according to https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100839
the real compiler flag enabling this bullshit isn't
-fexpensive-optimizations but -ffp-contract=fast which for some(*)
reason is default in optimized builds.
I think it's best to disabled that optimization globally in case it
also breaks other code (I really don't want to spend several days to
hunt such an idiot bug again). I really doubt it makes any measurable
performance difference.
As https://twitter.com/stephentyrone/status/1399424911328874499 says
that clang might also enable this in the future (though to =on instead
of =fast which should be a bit saner but would still break our code),
just set this option for all GCC-ish compilers incl. clang.
(*) the reason of course is that GCC developers don't develop GCC for
their users but to win idiotic SPEC benchmarks
Only happend if `ONATIVE` was enabled (or some other flag was set
that enables the FMA extension), the root cause was that the cross
product didn't return 0 when it should, but a small value < 0.
Caused some faces to be missing in maps compiled with dmap.
https://github.com/RobertBeckebans/RBDOOM-3-BFG/issues/436#issuecomment-851061826
has lots of explanation.
I think this is a compiler bug, this commit works around it.
fixes#147
It corrupted the stack when called with buffers allocated on the stack
and numSamples that are not a multiple of four, apparently, by writing
4 floats too many, at least in the 22KHz Stereo case..
This caused the crash described in
https://github.com/dhewm/dhewm3/issues/303#issuecomment-678809662
Now it just uses the generic C code, like all platforms besides MSVC/x86
already do.
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).
Sometimes memory was allocated with new[] but freed with delete instead
of delete[], which is wrong.
And there were some small memory leaks, too.
Furtunately clang's AddressSanitizer detected all that so I could easily
fix it.
(There seem to be some more small memory leaks which are harder to fix,
though)
Apart from some minor stuff, this changes the signature of some methods
of Parser and Token classes and of the (unused) Random2 class.
That no problem though, because the calling code uses normal ints
anyway.
The original implementation was pretty broken (but not used anyway),
it is now fixed and improved a bit (got rid of one inner loop).
This (at least part of the problem) was detected by PVS-Studio,
see http://www.viva64.com/en/b/0120/ Fragment 3
We want to use the SIMD functions of the base class if the
deriving class does not implement every overloaded variant.
Added missing idLight::SetColor(idVec3) which is declared in
idEntity.
Get rid of the 2 CPUID flags and combine them with SSE in one
implementation.
SSE flags can now be set on all x86 and x86_64 platforms -
independent of -ffast-math.
Helper defines borrowed from STREFLOP.
Don't include the lazy precompiled.h everywhere, only what's
required for the compilation unit.
platform.h needs to be included instead to provide all essential
defines and types.
All includes use the relative path to the neo or the game
specific root.
Move all idlib related includes from idlib/Lib.h to precompiled.h.
precompiled.h still exists for the MFC stuff in tools/.
Add some missing header guards.
Protect all SIMD implementations with the according defines and
let the compiler decide if it supports the intructions.
Linux will still use Simd_Generic because CPU feature runtime
detection is missing.
Get the idDrawVert member offsets directly instead of hardcoding
them and checking them via assert() at runtime.
Fixes compile error:
cast from pointer to smaller type 'int' loses information
variable set but not used
Removes some CollisionModel code under _DEBUG which was probably a
leftover, since it was completely useless (its done later anyways).
