When attaching a renderer (GL based specifically) to a window that was not created with the appropriate flags, the window would be destroyed and recreated to configure it for the desired rendering backend. While most of the issues with this have been mitigated over time, there can still be some undesirable side effects from doing so on certain platforms.
If the window was just created and was never configured for any graphics context, it is possible that the reconfiguration can be done without destroying the window first.
The Wayland implementation fixes an issue when creating a window with the fullscreen flag on wlroots based Wayland compositors, and can likely be extended to other platforms to avoid unnecessarily destroying/recreating a window in the very common case where a window is created, followed immediately by attaching a renderer.
If a bound keymap is being destroyed, unbind it to ensure that the keyboard state won't be left pointing to an invalid keymap. Backends now no longer need to manually do this when destroying keymaps during keyboard removal and shutdown.
In certain cases when moving fullscreen windows in scaled desktop configurations, the window origin might overlap two displays at once. Check if the window is at the origin of a specific display before falling back to the generic window rectangle check.
Fixes rare fullscreen window misplacement when moving fullscreen windows via a desktop shortcut while using the Wayland scale-to-display mode.
Trying to determine where a window should be made fullscreen from the size and position can be unreliable if the window exceeds the display bounds. Add additional checks with the following priority:
- If the window was positioned with a macro that explicitly passes a display ID, store and use the requested display as the explicit fullscreen target.
- Check if the window position is an exact match for any display origins, and use that display if found, as positioning a fullscreen window by moving it to the origin of the destination display is common behavior.
- Fall back to the existing center point check if the previous checks were not successful, as it is known behavior, and won't risk breaking existing clients that rely on it.
- Use modern Xkb functions where appropriate and cleanly separate the modern and legacy paths.
- Remove the deprecated XKeycodeToKeysym function in favor of directly querying the keymap on the legacy path.
- Look up virtual modifiers by name on the Xkb path to better handle remapping (equivalent to the modifier handling under Wayland).
- Optimize keymap creation on the Xkb path to cut keymap build times and enable fast group switching (equivalent to keymap handling on Wayland).
- Enable and handle Xkb events to handle changes to the group, mapping, and modifier states. This is more reliable than using the legacy events (group changes may not arrive if the window lacks pointer focus), and better handles cases where modifiers are latched, locked, or activated externally rather than physically pressed.
Prefer axes with the 'Rel X'/'Rel Y' labels, followed by 'Abs X'/'Abs Y', and only fall back to the old behavior of using the first two enumerated axes if no others are found.
Fixes a FIXME when determining which axes to use for relative motion.
In addition to hiding the border on bordered windows that will immediately become fullscreen, The combination of flags used in STYLE_BORDERLESS_WINDOWED will still show the borders on borderless windows if the initial window size exactly matches the desktop, so STYLE_BORDERLESS must be used instead.
If attempting to switch to an exclusive mode while a fullscreen spaces transition is active, wait until the transition is complete before trying to apply the changes, or the window can wind up in a weird, broken state if a mode switch occurs while in a fullscreen space.
This reverts commit ee8f2861e7.
It turns out that the problem is elsewhere, related to needing to block mode changes until spaces transitions are complete.
When skipping presentation due to the window being hidden, presentDataCount is not incremented on the command buffer, and subsequently the submitted command buffers will not be cleaned up as long as the window is hidden. This results in a lag spike when showing the window due to all previously submitted command buffers suddenly being cleaned up at once, and lag at shutdown due to an equivalent number of fences needing to be destroyed.
Instead of relying on presentDataCount to determine whether a command buffer should be cleaned up, use a flag, which is set under the appropriate circumstances.
There may have been a reason for not attaching a null buffer when destroying a popup at some point in the past, but that is unnecessary now, as is the comment about ShowWindow crashing, as ShowWindow assures that a null buffer is attached before (re)creating the window.
The previous calculation could result in a window whose original position was positioned exactly corner-to-corner with the parent not being adjusted to be adjoining, and thus subject to spurious closure.
Keymap construction is an expensive process, so keymaps are cached to facilitate fast switching, as they are static after initial construction, and do not need to be rebuilt every time.
