Re: [AD] malloc

[aj <aj@xxxxxxxxxx>]

> > As it stands at the moment, allegro can be severly performance hit, if
> > someone compiles it with SSE turned on, P4, and AMD CPUs both have SSE.
> >    
>
> Have you tried enabling SSE to test the performance hit? And if so, do you 
> know if it's not just from improper SSE code use, or due to improper memory 
> alignment? Do you know if GCC/MSVC don't already optimize malloc calls to 
> return aligned pointers when SSE is being used?
>  

no i dont know if mallocs auto align during SSE use, however all the 
documenation i've read including msdn stuff specifically tell you to 
align memory using  aligned_malloc()
surely they would not have made such a big deal of it if msvcr.dll does 
this already.

> Besides, I thought we already talked about this on IRC and such. Memory 
> bitmaps and samples and such won't benefit from aligned allocation since even 
> if the pointer is aligned, each value within will not be (eg. if
> bmp->line[0][0] is an aligned access, ((uint32_t*)bmp->line[0])[1] won't be, 
>  

i wouldn't be  accessing [0][1]    SSE is 128bits wide, pixels are 32, 
so i can process 4 pixels at once.
then the next load would be [0][4] (aligned again).


> etc). Plus, you can't gaurantee the alignment of video bitmaps since the 
> system allocates/maps them, not your program.
>  

video cards dont have SSE.
why would you write SSE code for video bitmaps?


> Personally, I don't see why GCC would generate SSE code to use on dereferenced 
> pointers (which it'll almost never know at compile time if they're aligned or 
> not). Instead, it probably would only use SSE code to use on the variables 
> themselves, where it'll know for certain if it will be properly aligned or 
> not.
>  


using array notation assists in most optimzing compilers abilities to 
use SSE effectively.



>  
>
> > Intel are now recommending that code favour SSE instead of MMX (cpu
> > manufacturers have already signalled the death of MMX)
> >    
>
> MMX won't disappear from processors. There are programs that require an 
> MMX-compatible CPU, and such programs would become unrunable if MMX went 
> away. And it's only natural that Intel would favour SSE over MMX. SSE is 
> newer, and probably better than MMX (and it gives users more of an incentive 
> to buy a newer CPU if they still have old ones).


nor will i386 instructions, so should we not bother with MMX either, as 
i386 is good enough.