Instead of connecting to the X server, let's run the Python tests with the testing backend, like we do for test-driver-rust, test-driver-cpp, etc.
When building the Python package, pass --features backend-testing to Cargo, to forward to enabling the backend-testing feature in i-slint-backend-selector. That consequently enables SLINT_BACKEND=testing.
This allows to store the result into a different cache property
than the one that comes out of solve_grid_layout, which fixes
a binding loop when the text of a widget depends on its row or col
value.
Build editable builds as cargo debug builds, not release. This is faster
(because we don't need LTO) and in theory gives additional coverage to
debug asserts.
The `rust_attributes` in langtype::Struct is a bit wierd as it is only
used for rust code generation, yet is stored in every Struct.
Instead, read the `node` while doing code generation like we do in C++.
The previous code was ignoring the error lexing the code in the
attribute, and now we add a compile_error!
Separate steps and remove nox, so that
1. We first set up the env and build slint **once**
2. We run the fast lint/ty/etc. checks
3. We run the python tests directly
The generated file offers the following two pieces of functionality:
- Convenient front-end to slint.load_file() by loading the .slint file
(replaces use of auto-loader)
- More importantly: Type information for exported properties, callbacks,
functions, and globals
On loading, the previously generated API is compared to what's in the
.slint file and if the .slint file was changed in an incompatible way,
an exception is thrown.
A Python test driver is added which performs three steps with each test
case:
1. Generate python stubs and checks that they can be loaded with the
python interpreter and that `uv ty check` passes.
2. Extract expected python API out of pyi sections and compares them.
3. Appends any python code from python sections and runs them (this is
combined with step 1)
Fixes#4136
- Make `FontHandle` opaque.
- Avoid memory allocations
- Reword documentation
- Avoid indirection to i_slint_core (prevent documentation to be
repeated)
ammends for https://github.com/slint-ui/slint/pull/9762
For the future Python generate, we need to know if a struct type is a
public Slint type, private perhaps, or custom. Instead of continuing
with string operations, an enum gives clarity what's what, and this also
emits removing some string matching to identify specific types.
Currently Ava is used to unit test the api/node project.
Vitest is already used elsewhere in the monorepo, integrates well with
Vite, is fast and automatically handles tests written in Typescript.
Replacing Ava would simplify how Typescript tests are performed.
This PR introduces Vitest to the project, sets it up to match Ava
(isolated process for each test) and ports just the window.spec.mts
file.
Later PR's will port the rest.
- Bundle translations for WASM/Android only (not for native builds)
- Load Noto Sans CJK fonts dynamically from GitHub at runtime
- Implement register_font_from_memory() with Result<FontHandle, RegisterFontError>
- Make API automatically available with std feature (no explicit feature flag needed)
- Export load_font_from_bytes() for WASM font loading from JavaScript
- Change from wasm_bindgen(start) to explicit main() call for better control
Fonts are loaded before app initialization to ensure proper CJK text
rendering across all major browsers without bundling font files.
Native builds load translations from lang/ directory at runtime.
WASM/Android builds bundle translations at compile time and detect
browser/system language automatically.
API is now available when std feature is enabled, since we always
have fontique with std. No need for experimental-register-font feature.
Coming from the QML world it was normal to just throw a bunch of QML
files into a single CMake target without any thinking. But with Slint, I
encounter into strange symbol conflict issues and thinking this was a
bug in our compiler.
In reality this was actually an acceptable limitation (see #2909) which
is fine... but this wasn't isn't mentioned anywhere in the CMake
documentation - so I never knew about this or the solution. I added a
new cautionary warning to let future developers know that in case of
symbol conflicts, they need to separate files into their own namespaces.
If you try to pass multiple files into slint_target_sources, some
generators may complain with a strange error like this:
ninja: error: build.ninja:226: multiple rules generate slint_generated_dialog_1.cpp
The reason for this was how the macro worked (and CMake scoping rules.)
While putting together the list of .cpp files to generate - which is 1
by default but is user-configurable - we append a list called cpp_num.
However we don't *reset* this list variable, so in each foreach
iteration we kept appending a new .cpp file and thus telling CMake we
were generating "slint_generated_dialog_1.cpp" twice, in addition to
whatever other file you had. We don't hit this issue with the other
variables like translation_domain_prop because they're set on every
iteration.
The fix for this is simple though, we can just set the variable to
empty.
* Add support for CSS conic-gradient 'from <angle>' syntax
Implement rotation support for conic gradients by adding the 'from <angle>'
syntax, which rotates the entire gradient by the specified angle.
