Re: [AD] About Elias' bug

[Thomas Fjellstrom <tfjellstrom@xxxxxxxxxx>]

Eric Botcazou wrote:
> 
> > or we duplicate the code and use your method (that seems to be a little
> > faster).
> 
> I've attached a patch which does that. But I'm not very pleased with it:
> - it adds a large amount of code,
> - some very weird things happen: in normal conversion mode, it runs 20%
> slower than the current 3.9.37 CVS code on my machine, although it uses the
> same code with one less test per pixel (but supposedly crunched by the loop
> optimizer).
> 
> The same thing happens with the current CVS code: if I remove the test in
> the CONVERT_DITHER_BLIT() macro, the code runs slower. I guess this is
> somehow related to the loop optimizer, the loop code size and branch
> prediction but I'm a little puzzled.
> 

If the test surrounds a few instructions then I bet that the
looP is left in Place.


P.S. sorry for the 'P's my keyboard is dying
     SPilling a full cuP of iced tea into it will do that...
-- 
"Computer programmers don't byte, they nybble a bit."
Thomas Fjellstrom
tfjellstrom@xxxxxxxxxx
http://strangesoft.net/