The string pointer used to write the contents of our clipboard data
offer was never updated, causing it to repeat parts of the beginning of
the string until the correct number of bytes had been written.
Emitting an error for one specific type of failure in retrieving the
correct name for a display is not very useful, especially when
initialization is otherwise unaffected.
There should be a path for information like that but this isn't it.
Fixes#1791
Operations that take an instance handle should be passed the handle of
whatever module we are inside instead of blindly passing the handle of
the executable.
This commit makes GLFW retrieve its own instance on initialization.
This makes the most difference for window classes, which are
per-instance. Using the executable instance led to name conflicts if
there were several copies of GLFW in a single process.
Note that having this is still a bad idea unless you know what things to
avoid, and those things are mostly platform-specific. This is partly
because the library wasn't designed for it and partly because it needs
to save, update and restore various per-process and per-session settings
like current context and video mode.
However, multiple simultaneous copies of GLFW in a single Win32 process
should now at least initialize, like is already the case on other
platforms.
Fixes#469Fixes#1296Fixes#1395
Related to #927
Related to #1885
Alt+PrtSc emits a different scancode than just PrtSc. Since the GLFW
API assumes each key corresponds to only one scancode, this cannot be
added to the keycodes array.
Instead we replace the scancode at the point of entry.
Fixes#1993
This change broke key names for dead keys, because that depends on
the effect on the global state of simulating two key presses.
Issue found by CTest tests.
Reverts #2018
The normal way of maximizing a window also makes it visible. This
implements window maximization manually for when the window passed to
glfwMaximizeWindow is hidden.
This will very likely not be forward-compatible and should be replaced.
The window content scale correction at creation overwrote the inital,
more pleasant placement of the window by CW_USEDEFAULT, if the window
was created with GLFW_MAXIMIZED set. This is because the translation
to screen coordinates was done using the current position, not the
position from the restored window rect.
A window created maximized and undecorated would cover the whole monitor
Windows placed it on instead of just that monitor's workarea.
This commit adjusts the maximized rect to cover just the workarea,
similar to how undecorated windows that become maximized are handled
during WM_GETMINMAXINFO.
Fixes#1806
The NetBSD sonames for X11 and related libraries is more stable than on
OpenBSD but the version numbers are still bumped more often than their
Linux counterparts, even excluding the one-time version bump across all
X11 related libraries.
This commit moves to using version-less sonames for X11 and related
libraries on NetBSD, which will hopefully be more forward-compatible
than hard-coding NetBSD-specific sonames.
This may not be the correct long-term solution but it runs now.
Binaries also appear to need an LD_LIBRARY_PATH or rpath entry of
/usr/X11R7/lib in order for the libraries to be found by dlopen.
Tested on NetBSD 9.2.
If the polling was interrupted by a signal or by incomplete or unrelated
data on any file descriptor, handleEvents could return before the full
timeout had elapsed.
This retries the Wayland prepare-to-read and poll until the full timeout
has elapsed or until any event was processed. Unfortunately, due to how
the Wayland client API is designed, this also includes the delete_id
for the frame callback created by eglSwapBuffers.
This means glfwWaitEvents* are still not fully functional on Wayland.
See #1911 for more details.
The display sync requests in glfwPostEmptyEvent could just accumulate as
the display was never flushed on secondary threads.
This adds a proper flush after each sync request.
Fixes#1520Closes#1521
This moves the X11 polling implementation to a separate file where it
can be used by either the X11 or Wayland backend or both.
This code should be POSIX compatible where necessary but will use the
lower latency but non-standard polling functions ppoll or pollts where
those are available.
This commit is based on work by OlivierSohn and kovidgoyal.
Fixes#1281Closes#1285
Cancel the prepared-to-read state on the calling thread before starting
to call back to user code.
Emitting close requests here is not a good choice but that is for
a future commit to address.
This uses ppoll for waiting on file descriptors with a timeout, where
that function has been available a while. On NetBSD, which will be
getting ppoll in the next release, the equivalent pollts is used.
This commit is based on work by OlivierSohn and kovidgoyal.
Related to #1281
Related to #1285
There is a seemingly unavoidable race condition when waiting for data on
the X11 display connection, as long as any other thread is also making
Xlib calls. The event data we are waiting for could be read by the
other thread as part of looking for the reply to its request, before our
poll has begun.
