Carlos
Angle supports GLES3 but when using these functions (SDL_CreateWindow and SDL_CreateRenderer), defaults again to GLES2.0.
A current workaround (hack) to retrieve a GLES3.0 context with Angle is:
1) set
SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 3);
SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 0);
after InitSDL AND after calling SDL_CreateWindow (before SDL_CreateRenderer)
2) Comment lines 2032-2044 in SDL_render_gles2.c, funtion GLES2_CreateRenderer
window_flags = SDL_GetWindowFlags(window);
if (!(window_flags & SDL_WINDOW_OPENGL) ||
profile_mask != SDL_GL_CONTEXT_PROFILE_ES || major != RENDERER_CONTEXT_MAJOR || minor != RENDERER_CONTEXT_MINOR) {
changed_window = SDL_TRUE;
SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_ES);
SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, RENDERER_CONTEXT_MAJOR);
SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, RENDERER_CONTEXT_MINOR);
if (SDL_RecreateWindow(window, window_flags | SDL_WINDOW_OPENGL) < 0) {
goto error;
}
}
This retrives a GLES3 context as confirmed using glGetString(GL_VERSION). This should be fixed by modifying a few if's.
Ozkan Sezer
Since changeset 11607:60cd425a2f14, I am getting the following
error upon quit. Running testsprite2, clicking the mouse, and
quiting it is enough to trigger it. This is on my old Fedora9
x86-Linux:
X Error of failed request: BadCursor (invalid Cursor parameter)
Major opcode of failed request: 2 (X_ChangeWindowAttributes)
Resource id in failed request: 0xb057340
Serial number of failed request: 905
Current serial number in output stream: 906
Reverting https://hg.libsdl.org/SDL/rev/60cd425a2f14 removes
the error.
The audioqueue thread needs to keep running, and processing the CFRunLoop
until the AudioQueue is disposed of, otherwise CoreAudio will hang waiting for
final data to feed the device.
At least, I think this is how it all works. It definitely fixes the bug here!
Since AudioQueueDispose() calls AudioQueueStop() internally, there's no need
for our thread to handle this, either, which is good because the AudioQueue
would be disposed by this point. So now the AudioQueue is disposed first, and
then our thread is joined, and everything works out okay.
Just in case, we mark the device "paused" before setting everything in motion,
so any further callbacks from CoreAudio will write silence and not fire the
app's audio callback again.
Fixes Bugzilla #3868.
Trent Gamblin
The documentation for SDL_TouchFingerEvent says that the x and y coordinates are normalised between 0-1. I've found that to be true on Windows, Android and iOS but on X11 they are in pixel coordinates. This patch fixes the issue. This was the cleanest way I could do it with what was available without changing things around a lot but you may know a better way.
(I thought padding size ranged from 5 frames to ~30 frames (based around
RESAMPLER_ZERO_CROSSINGS, which is 5), but it's actually between 512 and
several thousands (based on RESAMPLER_SAMPLES_PER_ZERO_CROSSING)). It gets
big fast when downsampling.
Manuel
I would like this small patch merged that adds support for my GreenAsia Inc. PSX to USB converter, so SDL_IsGameController() returns true when using this adaptor.
It's interesting because PSX/PS2 controllers connected using this model won't be detected as gamecontrollers otherwise, only as joysticks.
Sylvain
There are various YUV-RGB conversion coefficients, according to https://www.fourcc.org/fccyvrgb.php
I choose the first (from Video Demystified, with integer multiplication),
but the current SDL2 Dither functions use in fact the next one, which follows a specifications called CCIR 601.
Here's a patch to use the second ones and with previous warning corrections.
There are less multiplications involved because Chroma coefficient is 1.
Also, doing float multiplication is as efficient with vectorization.
In the end, the YUV decoding is faster: ~165 ms vs my previous 195 ms.
Moreover, if SDL2 is compiled with -march=native, then YUV decoding time drops to ~130ms, while older ones remains around ~220 ms.
For information, from jpeg-9 source code:
jpeg-9/jccolor.c
* YCbCr is defined per CCIR 601-1, except that Cb and Cr are
* normalized to the range 0..MAXJSAMPLE rather than -0.5 .. 0.5.
