This is needed to support CHERI, and thus Arm's experimental Morello
prototype, where pointers are implemented using unforgeable capabilities
that include bounds and permissions metadata to provide fine-grained
spatial and referential memory safety, as well as revocation by sweeping
memory to provide heap temporal memory safety.
The referential safety is enforced through the use of tagged memory, and
there is only a single tag bit per capability-sized word, meaning it is
impossible to store capabilities at unaligned locations, either getting
a trap on load/store or the validity tag being stripped when
round-tripepd through memory.
Since this is a new ABI for which SDL has never been compiled before, we
do not need to be concerned with this compatibility measure, so just
don't pack the struct for CHERI architectures.
This code is inherently rather dubious anyway; if MSVC and GCC disagree
on struct layout when targeting Windows then that is a bug in GCC, but
likely extends from the bogus #pragma pack directives for MSVC in
begin_code.h, which will force types to be *underaligned* (and is
attempting to work around something that is fundamentally a broken idea
to be doing). In particular 8-byte-aligned types will be underaligned to
4 bytes, but only on MSVC. Since that code is not used for GCC that is
probably the cause of the struct layout discrepancy, and there are
likely other instances of that throughout SDL. Moreover, the supposed
fix here is not in fact a fix, as now GCC will think SDL_AudioCVT is
only 1-byte-aligned but MSVC will think it's 4-byte or 8-byte-aligned,
meaning ABI incomatibility is introduced by this change. However,
removing it would break ABI compatibility for purely-GCC-compiled code
(as old binaries would see the struct as 1-byte-aligned and new binaries
would see the struct as 8-byte-aligned) so SDL is stuck with this until
it bumps its ABI.
This is needed to support CHERI, and thus Arm's experimental Morello
prototype, where pointers are implemented using unforgeable capabilities
that include bounds and permissions metadata to provide fine-grained
spatial and referential memory safety, as well as revocation by sweeping
memory to provide heap temporal memory safety.
The C standard does not guarantee that if two pointers compare equal
they are the same pointer, as C pointers have a notion of provenance,
and compilers have been known to exploit this during optimisation. For
CHERI, this becomes even more important, as in-place expansion can
result in realloc returning a capability to the same address but with
increased capability bounds, and so reusing the old capability will trap
trying to access outside the bounds of the original allocation.
In the case that ptr == mem, memdiff and ptrdiff should still be equal,
so the only overhead is a small amount of pointer arithmetic and a store
of the new pointer (which is required per the C standard in order to not
be undefined behaviour when next loaded).
This also fixes the calculation of oldmem to use uintptr_t rather than
size_t as casting the pointer to size_t on CHERI will strip the
capability metadata, including the validity tag, with the subsequent
cast back to void * resulting in a null-derived capability whose
validity tag is clear and thus cannot be dereferenced without trapping.
This is needed to support CHERI, and thus Arm's experimental Morello
prototype, where pointers are implemented using unforgeable capabilities
that include bounds and permissions metadata to provide fine-grained
spatial and referential memory safety, as well as revocation by sweeping
memory to provide heap temporal memory safety.
On most systems (anything with a flat memory hierarchy rather than using
segment-based addressing), size_t and uintptr_t are the same type.
However, on CHERI, size_t is just an integer offset, whereas uintptr_t
is still a capability as described above. Casting a pointer to size_t
will strip the metadata and validity tag, and casting from size_t to a
pointer will result in a null-derived capability whose validity tag is
not set, and thus cannot be dereferenced without faulting.
The audio and cursor casts were harmless as they intend to stuff an
integer into a pointer, but using uintptr_t is the idiomatic way to do
that and silences our compiler warnings (which our build tool makes
fatal by default as they often indicate real problems). The iconv and
egl casts were true positives as SDL_iconv_t and iconv_t are pointer
types, as is NativeDisplayType on most OSes, so this would have trapped
at run time when using the round-tripped pointers. The gles2 casts were
also harmless; the OpenGL API defines this argument to be a pointer type
(and uses the argument name "pointer"), but it in fact represents an
integer offset, so like audio and cursor the additional idiomatic cast
is needed to silence the warning.
This is needed to support CHERI, and thus Arm's experimental Morello
prototype, where pointers are implemented using unforgeable capabilities
that include bounds and permissions metadata to provide fine-grained
spatial and referential memory safety, as well as revocation by sweeping
memory to provide heap temporal memory safety.
If you continually poll for events it's possible that new events can come in while you're still processing the last one, delaying rendering. This is more likely with high update rate sensors.
When choosing an X11 Visual for a window based on its GLX capabilities, first
try glXChooseFBConfig (if available) before falling back to glXChooseVisual.
This normally does not make a difference because most GLX drivers create a
Visual for every GLXFBConfig, exposing all of the same capabilities.
