Re: [hatari-devel] UAE 68030 MMU + prefetch + instruction and data cache emulation

[Nicolas Pomarède <npomarede@xxxxxxxxxxxx>]

Le 30/07/2017 à 20:34, Laurent Sallafranque a écrit :
> Here it is :
>
> hg bisect --good
>
> The first bad revision is:
> changeset:   6571:6986f406b928
> user:        Nicolas Pomarede
> date:        Wed Apr 19 17:01:03 2017 +0200
> summary:     FPU/SoftFloat + small changes for 68020/30 prefetch cycles 
> (From WinUAE 3.4.1 b6)
>
>
> Laurent

Hi,

I try "_" demo, and I confirm it locks at start during DSP/CPU transfer 
; this is most likely the same problem as in many DSP/CPU demos where 
68030 cycle counting is not exact yet, which will make DSP and CPU sync 
not possible (cpu waits forever for dsp data)

The changeset 6571 contained some changes from Winuae to improve how 
68020/30 accesses memory, but this is still work in progress. Sometimes 
it will give better results, sometimes (as in these 2 demos), it will 
give not-accurate enough cycles count and the program will fail.

There's no solution at the moment for this, the memory "pipeline" of the 
68020/30 is really complex.

For now, we should look for regression in games/demos where DSP is not 
used, as DSP/CPU transfers don't have correct timings in many cases.

Nicolas