* The conversion equations to be implemented are therefore
* Y = 0.29900 * R + 0.58700 * G + 0.11400 * B
* Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
* Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
jpeg-9/jdcolor.c
* YCbCr is defined per CCIR 601-1, except that Cb and Cr are
* normalized to the range 0..MAXJSAMPLE rather than -0.5 .. 0.5.
* The conversion equations to be implemented are therefore
*
* R = Y + 1.40200 * Cr
* G = Y - 0.34414 * Cb - 0.71414 * Cr
* B = Y + 1.77200 * Cb
Sylvain
Few issues with YUV on SDL2 when using odd dimensions, and missing conversions from/back to YUV formats.
1) The big part is that SDL_ConvertPixels() does not convert to/from YUV in most cases. This now works with any format and also with odd dimensions,
by adding two internal functions SDL_ConvertPixels_YUV_to_ARGB8888 and SDL_ConvertPixels_ARGB8888_to_YUV (could it be XRGB888 ?).
The target format is hard coded to ARGB888 (which is the default in the internal of the software renderer).
In case of different YUV conversion, it will do an intermediate conversion to a ARGB8888 buffer.
SDL_ConvertPixels_YUV_to_ARGB8888 is somehow redundant with all the "Color*Dither*Mod*".
But it allows some completeness of SDL_ConvertPixels to handle all YUV format.
It also works with odd dimensions.
Moreover, I did some benchmark(SDL_ConvertPixel vs Color32DitherYV12Mod1X and Color32DitherYUY2Mod1X).
gcc-6.3 and clang-4.0. gcc performs better than clang. And, with gcc, SDL_ConvertPixels() performs better (20%) than the two C function Color32Dither*().
For instance, to convert 10 times a 3888x2592 image, it takes ~195 ms with SDL_ConvertPixels and ~235 ms with Color32Dither*().
Especially because of gcc vectorize feature that optimises all conversion loops (-ftree-loop-vectorize).
Nb: I put no image pitch for the YUV buffers. because it complexify a little bit the code and the API :
There would be some ambiguity when setting the pitch exactly to image width:
would it a be pitch of image width (for luma and chroma). or just contiguous data ? (could set pitch=0 for the later).
2) Small issues with odd dimensions:
If width "w" is odd, luma plane width is still "w" whereas chroma planes will be "(w + 1)/2". Almost the same for odd h.
Solution is to strategically substitute "w" by "(w+1)/2" at the good places ...
- In the repository, SDL_ConvertPixels() handles YUV only if yuv source format is exactly the same as YUV destination format.
It basically does a memcpy of pixels, but it's done incorrectly when width or height is odd (wrong size of chroma planes). This is fixed.
- SDL Renderers don't support odd width/height for YUV textures.
This is fixed for software, opengl, opengles2. (opengles 1 does not support it and fallback to software rendering).
This is *not* fixed for D3D and D3D11 ... (and others, psp ?)
Only *two* Dither function are fixed ... not sure if others are really used.
- This is not possible to create a NV12/NV12 texture with the software renderer, whereas other renderers allow it.
This is fixed, by using SDL_ConvertPixels underneath.
- It was not possible to SDL_UpdateTexture() of format NV12/NV21 with the software renderer. this is fixed.
Here's also two testcases:
- that do all combination of conversion.
- to test partial UpdateTexture
Aaron
As of 2.0.6, all of my games are failing with the following error:
process 31778: arguments to dbus_type_is_basic() were incorrect, assertion "dbus_type_is_valid (typecode) || typecode == DBUS_TYPE_INVALID" failed in file dbus-signature.c line 322.
This is normally a bug in some application using the D-Bus library.
D-Bus not built with -rdynamic so unable to print a backtrace
(patch by Ozkan Sezer)
Evgeny Kapun
Commit 490bb5b49f11 [1], which was a fix for bug #3790, introduced a new bug: now, calling SDL_FreeSurface(surface) deallocates surface->map even if there are other references to the surface. This is bad, because some functions (such as SDL_ConvertSurface) assume that surface->map is not NULL.
