This would have been a one-line patch to the documentation (specifying that
captures only work as long as the left mouse button is pressed), but I didn't
like that, so I got a little crazy about this instead.
There were several good arguments for this: it's how Windows works with
WM_NCHITTEST, SDL doesn't need to manage a list of rects, it allows more
control over the regions (how do you use rects to cleanly surround a circular
button?), the callback can be more optimized than a iterating a list of
rects, and you don't have to send an updated list of rects whenever the
window resizes or layout changes.
This should be a "long" which on a 64-bit system is likely to be > 32-bits,
causing XGetICValues() to write past the end of the variable (and stack).
Fixes Bugzilla #2513.
It seems like a net improvement in all the scenarios Sam and I could
think of, and looking at hg history it was added for fullscreen
window management specifically. Much of that code has changed since
then, but maybe it needs to stay there for that and simply be moved
to a fullscreen condition check.
It would solve this issue:
https://bugzilla.libsdl.org/show_bug.cgi?id=2439
As well as cases where on SteamOS, we hide/show specific Steam
overlay windows while expecting them to stay in the background, since
changing the window stacking order really angers the NVIDIA driver.
CR: Sam.
(Most of its complaints here are that these ints can be negative, although
they wouldn't been in sane cases. Checking sanity is what assertions do, and
it placates the analyzer appropriately.)
These scenarios can happen when a GPU is switched, its driver updated, or in
some virtual machines (such as Parallels) are suspended and then resumed. In
these cases, all GPU resources will already be lost, and it's up to the app to
recover.
For now, SDL's D3D11 renderer will handle this by freeing all GPU resources,
including all textures, and then sending a SDL_RENDER_TARGETS_RESET event.
It's currently up to an app to intercept this event, destroy all of its
textures, then recreate them from scratch.
SDL_WINDOWEVENT_FOCUS_LOST is now sent when an app's native window is hidden.
Likewise, SDL_WINDOWEVENT_FOCUS_GAINED is sent when an app's window is shown.
This mimicks behavior seen on iOS and Android.
SDL_WINDOWEVENT_MINIMIZED and SDL_WINDOWEVENT_RESTORED are now sent when the
app's native window is hidden and shown. Previously, these were sent when an
app was suspended and resumed. On Windows 8.x/RT, an app may be sent to the
background without being suspended, which previously meant that
SDL_WINDOWEVENT_MINIMIZED might never have been sent. (On Windows Phone 8,
however, this seems to be different, whereby apps sent to the background appear
to always get suspended.)
SDL_APP_WILLENTERBACKGROUND is now sent as soon as the app is told that it is
about to go to the background. SDL_APP_DIDENTERBACKGROUND is sent via a WinRT
'deferral operation', which is how WinRT gives apps a bit of extra time
(multiple seconds worth) to prepare for an app-backgrounding.
The distinction may be important as the deferral operation's code is always run
in a separate thread. For Direct3D-only apps, this means that between the
two SDL app-backgrounded events, SDL_APP_WILLENTERBACKGROUND will be the only
one run from the main thread. Given that some WinRT operations can only be done
on the main thread (operations to the CoreWindow fall into this category), this
could be important.
It is important to note that pre-deferral code may only have a very short bit of
time to execute code, less so than code run in the deferral operation (where
SDL_APP_DIDENTERBACKGROUND is sent from), which usually gets several seconds to
run.
SDL/WinRT did have support for OpenGL ES 2 via an older version of ANGLE/WinRT,
however its API changed a few months ago, and SDL/WinRT would crash when trying
to use it. It would also occasionally crash when using the older version.
This changeset should make SDL/WinRT work with the latest version, as
available via MS Open Tech's git repository of it at
https://github.com/msopentech/angle
Older versions of ANGLE/WinRT (from either https://github.com/stammen/angleproject
or https://bitbucket.org/DavidLudwig/angleproject) will need to be updated to
MS Open Tech's latest version.
It's been hardcoded out forever now, but I've now forcibly removed it with
the preprocessor so static analysis doesn't complain about it for now.
Eventually I want to rewrite or remove this code.
This is actually a false-positive, in this case, since Clang doesn't know
that SDL_SetError() only ever returns -1. Feature request to improve that,
with explanation about these specific SDL patches, is here:
http://llvm.org/bugs/show_bug.cgi?id=19208
Rotation detection and handling should now work across all, publicly-released,
WinRT-based platforms (Windows 8.0, Windows 8.1, and Windows Phone 8.0).
