The `links` annotation in `Cargo.toml` is intended to ensure that in the
crate graph there's at most one crate that's an implementation of
some sort concept.
This can make sense in some scenarios, most prominent of which is when
the crate defines `#[no_mangle]` symbols (e.g. by compiling a vendored C
library.) In that situation linking a binary that depends on two
versions of the library cannot work because of colliding symbol names.
There does not appear to be a similar reason to impose such a
restriction on the users of `gstreamer-sys` and similar, however. All of
these crates link to a system library, they do not define any
`#[no_mangle]` symbols nor they vendor and build C libraries as part of
their build process. All they do is linking to a system library. Most
likely all the different versions of the bindings will link to the exact
same library too.
I haven't seen any global resources that these bindings use to ensure
soundness of the library, either.
Gir now uses analyzed objects to generate documentation, and to know
exactly what is available. Additionally, this allows more bindings to
be generated.
It is common for GStreamer crates to reexport all their
(GStreamer-related) dependencies, and the GL windowing crates should not
be an exception to this rule. Furthermore EGL documentation references
`crate::gst_gl::GLDisplayType` in the crate root which would be
unreachable otherwise.
This version adds a `--strip-docs` flag to `generator.py`, used in
conjunction with `--strip-docs --embed-docs` to clean documentation
first before re-embedding it (otherwise the same text would show up
multiple times). It is also used in the CI to check that no
documentation disappears on stripping, ie. all documentation is properly
annotated with `// rustdoc-stripper-ignore-next`.
When importing the prelude of a crate like `gst` the `glib` prelude is
provided too. Shedding these imports saves quite a few lines and
adheres to keeping it simple; we're not reexporting base/parent preludes
for no reason :)
For cleanliness the prelude module only needs to reexport preludes from
direct, "top-most" crates, which themselves take care of reexporting
preludes from its dependencies again. This shaves off some code while
maintaining the same set of exports.
This was temporarily allowed by Value trait refactoring, but the root
cause of the unused imports has been found (`glib::ToValue` for property
getters) and fixed in https://github.com/gtk-rs/gir/pull/1117.
In gir it was brought up [1] that some traits (in particular
`*ExtManual`) are exported from the crate root in addition to the
prelude, cluttering the environment unnecessarily. This commit removes
all these reexports, leaving those in prelude (that were already there)
only.
After this commit everything matching `Ext(Manual)?\b` in `lib.rs` sits
within `pub mod prelude {};`.
[1]: https://github.com/gtk-rs/gir/pull/1111
The `lgpl-docs` and documentation embedding step is now solely invoked
from `./gir/generator.py` in the CI, and does not need the embed/purge
build features anymore.
The `lgpl-docs` crate and all precompiled (easy to get outdated!)
docmentation files are removed in favour of being generated at runtime
by the improved generator, both during local development as well as in
the CI.
The previous commit reinstantiated `girs_dir`, and this commit applies
the rename to `girs_directories` and adds the new/missing
`gst-gir-files` directory where GStreamer-specific `.gir` files live.
This reverts commit 7f9fcb09e2.
`generator.py` is in the process of being moved to a generic place in
the `gir` repository for reuse across crates. This means `-d` cannot be
passed for our GStreamer-specific `gst-gir-files` directory anymore, and
should be configured from `Gir.toml` instead.
All these params are marked const in gstreamer-base 1.20 but their
results will only show here when the resulting g-ir files are imported.
For now the mutability is overridden in Gir.toml, and this commit should
be reverted when the 1.20 bindings are imported.
This includes mut overrides to prevent the caller from borrowing memory
as mutable. That enforces all kinds of ownership restrictions inside
the miniobject logic when only the underlying data for this memory needs
to be mutable.
Copy functions have been marked unsafe because tex_id might not be valid
and result in UB.
GLMemory input/output objects cannot be passed as mutable:
to_glib_none_mut checks if the miniobject is writable, even though the
underlying implementation should really only care about whether the
_data_ owned by this miniobject is.
(This temporary GLMemory object references its parent buffer object
that is also referenced externally, resulting in two references making
the object itself immutable)
Furthermore take care of automatically calling `add_rgba_pad_templates`
in `class_init` if `ADD_RGBA_PAD_TEMPLATES` is set to `true` (the
default).
Also enable all dependencies used by this type. Copying
AllocationParams is possible without borrowing the source mutable.
The offset and size parameters of `memcpy` are unchecked and may result
in out-of-bounds reads/writes, which is why this function is marked as
`unsafe`.
This class only contains two functions with a version requirement of at
least 1.16, but the class is used by GLFilter without any version
restrictions.
As collected with:
xmlstarlet sel -t -v "//_:field[@name = \"parent_class\"]/_:type/@name" gir-files/GstGL-1.0.gir | uniq
This adds GstBase.BaseTransform for GLBaseFilter and GLFilter, and
Gst.Element for gst::panic_to_error.
Gir now prints all directories and their hashes in the version file and
comments; useful now that gstreamer-rs is being generated from both
gir-files/ and gst-gir-files/ submodules.
We were already using `gir -d` and especially now that our files are
separated across two directories that are relative to the directory
containing Gir.toml this only becomes cumbersome. Besides `gir` lacks
functionality to normalize the path, leading to ie.
gstreamer-gl/egl/sys/../../../gir-files in the version comment as a
result.
These functions have been checked to never return NULL. The overrides
can hopefully be gone again when C contains better annotations and
trust_return_value_nullability is enabled.