- Add `from_angle` field to ConicGradient expression
- Parse 'from <angle>' syntax in compiler (defaults to 0deg when omitted)
- Normalize angles to 0-1 range (0.0 = 0°, 1.0 = 360°)
- Add `ConicGradientBrush::rotated_stops()` method that:
* Applies rotation by adding from_angle to each stop position
* Adds boundary stops at 0.0 and 1.0 with interpolated colors
* Handles stops outside [0, 1] range for boundary interpolation
- Update all renderers (Skia, FemtoVG, Qt, Software) to use rotated_stops()
The rotation is applied at render time by the rotated_stops() method,
which ensures all renderers consistently handle the gradient rotation.
* Add screenshot to rotated conic gradient docs example
Wraps the rotated conic gradient example in CodeSnippetMD to automatically
generate and display a visual screenshot of the gradient rotation effect.
This makes it easier for users to understand how the 'from' parameter rotates
the gradient.
* Make ConicGradientBrush fields private
The from_angle and stops fields don't need to be pub since:
- Rust code in the same module can access them without pub
- C++ FFI access works through cbindgen-generated struct (C++ struct members are public by default)
* Optimize ConicGradientBrush::rotated_stops to avoid allocations
- Changed return type from Vec to SharedVector
- When from_angle is zero, returns a clone of internal SharedVector
(only increments reference count instead of allocating new Vec)
- Removed break from duplicate position separation loop to handle
all duplicate pairs, not just the first one
- Updated documentation to match actual implementation
* Remove automatic sorting in ConicGradientBrush::new() to match CSS spec
- CSS conic-gradient does not automatically sort color stops
- Stops are processed in the order specified by the user
- Changed boundary stop interpolation logic to use max_by/min_by
instead of relying on sorted order
- This allows CSS-style hard transitions when stops are out of order
* Move conic gradient rotation processing to construction time
Major changes:
- ConicGradientBrush::new() now applies rotation and boundary stop
processing immediately, instead of deferring to rotated_stops()
- Removed rotated_stops() method - backends now use stops() directly
- Changed to LinearGradientBrush pattern: store angle in first dummy stop
- Added angle() method to retrieve the stored angle
- Maintained #[repr(transparent)] by removing from_angle field
- All backends updated: rotated_stops() -> stops()
- Qt backend
- Skia renderer
- femtovg renderer
- Software renderer
C++ API changes:
- Added FFI function slint_conic_gradient_new() for C++ to call Rust's new()
- Updated make_conic_gradient() to call FFI function instead of manually
constructing SharedVector
- Ensures C++-created gradients get full rotation processing
Benefits:
- Eliminates per-frame rotation calculations
- Reduces memory usage (no from_angle field)
- Consistent with LinearGradientBrush design
- C++ and Rust APIs now produce identical results
* Change ConicGradientBrush::new() from_angle parameter to use degrees
- Changed from_angle parameter from normalized form (0.0-1.0) to degrees
- Matches LinearGradientBrush API convention (angle in degrees)
- Updated internal conversion: from_angle / 360.0 for normalization
- Stores angle as-is in degrees in the first dummy stop
- FFI function slint_conic_gradient_new() passes degrees directly
Example usage:
ConicGradientBrush::new(90.0, stops) // 90 degrees
LinearGradientBrush::new(90.0, stops) // 90 degrees (consistent)
* Fix ConicGradient color transformation methods to preserve angle
Changed brighter(), darker(), transparentize(), and with_alpha() methods
to clone and modify the gradient in-place instead of calling new().
- Clones the existing gradient (preserves angle and rotation)
- Modifies only color stops (skips first stop which contains angle)
- Avoids re-running expensive rotation processing
- Maintains the original angle information
Before: ConicGradientBrush::new(0.0, ...) // Lost angle information
After: Clone + modify colors in-place // Preserves angle
* Use premultiplied alpha interpolation for conic gradient colors
- Changed interpolate_color() to use premultiplied RGBA interpolation
- Updated signature to match Color::mix convention (&Color, factor)
- Added documentation explaining why we can't use Color::mix() here
(Sass algorithm vs CSS gradient color interpolation)
- Reference: https://www.w3.org/TR/css-images-4/#color-interpolation
This ensures correct visual interpolation of semi-transparent colors
in gradient boundaries, following CSS gradient specification.
* Run rustfmt on conic gradient code
* Fix ConicGradientBrush edge cases and add comprehensive tests
- Handle stops that are all below 0.0 or all above 1.0
- Add default transparent gradient when no valid stops remain
- Add 7 unit tests covering basic functionality and edge cases
* Apply clippy suggestion: use retain() instead of filter().collect()
* Fix radial-gradient parsing to allow empty gradients
Allow @radial-gradient(circle) without color stops, fixing syntax test
regression from commit 820ae2b04.
