This extension allows the user to specify whether a full flush is performed when making a context not current.
The only way to set this currently is at context creation, so this patch provides that functionality.
Defualt behaviour is set at FLUSH, as per the spec.
This patch does not contain the changes to WGL, appleGL or other platforms as I do not have access to GL 4.5 hardware on those platforms.
Full details on the use of KHR_context_flush_control can be found here:
https://www.opengl.org/registry/specs/KHR/context_flush_control.txt
The data produced by the callback is just thrown away and the audio thread
delays as if it's waiting for the hardware to drain.
This lets apps that rely on their audio callback firing regularly continue
to make progress to function as properly as possible in the face of disaster.
Apps that want to know that the device is really gone and deal with that
scenario can use the new hotplug functionality.
Device enumeration now happens at startup and then is managed exclusively
through hotplugging instead of full redetection. The device name list now has
a unique "handle" associated with each item and SDL will pass this to the
backend so they don't have to figure out how a human readable name maps to
real hardware for a second time.
Other cleanups, fixes, improvements, plus all the audio backends updated to
the new interface...largely untested at this point, though.
This fills in the core pieces and fully implements it for Mac OS X.
Most other platforms, at the moment, will report a disconnected device if
it fails to write audio, but don't notice if the system's device list changed
at all.
Various constants in ANGLE/WinRT, both in MSOpenTech's ms-master branch, and
in Google's branch, were changed again. This change makes SDL/WinRT work with
them.
To note, the ms-master branch (of ANGLE) was updated via this merge:
bbd2eb0a9c (diff-d1377fbe747de154e1bfcf7221d3de67)
If the function WIN_ConvertUTF32toUTF8() failed (should currently not be
possible) a not terminated string would have been sent as text input event.
This also fixed converting characters more often than needed on key repetition.
If the function Emscripten_ConvertUTF32toUTF8() failed (should currently not be
possible) a not terminated string would have been sent as text input event.
The key event callbacks were always removed from the same target although it is
possible to set them to different targets using the hint. This is only a partial
fix because it assumes that the hint is not changed to a different value later.
- fix crash on OSX when removing a device. If the remove happened due to the CFRunLoopRunInMode call in SDL_SYS_JoystickDetect then we would delete the device right away, before SDL_SYS_JoystickUpdate could clean it up. So move the CFRunLoopRunInMode to after the cleanup logic, preventing this case. This does mean that adds and removes of joysticks now take 1 extra frame to show up.
After disconnecting a joystick the remaining kept their original device index.
This was not correct because the device index must be a number between 0 and
SDL_NumJoysticks(). It was fixed with ideas from SDL's joystick implementation
for Android. Some range checks were removed as the caller already checks them.
SDL_SYS_JoystickUpdate() was implemented incorrectly. For every call to it all
attached joysticks were checked. But actually only the given SDL_Joystick should
be checked then. This allowed sending broken events for attached but not opened
joysticks. It also checked the opened joysticks more often than actually needed.
This allows for this kind of code in an application:
int monitorID = 1; // the second monitor!
SDL_SetWindowPosition(sdlWin,
SDL_WINDOWPOS_CENTERED_DISPLAY(monitorID),
SDL_WINDOWPOS_CENTERED_DISPLAY(monitorID));
Fixes Bugzilla #2849.
Yes this seems confusing as on mac+linux Long is either 32 or 64bits depending on the architecture, but this is how the X11 protocol is defined. Thus 5 is the correct value for the nelts here. Not 5 or 10 depending on the architecture.
More info on the confusion https://bugs.freedesktop.org/show_bug.cgi?id=16802
- Added new custom launch screen code. It uses the launch screen nib when available on iOS 8+, the launch images dictionary if the launch screen nib isn't available, and the old standard image names if the launch image dictionary isn't in the plist.
The launch screen is now hidden during the first call to SDL_PumpEvents rather than SDL_CreateWindow so apps can have the launch screen still visible if they do time-consuming loading after creating their window. It also fades out in roughly the same way as the system launch screen behavior.
