Originally Posted by Toni Wilen
In my opinion biggest problem is planar data getting "lost" inside graphics subsystem and use of routines and algorithms that are not planar friendly.
there is more to that. as thor already indicated over at a1k, there is problem with layers. then the uae hidd is too rudimentary, it doesnt seem to notice when the vmem is going out, aros cgx doesnt seem to support swapping memory out the graphics memory to the main memory like p96 does and so on.
Using assembly outside of performance critical inner loops does not fix it. One exception is exec task switching and bigger parts of interrupt handling and CIA timers, it is very performance critical part (any time used for housekeeping is time taken from programs) and should be in assembly.
aros exec is not that bad. time critical parts may be solved by asm inlines within ifdefs. i have timed aros exec function for function in comparison to amiga exec on the same testcases using librarytimer in the past and im going to do it again when necessary.
Anyway, I am still not doing anything until compiler is fixed
so, why you even speak up here?
its not in my or olafs expertise field to do that. all i can do is try to replace gcc with vbcc building some small module to see if it even works. then one would probably have to adopt the compiler options to get any advantage of it. then probably parametrizing macros need to be adopted to vbcc.
most of that are things i have not much clue about, but how do you exactly would proceed, step by step?
now, apart of compiler discussion, on practical matter there are my cxurrent priorities:
1. get a4000 (and a600 by the way) ide (ata hidd) working in order to make possible regular testing on these machines. a4k is fine with scsi bus, but in order to have minimal setup and space for expansions like rtg or lan card it would be mandatory. fastata conflicts with too many other zorro cards also on genuine kickstart.
2. get uaehidd to recognize and correctly handle particular pixel formats and restricted vmemory a real rtg card offers. the prefs screenmode should only offer resolutions masked out of what the driver offers on these. the maximal resolutions may be simply calculated from the available ram.
3. get hunk binaries loading as fast as elf.
4. check where exec can be sped up.
5. while that all, port or fix compiling of aros contributions for 68k (i have done number of things in this direction already with tremendous support of deadwood, most notably scalos works, some additional mui classes work, calculator tool works, lunapaint opens but doesnt work correctly yet)