Re: More Quesa optimisation

[James W. <ja...@fr...>]

Roger Holmes wrote:
> e3group_submit_render is used a great deal, recursing very deeply in  
> Microspot's applications, so Shark shows it, and the routines it  
> calls use a lot of CPU time. I have a couple of optimisations for it,  
> but lets keep it simple and do one at a time for now. It calls  
> OpaqueTQ3Object::GetClass, which (in the disassembled code)pushes a  
> register, checks that its 'this' pointer is NULL and sets the return  
> register to a field of 'this', or to NULL if 'this' was NULL, then  
> pops the register back off and actually returns.
> 
> e3group_submit_render then ignores whether the return value was NULL  
> and dereferences it.
> 
> An optimisation for this is to have an inline   
> OpaqueTQ3Object::FastGetClass which just returns the field of the  
> class. This avoids the register pop, the call, the test for NULL and  
> may allow e3group_submit_render to make better use of its registers,  
> which are somewhat more limited on Intel than they were on PowerPC.

Is there any case in which it is not a bug for 'this' to be NULL?  I 
can't find one.  I suggest that instead of adding FastGetClass, we just 
change the implementation of OpaqueTQ3Object::GetClass to

Q3_ASSERT( this != NULL );
return theClass;

so that in non-debug builds it will just return theClass.  Also GetClass 
could be made inline.

-- 
   James W. Walker, Innoventive Software LLC
   <http://www.frameforge3d.com/>