It can be disabled by setting the SDL_IPHONE_LAUNCHSCREEN define in SDL_config_iphoneos.h to 0.
- A blank UIView is now created and displayed when the window is first created. The old behavior was to defer creating any view until SDL_GL_CreateContext, which prevented rotation, touch events, and other windowing-related things from working until then. This also makes it easier to use SDL_GetWindowWMInfo after creating a window.
- Moved the keyboard and animation callback code from SDL's UIView subclasses to its UIViewController subclass, which lets them work properly in all cases when a SDL window is valid, even before SDL_GL_CreateContext is called and after SDL_GL_DeleteContext is called.
- SDL_GL_CreateContext, SDL_GL_SwapWindow, SDL_GL_MakeCurrent, and SDL_GL_DeleteContext are more robust.
- Fixed some edge cases where SDL windows weren't rotating properly or their reported sizes were out of sync with their actual sizes.
- Removed all calls to [UIApplication setStatusBarOrientation:]. It doesn't seem to work as expected in all cases in recent iOS versions.
- Some code style cleanup.
The callbacks used to receive the HTML events were not removed if the joystick
subsystem initialization failed or if the joystick subsystem was quit. Also, the
already connected joysticks were not deleted if the initialization failed later.
It's slightly faster than failing later, after a strchr() call, since this
will get called multiple times with a NULL string if the system totally
fails elsewhere.
This change integrates initialization settings for ANGLE/WinRT, as suggested in
MSOpenTech's latest ANGLE template-projects (for MSVC).
This should fix some OpenGL initialization issues on WinPhone 8.1 on ARM, and
on the 1st-generation Surface RT.
SDL_JOYDEVICEADDED events must contain the device index which is a value between
0 and the number of connected joysticks. The old implementation included a value
based on the instance id instead. It worked in some cases because the values are
similar initially. But after disconnecting joysticks this is no more the case
and the always increasing instance id becomes larger than number of joysticks.
Windows Phone does not appear to allow VSync to be turned off. Doing so appears
to either result in content not getting drawn (when the D3D debug runtime is
turned off), or forcing VSync back on and logging an error (when the D3D debug
runtime is turned on).
VSync had been getting turned on anyways, this change just notes such:
- via the WinRT README
- by always setting the SDL_RENDERER_PRESENTVSYNC flag when creating an
SDL_Renderer on Windows Phone
- Don't create a temporary context first; this was probably due to Windows
needing one to get the address of wglCreateContextAttribsARB(), but that's
a unique quirk of WGL, and doesn't apply to glX. The glX spec explicitly says
you have to get a function pointer that works with any context from
glXGetProcAddress(), including when no context exists.
- Properly check for the GLX_ARB_create_context instead of just looking for a
non-NULL glXCreateContextAttribsARB()...some implementations, like Mesa,
never return NULL for function lookups (Mesa returns pointers into a jump
table that is filled out when the GL is initialized; since you can look up
functions before you have a valid context, it can't definitely say a function
isn't valid at that point).
Normally there's a 200 millisecond delay on all focus events in case there
was a vidmode change, now we note the last vidmode change and only impose this
delay if a change happened extremely recently.
Thanks to Epic Games for reporting this issue.
Jonas Kulla
The configure script didn't differentiate between Linux and Android, unconditionally compiling in the unix implementation of SDL_sysfilesystem.c.
I'm probably one of the very few people building SDL for android using classic configure + standalone toolchain, so this has gone undetected all along.
This is a little macro magic to use malloc() directly instead of SDL_malloc(),
etc, so static analysis tests that know about the C runtime can function
properly, and understand that we are dealing with heap allocations, etc.
This changed our static analysis report from 5 outstanding bugs to 30.
5x as many bugs were hidden by SDL_malloc() not being recognized as malloc()
by the static analyzer!
