* Added support for mini-gamepad mode for Joy-Con controllers, matching the mapping for hid-nintendo on Linux and iOS 16
* Added the ability to merge left and right Joy-Con controllers into a single Pro-style controller
* Added the hint SDL_HINT_JOYSTICK_HIDAPI_SWITCH_COMBINE_JOY_CONS to control this merging functionality
* Removed the hint SDL_HINT_JOYSTICK_HIDAPI_JOY_CONS
* Added support for mini-gamepad mode for Joy-Con controllers, matching the mapping for hid-nintendo on Linux and iOS 16
* Added the ability to merge left and right Joy-Con controllers into a single Pro-style controller
* Added the hint SDL_HINT_JOYSTICK_HIDAPI_SWITCH_COMBINE_JOY_CONS to control this merging functionality
* Removed the hint SDL_HINT_JOYSTICK_HIDAPI_JOY_CONS
- Touch events may be translated to mouse movement events without the normal Xinput2 raw motion events
being sent. Not all touch events will necessarily move the mouse but this ensures we update the global
mouse state just in case.
- Fix up some formatting
CR: saml
Cache off NSWindow's windowNumber in SDL_WindowData on setup and use that in `Cocoa_SendWakeupEvent` to prevent accessing windowNumber off the main thread.
Unopened devices, if removed, now send SDL_AUDIODEVICEREMOVED events with
a `which` field set to zero. Apps can use this to decide if they need to
refresh a list of devices being shown in an options menu, etc.
It's safe to call SDL_CloseAudioDevice(0), so even if they try to clean
up this bogus id, it should be safe.
Fixes#5199.
This allows setting the brightness of the home LED on Nintendo Switch Pro controllers, in the range 0.0 - 1.0.
This can be updated at runtime by setting the hint dynamically.
Fixes https://github.com/libsdl-org/SDL/issues/3787
SDL_events_need_periodic_poll() and SDL_events_need_polling() are intended to allow the event loop to update joysticks and/or sensors if needed, however those systems only update when the SDL_update_joysticks and/or SDL_update_sensors variables are true. This change brings the behavior of these functions in line with if work will actually need to be performed.
This change allows the hints for AUTO_UPDATE to influence the polling behavior of the event loop such that an app can choose to update joysticks/sensors itself and avoid the expense of constantly sleeping and waking the event loop. Additionally in makes these functions marginally faster in some situations by not searching the active events.
Hint callbacks are called before the actual value in the hint is changed, so the functions SDL_AutoUpdateJoysticksChanged and SDL_AutoUpdateSensorsChanged were not actually properly updating their respective variables in repsonse to their auto update hint changing.
Instead, we pull the new hint value out of the value passed into the callback and use that to update the variables. Assume true on a null value as that was the previous behavior and it matches with the default values of SDL_update_joysticks/SDL_update_sensors.
Events to handle controller touchpads and sensors were added to the library but not added in `SDL_GameControllerEventState()`. This change adds the missing events.
- CMD_CHARGE_STATE was checking the seqnum instead of the payload
- Off-by-one error in size validation for command payload
- Unused payload space was left uninitialized in output report
As discussed in PR review, there may be an off-chance that the index
returned doesn't match up with SDL's display indexing.
This change ensures that the indices match and adds a safety check for
off-screen windows.
With the introduction of this function, it is possible that for certain
monitor and window configurations, creating an SDL window will cause a
native crash.
```
Crashed Thread: 0 Dispatch queue: com.apple.main-thread
Exception Type: EXC_BAD_ACCESS (SIGSEGV)
Exception Codes: KERN_INVALID_ADDRESS at 0x0000000000000050
Exception Codes: 0x0000000000000001, 0x0000000000000050
Exception Note: EXC_CORPSE_NOTIFY
Termination Reason: Namespace SIGNAL, Code 11 Segmentation fault: 11
Terminating Process: exc handler [56627]
VM Region Info: 0x50 is not in any region. Bytes before following region: 140737486737328
REGION TYPE START - END [ VSIZE] PRT/MAX SHRMOD REGION DETAIL
UNUSED SPACE AT START
--->
VM_ALLOCATE 7fffffe75000-7fffffe76000 [ 4K] r-x/r-x SM=ALI
Thread 0 Crashed:: Dispatch queue: com.apple.main-thread
0 libSDL2.dylib 0x10247f665 SDL_UpdateFullscreenMode + 357
1 libSDL2.dylib 0x10247ec70 SDL_CreateWindow_REAL + 1504
2 ??? 0x111262de8 ???