This commit replaces the X11 event sent by glfwPostEmptyEvent with
writing to an unnamed pipe. The race condition remains if other Xlib
calls are made on other threads, but glfwPostEmptyEvent should now be
race-free.
This commit is based on work by pcwalton, OlivierSohn, kovidgoyal and
joaodasilva.
Closes#2033
Related to #379
Related to #1281
Related to #1285
The data available on the X11 connection may be a reply or an internal
event for an X11 extension. Previously the check for whether an event
was available for us was done outside waitForEvent. This prevented data
available on other file descriptors from breaking the outer wait loop.
This commit moves the check for whether an event is available into the
wait functions, where there is enough knowledge to limit the check to
the X11 connection.
Related to #932
On Linux, the inotify descriptor was included in the set used for
select, but could not break the outer loop, leading to busy waiting
until timeout or the correct X11 event arrived.
This commit adds a new function for waiting just on X11 events.
Fixes#1872
We switched to kUTTypeURL when NSURLPboardType was deprecated, as the
official replacement symbol NSPasteboardTypeURL was not available on
every version of macOS supported by GLFW.
kUTTypeURL has now also been deprecated.
This commit moves to a compile-time choice between NSURLPboardType and
NSPasteboardTypeURL depending on the minimum targeted macOS version.
Fixes#2003
The Wayland backend now requires xkbcommon-compose, which was added in
version 0.5.0. xkbcommon 0.5.0 was released in 2014.
This removes the non-composing fallback path for text input.
By definition a hidden window on Wayland does not have valid framebuffer
contents.
This adds a window damage (refresh) event when a window is shown, to
request an initial frame for the now visible window.
A window created with GLFW_VISIBLE set was not made visible by the
initial buffer swap during context attribute refresh.
Regression introduced by @elmindreda in
094aa6d3c7.
Platform code should not generate key events with GLFW_REPEAT.
GLFW_PRESS is translated into GLFW_REPEAT by shared code based on the
key state cache.
This confused the automatic key release logic into not generating an
event with GLFW_RELEASE for a key being repeated when the window lost
input focus.
We were previously storing the pointer position only when on the main
window, so when the user clicked on a fallback decoration it would use
the last position of the cursor on the main window, instead of the
position in the decoration surface.
Fixes part of #1991.
The OpenBSD ports tree assigns its own soname version numbers, so the
hardcoded sonames GLFW uses to load libraries on non-macOS Unices are
often incorrect. Instead OpenBSD recommends that run-time loading
should leave out the version numbers entirely. The OpenBSD ld.so then
finds the correct library.
This upstreams the ports tree fixes for Xcursor and EGL, and adds the
corresponding fix for all other run-time loaded library sonames.
Tested on OpenBSD 7.0.
This issue was initially reported on IRC.
Corrects the protocol violation when creating an xdg_surface from a
wl_surface that already has a buffer due to EGL buffer swaps.
This commit is based on PR #1731 by @ghost, but adapted and altered:
- The XDG surface and role are now only created when a window is shown
to prevent application lists from showing command-line applications
with off-screen-only windows
- The special case of Wayland+EGL buffer swap is now in the EGL code
to mirror how X11 is handled
- Adaption to run-time platform selection and separate credits file
Fixes#1492Closes#1731
This extensions allows GLFW to instruct the driver to ignore the alpha
bits, even in formats which contain them. This makes it possible to use
the alpha bits as extra storage, without it affecting the end result
getting displayed to the user.
Fixes#1434Fixes#1803
The monitor handle could have become invalid just before the call to
GetDpiForMonitor. It was possible for both window and monitor content
scale queries.
This ensures both that an appropriate error is emitted and that the
retrieved values are zero on error.
Fixes#1615
The conversion of window icon image data involves unsigned char color
values being promoted to int and then shifted to the left by 24. For
32-bit ints this is just far enough to trigger undefined behavior.
It worked by accident because of how current compilers translate this
piece of code.
This was caught by @slimsag while working on [Zig bindings for GLFW][1],
and diagnosed together with @Andoryuuta, as described [in an
article][2]. Zig has UBSan enabled by default, which caught this
undefined behavior.
[1]: https://github.com/hexops/mach-glfw
[2]: https://devlog.hexops.com/2021/perfecting-glfw-for-zig-and-finding-undefined-behavior#finding-lurking-undefined-behavior-in-6-year-old-glfw-code
Thanks to Maato, martinhath, dcousens, drfuchs and Validark for helping
to refine the solution.
