The code from 748be364e5 incorrectly accepted -0x100000000 on x86_64.
Also if ERANGE is returned by strtoul(), reject the parsed value regardless
of what that value is. This fixes invalid values being returned when compiling
with Visual C++. Thanks to "cotton" for testing this.
The following changes were made:
Restricted the "-march=core2" option to i386 because the first Intel Macs
had Intel Core CPUs, not Core2.
Removed the "-mdynamic-no-pic" flag as GCC lists it as a PPC specific flag.
Removed "-Wl,-read_only_relocs,suppress" because it seems to be related
to "-mdynamic-no-pic" and I see no need for it.
Removed "-Wextra-tokens -Wnewline-eof" because they are GCC specific and
not OS X specific.
This allows us to add keys that don't exist in the CMake template.
I added the keys from the Info.plist that was generated by our SCons build
to the new template.
Previously, there was just one list of frameworks regardless of which part
of the code depended on which frameworks. Now we keep separate lists for
the Dolphin core, the Dolphin GUI and internal use by wxWidgets.
Flags in CMAKE_CXX_FLAGS are passed to both compile and link commands.
A cleaner solution would be to use set_source_files_properties().
However, currently there are headers (StdThread.h, maybe more) that contain
Objective syntax. So it is not easy to determine exactly which source files
should be compiled as Objective C/C++ and that set can quickly change when
certain #include directives are modified. The solution for that would be to
move all uses of Objective syntax to implementation (.cpp) files and then
apply set_source_files_properties() to those.
If the systemwide wxWidgets is too old, we can fall back to the version
from Externals instead of letting the build fail.
Note that while the version in Externals appears to be wxWidgets 2.8,
it has post-2.8 features manually added to it (wxAuiToolBar), so the
version check only accepts 2.9 or higher.
To enforce SM2.0 compatibility, the OpenGL plugin was made to crash when
compiling a shader which does not fit in the SM2.0a limits. However, on some
combinations of OS/drivers/GPU, our shaders already do not fit in these limits,
causing artificial failures only to try to keep a non existant SM2.0a compat.
Basically, this sucks.
This commit increases the artificial limit to SM3.0. If you're using a GPU
which does not support SM3.0 and Dolphin works properly, this should not cause
any problem at all.
On x86-64, "unsigned long" is 64 bits wide, so it is possible for a number
to not trigger a range error on strtoul() but still not fit inside an u32.
An extra check is added to ensure that 32-bit and 64-bit builds will accept
the same numbers.
The previous computation was very likely to go out of array bounds,
which could result in crashes on EFB access.
Also, the cache size was rounded down instead of up. This is a problem
since EFB_HEIGHT (528) is not a multiple of EFB_CACHE_RECT_SIZE (64).
Passing MAP_FIXED to mmap causes already mapped pages in the requested
region to be replaced. On Mac OS X this caused pages for JIT-generatd
code to appear in the memory range previously auto-allocated for the RAM
of the emulated machine. This led to a hang at boot time. The same problem
can probably occur on FreeBSD, but not on Linux since MAP_32BIT is used
there instead of MAP_FIXED.
The solution is to not use MAP_FIXED, but instead rely on the OS honoring
the hinted address which is below 4 GB: we don't need an exact match,
just a low address.
