There is a possbile race that _glapi_Context is reset by another thread
after it is tested in GET_CURRENT_CONTEXT but before it is returned. We
definitely do not want a lock here to solve the race. To have correct
results even under a race, no other threads should reset _glapi_Context
(or _glapi_Dispatch).
This patch adds a new global variable _glapi_SingleThreaded. Since
_glapi_Context or _glapi_Dispatch are no longer reset,
_glapi_SingleThreaded is tested instead, before accessing them.
DRI drivers compiled with this patch applied will not work with existing
libGL.so because of the missing new symbol. If this turns out to be a
real problem, this patch should be reverted.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
If multiple threads set/get a TSD at roughly same time for the first
time, glthread might (wrongly) initialize it more than once. This patch
solves the race by initializing context/dispatch TSDs early.
Acked-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Multiple threads might call _glapi_check_multithread at roughly the same
time. It is possbile that all of them are wrongly regarded as firstCall
if there is no mutex. This bug causes xeglthreads to crash sometimes.
Acked-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
When translating an incoming shader (rather than building one from scratch)
it's preferable to be able to call a single, generic instruction emitter
rather than figuring out which of the opcode-specific functions to call.
Not doing so seems to cause lock-ups or rendering problems
on some chips. I think there is an logic issue related to
CB and VGT on some chips. We ran into similar issues in
r600_demo IIRC.
This patch fixes some "implicit declaration of function" compilation
errors/warnings on the new shader code.
Signed-off-by: Renato Caldas <seventhguardian@gmail.com>
Part of this code is disabled since no performance gains
where detected with it enabled.
This code only detects if it is a pixmap that it is
rendering to on the st/xorg DDX since it sets the fake
front to the same handle as front.
The real difference is that the driver suffix is now appended. This
also fixes an annoying bug that EGL_DRIVER could not specify the path to
a driver because a suffix was always appended.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
libX11 is used to determine the screen number, which is in turned used
to determine the DRI driver. However, the sysfs interface for
determining the DRI driver is gone, and no working driver depends on
this mechanism.
Display string parsing is moved to a new function,
_eglSplitDisplayString.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
