When changing the window state from non-floating to floating (e.g. leaving fullscreen), libdecor can send bogus content sizes that are +/- the height of the window title bar and start 'walking' the window height in one direction or the other with every transition.
The floating window size is known, so use the cached value instead of the size reported by libdecor when restoring the floating state.
Some applications (and embarrassingly, testime is one of them) call
SDL_StartTextInput() or SDL_SetTextInputRect() every frame. On KDE/KWin
with fcitx5, this causes there to be several preedit events every frame
(particularly given some of the workarounds in Wayland_StartTextInput),
which slows testime down to an unusable crawl.
Instead, make SDL_StartTextInput() a no-op if text input is already
enabled, and cache the input rect, only changing it when the new rect is
actually different.
With these changes, we only get preedit events (and hence
SDL_TEXTEDITING events) when the preedit string actually changes. This
matches the behaviour under XWayland, and works very smoothly.
Most of these are probably harmless, but the changes to SDL_immdevice.c and SDL_pixels.c appear to have fixed genuine bugs.
SDL_audiocvt.c: By separating the calculation of the divisor, I got rid of the suspicion that dividing a double by an integer led to loss of precision.
SDL_immdevice.c: Added a missing test, one that could have otherwise led to dereferencing a null pointer.
SDL_events.c, SDL_gamecontroller.c, SDL_joystick.c, SDL_malloc.c, SDL_video.c: Made it clear the return values weren't used.
SDL_hidapi_shield.c: The size is zero, so nothing bad would have happened, but the SDL_memset() was still being given an address outside of the array's range.
SDL_dinputjoystick.c: Initialize local data, just in case IDirectInputDevice8_GetProperty() isn't guaranteed to write to it.
SDL_render_sw.c: drawstate.viewport could be null (as seen on line 691).
SDL.c: SDL_MostSignificantBitIndex32() could return -1, though I don't know if you want to cope with that (what I did) or SDL_assert() that it can't happen.
SDL_hints.c: Replaced boolean tests on pointer values with comparisons to NULL.
SDL_pixels.c: Looks like the switch is genuinely missing a break!
SDL_rect_impl.h: The MacOS static checker pointed out issues with the X comparisons that were handled by assertions; I added assertions for the Y comparisons.
SDL_yuv.c, SDL_windowskeyboard.c, SDL_windowswindow.c: Checked error-result returns.
This reverts commit 04b50f6c6b.
It turns out OpenGL vsync has broken again in macOS 12, so
we're reintroducing our CVDisplayLink code to deal with it,
again.
Reference Issue #4918.
SDL is built around the concept of keyboards having a fixed layout with scancodes that correspond to physical keys no matter what linguistic layout is used. Virtual keyboards don't have this concept and can present an arbitrary layout of keys with arbitrary scancodes and names, which don't fit the SDL model. When one of these keyboards is encountered, it requires special handling: use the keysym of the pressed keys to derive their ANSI keyboard scancode equivalents for control keys and ASCII characters. All other characters are passed through as text events only.
It turns out that we can safely create a Metal view on an existing window, and that avoids issues with the window being recreated with the wrong orientation in iOS 16.
Fixes https://github.com/libsdl-org/SDL/issues/6289
Clamp the wl_seat max version to 5 if being built against a version of libwayland below 1.21.0, or containers that bundle newer versions of SDL with older versions of libwayland can break if the compositor advertises support for a protocol version above 5.