Only use the specified output if an exclusive mode is being used, or a position was explicitly requested before entering fullscreen desktop. Otherwise, let the compositor handle placement, as it has more information about where the window is and where it should go, especially if fullscreen was requested before the window was fully mapped, or the window spans multiple outputs.
When changing the text input mode on a window, only update the seats that currently hold keyboard focus on that window, otherwise, text input might be inadvertently enabled or disabled on a seat focused on another window.
Build all the available keyboard layouts at once; this adds a negligible bit of overhead when initially handling the keymap (which was already significantly lowered by previous commits), but reduces the later cost of changing layouts to just swapping a pointer.
Additionally, handling of unknown keysyms, particularly when dealing with virtual keyboards, is improved, as keys generating valid Unicode values with no corresponding scancode will be dynamically added to the keymap with reserved scancodes, allowing for proper round-trip lookup behavior.
Initializing the compose table is a very expensive operation, and only necessary if the locale envvar changed, which is often not the case when just changing the keymap, so don't destroy and recreate it whenever the keymap changes. The state only needs to be reset in this case.
Wayland previously didn't specify that the seat name preceded the capabilities, but it is now specified that the name event must always come first. Remove the 'SDL_Set<device>Name()' functions that were only added to accommodate the case of compositors sending the name after the seat capabilities, as this clarification means that they are no longer needed.
The internal key repeat mechanism already disables itself if the key repeat interval is 0, and SDL tracks and handles the flagging of repeated keys itself, so just map the 'repeated' event to 'pressed'.
The spec states that xdg_surface must have seen an initial configure event before attaching a buffer, however, this was only being done when initially showing the window, and not after show->hide->show cycle.
Always wait for the initial configure event when (re)creating an xdg_surface as part of the show window sequence.
If the pointer should be confined, keep trying until a confine/lock signal is received. This considerably improves locking/confinement reliability on compositors where confining can be a racy operation, or may not take effect until the pointer is actually in the confinement region.
A pointer lock is used to special-case 1x1 confinement regions, as otherwise, the pointer can still exhibit jitter at the subpixel level, particularly on scaled desktops.
Pointer events in seat version 5 and higher should be grouped and sent together when a pointer frame event occurs. Store pending pointer motion events and dispatch them together when a frame event is received.
This helps avoid spurious pointer motion events that some compositors generate with no associated frame (e.g. from some tablet events).
Always update the border sizes on frame extent events, or they can incorrectly still be zero if followed by a PropertyNotify event when leaving fullscreen.
Fixes sending the correct restored window size when leaving fullscreen in fvwm.
XMonad ignores size hints and shrinks the client area to overlay borders on fixed-size windows, even if no borders were requested, resulting in the window client area being smaller than requested. Calling XResizeWindow after mapping seems to fix it, even though resizing fixed-size windows in this manner doesn't work on any other window manager.
When querying the keycode produced by a scancode with a certain set of modifiers, it would fall back to defaults if a key hash value with the exact set of modifiers wasn't found, which resulted in certain modifier combination returning incorrect keycodes on non-ANSI keyboard layouts. For example, querying SDL_SCANCODE_Y with the alt modifier on a QWERTZ layout returns SDLK_Y instead of SDLK_Z on most platforms, as the backends don't generate a specific entry for this key + modifier combo, so the lookup would fall back to the default ANSI layout.
Adding additional key+modifier combinations when building the keymap is one solution, but it makes an already expensive operation even more so, pushing the time needed to build the keymap into double-digit milliseconds in some cases due to the large amount of key combos that need to be queried, most of which are redundant.
Instead, falling back to searching through the shift levels for the given modifier state when querying the keymap will ensure that the most appropriate keycode is returned. This does add some overhead to lookups if the key doesn't have an entry with the exact set of modifiers, but it is minimal as hash table lookups are an inexpensive operation, and unnecessary lookups are avoided. In my own testing of an optimized build, the difference between best-case and worst-case performance (the latter of which is highly unlikely in real-world usage) is only a few hundred nanoseconds. Additionally, the unmodified keys are queried when pumping events, so there is no additional overhead in that case.
Windows seems to implicitly enable IME text input on windows created while an IME is active, which causes the IME suggestion window to pop up when keys are pressed, even if a client never explicitly enabled it. Ensure that IME support is initially disabled on new windows; SDL will enable it at a later time, if required.