Robert Turner
SDL_windowsevents.c contains code to retrieve the x and y coordinate for a requested hit test. It does this as follows:
POINT winpoint = { (int) LOWORD(lParam), (int) HIWORD(lParam) };
LOWORD(lParam) does not correctly mask off high bits that are set if the point is on a second (or third, etc.) monitor. This effectively offsets the x-coordinate by a large value.
MSDN documentation suggests that LOWORD() and HIWORD() are the wrong macros for the task, instead suggesting we should be doing something like the following:
POINT winpoint = { GET_X_LPARAM(lParam), GET_Y_LPARAM(lParam) };
Testing this change on my Windows 10 machine with 2 monitors gives the correct results.
- Fixing rendering borderless window. Need to force windows to send a WM_NCCALCSIZE then return 0 for non-client area size.
- Adding WS_CAPTION | WS_SYSMENU | WS_MINIMIZEBOX to borderless windows, for reasons noted in comments.
- Fix SetupWindowData() setting SDL_WINDOW_BORDERLESS. This was being cleared at window creation, causing hanlding for the first WM_NCCALCSIZE message to fail
Previously, the padding was silence, which was a problem when streaming since
you would sample a little bit of this silence between each buffer.
We still need a means to get padding data for the right hand side, but this
patch makes the resampler output more correct.
Anthony
This is what's making the software renderer crash with rotated destination rectangles of w or h = 0:
SDL_SetHint(SDL_HINT_RENDER_SCALE_QUALITY, "2");
This time it's using real math from a real whitepaper instead of my previous
amateur, fast-but-low-quality attempt. The new resampler does "bandlimited
interpolation," as described here: https://ccrma.stanford.edu/~jos/resample/
The output appears to sound cleaner, especially at high frequencies, and of
course works with non-power-of-two rate conversions.
There are some obvious optimizations to be done to this still, and there is
other fallout: this doesn't resample a buffer in-place, the 2-channels-Sint16
fast path is gone because this resampler does a _lot_ of floating point math.
There is a nasty hack to make it work with SDL_AudioCVT.
It's possible these issues are solvable, but they aren't solved as of yet.
Still, I hope this effort is slouching in the right direction.
This would cause playback problems in certain situations, such as on the
Raspberry Pi. The device that the wait was added for seems to not benefit from
it in modern times, and standard desktop Linux seems to do the right thing
when a USB device is unplugged now, without this patch.
Fixes Bugzilla #3599.
Andreas Falkenhahn
My app opens a 640x480 window. When I click on the window's maximize button, the window correctly fills the entire screen and loses its borders. But clicking on the restore button now doesn't restore the window to its original 640x480 size. Instead, the window size is identical to the screen size now. The only difference to the previous state is that the window now has borders again but it isn't restored to 640x480.
bastien.bouclet
The window is now resized to its specified size, but it moves to the top left corner of the screen. That is unexpected because neither the user nor the program moved it there. Test program attached (the same one as before).
Simon Hug
When RWops seeks with fseek or fseeko it uses the types long or off_t which can be 32 bits on some platforms. stdio_seek does not check if the 64-bit integer for the offset fits into a 32-bit integer. Offsets equal or larger than 2 GiB will have implementation-defined behavior and failure states would be very confusing to debug.
The attached patch adds range checking by using the macros from limits.h for long type and some bit shifting for off_t because POSIX couldn't be bothered to specify min and max macros.
It also defines HAVE_FSEEKI64 in SDL_config_windows.h so that the Windows function gets picked up automatically with the default config.
And there's an additional error message for when ftell fails.
Andreas Falkenhahn
When compiling SDL for the Raspberry Pi, I have to use the --host parameter to enable compilation of the native Raspberry Pi video driver, like so:
--host=arm-raspberry-linux-gnueabihf
It took me a while to figure out that this was necessary in order to have the native Raspberry Pi video driver compiled in. I think it would be better if there was an option like --enable-video-rpi that could be passed to configure and that would also show up when saying configure --help. Currently, it?s rather difficult to figure out that you have to use the --host parameter with arm-raspberry-linux-gnueabihf in order to get Raspberry Pi video support. It?s also somewhat inconsistent because most other video drivers can in fact be enabled/disabled through specific configure parameters but there is no such parameter for the native Raspberry Pi video driver.
