t.grundner@goto3d.de 2011-09-01 03:59:17 PDT
I figured out what is going on. GCC 4.5.2 assumes the stack is 16 byte aligned
by default. Therefore there are no AND alignment corrections necessary if we
wish to align a stack variable to a 16 byte boundary. That is bad if your OS
ABI is not 16 byte aligned. Windows 32 bit stacks are 4 byte aligned. This
results in the above mentioned SIGSEGV. This is also no problem if I compile
both SDL.dll and my app with MingW because MinGW/GCC inserts a
andl $-16, %esp
instruction right in the beginning of the main function. So at least the stack
of the thread calling the main function is 16 byte aligned. But as soon as I
start to use the SDL.dll from an application not compiled by MinGW there is no
ANDL safing my app.
However there is a GCC option that can change the default stack alignment:
-mpreferred-stack-boundary=num
Setting num=2 assumes a the stack is aligned to a 4 byte boundary. This results
in GCC inserting the necessary
andl $-16, %esp
into SDL_FillRect. Rebuilding SDL with
./configure "CFLAGS=-mpreferred-stack-boundary=2 -g -O3"
solved the problem.
IMHO this should also be a problem on Solaris.
The following links contain further information:
http://gcc.gnu.org/onlinedocs/gcc-4.5.2/gcc/i386-and-x86_002d64-Options.html#i386-and-x86_002d64-Optionshttp://www.agner.org/optimize/calling_conventions.pdf
attached is a working directfb driver diff which works with the current
changes. There are a number of changes around it as well, e.g.
configure.in.
The directfb renderdriver right now still depends on a some "includes"
from src/video/directfb. That's why it is not yet moved to the new
render folder.
when .cc files are used due to bad sed substitution on multiple passes:
$(objects)/SDL_BeApp.lo:
$(objects)/SDL_BeApp.lo: ./src/main/beos/SDL_BeApp.c
$(LIBTOOL) --mode=compile $(CC) $(CFLAGS) $(EXTRA_CFLAGS) -c $< -o $@c
$(LIBTOOL) --mode=compile $(CC) $(CFLAGS) $(EXTRA_CFLAGS) -c $< -o $@
Signed-off-by: François Revol
The configure script was breaking the substitutions into multiple fragments, breaking them across the substitution for the build rules. This of course totally hosed the process.
I switched to using a more modern usage of AC_OUTPUT and added a post-process step that appends the build rules to the Makefile.