This changeset prevents IDXGISwapChain::ResizeBuffers from being invoked on
Windows Phone 8, a function that isn't available on the platform (but is
available on other Windows platforms). The call would fail, which ultimately
led to a crash.
This changeset also attempts to make sure that the D3D11 swap chain is created
at the correct size, when using Windows Phone 8.
Still TODO: make sure rotation-querying works across relevant Windows
platforms (that support Direct3D 11.x).
Eventually we'll probably move the version checking to higher level code and report the actual version of context that got created, but to avoid breakage we'll leave it like this for now.
The D3D11 renderer is now slightly faster than D3D9 on my Windows 8 machine (testsprite2 runs at 3400 FPS vs 3100 FPS)
This will need tweaking to fix the Windows RT build.
klose
File: SDL_gesture.c
Method: SDL_GestureAddTouch
When a new SDL_GestureTouch element is added to the global SDL_gestureTouch array the variable 'centroid' of the new element is not initialized.
The problem is that this variable is read isndie SDL_GestureProcessEvent when a SDL_FINGERDOWN event occurs.
- SDL_WINDOW_FULLSCREEN works as always (change resolution, lock to window).
- SDL_WINDOW_FULLSCREEN_DESKTOP now puts the window in its own Space, and
hides the menu bar, but you can slide between Spaces and Command-Tab between
apps without the window minimizing, etc.
- SDL_WINDOW_RESIZABLE windows will get the new 10.7+ "toggle fullscreen"
window decoration and menubar item. As far as the app is concerned, this is
no different than resizing a window, but it gives the end-user more power.
- The hint for putting fullscreen windows into the Spaces system is gone,
since Spaces can't enforce the requested resolution. It's a perfect match
for FULLSCREEN_DESKTOP, though, so this is all automated now.
SDL was expected that each SDL_DisplayMode had a driverdata field that was SDL_malloc'ed, and was calling SDL_free on them. This change moves WinRT's driverdata content into a SDL_malloc'ed field.
Yamagi
A customer of mine had the strange problem, that SDL_SetRelativeMouseMode() was failing for him on Windows 7. Luckily he was willing to provide some debug informations. We could track this problem down to RegisterRawInputDevices() failing due to security software running on his system (Norton Internet Security to be precise, but there are reports of similar problems with other products. For example [1]). Working around this issue with SDL_WarpMouseInWindow() is easy, and while I don't think that SDL2 can provide an internal workaround it would be really nice and helpfull if this could be documentated somewhere.
1: http://forums.codeguru.com/showthread.php?498374-How-to-run-a-very-long-SQL-statement
Coriiander
In src\timer\windows\SDL_systimer.c there is an error with regards to timeBeginPeriod and timeEndPeriod. These functions typically get called when no high resolution timer is available, and GetTickCount is not used.
According to MSDN (link: http://msdn.microsoft.com/en-us/library/windows/desktop/dd757624(v=vs.85).aspx), for every call to timeBeginPeriod a subsequent call to timeEndPeriod is required. While SDL is currently doing this, it fails to call timeEndPeriod when cleaning up/shutting down SDL. Please note that these functions affect things on a system level. Failing to call timeEndPeriod, disables applications for using WINMM-timers after usage&shutdown of SDL, as effectively they the mechanism is now broken.
Solution:
Ensure this code gets called when shutting down the timer subsystem:
#ifndef USE_GETTICKCOUNT
if (!hires_timer_available)
{
timeSetPeriod(0);
}
#endif
This makes it possible to move windows by their title bar, even if they're in
relative mode, if you click the title bar when the window does not have focus.
Bug: https://bugzilla.libsdl.org/show_bug.cgi?id=2396
Previously we were postponing our -[NSOpenGLContext update] call to the next
SDL_GL_SwapWindow, even if the context was current on the current thread. This
changes it so that we will do the update immediately if it's the current
context.
If you're rendering on another thread, you need to call SDL_GL_SwapWindow once
after a resize event to ensure your drawable will produce non-garbage data.
Bug: https://bugzilla.libsdl.org/show_bug.cgi?id=2339
If the window has been created with values for SDL_GL_CONTEXT_PROFILE_MASK,
SDL_GL_CONTEXT_MAJOR_VERSION and SDL_GL_CONTEXT_MINOR_VERSION not matching those
required by the renderer, attempt to recreate the window.
This is needed on platforms where both GL and GLES 1/2 surfaces are supported
by the video backend, requiring that the window be recreated when switching
between context types.
