, by finding a completely unrelated python
installation installed to a system location, which isn't the one we are
actually building for.
Instead, return early so that we use the system dependency class
fallback.
While we are at it, add back the debug messages from #3989 which got
removed.
Virtualenvs do not have their Python DLLs etc in the `sys.prefix`
directory, but in the `sys.base_prefix` directory. This directory is
the same as `sys.prefix` if not in a virtualenv, so is safe for either
case:
https://docs.python.org/3/library/sys.html#sys.base_prefix
Hardcoding the name is fragile, and enabling it based on the existence of
/etc/debian_version (as the is_debianlike helper does) will result in
incorrect paths if the Python binary is not provided by Debian.
Using the deb_system distuils scheme instead makes sure we use the
install path from the current interpreter, which Debian could change
between releases, and gives us the correct value on Python installations
that are not provided by Debian (eg. deadsnakes, Github Action Python,
etc.)
Do notice, though, that there is still no guarantee that these are the
correct paths, as they assume all schemes paths have the install prefix
as a base, see #9284.
Signed-off-by: Filipe Laíns <lains@riseup.net>
The sysconfig paths are, by default, correct for every OS -- they are
supposed to follow the scheme that python knows about per default.
For some reason, this overrode the scheme to posix_prefix, which is the
default for posix OSes like linux and macOS, but wrong on Windows.
Simply deleting this entirely makes everything that used to work, still
work, and a couple new things start working.
rsplit(..., 1) always produces exactly one split, by design, there's no
need to then join a 1-element list via a generator comprehension after
extracting the end of it via pop. If this commit message sounds
confusing, then so was I when trying to figure out what this actually
did and if it needed extracting to PythonExternalModule.
These were spotted by mypy and pyright. One is a string where a Path is
expected, another other is a possibly unbound variable, and the third is
bound but unused variables.
Including not calling back into `Interpreter.func_*`, which is not a
good idea both from a type saftey and perforamance point of view.
Instead there's now a shared _impl method
mesonbuild.dependencies.__init__ exposes configtool, pkgconfig, cmake
and more in __init__.py, so there's no reason we should be tying
SystemDependency to the internal organization implementation of the
subpackage!
In the 2nd previous commit it took quite some effort to figure out that
the python module "does not exist" because of import errors while
refactoring something completely different.
The only advantage they have is they have the interpreter in arguments,
but it's already available as self.interpreter. We should discourage
usage of the interpreter API and rely on ModuleState object instead in
the future.
This also lift the restriction that a module method cannot add build
targets, but that was not enforced for snippet methods anyway (and some
modules were doing it) and it's really loose restriction as it should
check for many other things if we wanted to make it consistent.
Eli Schwartz
Chun-wei Fan
Matti Picus
Jussi Pakkanen
Xavier Claessens
Ralf Gommers
Matthew Brett
Thomas Heijligen
Christian Clauss
Filipe Laíns
Dylan Baker
Daniel Mensinger