Simon Hug
These are the remaining compiler warnings I see in the current tip cb049cae7c3c.
- SDL_test_log.c defines _CRT_SECURE_NO_WARNINGS without checking if it was already set.
- SDL_windowskeyboard.c converts integers to pointers without going over the (U)INT_PTR types. That bothers MSVC.
Now we try the new (hardware-specific) pathnames first, and if those fail to
load, we'll try the more generic names that earlier versions of Raspbian used.
Fixes Bugzilla #3800.
Mart?n Golini
I'm having a very slow initialization of the video subsystem that locks the window creation for about 500 ms ( tested in at least 4 different systems ). What i found is that X11_InitModes_XRandR is using XRRGetScreenResources, that explicitly ask to poll the hardware for changes. This is not really necessary since if the data is already available you can use XRRGetScreenResourcesCurrent.
I attached a tentative patch that fix this issue. With the patch there's no lock when the subsystem is initialized and the window creation is instant in my applications. The patch only uses XRRGetScreenResourcesCurrent in X11_InitModes_XRandR but it could be potentially used in X11_GetDisplayModes and X11_SetDisplayMode.
bastien.bouclet
When creating two surfaces and blitting them onto the other, SDL's internal reference counting fails, and one of the surfaces is not freed when calling SDL_FreeSurface.
Example code :
SDL_Surface *s1 = SDL_CreateRGBSurfaceWithFormat(0, 640, 480, 32, SDL_PIXELFORMAT_ARGB8888);
SDL_Surface *s2 = SDL_CreateRGBSurfaceWithFormat(0, 640, 480, 32, SDL_PIXELFORMAT_ARGB8888);
SDL_BlitSurface(s1, NULL, s2, NULL);
SDL_BlitSurface(s2, NULL, s1, NULL);
SDL_FreeSurface(s2);
SDL_FreeSurface(s1);
With this example, s1 is not freed after calling SDL_FreeSurface, its refcount attribute is still positive.
This also seems to fix the follow-up issue in bug #3719, whereby the initial fix caused the SDL window to move, after transitioning from fullscreen to windowed-mode
This is necessary because the Raspberry Pi is a strange beast, that believes
it has OpenGL support (through glX?) but generally has GLES2 support.
So when using the raspberry video target, we need to force this to default
to a GLES2 context, or by default SDL_CreateWindow() will fail, deep down
when it tries to load the proper GL library.
Fixes testsprite2 (and basically everything else that wasn't testgles2) when
run on a Raspberry Pi without a X server.
Please note that other targets might also need this filled in, the Raspberry
Pi is just the most prominent and readily-available System-On-A-Chip style
thing on my desk. :)
Romain Tisserand
Using KMS/DRM driver from WIP SDL2.0.6 on Linux/ARM SoC RockChip RK3328 (ARM Mali 450 MP2 GPU).
The current code is using GBM_BO_FORMAT_XRGB8888 as GBM buffer format specifier.
The Mali driver (it has been confirmed some other vendor implementations too) expects GBM_FORMAT_XRGB8888.
The Mesa implementation is actually handling both values as the same, but it's not implemented like this into every gbm.h vendor header.
https://github.com/ideak/mesa/blob/master/src/gbm/backends/dri/gbm_dri.c
So with stock SDL2 on my card (Mali vendor implementation), it does not work, eglCreateWindowSurface fails, and gbm_is_format_supported fails too (with the BO variant).
It runs fine with GBM_FORMAT_XRGB8888.
Here is a link of the gbm.h from Mali user-space driver :
https://github.com/rockchip-linux/libmali/blob/rockchip/include/gbm.h
This is necessary because we need to see if GLES compat extensions exist.
All of this code (including ShouldUseTextureFramebuffer()) should be
revisited after 2.0.6 ships; ideally we don't make throwaway contexts if
we can avoid it...but maybe we can't. I hear Vulkan is pretty cool.