Use the proper types for xkb mod masks and layout indices, rename the mod masks to reflect that they are bitmasks and not indices, and use the 'layout' nomenclature instead of the deprecated 'group' nomenclature.
xkbcommon 1.10.0 declared certain modifier names to be deprecated, and the current plan is to remove them in 1.12.0. Use the new recommended names and modifier mask retrieval function when building against version 1.10.0 and higher.
Using processed timestamps can result in anomalies that cause excessive repeat events, and hard caps can cause issues.
In the key event handler, use the raw elapsed time to calculate any remaining repeat events to avoid the artifacts that can result from using processed timestamps.
The Wayland key repeat rate ranges from 0 to 1000 per second, so millisecond resolution doesn't lose any precision.
More non-compositing window managers than just openbox seem to need this, so always force sending the window position and dimensions when entering/leaving fullscreen. If they are wrong, they will be immediately overwritten by the correct dimensions from a subsequent ConfigureNotify event.
This is disabled on XWayland, as it seems to cause hitching on some compositors.
Cap the elapsed time to something sane in case the compositor sends a bad timestamp, which can result it in it looking like the key has been pressed for a *very* long time, bringing everything to a halt while it tries to enqueue all the repeat events.
The function SDL_GetCurrentKeymap() would return null instead of the actual bound keymap for non-latin layouts if certain mapping options were set. Add a parameter to ignore the keymap options and always return the actual bound keymap, which is needed in the Wayland backend to avoid spamming keymap changed events on every keystroke with certain layouts.
The text input cursor should reflect the cursor position, not the entire text input rect. Set it correctly so that IME chooser dialogs appear in the correct location.
This can be toggled per-window, so use the individual window flags instead of the global toggle to selectively enable it only for the relevant window in a multi-seat scenario, as is already done with keyboard and pointer grabs.
The main view always reflects the size of the output, so don't use the dimensions of the currently bound render target texture when updating it, or it will reflect an incorrect size when the render target texture is unbound.
This and its related commits introduced some bugs and quirks such as duplicated and delayed/missed key events that can't be easily worked around, so revert this for now.
This reverts commit 47162a4168.
Compositors typically send keyboard, pointer, touch, and tablet leave events when a window is destroyed, however, this is not guaranteed behavior, and at least one compositor in widespread use doesn't always send pointer leave events immediately upon destroying a window. Ensure that all references held by seats to a focused window are removed before the underlying window surface and structs are destroyed to prevent potential segfaults if the seats are immediately destroyed after the window.
IME text events can result in sending duplicate key press events, which will result in undesired repeated key presses. Since the events are exact duplicates, compare the serials to filter out redundant key down events.
The xserver will still send EnterNotify events while the pointer is captured, and the grab shouldn't be updated in these cases, as it will cause the capture to be lost.
If it is known that the window will immediately enter fullscreen upon being shown, set the borderless style when showing the window to hide the borders, or they may linger in the background if the client takes some time to draw the first frame.
Unnecessarily calling ShowWindow with SW_RESTORE when applying the window flags must be suppressed in this case, or the borders can reappear in a weird, partial state.
sdl2-compat will call SDL_GetDisplayForWindow() when querying the display to use for fullscreen, which won't always be correct if the backend can't actually reposition the window. When calling this function, get the ideal fullscreen display and store it in a property for retrieval by sdl2-compat.
Mouse button events that trigger a hit test are not passed to the client, but the client may still query the global mouse button state from within the hit test handler, so the reported buttons need to be accurate. Query the buttons directly from the seat instead of the higher global mouse state to match the behavior of other platforms.
By default, popups are automatically constrained to be completely within display bounds, so as not to cut off information and result in an unusable menu, or unreadable tooltip. In some cases, however, this is not wanted, so a property to toggle this behavior is added.
There are also cases where the client may not want a popup menu to implicitly grab the keyboard focus, as is the default behavior, so popup menus now respect the focusable flag/property, as well as being able to toggle focus grabbing via SDL_SetWindowFocusable().
When showing or hiding a popup menu, manually check and set the focus if the new topmost window under the cursor is an SDL window. Otherwise, the focus won't be updated until the cursor is actually moved.