3 ??? 0x110c39fff ???
4 libcoreclr.dylib 0x101fdf2a9 CallDescrWorkerInternal + 124
```
Tracking thread from our end: https://github.com/ppy/osu-framework/issues/5190
Regressed with: https://github.com/libsdl-org/SDL/pull/5573
In testing, the window would not find a valid screen if created
"hanging" off a primary display with a secondary display below it. In
checking why this was the case, the `display_centre` was being
calculated with a negative y origin, causing a final negative value
falling outside all display bounds:
```
SDL error log [debug]: display_centre.y = -1296 + 1296 / 2
SDL error log [debug]: Display rect 0: 0 0 2560 1440
SDL error log [debug]: Display rect 1: 2560 -625 1080 2560
SDL error log [debug]: Display rect 2: 0 1440 1728 1296
```
The method that was being used to find the current window using the frame
origin/size seems unreliable, so I have opted to replace it with with a
tried method (https://stackoverflow.com/a/40891902).
Initial testing shows that this works with non-standard DPI screens, but
further testing would be appreciated (cc @sezero / @misl6 from the
original PR thread).
We actually request CSD mode with xdg-decoration for borderless
windows, so we get what we wanted there and there's no point in going
into fallback paths.
The 5.1 versions didn't use the new algorithm, and making that new
algorithm work took so many permutes that it was significantly slower
than just using the scalar versions.
However, mono-to-stereo is an extremely common conversion, and it's
trivial to accelerate it with plain SSE, so that was added!
Currently, the SDL_WINDOWEVENT_RESIZED event is sent before the actual
window is resized (and various internal state, such as the desired
GL/Vulkan backbuffer size, are updated). This makes sense, as SDL will
discard a no-op resize, which would be the case if we had resized before
sending the event (indeed, there are existing hacks to prevent this).
However, this means that SDL_{GL,Vulkan}_GetDrawableSize() will still
use the old size in the SDL_WINDOWEVENT_RESIZED handler. In the case of
SDL_Renderer, this means the drawable size it uses will be wrong, and
the viewport will get "updated" to the old value.
This then results in bug #5899.
Pipewire 0.3.44 introduced PW_KEY_TARGET_OBJECT, which is to be used to specify target connection nodes for streams. This parameter takes either a node path (PW_KEY_NODE_NAME) or serial number (PW_KEY_OBJECT_SERIAL) to specify a target node. The former is used in this case since the path is already being retrieved and stored for other purposes.
The target_id parameter in pw_stream_connect() is now deprecated and should always be PW_ID_ANY when PW_KEY_TARGET_OBJECT is used.
Make the default device metadata node persist for the lifetime of the hotplug loop so the default source/sink devices will be updated if they change during runtime.
With Pipewire now requiring a minimum version 0.3.24, the PW_KEY_CONTEXT_PROFILE_MODULES value is no longer required for legacy compatability and can be safely removed.
* Fix assumption that DRI_DEVNAME begins at 0
The existing logic of the code was to count every possible entry in
KMSDRM_DRI_PATH. After this a for loop would start trying to open
filename0, filename1, filename2, etc. In recent Linux kernels (say
5.18) with simpledrm, the lowest KMSDRM_DRI_DEVNAME is often
/dev/dri/card1, rather than /dev/dri/card0, causing the code to fail
once /dev/dri/card0 has failed to open. Running:
modprobe foodrm && modprobe bardrm && rmmod foodrm
before you try to run an application with SDL KMSDRM would have also
made this fail.
* Various changes from review
- Removed newline and period from SDL error
- Explicitely compare memcmp to zero (also changed to SDL_memcmp)
- Changed memcpy to strncpy
- Less aggressive line wrapping
* Various changes from review
- strncpy to SDL_strlcpy
- removed size hardcodings for KMSDRM_DRI_PATHSIZE and
KMSDRM_DRI_DEVNAMESIZE
- made all KMSDRM_DRI defines, run-time variables to reduce bugs caused
by these defines being more build-time on Linux and more run-rime on
OpenBSD
- renamed openbsd69orgreater variable to moderndri
- altered comment from "if on OpenBSD" to add difference in 6.9
* Various changes from review
- Use max size of destination, rather than max size of source
- Less hardcodings
When building on macOS without gcc (e.g. clang) where HAVE_GCC_ATOMICS
is not defined, `SDL_AtomicTryLock` will call
`OSAtomicCompareAndSwap32Barrier` which is not yet declared.