Pressing the hardware back button on a Windows Phone 8 device will now cause SDL to emit a pair of key-down and key-up events, with the SDL scancode, SDL_SCANCODE_AC_BACK.
By default, if WinRT's native back-button-press events are not explicitly marked as 'handled', then Windows Phone will terminate the app. More details on Microsoft's reasoning behind this can be found on MSDN, at http://msdn.microsoft.com/en-us/library/windowsphone/develop/jj247550(v=vs.105).aspx
To mark back-button-press events as 'handled', set SDL_HINT_WINRT_HANDLE_BACK_BUTTON to 1. Setting it to anything else will cause these events to not be marked as 'handled'.
Due to limitations in Windows Phone's APIs, SDL will emit a virtual key-up event immediately after the back button's key-down event is registered. Unfortunately, Windows Phone 8 only allows one to register for back-button-press events, and not back-button-release events.
This change is only relevant for Windows 8, 8.1, and RT apps, and only for those that are network-enabled. Such apps must feature a link to a privacy policy, which must be displayed via the Windows Settings charm. This is needed to pass Windows Store app-certification.
Using SDL_SetHint, along with SDL_HINT_WINRT_PRIVACY_POLICY_URL and optionally SDL_HINT_WINRT_PRIVACY_POLICY_LABEL, will cause SDL/WinRT to create a link inside the Windows Settings charm, as invoked from within an SDL-based app.
Network-enabled Windows Phone apps do not need to set this hint, and should provide some sort of in-app means to display their privacy policy. Microsoft does not appear to provide an OS-integrated means for displaying such on Windows Phone.
The destination target's alpha wasn't getting set correctly in many cases. Among other problems, this prevented some alpha-blended textures from displaying correctly in Windows Phone 8's multitasking screen.
The d3d11 renderer now uses the same blending settings found in the d3d9 renderer.
The projection and view matrices are now computed ahead of time, as they both get computed in the same spot, and typically not often. If this does, however, become a performance problem later on, this change can always be reverted.
Previously, the shaders would get compiled separately, the output of which would need to be packaged into the app. This change should make SDL's dll be the only binary needed to include SDL in a WinRT app.
Due to the new "tap and hold" IME in Mountain Lion and above, we were getting
inconsistent repeat of SDL_TEXTINPUT events. Disabling that functionality (since
you can't see the popover anyway) solves this.
Bug: https://bugzilla.libsdl.org/show_bug.cgi?id=2387
The DualShock 4 has all elements listed twice: once in the top-level list of
elements, and once in an "Application Collection" element at the top-level.
Each element has a proper cookie with a unique value, so now we descend into
each element collections, but before we add an element to the device's list,
we make sure we don't already have one with that cookie, probably from
another collection or a buggy device.
Tiemo Jung
All CPU detection functions SDL_Has* will return false, even if it is supported by the CPU, if SDL is compiled with MSVC and the target is x64.
The reason for this is that 'CPU_haveCPUID' will return 0 and macro 'cpuid' in SDL_cpuinfo.c is the fallback implementation, which sets all params to zero.
It is safe to assume that cpuid is supported on a CPU that runs windows x64, so CPU_haveCPUID can just return 1, and the empty macro can be replaced with a small wrap around the __cpuid intrinsic.
Felix Geyer
Starting from version 2.0.1 libsdl fails to build on Debian GNU/kFreeBSD in SDL_cpuinfo.c.
GNU/kFreeBSD defines __FreeBSD_kernel__ but not __FreeBSD__.
The #ifdef __FreeBSD__ check should be extended for __FreeBSD_kernel__, see the attached patch.
