G
macOS: keep GTK frames flowing during native window zoom What does this change do? On macOS, -[NSWindow zoom:] enters an AppKit transition in which GTK4's normal frame clock does not submit intermediate content frames. The window outline moves, but the application content can remain at its old size until the transition finishes. Handle windowShouldZoom:toFrame: in GdkMacosWindow with a nonblocking NSViewAnimation. Existing windowDidResize notifications continue to update the GDK surface and…
G
Thanks for reading the guidelines, and trying to follow them.
G
This branch couldn't be merged: CI failed! See pipeline https://gitlab.gnome.org/GNOME/gtk/-/pipelines/1144636.
G
Benjamin Otte (3d874c9c) at 11 Nov 21:08 gpu: Implement get_node_as_image() for repeat nodes
G
See individual commits for more details. !9965 (closed) shouldn't be needed anymore with this. Particular glitch in question: The Vulkan renderer is broken (at least on my machine). The MR only fixes the same kind of issue for the Vulkan path, everything else was already broken before. CC: @amolenaar
G
Looks good. Thanks @haecker-felix :)
G
Benjamin Otte (1ffd3ca6) at 11 Nov 20:24 gpu: Implement get_node_as_image() for repeat nodes
G
I could help if some git-fu is needed
G
Unfortunately I don't know this code well enough, and the guidelines strictly forbid using AI in a way that would let me help further here myself. When I made the commit, I hadn't read CONTRIBUTING.md far enough. If I may still bring in the AI's expertise, its assessment is this: Yes, that is correct: in the path without swapchain-maintenance, wait_present() and release_presents() set vk_swapchain to VK_NULL_HANDLE without touching present->frame. With the patch, the frame still gets completed:…
G
AppKit runs the zoom animation in a nested run loop. The GLib main loop doesn't run there, so the window content only updates once the animation finished. Produce a frame directly in windowDidResize: instead, so the content follows the window size. Skip if the frame clock is already in a frame or the frame clock is stopped. I'm not 100% sure this is the right approach, but it fixes the visible glitch during the zoom animation, and I haven't noticed any unwanted side effects so far. Switched to…
G
I've rebased on latest main. 9ef4e7d4, the cause of the conflict, fixes the adjustment being overwritten during allocation so I've been able to drop one of my commits. Adding a gtk_adjustment_animate_deceleration method possibly still the right approach but requires invasive changes that I'm not sure I can get right. @carlosg - Are you happy to have a look?
G
For this sort of stuff (where the first pass can be unrolled), i very much prefer simple and let the compiler figure it out. I'd be very surprised if it didn't figure this out at -O2 or -O3 with LTO. And if it doesn't, we should just make rbtree static inlines similar to from BSD has done for coming up on 25 years.
G
there's no issue. But these are hard to track down, or else they would have been fixed already
G
This is not part of my other patch series related to buffer overreads in CSS tokenizing.
G
I was doing some other work on gtk's css system and ran into a problem where a script I was using couldn't match an error message as expected. Turns out the error message from gtk was getting corrupted by a double unref in filter URL parsing and causing my other work to fail ("Filter url contahns non-SVG data", notice the word "contains" is corrupted). When a filter url() is a data: URL with a mime type other than image/svg+xml, the bytes are unreffed and then unreffed again via…
G
Hello, here is a minimal reproducer (90 lines instead of 197, same scenario and same stack), as suggested in #6758. It needs a window manager (the crash happens when the main window becomes active again). The view has the focus; the dialog deactivates the window; OK clears and refills the store, then destroys the dialog. On reactivation GTK restores the focus to the remembered cell, whose row is gone: its parent is NULL and gtk_column_view_row_widget_get_column_view() reads it. Valgrind, GTK…
G
Is there an issue on file about the flaky tests? Maybe I can take a look.
G
Marge Bot (aaf281e5) at 07 Nov 04:34
G
This is the default filter Gimp is using for scaling images. It's mainly intended for upscaling by scale factors > 2 because at that scale the artifacts of linear scaling become visible. The other reason I'm looking at this is that I need the shader (just with a different kernel) for !10436.
G
Marge Bot (48232bcb) at 07 Nov 04:34 Merge branch 'wip/otte/look-at-my-quads' into 'main' ... and 10 more commits
G
We unfortunately have some flaky tests in ci. Retrying usually gets them to pass
G
I don't understand the CI failure. Doesn't seem related to the patch here?
G
Marge Bot (aaf281e5) at 07 Nov 04:10 testsuite: Test that bicubic is a no-op for 1:1 ... and 30 more commits
G
This branch couldn't be merged: CI failed! See pipeline https://gitlab.gnome.org/GNOME/gtk/-/pipelines/1141754.
G
I still haven't figured out why the test failed on GL when it included transparency, but I now used a black background and it passes, so I'm happy.
G
Modifying the counter in the condition just smells, for my taste
G
I don't do that anymore (just like using signed integers) because I get things wrong in the loop and then I have to debug it later. And I haven't gotten my version wrong yet, so it wins just because of that, not ebcause I think it looks beautiful. It does look more beautiful than a while loop though IMO.
G
I usually do the more cumbersome for (unsigned int i = n; i > 0; i--) { /* use i - 1 */ } for a loop that runs from n - 1 to 0.
G
It's the best way I've found how to write it with an unsigned variable. Do you have an alternative you like better?
G
Benjamin Otte (0ca9bb8e) at 07 Nov 03:26 testsuite: Test that bicubic is a no-op for 1:1