Including OSAtomic.h on OSX resolves this error/warning:
SDL_spinlock.c:125:12: error: implicit declaration of function
'OSAtomicCompareAndSwap32Barrier' is invalid in
C99 [-Werror,-Wimplicit-function-declaration]
return OSAtomicCompareAndSwap32Barrier(0, 1, lock);
This was reported in issue #3885 but marked Invalid and closed - possibly
because the default CMake build uses gcc instead of clang.
The documentation doesn't state that the argument is ever modified,
and no implementation does so currently.
This is a non-breaking change to guarantee as much to callers.
strchr and strrchr return a pointer to the first/last occurrence of a
character in a string, or NULL if the character is not found. According
to the C standard, the final null terminator is part of the string, and
it should thus be possible to get a pointer to the final null with
these functions. The fallback implementations of SDL_strchr and
SDL_strrchr would always return NULL if trying to find '\0', and this
commit fixes that.
* Xbox GDK support (14 squashed commits)
* Added basic keyboard testing
* Update readme
* Code review fixes
* Fixed issue where controller add/removal wasn't working (since the device notification events don't work on Xbox, have to use the joystick thread to poll XInput)
Also added test functions for multi-line debug text display
Currently this only supports ASCII, as the font doesn't have the correct Latin-1 characters
The HIDAPI joystick driver doesn't properly reset the change counter
it uses to track if re-enumeration is needed when the joystick
subsystem is quit and then reinitialized.
The first SDL_Init(SDL_INIT_JOYSTICK) will result in the expected
HIDAPI joysticks appearing, but subsequent calls will result in no
joysticks being enumerated until another HIDAPI joystick is added
or removed from the system.
Use SDL_lroundf() to round fractional backbuffer sizes halfway away from zero, as this is the rounding method recommended by the forthcoming Wayland fractional scaling protocol.
We might want to use ssize_t as @Guldoman suggested, but that's a larger internal API change, and still requires casting of the SDL_utf8strnlen() result.
Fixes https://github.com/libsdl-org/SDL/pull/5821
This allows us to handle controllers that use the Xbox protocol but look like Nintendo Switch or Playstation controllers, like the Qanba Dragon Arcade Stick in PC mode
This is done such that we can disable LTO for these 2 functions when
building with MSVC.
This is due to a limitation of Link Time Code Generation (LTCG).
Code generation might generate a new reference to memset after linking
has started. The LTCG must make assumptions about where memset is
defined which is normally the C runtime.
Some compositors will provide 'nicer' / 'human readable' output descriptions via the xdg-output protocol. Use these description strings, when available, instead of the model name provided by wl-output. On compositors such as GNOME where this is provided, the display names provided to applications by SDL will now match those in the desktop display settings panel. On compositors where this data isn't provided, the old behavior of using the model string provided by wl-output will remain unchanged.
Additionally, per the protocol spec, output data provided by xdg-output should supersede wl-output data, so this is the recommended behavior in general.
If the width is sufficiently ludicrous, then the calculated pitch or
the image size could conceivably be a signed integer overflow, which
is undefined behaviour. Calculate in the unsigned size_t domain, with
overflow checks.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Adds hint "SDL_WINDOWS_DPI_SCALING" which can be set to "1" to
change the SDL coordinate system units to be DPI-scaled points, rather
than pixels everywhere.
This means windows will be appropriately sized, even when created on
high-DPI displays with scaling.
e.g. requesting a 640x480 window from SDL, on a display with 125%
scaling in Windows display settings, will create a window with an
800x600 client area (in pixels).
Setting this to "1" implicitly requests process DPI awareness
(setting SDL_WINDOWS_DPI_AWARENESS is unnecessary),
and forces SDL_WINDOW_ALLOW_HIGHDPI on all windows.
If the move results in a DPI change, we need to allow the window to resize (e.g. AdjustWindowRectExForDpi frame sizes are different).