This commit message was rewritten by @elmindreda to hopefully reflect
the conclusions of the pull request thread.
Related to hexops/mach#20
Closes#1986
Looking into the definition of kIOMainPortDefault, the following
description can be found:
When specifying a main port to IOKit functions, the NULL argument
indicates "use the default". This is a synonym for NULL, if you'd
rather use a named constant.
Thus, we do not have to utilize an external symbol for the identifier
of the default main IOKit port, but MACH_PORT_NULL suffice. This
simplifies compatibility between macOS versions as the symbol was
renamed with macOS 12.0.
Fixes#1985Closes#1994
This removes the GLFW_VULKAN_STATIC CMake option and the
_GLFW_VULKAN_STATIC configuration macro and replaces them with the
glfwInitVulkanLoader function, allowing a single library binary to
provide both behaviors.
This is based on the design from PR #1374 by @pmuetschard.
Closes#1374Closes#1890
This removes the global struct member for
vkEnumerateInstanceExtensionProperties, which is only used in the same
function that fetches the function from the loader.
The pattern is now more in line with how other single uses of Vulkan
functions are structures, such as window surface creation.
C does not allow conversions between data pointers and function
pointers.
Yes, the name of the macro is reserved. That's something for a future
commit to fix.
Fixes#1703
This hack breaks when switching a window to fullscreen, if the OpenGL
ICD detects this and switches its swapchain to exclusive mode.
This limits the hack to Windows Vista and 7. This hack was added
because of vsync jitter under DWM on Windows 7 and I have been unable
to reproduce it on any later version.
Is this change causing any problems on any version of Windows? Please
open an issue!
Fixes#1072
This adds compile-time support for multiple platforms and runtime
detection of them. Window system related platform functions are now
called from shared code via the function pointer struct _GLFWplatform.
The timer, thread and module loading platform functions are still called
directly by name and the implementation chosen at link-time. These
functions are the same for any backend on a given OS, including the Null
backend.
The platforms are now enabled via CMake dependent options following the
GLFW_BUILD_<platform> pattern instead of a mix of automagic and ad-hoc
option names. There is no longer any option for the Null backend as it
is now always enabled.
Much of the struct stitching work in platform.h was based on an earlier
experimental branch for runtime platform selection by @ronchaine.
Every platform function related to windows, contexts, monitors, input,
event processing and Vulkan have been renamed so that multiple sets of
them can exist without colliding. Calls to these are now routed through
the _glfw.platform struct member. These changes makes up most of this
commit.
For Wayland and X11 the client library loading and display creation is
used to detect a running compositor/server. The XDG_SESSION_TYPE
environment variable is ignored for now, as X11 is still by far the more
complete implementation.
Closes#1655Closes#1958
The performance counter API is guaranteed to succeed on Windows XP and
later so there is no need for a fallback.
This removes our last dependency on winmm.
These headers come from when GLFW only supported a single context
creation API at a time, chosen at compile-time.
Since EGL and OSMesa are now always enabled and the platform-specific
APIs are tied to their specific window system, there is no reason to
keep these declarations separate from the internal.h and *_platform.h
header files.
This is done in preparation for runtime platform selection, to make sure
every platform can build with EGL enabled.
It may be possible to add support for things like the ANGLE null
platform later.
The native access functions for context handles did not verify that the
context had been created with the same API the function was for.
This makes these functions emit GLFW_NO_WINDOW_CONTEXT on API mismatch.
There is a suppression interval (0.25 seconds by default) after a call
to CGWarpMouseCursorPosition, during which local hardware events
(keyboard and mouse) are ignored. GLFW already calls
CGEventSourceSetLocalEventsSuppressionInterval with a value of 0.0, but
it doesn't help in this case, there is still a short delay before the
cursor can be moved. Moving the CGAssociateMouseAndMouseCursorPosition
call after the cursor position has been restored, fixes the issue.
Closes#1962
The Wayland protocol spec[1] states that set_cursor must be called
with the serial number of the enter event. However, GLFW is passing in
the serial number of the latest received event, which does not meet the
protocol spec.
[1] https://wayland.freedesktop.org/docs/html/apa.html#protocol-spec-wl_pointer
As a result, set_cursor calls were simply ignored by the compositor.
This fix complies with the protocol more closely by specifically caching
the enter event serial, and using it for all set_cursor calls.