Fixes Bugzilla #3725.
benjamin.feng
Probable underlying cause: https://bugzilla.libsdl.org/show_bug.cgi?id=3124#c5
"If you download and build the HID Calibrator sample you can see that these are totally legitimate HID devices (except for inverting the Y-axis of joysticks, which is contrary to the HID specification but does make them more compatible with games compiled expecting XBOX controllers)."
Colin Barrett
Using the pre-built x86 devel libs from here:
https://www.libsdl.org/release/SDL2-devel-2.0.5-VC.zip
If I have:
SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_ES);
SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 2);
SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 0);
SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
and I'm using ANGLE/(a GL driver that doesn't provide an ES2 context) such that SDL_EGL_CreateContext is called by SDL_GL_CreateContext, I get the error "Could not create EGL context (context attributes are not supported)" and no context is created.
Looking at the code in SDL_EGL_CreateContext - if gl_config.flags is non-zero, it looks like the code in the section guarded with "#ifdef EGL_KHR_create_context" should be executed - but it apparently isn't.
Is it possible this section hasn't been compiled into the pre-built libraries? If I build SDL2.dll myself using the Visual C++ solution (VS2015 Community Update 3) then the call succeeds as I expect
bastien.bouclet
When exiting a "fullscreen space" on OS X, windows don't go to their defined "windowed mode size", but go back to their previous size.
Steps to reproduce:
1. Create a windowed mode SDL window
2. Toggle it to fullscreen with the SDL_WINDOW_FULLSCREEN_DESKTOP flag
3. While in fullscreen, change the windowed mode size using SDL_SetWindowSize
4. Toggle the window back to windowed mode
Expected result:
- The window has the size specified during step 3.
Actual result:
- The window has the size specified when creating the window in step 1.
Attached is a minimal reproduction test case.
The attached test case works as expected on X11 and Windows.
Ozkan Sezer
fix windows build after revision 11382: commit 2026e42e377a renamed
_SDL_msctf_h to SDL_msctf_h_ . SDL_windowskeyboard.c relies on that
macro, so update it accordingly.
Ozkan Sezer
Since the Vulkan merge, building against a Mac OS X SDM older than
10.11 fails in SDL_cocoametalview.m because Metal.framework is not
present. There is no conditional compiling in SDL_cocoametalview.m
either, so --disable-video-vulkan doesn't help with anything. (The
configury doesn't check darwin for x86_64 either, but it's another
story.)
I cross-build against 10.8 SDK on linux using clang-3.4.2 and this
is a problem for me. Will this be fixed?
Simon Hug
This issue actually raises the question if this API change (requirement of initialized audio subsystem) is breaking backwards compatibility. I don't see the documentation saying it is needed in 2.0.5.
"Major changes, roughly in order of appearance:
- Use float math everywhere, instead of promoting to double and casting back
all the time.
- Conserve sound energy when downmixing any channel into two other channels.
- Add a QuadToStereo filter. (The previous technique of reusing StereoToMono
never worked, since it assumed an incorrect channel layout for 4.0.)
- Add a 71to51 filter. This removes just under half of the cases the previous
code would silently break in.
- Add a QuadTo51 filter. More silent breakage fixed.
- Add a 51to71 filter, removing another almost-half of the silently broken
cases.
- Add 8 to the list of values SDL_SupportedChannelCount will accept.
- Change SDL_BuildAudioCVT's channel-related logic to handle every case, and
to actually fail if it fails instead of silently corrupting sound data and/or
crashing down the road."
(Note that SDL doesn't otherwise support 7.1 audio yet, but hopefully it will
soon and the 7.1 converters are an important piece of that. --ryan.)
Fixes Bugzilla #3727.
David Brady
When I attempted to make a mapping file for Android gamepads, I quickly discovered that most of the ones that I have here show up as the same device (Broadcom Bluetooth HID), meaning that it was impossible to make mappings on Android, since every device looked the same.
This patch will check for the existence of the getDescriptor function added in Jelly Bean, and use it if it's there. The Android Dashboard says that the majority of Android phones should support this function, and doing it this way will not force us to bump up our API version.