- WM_DPICHANGED: Don't assume WM_GETDPISCALEDSIZE is always called for PMv2 awareness - it's only called during interactive dragging.
- WIN_AdjustWindowRectWithStyle: always calculate final window size including frame based on the destination rect,
not based on the current window DPI.
- Update wmmsg.h to include WM_GETDPISCALEDSIZE (for WMMSG_DEBUG)
- WIN_AdjustWindowRectWithStyle: add optional logging
- WM_GETMINMAXINFO: add optional HIGHDPI_DEBUG logging
- WM_DPICHANGED: fix potentially clobbering data->expected_resize
Together these changes fix the following scenario:
- launch testwm2 with the SDL_WINDOWS_DPI_AWARENESS=permonitorv2 environment variable
- Windows 10 21H2 (OS Build 19044.1706)
- Left (primary) monitor: 3840x2160, 125% scaling
- Right (secondary) monitor: 2560x1440, 100% scaling
- Alt+Enter, Alt+Enter (to enter + leave desktop fullscreen), Alt+Right (to move window to right monitor). Ensure the window client area stays 640x480. Drag the window back to the 125% monitor, ensure client area stays 640x480.
The hint allows setting a specific DPI awareness ("unaware", "system", "permonitor", "permonitorv2").
This is the first part of High-DPI support on Windows ( https://github.com/libsdl-org/SDL/issues/2119 ).
It doesn't implement a virtualized SDL coordinate system, which will be
addressed in a later commit. (This hint could be useful for SDL apps
that want 1 SDL unit = 1 pixel, though.)
Detecting and behaving correctly under per-monitor V2
(calling AdjustWindowRectExForDpi where needed) should fix the
following issues:
https://github.com/libsdl-org/SDL/issues/3286https://github.com/libsdl-org/SDL/issues/4712
I was looking at how errors are handled by SDL and came across this #define SDL_ERRBUFIZE which looks like a typo for SDL_ERRBUFSIZE, but either way, it looks like this isn't being used anywhere anymore because it was getting reported whenever I compile SDL with -Wunused-macros, and the last time it was mentioned in the code was from commit 09ca66b.
XDG-toplevel min/max size values are double-buffered data and must be committed before entering the fullscreen state, or a max window size value smaller than the display dimensions may cause the compositor to incorrectly configure the fullscreen window size. This fixes windowed->fullscreen transitions on GNOME, where, previously, certain combinations of window flags and min/max size values could cause entering fullscreen mode to fail with odd window sizes and/or offsets due to the new max size values not being committed before entering fullscreen, causing the compositor to clamp to the old values.
In the case of libdecor, it has its own layer of buffering on top of the xdg-toplevel surface for the min/max window dimensions, so both a frame commit and surface commit are required to set the state properly.
Previously, the surface damage region was being set in the same callback used for preventing render hangs in the GL backend when the surface was not visible. This was not ideal, as the callback was never fired in the case of using a different render backend or having a swap interval of 0. Use a separate frame callback for setting the surface damage region to ensure that it fires reliably, regardless of the backend being used or swap interval.
Some compositors (GNOME for example) don't set the transform flag when dealing with portrait mode displays, so the video modes won't have the width/height swapped in all cases where they should be. Check for both the 90/270 degree transform flag and if the display is taller than it is wide when determining whether to swap the width and height of the emulated video modes, and adjust the comparison logic when size testing against the native mode to account for this.
Add the hint "SDL_VIDEO_WAYLAND_MODE_EMULATION", which can be used to disable mode emulation under Wayland. When disabled, only the desktop and/or native display resolution is exposed.
Previously, scale values used by the displays and surfaces were always integers, with fractional scale values only being calculated when the backbuffer and viewport sizes were being determined. Now, if xdg-output is available, the fractional scale of output displays is calculated when the displays are enumerated and the true scale values of the output devices are used whenever possible.
This unifies the integer and fractional scaling systems, allows for the use of more accurate scale values that minimize overdraw when windows straddle multiple outputs, and lays the groundwork for the pending Wayland scaling protocols that will report the preferred scale values per-surface instead of per-output.
Compartmentalize the fullscreen mode emulation code blocks, unify the windowed/fullscreen viewport logic, consolidate all window geometry code into a central function to eliminate blocks of duplicate code and rename related variables and functions to more explicitly reflect their purpose.