Fixes#1706Closes#1899
According to the libxkbcommon documentation[1], xkb_keymap_key_repeats
requires keymap and keycode as input:
int xkb_keymap_key_repeats( struct xkb_keymap * keymap,
xkb_keycode_t key)
However, in inputChar in wl_input.c we are passing in xkb_keysym_t,
which was a type mismatch.
This results in some keys not repeating when they should and vice versa.
[1] https://xkbcommon.org/doc/current/group__components.html#ga9d7f998efeca98b3afc7c257bbac90a8Closes#1908.
This adds the glfwInitAllocator function for specifying a custom memory
allocator to use instead of the C runtime library.
The allocator is a struct of type GLFWallocator with fields
corresponding to malloc, realloc and free, while the internal API
corresponds to calloc, realloc and free.
Heap allocation calls are filtered before reaching the user-provided
functions, so deallocation of NULL and allocations of zero bytes are not
passed on, reallocating NULL is transformed into an allocation and
reallocating to size zero is transformed into deallocation.
The clearing of a new block to zero is performed by the internal
calloc-like function.
Closes#544.
Fixes#1628.
Closes#1947.
This adds the GLFW_LIBRARY_TYPE CMake cache variable, which allows users
and higher-level projects to set what type of library GLFW is built as.
When not empty, this value overrides the standard BUILD_SHARED_LIBS
option for GLFW while still allowing it to control the type of other
libraries in a larger project.
This also allows building GLFW as an object library without adding dummy
source files (as required by Xcode) or producing unused library
binaries.
Projects using CMake 3.12 or later can link the resulting GLFW object
library normally using target_link_libraries.
Fixes#279.
Related to #1307.
Closes#1497.
Closes#1574.
Closes#1928.
The insight to use wayland.xml to resolve the difficult-to-redirect
interface symbols was gleaned from SDL.
Instead of compiling the code output of wayland-scanner separately it is
made part of the wl_init compilation unit. This lets us do things like
transparently rename our copies of Wayland globals.
The OS version of wayland-client-protocol.h is no longer used by GLFW,
but it is presumably ABI compatible with the output of wayland-scanner.
Closes#1174.
Closes#1338.
Related to #1655.
Closes#1943.
POSIX.1-2008 deprecated gettimeofday, which we used as a fallback if the
monotonic clock was unavailable.
This replaces that fallback with the non-monotonic real-time clock.
Because of the Gordian knot of feature test macros across Unices, this
also includes the shift from some platform source files defining
_POSIX_C_SOURCE to various values to _DEFAULT_SOURCE being defined for
all source files on Linux. This is because -std=c99 on Linux disables
_DEFAULT_SOURCE (POSIX 2008 and extensions).
Once runtime platform selection comes in, this kind of platform-specific
preprocessor logic can be moved into the platform glue files and won't
need to be replicated by third-party build setups, but for now, sorry.
The outer glfwUpdateGamepadMappings function is now bypassed when
parsing the default gamepad mappings. The data in mappings.h does not
contain any comments or line breaks. Also this is done before joystick
support has been initialized, so there is no need to look for matching
devices.
Finally, the array of default mappings is pre-allocated. This has no
measurable performance impact but does generate a lot of calls, which
won't be nice for a user provided custom allocator to deal with.
This avoids generating a macro redefinition warning when the
GL_SILENCE_DEPRECATION macro has been defined for the whole of a larger
project.
Closes#1840.
The contents scale of the hosted CAMetalLayer created for MoltenVK was
updated only after the GLFW content scale and framebuffer size events
were emitted, causing the layer to get out of sync with the monitor the
window was on.
Because there are controllers in the wild using the same hardware ID
despite having different numbers of buttons and axes, an error message
was triggered that was only expected for a corrupted mapping database.
This removes the error for now, in preparation for better error handling
for gamepad mappings overall.
Fixes#1763.
The EGL backend ignored the state of GLFW_DOUBLEBUFFER and always
created a double-buffered EGL window.
This sets the EGL_RENDER_BUFFER attribute at EGL window creation
appropriately.
Fixes#1843.
This skips the buffer swap after the initial glClear performed during
window creation, if the window is single-buffered. This call confused
apitrace into thinking the window was double-buffered.
Fixes#1873.
The /clang: suffix passed to Clang-CL was accidentally also passed to
the regular standalone Clang, which caused compilation to fail. We now
pass /W3 to Clang-CL, which it interprets as -Wall.
The _CRT_SECURE_NO_WARNINGS macro is now defined for both Clang and
Clang-CL.