Clayton Craft
The default path used by directfb for libGL is different than the default path used by x11 in SDL2:
./src/video/directfb/SDL_DirectFB_opengl.c:
path = "libGL.so";
./src/video/x11/SDL_x11opengl.c:
#define DEFAULT_OPENGL "libGL.so.1"
On at least one distro (Alpine Linux), libGL.so is not created (or more accurately the symlink to libGL.so.1 is not created). For consistency, the 'path' variable in SDL_DirectFB_opengl.c should patch the DEFAULT_OPENGL in SDL_x11opengl.c ("libGL.so.1")
Error message was:
[mvk-info] MoltenVK version 0.18.2. Vulkan version 1.0.51.
[***MoltenVK ERROR***] VK_ERROR_INITIALIZATION_FAILED: On-screen rendering requires a view that is backed by a layer of type CAMetalLayer.
2017-08-28 02:17:29.579 testvulkan[95627:1716939] ERROR: SDL_Vulkan_CreateSurface(): vkCreateMacOSSurfaceMVK failed: VK_ERROR_INITIALIZATION_FAILED
David Ludwig
I've created a new set of patches. I am happy to create more, if it would help.
One version only copies 'size'.
A second version copies both 'size' and 'silence'. When looking over the documentation for SDL_OpenAudio in SDL_audio.h, it mentioned that both 'size' and 'silence' were things that SDL_OpenAudio would calculate.
Regarding *both* patches, I did notice that SDL 1.2 appears to have always modified desired's size and silence fields. The SDL wiki, at https://wiki.libsdl.org/SDL_OpenAudio#Remarks , does note:
Carlos
We would like to add a switch (define) that allows us to compile Angle statically with SDL. That is, getting rid of the OpenGL DLL. Usually you need OpenGL to be loaded dynamically as DLL because implementation is provided by the system but no need with Angle.
Only 2 files need modification and it shouldn't affect current behaivor:
include/SDL_egl.h and src/video/SDL_egl.c, as in here
https://github.com/native-toolkit/sdl/pull/10/files
The flag name could be SDL_VIDEO_STATIC_ANGLE (instead of NATIVE_TOOLKIT_STATIC_ANGLE) as discussed here https://github.com/native-toolkit/sdl/pull/10
We have tested this with both Windows and UWP, using NME engine (https://github.com/haxenme/nme).
Releated issue: https://bugzilla.libsdl.org/show_bug.cgi?id=1820
Sylvain
Hi! here's a patch for that with two class loaded regarding API level.
Test both case : before API 11 and after.
I also remove now unused GetSystemServiceFromUIThread() and minor clean-up (haptic warning prototype).
This fixes a strange corner case (notes appended below), and should be
safe to do anyhow.
Fixes Bugzilla #3674.
"I did more tests.
It appears the bug only happens if there is
another window on the screen that has "always
on top" property. For me it is xawtv - it is
always opened in a screen corner. Closing
xawtv or removing "always on top" property
from it makes the problem to go away.
Plus, it doesn't appear like the buttons are
not delivered at all. It appears that instead
the button presses are delivered on some mouse
positions, but not delivered when you move the
mouse to other part of the window... So this is
really weird and is likely somewhere deep in the
Xorg.
Maybe somehow it happens that the cursor is
actually above the xawtv window, but, because
my app uses grab, it is not visible there, and
in that case the events are not delivered to
my app?
But with my patch the button events are
always delivered flawlessly, it seems.
Hmm, and that indeed seems to explain my problem:
if the mask is set properly and my app uses
grab, then, even if the mouse is above some
other window, the events would still be delivered
to the grabbing app, which is what actually wanted
because my app uses relative mouse mode, so it
doesn't know the pointer can cross some other window
(my app draws the pointer itself).
So my current theory is that my patch only enforces
the mouse grab, which otherwise can be tricked by
some other window preventing the button events
delivery (but motion events are still delivered
via xinput2, which makes it all look very obscure)."
This patch was originally written by Marc Di Luzio for glX and enhanced by
Maximilian Malek for WGL, etc. Thanks to both of you!
Fixes Bugzilla #3643.
Fixes Bugzilla #3735.