Because we were sending multiple chunks of preedit strings,
`SDL_SendEditingText` was using the old `SDL_TEXTEDITING` event only.
Now if `SDL_HINT_IME_SUPPORT_EXTENDED_TEXT` is enabled, we send the full
string and correctly set the cursor position and selection size.
- SDL_JoystickGUID -> SDL_GUID (though we retain a type alias)
- Operations for GUID <-> String ops are now in
src/SDL_guid.c and include/SDL_guid.h
- The corresponding Joystick operations delegate to SDL_guid.c
- Added test/testguid.c
As of #5703, we call libdecor_dispatch() in Wayland_WaitEventTimeout(),
but this will crash if we don't load libdecor, as
SDL_VideoData::shell.libdecor will be NULL.
Since we don't load libdecor if we don't intend to use it (i.e., if
should_use_libdecor returns false), this results in a crash under KDE in
almost all circumstances.
The MinGW-64 header defines the parameters as ABI::Windows::Foundation::IReference<INT32 > **, but the Windows header defines the parameters as __FIReference_1_int**
Since #5602, SDL is intended to have the same ABI across the whole
major-version 2 cycle, so we should not check that the minor version
matches the one that was used to compile an application.
There are two checks that could make sense here.
The first check is that the major version matches the expected major
version. This is usually unnecessary and is not usually done (if we're
calling into the wrong library we'll likely crash anyway), but since we
have the information, we might as well continue to use it.
The second check is whether the version provided by the caller is
equal to or greater than a threshold version at which additional fields
were added to the struct. If it is, we should populate those fields;
if it is not, then we cannot. This is only useful on platforms where
additional fields have genuinely been added during the lifetime of
SDL 2, like Windows and DirectFB (but not X11).
This commit changes the first check to be consistent about only looking
at the minor version, while leaving the second check using SDL_VERSIONNUM
(which will be removed or widened in SDL 3, but it's fine for now).
Resolves: https://github.com/libsdl-org/SDL/issues/5711
Fixes: cd7c2f1 "Switch versioning scheme to be the same as GLib and Flatpak"
Signed-off-by: Simon McVittie <smcv@collabora.com>
Otherwise we ignore the Configure/etc events when they come in because
the window is already in an identical state as far as SDL is concerned.
Fixes#5593.
May also fix:
Issue #5572.
Issue #5595.
This reverts commit 3fcc2cb500.
Button4 and Button5 are for the scrollwheel, not the extended buttons.
I don't know of a way to query the state of the extended buttons using X11.
These try to pull from the .pdf files that are installed with
macOS, which fit our needs better, and fall back to the most
reasonable defaults available from NSCursor if we can't load
them.
Since these are installed under /System, they should be sandbox
accessible, and if this totally fails, it should still go on,
albeit with a less good cursor.
Reference Issue #2123.
On Gnome (and hopefully others!), this produces something that actually
matches SDL_SYSTEM_CURSOR_SIZENWSE/SDL_SYSTEM_CURSOR_SIZENESW. On
other desktop enviroments, it probably fits the spirit better than
XC_fleur in any case.
Reference Issue #2123.
Added the ability to specify a name and the product VID/PID for a virtual controller
Also added a test case to testgamecontroller, if you pass --virtual as a parameter
Don't be fooled by the diff size - this ended up being a big refactor of the
shell surface management, masked only by some helper macros I wrote for the
popup support.
This change makes it so when xdg_decoration is supported, but CSD is requested,
the system bails on xdg support entirely and resets all the windows to use
libdecor instead. This transition isn't pretty, but once it's done it will be
smooth if decorations are an OS toggle since libdecor will take things from
there.
In hindsight, we really should have designed libdecor to be passed a toplevel,
having it manage that for us keeps causing major refactors for _every_ change.
Move rendering of the assert message into a separate
function so we can remove the ugly loop construction.
Changes the logic such that allocation failure no longer
immediately returns SDL_ASSERTION_ABORT, instead we
fall back to the truncated message.
If an error is indicated from SDL_snprintf, then we do
abort with SDL_ASSERTION_ABORT.
* Add changes from code review by @ccawley2011, #5597, overall cleanup
* Update N-Gage README, minor cleanup and rephrasing
* Call SDL_SetMainReady() before calling SDL_main, return SDL_main instead of main
Change Cocoa SDL_VideoData and SDL_WindowData implementations from C structs to Objective-C objects, since bridging between C and ObjC is easier that way.