The /entry: flag passed to link.exe is now also passed to lld-link,
letting the windows subsystem tests and examples link.
Fixes#1807.
Closes#1824.
Closes#1874.
The GPU driver only looks in the executable for the symbol requesting
the high-performance GPU, so enabling them when buidling GLFW as a DLL
will have no effect.
This removes the dependency on the (unspecified) ordering of geometry
and mode events in wl_output.
Based on feedback from @linkmauve and @caramelli.
Related to #1792.
GLFW_SCALE_TO_MONITOR had no effect on Windows 8.1 up to and including
Windows 10 version 1607 (Anniversary Update), despite those having
support for per-monitor DPI.
That done was to avoid handling systems that have non-client scaling,
introduced in Windows 10 version 1607, without reliable overriding of
the new window size, introduced in Windows 10 version 1703 (Creators
Update). Both are needed to keep the content area at a fixed size for
windows that have GLFW_SCALE_TO_MONITOR disabled.
This change enables window rescaling on Windows 8.1 and all later
versions but disables non-client scaling for unscaled windows on Windows
10 version 1607. Versions after 1607 are unaffected.
Fixes#1511.
The install name was incorrectly set to a relative path.
This change leaves the install name of the installed dylib as
@rpath/soname. Those who wish to override this can set the
CMAKE_INSTALL_NAME_DIR variable.
Closes#1504.
On Apple Silicon, IOKit is deprecated and there will be no
matching io_service that we can query for name. Luckilly,
NSScreen got an API to fetch the display name in 10.15.
This is a blocker to get glfw running on Apple Silicon.
Fixes#1809.
Closes#1833.
NSApp setActivationPolicy: was being called too soon when the app was
not bundled and launched from the command line.
This fix is based on #1802 by @richardwilkes.
Fixes#1648.
Closes#1802.
On Wayland we implement Client-Side Decorations if the compositors do not
implement SSDs. In that case, the destructors of the surfaces were called
in the wrong order, leading to a dereference of an already freed object.
We need to first destroy the subsurface before destroying the parent surface.
Related PR on kitty: https://github.com/kovidgoyal/kitty/pull/3066
Related issue on kitty: https://github.com/kovidgoyal/kitty/issues/3051Closes#1798.
CMake 3.19 adds -xc when the LANGUAGE file property is C, breaking our
workaround for CMake 3.15 and earlier not understanding the .m suffix.
Fixes#1787.
Many years ago this value was changed from F13 to Print Screen with
1ae9ce1e0a. However, this was later
reverted to F13 with 5759d0fdf2.
This changes it back to Print Screen for consistency with other
platforms.
Closes#1786.
An incorrect value for _WIN32_WINNT_WINBLUE caused GLFW to think Windows
8 was 8.1 or greater. This led to a crash when attempting to call
SetProcessDpiAwareness.
Closes#1775.
On Windows 7, when GLFW framebuffer transparency and the DWM are enabled
but DWM transparency is disabled (i.e. when the Transparency setting is
disabled under Personalization > Color), the contents of the framebuffer
is combined with the last frame using additive blending instead of
replacing the previous contents.
This commit limits GLFW framebuffer transparency on Windows 7 to when
DWM transparency is enabled, removing the previous workaround of setting
a layered window color key that led to rendering artifacts.
Fixes#1512.
This makes joystick support initialize the first time a joystick
function is called, including those gamepad functions that are layered
on top of joystick functions.
Related to #1284.
Related to #1646.
Returning HTTRANSPARENT from WM_NCHITTEST does cause the window to be
transparent for some hit-testing APIs but does not make it pass mouse
input through to whatever window is below it.
For that to work on modern Windows, the window needs to be both layered
and extended-window-style-transparent.
Additional logic changes to ensure mouse input passthrough, framebuffer
transparency and window opacity are mindful of one another when
modifying WS_EX_LAYERED.
Related to #1568.
The client clip region was left in place when mouse passthrough was
disabled, leading to missing mouse input if the window grew beyond it.
Related to #1568.
This adds the GLFW_MOUSE_PASSTHROUGH window hint and attribute for
controlling whether mouse input passes through the window to whatever
window is behind it.
Fixes#1236.
Closes#1568.
It seems it has not been possible to compile GLFW on systems without the
UTF-8 extensions to Xlib since a0a5cc57df
was merged five months ago. The UTF-8 extension functions were
introduced with XFree86 4.0.2 in December 2000 and are likely widely
available at this point.
