Re: [hatari-devel] Issues with cache hits/misses?

[Nicolas Pomarède <npomarede@xxxxxxxxxxxx>]

Le 31/05/2015 23:35, Nicolas Pomarède a écrit :

> I don't know yet, but we can clearly see that even during the early
> memory tests at boot in falcon mode, the screen refresh is very slow (we
> see red and yellow Atari logo, because all 4 planes are not copied fast
> enough to get the 'black' color in one go).
> So, this looks more like a 10 times slowdown than just a 10-20% overhead.

I made more tests in falcon mode, and the big part of the slow down 
comes from this in read_dcache030  :

        // only one long fetch needed?
        if (size == 0) {
//                                return get_byte (addr+aligned);
                v1 >>= (3 - aligned) * 8;
                return v1;
        } else if (size == 1 && aligned <= 2) {
//                                return get_word (addr+aligned);
                v1 >>= (2 - aligned) * 8;
                return v1;
        }

In most cases, when just opening one of the top menu in the desktop gem, 
the data read from the cache are mostly bytes and aligned words.

If I bypass the "v1 >> .... * 8" line and do a return of the data 
immediately, emulation is nearly as fast as when 68030 cached is disabled.

But it seems that the right shift to extract the correct byte/word from 
the long word is taking a huge time (even with gcc -O3). Maybe some more 
recent cpu are better at doing right shift, but I would have expected it 
took only 1 cycle (in my case, it's an intel core 2 duo @ 3 GHz)

Maybe by splitting this line in several sub-parts it will be faster, 
could be gcc that does bad optimisations.

Nicolas