For GLX render offload configurations (also know as "PRIME") where one GPU is
providing GLX rendering support for windows on an X screen running on a
different GPU, the GPU doing the offloading needs to use the Visuals that were
created by the host GPU's driver rather than being able to add its own. This
means that there may be fewer Visuals available for all of the GLXFBConfigs the
guest driver wants to expose. In order to handle that situation, the NVIDIA GLX
driver creates many GLXFBConfigs that map to the same Visual when running in a
render offload configuration.
This can result in a glXChooseVisual request failing to find a supported Visual
when there is a GLXFBConfig for that configuration that would have worked. For
example, when the game "Unnamed SDVX Clone" [1] tries to create a configuration
with multisample, glXChooseVisual fails because the Visual assigned to the
multisample GLXFBConfigs is shared with the GLXFBConfigs without multisample.
Avoid this problem by using glXChooseFBConfig, when available, to find a
GLXFBConfig with the requested capabilities and then using
glXGetVisualFromFBConfig to find the corresponding X11 Visual. This allows the
game to run, although it doesn't make me any better at actually playing it...
Signed-off-by: Aaron Plattner <aplattner@nvidia.com>
Fixes: https://forums.developer.nvidia.com/t/prime-run-cannot-create-window-x-glxcreatecontext/180214
[1] https://github.com/Drewol/unnamed-sdvx-clone
As of [1], SDL now compiles with a warning in SDL_waylandevents.c on
32-bit systems under gcc 10.3.0:
/tmp/SDL/src/video/wayland/SDL_waylandevents.c: In function 'seat_handle_capabilities':
/tmp/SDL/src/video/wayland/SDL_waylandevents.c:958:22: warning: cast from pointer to integer of different size [-Wpointer-to-int-cast]
958 | SDL_AddTouch((SDL_TouchID)seat, SDL_TOUCH_DEVICE_DIRECT, "wayland_touch");
| ^
/tmp/SDL/src/video/wayland/SDL_waylandevents.c:964:22: warning: cast from pointer to integer of different size [-Wpointer-to-int-cast]
964 | SDL_DelTouch((SDL_TouchID)seat);
| ^
This is due to SDL_TouchID always being 32-bit, but seat being a pointer
which is (obviously) only 32-bit on 32-bit systems. The conversion is
therefore harmless, so silence it with an extra cast via intptr_t.
This is what the cocoa backend does (and is similar to what the Win32
backend does, except with size_t).
Fixes: 03c19efbd1 ("Added support for multiple seats with touch input on Wayland")
[1]: 03c19efbd1
When wayland is not dynamically loaded (--enable-wayland-shared=no)
libdecor.h is not included unless SDL_VIDEO_DRIVER_WAYLAND_DYNAMIC
is set, so it fails to build. We can't simply move the libdecor.h
include above the #ifdef SDL_VIDEO_DRIVER_WAYLAND_DYNAMIC block, as
libdecor.h itself #includes wayland headers we need to replace with
#defines. Instead, duplicate the #include.
Fixes https://github.com/libsdl-org/SDL/issues/4543
Note that this doesn't fix any of the underlying issues of libdecor
being treated as part of wayland, it just fixes the build. A better
solution would probably be to decouple the wayland dynamic loading
from the libdecor dynamic loading completely, though that is a lot
more work...
Each window can have at most one zxdg toplevel decoration, but as of
[1], we accidentally create two. (If libdecor is not in use). This
causes wayland windows with server-side decorations (e.g. on KDE/KWin)
to crash with the message:
zxdg_decoration_manager_v1@7: error 1: decoration has been already constructed
This extra zxdg_decoration_manager_v1.get_toplevel_decoration() call was
introduced while deprecating wl-shell and xdg-shell-stable[1] support,
and possibly was a bad interaction with [2], which moved the decoration
creation around.
Fixes: 6aae5b44f8 ("Remove wl-shell and xdg-shell-unstable-v6 support (#4323)")
[1]: https://github.com/libsdl-org/SDL/pull/4323
[2]: https://github.com/libsdl-org/SDL/pull/4374
WASAPI_WaitDevice is used for audio playback and capture, but needs to
behave slighty different.
For playback `GetCurrentPadding` returns the padding which is already
queued, so WaitDevice should return when buffer length falls below the
buffer threshold (`maxpadding`).
For capture `GetCurrentPadding` returns the available data which can be
read, so WaitDevice can return as soon as any data is available.
In the old implementation WaitDevice could suddenly hang. This is
because on many capture devices the buffer (`padding`) wasn't filled
fast enough to surpass `maxpadding`. But if at one point (due to unlucky
timing) more than maxpadding frames were available, WaitDevice would not
return anymore.
Issue #3234 is probably related to this.
On modern CPUs, there's no penalty for using the unaligned instruction on
aligned memory, but now it can vectorize unaligned data too, which even if
it's not optimal, is still going to be faster than the scalar fallback.
Fixes#4532.
Ryan C. Gordon
Christian Rauch
Sam Lantinga
Ethan Lee
Jessica Clarke
SDL Wiki Bot
Brick
Aaron Plattner
Ozkan Sezer
Adam
David Gow
Simon Zeni
Splamy