The previous logic required a comma after 'circle', but it should only
error if there's something that is NOT a comma.
* Fix conic-gradient syntax test error markers
Update error markers to match actual compiler error positions.
The 'from 2' case produces two errors:
- One at the @conic-gradient expression level
- One at the literal '2' position
Auto-updated using SLINT_SYNTAX_TEST_UPDATE=1.
* Refactor ConicGradientBrush epsilon adjustment and update tests
- Move epsilon adjustment for first stop into rotation block
(only needed when rotation is applied)
- Update property_view test to reflect boundary stops added by
ConicGradientBrush::new()
* Update conic-gradient screenshot reference image
Update the reference screenshot to match the current rendering output.
The small pixel differences (1% different pixels, max color diff 3.46)
are due to minor rounding differences in the conic gradient implementation.
* Fix ConicGradientBrush C++ FFI to avoid C-linkage return type error
Refactored ConicGradientBrush construction to match LinearGradientBrush
pattern, fixing macOS Clang error about returning C++ types from extern "C"
functions.
Changes:
- Rust: Split ConicGradientBrush::new into simple construction + separate
normalize_stops() and apply_rotation() methods
- Rust: Added FFI functions slint_conic_gradient_normalize_stops() and
slint_conic_gradient_apply_rotation() that take pointers (no return value)
- C++: Construct SharedVector directly in make_conic_gradient(), then call
Rust functions via pointer (matching LinearGradientBrush pattern)
- Optimized both methods to only copy when changes are needed
This resolves the macOS Clang error:
"'slint_conic_gradient_new' has C-linkage specified, but returns incomplete
type 'ConicGradientBrush' which could be incompatible with C"
The new approach maintains ABI compatibility while keeping complex gradient
processing logic in Rust.
* Fix C++ header generation to avoid GradientStop redefinition error
Resolves the macOS CI compilation error where GradientStop and
ConicGradientBrush were being defined in multiple headers
(slint_color_internal.h, slint_image_internal.h, and slint_brush_internal.h).
Changes:
- cbindgen.rs: Add ConicGradientBrush and FFI functions to slint_brush_internal.h include list
- cbindgen.rs: Add GradientStop, ConicGradientBrush, and FFI functions to exclude list for other headers
- slint_color.h: Add forward declaration for ConicGradientBrush
- slint_color.h: Add friend declaration for ConicGradientBrush to allow access to Color::inner
Root cause: After adding extern "C" functions in graphics/brush.rs,
cbindgen automatically detects and tries to include them in all headers
that use graphics/brush.rs as a source. The exclude list + filter logic
ensures these types only appear in slint_brush_internal.h.
This fixes the C++ compilation errors:
- "redefinition of 'GradientStop'"
- "ConicGradientBrush does not name a type"
- "Color::inner is private within this context"
* Prepare ConicGradientBrush FFI for Rust 2024 edition
Update FFI functions to use the new `#[unsafe(no_mangle)]` attribute
syntax and safe function signatures in preparation for Rust 2024 edition.
- Add `#![allow(unsafe_code)]` to graphics module for `#[unsafe(no_mangle)]`
- Add `#[cfg(feature = "ffi")]` to conditionally compile FFI functions
- Change from raw pointers to safe references (&mut)
- Remove manual null checks and unsafe blocks
* Change Color type to be f32 internally but no other changes
* Add missing clamp
* Add round function as we're not on MSRV 1.90+
* Prepare for being able to switch between u8s and f32s
* [autofix.ci] apply automated fixes
* Fix Display impl
* Add feature flag for 8-bit color values. Name to be bikeshed
* Fix brighter-darker test
* Update test screenshot
* Remove unused Float trait import
* Add cfg thing for not(cbindgen)
* Change Channel to float in cbindgen
* Change the cpp color type for uint8_t for now
* Opt cpp into 8-bit-color
* Switch feature around
---------
Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
We call Python::attach in a few places. When invoked from the Slint event loop, those should be fine. But there are some that may fail, for example during gc. Fall back to using try_attach and return default values instead of causing a panic.
Fixes#9984
napi-build 2.3 seems to require a newer major version of the napi crates
to be used, which we don't do yet. So pin the old napi-build version for
now.
See also https://github.com/napi-rs/napi-rs/issues/2999
* Add an open-url function
* Add cpp implementation
* Add workspace dep
* Add core::open_url function
* Remove pub use webbrowser
* Put open_url behind a flag
* [autofix.ci] apply automated fixes
---------
Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
No extra safety constraints that is not expressed in the type system
(eg, if they don't take any raw pointers)
Fix UB in slint_interpreter_value_to_array (The vector was already constructed in C++)
