[ODE] Report: performances, OPCODE, x86-64, GCC 4.0, etc.

[Tanguy Fautre]

Adam D. Moss wrote:
> 
>> So here are the new results:
>> 32 bits GCC 3.3: 4.2 fps
>> 64 bits GCC 3.3: 2.2 fps
>> 64 bits GCC 4.0: 7.0 fps
> 
> That's amazing.  32 bits GCC 4.0 would be an interesting test
> to see how much of that boost is 64-bit specific!
> 

But keep in mind that the 32 bits GCC 3.3 build is directly from the 
Debian package (which is a shared object library), while I compiled the 
others from the CVS. Hence it is not a fair apples-to-apples comparison.

For the moment, I think that GCC 4.0 is simply doing a better job on 
OPCODE with -O1 than GCC 3.x. I don't think the boost itself is 64 bits 
specific.

To be tested. ;-)


> I agree that this bug seems to have been prevailant through
> enough big GCC releases now to indicate that this may be an OPCODE
> problem.  The specific OPCODE code module which has the problem
> has been isolated in the past (and I thought I added a GCC flag
> workaround for that module, but now I can't find it; that may
> have been clobbered when OPCODE ceased to need to be built
> separately).  Anyway, a deep list archives search should
> hopefully turn it up...

Here is what ODE Makefile says:

# GCC versions 3.X generate bad code for some modules
# with -O2, so we use -O1 until we get this tied down a bit more.
# There are at least two problematic code modules with -O2, one of
# which is OPC_TreeCollider (if you identify the other, which causes
# problems with trimesh-trimesh collisions, then let us know).


Once you found both modules, good luck finding where the problem comes 
from. :-(


PS: Looking at OPCODE code tells me we should clean it up once. There is 
way too much noise (defines mainly) for it to be understandable and 
maintainable.


Regards,

Tanguy