This removes the locale-dependent fallback paths and uses the UTF-8
extension functions where available.
Background: The IM will filter out key events, instead sending exact
duplicate events that are not filtered. It does not send these for
every event, however, so the duplicate events cannot be relied on for
key input. Instead we need to identify and discard them. Since they
are identical, they have the same timestamp as the originals.
The previous duplicate event detection would consume unrelated key
events if the keys were pressed simultaneously, as it only tracked
a single timestamp.
This fixes that issue for any combination of keys, at the expense of
a 1 KB array per GLFW window.
This fix is a stopgap until explicit IME support is done.
Based on #1472 by @LucaRood.
Fixes#1112.
Fixes#1415.
Fixes#1616.
Fixes#1663.
Closes#1472.
This adds GLFW_CONTEXT_DEBUG as a preferred alias for
the GLFW_OPENGL_DEBUG_CONTEXT window hint, as debug contexts are
defined for both OpenGL and OpenGL ES.
Related to #1720.
This adds support for EGL_EXT_platform_base and its associated X11 and
Wayland extensions, allowing us to explicitly tell EGL which window
system we are using.
This is based on work by @linkmauve in #1691.
Closes#1691.
This makes the null platform behave more like an actual stub. More
queryable state is tracked and there is even a fake monitor with one
whole fake video mode.
This fixes a race between the key repeat logic and the surface leave
event handler, which could result in repeated keys being reported with
a window of NULL.
Fixes#1704.
This adds two new error codes: GLFW_FEATURE_UNAVAILABLE for when
a GLFW feature cannot be reasonably implemented on that platform, and
GLFW_FEATURE_UNIMPLEMENTED for when it can be but has not been yet.
This replaces the current situation where the Wayland code emitted
GLFW_PLATFORM_ERROR in both cases while the macOS code silently did
nothing.
If your application exits on any GLFW error, these error codes should at
least be easy to filter out from that behavior.
Ideally, GLFW_FEATURE_UNAVAILABLE should be rare and
GLFW_FEATURE_UNIMPLEMENTED should never be emitted at all.
Fixes#1692.
These are harmless errors but the code was worth fixing just to reduce
confusion and be more explicit. E.g. using a different variable name
for a new variable of a different type in win32_joystick.c.
Closes#1700.
When compiling with `-Wall` and `-pedantic-errors`, gcc complains with
```
warning: ISO C does not support the '%m' gnu_printf format [-Wformat=]
```
because the `%m` conversion specifier is a GNU extension.
Closes#1702.
The only two EGL implementations on macOS are Swiftshader and ANGLE.
While Swiftshader supports both `NSView` and `CALayer` as
`EGLNativeWindowType`, ANGLE supports only `CALayer`. Furthermore
Swiftshader's OpenGL ES frontend is deprecated in favor of using ANGLE's
Vulkan backend on top of Swiftshader's Vulkan frontend.
This means that on macOS `EGLNativeWindowType` should be a `CALayer` for
compatibility with ANGLE.
Fixes#1169.
Closes#1680.
A regression introduced by b889aa7841
broke the special handling of numpad keys for the non-XKB fallback path.
The non-functional remains were later removed. This restores the
original behavior.
This fixes the issue where function keys would be reported as
GLFW_KEY_UNKNOWN if XKB was available and one of the configured keyboard
layouts was Arabic.
This is only part of #1598, because the full patch removed parts of the
fallback path for when XKB is unavailable.
Closes#1598.
The non-root parent window owned by the WM could be destroyed before we
process the ConfigureNotify event using the cached parent handle.
Bug was found by unmapping a decorated window.
This like all uses of the Xlib error handler is not thread safe and
there is nothing we can do about that short of moving to XCB.
Fixes#1633.
This adds support for the XIM instantiate and destroy callbacks, letting
GLFW detect both when the current input method disappears and when a new
one is started.
Tested with ibus.
This moves the remaining bits of NSApplication initialization into
_glfwPlatformInit. As a side-effect of this, any command-line program
initializing GLFW will get a menu bar, which is not ideal.
If this has happened to you and a bisect led you here, please see the
GLFW_COCOA_MENUBAR init hint introduced in GLFW 3.3.
If this patch is a terrible idea, please get in touch in the 3.4 release
timeframe.
This is a replacement for 6e6805000a,
which attempts to preserve the existing menu bar creation behavior for
the 3.3-stable branch.
Fixes#1649.