Build log:
libtool: compile: gcc -g -O2 -fstack-protector --param=ssp-buffer-size=4 -Wformat -Werror=format-security -D_FORTIFY_SOURCE=2 -DUSING_GENERATED_CONFIG_H -Iinclude -I/?BUILDDIR?/libsdl2-2.0.1+dfsg1/include -mmmx -m3dnow -msse -msse2 -fvisibility=hidden -D_REENTRANT -I/usr/include/dbus-1.0 -I/usr/lib/x86_64-kfreebsd-gnu/dbus-1.0/include -DHAVE_USBHID_H -DUSBHID_NEW -D_REENTRANT -Wall -c /?BUILDDIR?/libsdl2-2.0.1+dfsg1/src/cpuinfo/SDL_cpuinfo.c -fPIC -DPIC -o build/.libs/SDL_cpuinfo.o
/?BUILDDIR?/libsdl2-2.0.1+dfsg1/src/cpuinfo/SDL_cpuinfo.c: In function 'SDL_GetSystemRAM':
/?BUILDDIR?/libsdl2-2.0.1+dfsg1/src/cpuinfo/SDL_cpuinfo.c:632:35: error: 'HW_MEMSIZE' undeclared (first use in this function)
int mib[2] = {CTL_HW, HW_MEMSIZE};
^
/?BUILDDIR?/libsdl2-2.0.1+dfsg1/src/cpuinfo/SDL_cpuinfo.c:632:35: note: each undeclared identifier is reported only once for each function it appears in
make[2]: *** [build/SDL_cpuinfo.lo] Error 1
Tim McDaniel
On OSX, an SDL app forces a system shutdown to be cancelled. This happens because [SDLAppDelegate applicationShouldTerminate] returns NSTerminateCancel. A better approach is to subclass NSApplication and override terminate to do nothing except call SDL_SendQuit. In response to a system shutdown notification, this allows the normal SDL Quit event processing to occur, and if the app then terminates, system shutdown occurs normally. Please see the attached patch, based on SDL 2.0.1.
Tim McDaniel
On OSX, the app name in the menu bar is not localized. This can be fixed using the following implementation for GetApplicationName in SDL_cocoaevents.m:
static NSString *
GetApplicationName(void)
{
NSDictionary *dict;
NSString *appName = 0;
appName = [[NSBundle mainBundle] objectForInfoDictionaryKey:@"CFBundleDisplayName"];
if (!appName)
appName = [[NSBundle mainBundle] objectForInfoDictionaryKey:@"CFBundleName"];
if (![appName length])
appName = [[NSProcessInfo processInfo] processName];
return appName;
}
- Use the SDL_AUDIO_MASK_DATATYPE bit when selecting an implementation
where it matters. Previously two existing AUDIO_F32 cases had been
written, but were unreachable.
- Add AUDIO_F32 case for SDL_ConvertSurround_4.
- Fix incorrect pointer arithmetic causing the 2 to 6 channel
conversion for 4 byte audio formats to read and write beyond the end
of the buffer.
Leonardo
Event watchers are being executed on the inverse order they are added because they are added to the head of the SDL_event_watchers list.
Since watchers are allowed to change events before they are reported (they shouldn't, imo), this breaks code that rely on watcher execution order (such as distributed event handling).
An easy scenario to see this behaving weird to the user is if you add an event watcher to check mouse coordinates and check them again in your event loop. If you add the watcher after renderer's one (which always happens after you have initialized renderer), you get the same event but different coordinates.
The proposed patch adds the event watcher in the tail of the list, not in the beginning, and correctly fixes this problem.
Alex Szpakowski
Currently the UIKit/EAGL backend for SDL's OpenGL context creation API doesn't support OpenGL ES 3, despite iOS 7+ being capable (on devices with the necessary hardware.)
I have attached a patch to add support. It's also slightly more future-proof, so eventual OpenGL ES 4+ capability on iOS should hopefully work without requiring changes to SDL's UIKit/EAGL backend.
Alex Szpakowski
On my Mac OS X system (10.9.1), the SDL_MOUSEWHEEL event reports negative X values when my trackpad scrolls to the right, and positive X values when my trackpad scrolls to the left. This is backwards from what I'd expect, and I don't think it matches the Windows wheel events.
The vertical scroll values are what I'd expect though, and are consistent what gets reported on Windows (positive Y for scrolling up, negative Y for scrolling down.)
This is with "scroll direction: natural" disabled in the OS X trackpad settings (i.e. my scroll direction in non-SDL OS X programs matches what happens in Windows and Linux.)
I also tested with the horizontal scroll on a real mouse (Logitech G500 without custom drivers), and the horizontal scroll values in SDL are still flipped.
I "solved" the issue for myself by changing this line in the Cocoa_HandleMouseWheel function:
float x = [event deltaX];
to this:
float x = -[event deltaX];
I believe it should work fine with that change - I found something similar in another codebase while looking online for my issue - but I haven't tested on anything below Mac OS 10.8.
Besides being a little more simple to use than the earlier IOKit HID API, and
less likely to be deprecated, it also has the added benefit of working with
the Sony DualShock 4 controller in Bluetooth mode out of the box, whereas
the previous API has a bug that makes it report bad data for the
controller.
Cleaned up several other things in this code, having gone over every line of
it. The remaining deprecated calls are also gone.
