turns out that __builtin_alloca_with_align() might releases the
allocated memory at the end of the block it was allocated in, instead
of the end of the function (which is the behavior of regular alloca()
and __builtin_alloca()): "The lifetime of the allocated object ends at
the end of the block in which the function was called. The allocated
storage is released no later than just before the calling function
returns to its caller, but may be released at the end of the block in
which the function was called."
https://gcc.gnu.org/onlinedocs/gcc/Other-Builtins.html#index-_005f_005fbuiltin_005falloca_005fwith_005falign
Clang also supports __builtin_alloca_with_align(), but always releases
the memory at the end of the function.
And it seems that newer GCC versions also tend to release it at the
end of the function, but GCC 4.7.2 (that I use for the official Linux
release binaries) didn't, and that caused weird graphical glitches.
But as I don't want to rely on newer GCC versions behaving like this
(and the documentation explicitly says that it *may* be released at
the end of the block, but will definitely be released at the end of
the function), I removed all usage of __builtin_alloca_with_align().
(Side-Note: GCC only started documenting the behavior of
__builtin_alloca and __builtin_alloca_with_align at all with GCC 6.1.0)
The HD version of the "Birdman" Hellknight model crashed the game,
because on Linux it triggered the 600KB _alloca() assertion and on
Windows (that didn't have such an assertion up to this commit) it
overflowed the stack.
Thde default stacksize on Windows is 1MB, increasing that to 8MB fixes the
problem (in Doom3 1.3.1 it was 4MB).
I also implemented the _alloca() size check assertion for Windows, and
increased the allowed size to 2MB (so it can be safely called multiple
times per function)
Model: https://www.moddb.com/mods/birdman-doom3/downloads/birdman-doom3-v11
For some reason MSVCs integrated CMake doesn't set CMAKE_GENERATOR_PLATFORM
so my CPU arch detection magic in CMakeLists.txt didn't work in that case..
Now (when using MSVC) the CPU architecture (D3_ARCH) is set in
neo/sys/platform.h instead.
- Fix build with SDL <=2.0.3
SDL_GetGlobalMouseState was introduced in 2.0.4
(which doesn't support OSX 10.5 or older)
- Don't include execinfo.h on Mac OS X 10.4
This file isn't included in the 10.4 SDK
- Use custom typedef for PFNGLSTENCILOPSEPARATEPROC on OSX 10.4/10.5
because the system OpenGL headers for those versions don't have it
On Windows, ID_INLINE does __forceinline, which is bad if the function
calls alloca() and is called in a loop..
So use just __inline there so the compiler can choose not to inline
(if called in a loop).
This didn't cause actual stack overflows as far as I know, but it could
(and MSVC warns about it).
