@PoP:
I noticed 
.
I'm not sure if that's a quirk of AA (since it's done over 256 shades of grey, not 16, and I haven't checked if it's "stable"), of the blending, or of the color decimation (I'm loath to use quantization here, since what I've seen it do looks more like decim than proper quant ;p) done by the controller (especially since it does it on 32bpp fbs, too, where we do read back full-range 8-bit values from the fb, values that have almost no bearing as to what's actually displayed by the screen).
It is a tiny bit more even on >4bpp fbs, because the crappy 4bpp layout kind of makes it worse (which leads to artifacts that looks a tiny bit like aliasing here). Increased screen density also helps, and as 4bpp screens are all 800x600 @ 167dpi, that's another strike against 'em

.
I didn't look into it much further, because the "double -o" usage pattern is kind of anecdotal