Originally, macOS had opposite Y axis inversion as every other platform, likely to correct for an issue with the virtual gamepad reported by the old 360Controller driver.
Wired Xbox 360 controllers using native macOS drivers were first reported to be broken in https://github.com/libsdl-org/SDL/issues/11002. The inversion was removed in 7da728a642, presumably breaking 360Controller usage, but fixing wired 360 controller using the new native support in macOS 15 and above. However, this change was reverted without explanation in d7b1ba1bfc which added explicit support for the Steam Virtual Gamepad. Presumably, Steam on macOS reports inverted Y axes to match what SDL expected on the platform. However, this reversion broke the native macOS controller support. The incorrect inversion also breaks using off-brand 360-class gamepads via the libusb backend of HIDRAW.
Backport of 157c839139 (https://github.com/libsdl-org/SDL/pull/15792) to SDL2
Mesa's eglGetDisplay() chooses the native EGL platform from
EGL_PLATFORM, limited autodetection, or the build-time default. If that
selects Wayland while SDL is using the X11 video backend, Mesa can treat
the X11 Display pointer as a wl_display and crash during
eglInitialize().
See: https://gitlab.com/qemu-project/qemu/-/work_items/3540
On some platforms, SDL_MemoryBarrierRelease() is defined to
SDL_CompilerBarrier(). If SDL_CompilerBarrier() is also defined to
the fallback spinlock acquire/release, then we will infinitely
recurse in SDL_UnlockSpinlock(). Avoid this by not unlocking the
temporary spinlock we create.
- Add missing SDL_MemoryBarrierRelease() in the generic codepath
- Remove Watcom and MSVC x86/x64 cases which are now identical to the generic codepath
- Fix Solaris barrier to ensure prior stores are visible before unlocking
This avoids requiring inline assembly for each architecture.
It also fixes some weakly ordered architectures which lacked
said inline assembly (RISC-V, MIPS, LoongArch, etc) and were
thus disasterously broken.
__sync_lock_test_and_set() is designed for creating locks, not as
a general atomic exchange function. As a result, it only provides
an acquire memory barrier and isn't guaranteed to actually store
the provided value (though it does on architectures we care about).
__atomic_exchange_n() is supported on GCC/Clang for the last ~10
years, so let's use that instead if available. We will keep the
__sync_lock_test_and_set() fallback around for ancient platforms,
but add a full memory barrier to match the documented behavior.
Someone mentioned that Claude sees this file's previous text and adds "look at
AGENTS.md" to it's TODO list instead of treating it as the thing it should do
first.
If everything decides to prefer AGENTS.md at some point, we'll delete this
file.
A symlink would apparently also work, but it would be the only symlink in the
SDL git repo, so it seemed less disruptive to just copy the file over for now.
(cherry picked from commit 847fc72b1b)
Apparently Claude Code won't read AGENTS.md by default at this moment, but
since that's the direction we're moving, let's just add the file Claude
currently looks for and tell it that the actual instructions are in AGENTS.md.
I thought these things were supposed to be less complicated that interacting
with humans, lol.
(cherry picked from commit 1124e44e4c)