There is a quirk with XInput2 mouse capture that causes a leave event to be sent if the pointer moves out->in->out, which breaks mouse tracking outside the window. If the mouse leaves the window with buttons pressed, continue tracking it until the buttons are released.
Xfce, unlike every other window manager in existence, sends ConfigureNotify events before PropertyNotify events when toggling the fullscreen and maximized window state. Check the window state when handling ConfigureNotify events, and defer emitting SDL size/position events until the corresponding PropertyNotify event arrives, since SDL and clients expect to get the window state before the new size and position.
This warning is actually wrong, as the variable is only used if the string pointer is valid, in which case it was initialized, but this is needed to quiet MSVC.
Wayland compositors may send recursive clipboard offers to the client, which need to be filtered out to avoid clearing local data. Previously this was worked around with a hack, but this caused the ownership flag to be set incorrectly, which broke some clients.
This introduces a metadata MIME type of application/x-sdl3-source-id to be sent with SDL3 selection offers, which contains a string that is a unique identifier for the instance, and can be used to detect if a received selection offer is originating from the same instance that generated it.
If DBus is available, the unique identifier string is the unique name of the connection, otherwise, the process ID is used.
If a window isn't resizable from specific directions, the compositor can inform clients of the current edge constraints, so they don't display resize cursors for non-resizable edges.
Wayland environments can expose more than one seat for multiple collections of input devices, which can include multiple, simultaneously active, desktop pointers and keyboards with independent layouts. The Wayland input backend previously presumed that only one seat could exist, which caused broken behavior if the compositor exposed more than one, which is possible on wlroots based compositors such as Sway. This introduces support for handling multiple seats, including proper handling of dynamically added and removed seats and capabilities at run time.
The SDL Wayland input system was accreted over time, and the assumption that only one seat will ever exist resulted in state and related objects not always being tied to their most appropriate owner in a multi-seat scenario, so refactoring was required to manage several bits of state per-seat, instead of per-window or globally.
As Wayland keyboards can have per-seat layouts, fast keymap switching is required when multiplexing input from multiple seats to the global SDL keyboard device. A parameter was added to the keymap creation function to specify if the keymap lifetime should be externally managed to facilitate keymap reuse, and some layout info was moved from the global keyboard state to the keymap state to avoid unnecessarily redetermining it whenever a reused keymap is bound. This reduces the overhead of switching keymaps to setting a single pointer.
Multiple seats also means that multiple windows can have keyboard and/or mouse focus at the same time on some compositors, but this is not currently a well-handled case in SDL, and will require more work to support, if necessary.
The window may be initially maximized or made fullscreen by the window manager for various reasons, such as automatically declaring a window that precisely fills the usable desktop space as maximized, or a "kiosk-mode" automatically making the window fullscreen.
Don't redundantly make restored or unset fullscreen calls when initially showing a window, or the expected state can be unset.
It doesn't work well with Vulkan, and Vulkan windows may not have the Vulkan flag set in all circumstances, so only enable it if we explicitly know that OpenGL is being requested instead.
On non-compositing window managers, XIconifyWindow can trigger map/unmap events, which would toggle the window hidden/shown state. This should not be done, as a hidden window in SDL is equivalent to a withdrawn window in X, and SDL will try to set/reset state when it is shown again.
Unless the window is explicitly being withdrawn via X11_HideWindow(), set unmapped windows to the minimized/occluded state, and send a restored event when mapped again.
The window failing to enter/exit fullscreen notifications don't necessarily imply anything about the actual state of the window. On failure, dump pending events, and don't presume anything about the current window state, as it will be handled by subsequent enter/leave notifications.
The accelerated relative coordinates are always relative to the desktop resolution, and need to be scaled with emulated fullscreen resolutions for mouse movement to remain stable. Otherwise, pointer speeds will be too fast on emulated resolutions below the desktop resolution, and too slow on those above.
If the window is flagged with SDL_WINDOW_TRANSPARENT, create the associated swapchain with a composite alpha value that supports blending, if available.
Fixes transparent windows on the Vulkan backend, and prevents possible validation errors by ensuring that the composite alpha value is always a valid bit in VkSurfaceCapabilitiesKHR::supportedCompositeAlpha.
Frank Praznik