If the size to be allocated is very large and untrusted, then adding
the padding etc. might be enough to cause unsigned overflow, after
which a very small amount of memory will be allocated.
Signed-off-by: Simon McVittie <smcv@collabora.com>
If we're strict about applying something resembling semantic versioning
to the "marketing" version number, then we can mechanically generate
the ABI version from it.
This limits the range of valid micro versions (patchlevels) to 0-99.
Signed-off-by: Simon McVittie <smcv@collabora.com>
For stable releases, this gives us the ability to make bugfix-only point
releases such as 2.24.1 if we want to, and distinguish between them
programmatically. For example, this ability could have been useful after
2.0.16 to fix Xwayland regressions, and after 2.0.18 to fix event loop
regressions.
For development releases, this gives us the ability to make multiple
prereleases during the same feature cycle, and distinguish between them
programmatically. For example, this would have been useful during 2.0.22
development, which went through three prereleases before reaching the
final release.
Signed-off-by: Simon McVittie <smcv@collabora.com>
* Add initial support for the Nokia N-Gage
* N-Gage: disable clipping for the time being, issue needs to be resolved later
* Move va_copy definition to SDL_internal.h
* Move stdlib.h include to SDL_config_ngage.h, much cleaner this way
* Remove redundant include, add HAVE_STDLIB_H
* Revert "N-Gage: disable clipping for the time being, issue needs to be resolved later"
This reverts commit 4f5f0fc36cc7f34fad05e45671dfa7b8dc32fd51.
* N-Gage: fix clipping issue by providing proper math functions
Enabling GCController.shouldMonitorBackgroundEvents to read background events
for MFi controllers before receiving the first GCControllerDidConnectNotification
is apparently a no-go on macOS (12.3.1 for me), and would crash on attempt.
Apple's documentation is... not great, and doesn't point this out.
This waits for IOS_AddMFIJoystickDevice() to get called down the chain from GCControllerDidConnectNotification, and enables GCController.shouldMonitorBackgroundEvents
if it hadn't been already.
On iOS and tvOS, GCController.shouldMonitorBackgroundEvents is ignored, so
there's no need to check their versions.
Ensure that we're not trying to call SDL_small_alloc()
with a count of zero.
Transforming the code like this fixes a
-Wmaybe-uninitialized warning from GCC 12.0.1
For short messages, use a stack buffer that is
significantly smaller than SDL_MAX_LOG_MESSAGE.
The rationale for this is that we don't want to risk
blowing the stack, while at the same time we would
like to not put pressure on the memory allocator unless
absolutely necessary.
This is the one that splits the "left wing" into two for loops to
bubble out the conditional that decides if it should read from the
left padding or the input buffer.
I still believe the optimization is good, but the basic logic of it
was incorrect, and needs to be reexamined and fixed before going
back into revision control.
We can get _some_ of the info we need out of standard Xlib and report a
single display (which might actually be multiple physical displays mushed
into a single desktop). This is better than nothing, but you should really
just build with XRandR support and get a better X server. :)
- Calculate `j * RESAMPLER_SAMPLES_PER_ZERO_CROSSING` once per loop
iteration since we use it multiple times.
- Do the left-wing loop in two sections: while `srcframe < 0` and then
the remaining calculations when `srcframe >= 0`. This bubbles a conditional
out of every iteration of a tight loop, giving us a boost. We could
_probably_ do this to the right-wing loop too, but it's less straightforward
there.
- The real win: Use floats instead of doubles. This almost doubles the speed
of the entire function on Intel CPUs, and for embedded things without
hardware-level support for doubles, the speedup is enormous. This in
theory might reduce audio quality, though, and I had to put a check in
place to avoid a division-by-zero that we avoided at higher precision, but
this is likely to be worth keeping for at least the Sony PSP and other
smaller platforms, if not everyone.
Instead of waiting until the entire buffer from the SDL callback is ready
to be accepted by PulseAudio, we use pa_stream_set_write_callback and
feed some portion of the buffer as callbacks come in asking for more.
This lets us remove the halving of the buffer size during device open,
and also (hopefully) solves several strange hangs that happen in unusual
circumstances.
Fixes#4387Fixes#2262