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

[Eero Tamminen <oak@xxxxxxxxxxxxxx>]

Hi,

On 07/30/2017 04:38 PM, Laurent Sallafranque wrote:
> Tested with hatari 2.0.0, MMU on, cycle exact and prefetch mode ON, 14 
> Mo, Tos 4.04 FR, 24 bits mode addressind ON, no Patch timer_D, no boot 
> faster by patching TOS...
>
> I've removed my hatari.cfg before testing.
>
>
>     X_tasie works well with 2.0.0 and MMU activated.
>
>     _ (underscore) demo runs well too (until the 4th screen, but that's 
> another problem.
>
>
> With the latest source, both demos freeze (probably because of cpu <-> 
> dsp timings as I can see when I debug).
>
>
> To be noticed : these 2 demos runs only in MMU mode.

2.0 is a bad comparison point for MMU & non-MMU compatibility because
it had broken instruction cache emulation.

Because of this, *many* games and demos happened to work with old MMU
instead of non-MMU mode, just because MMU mode didn't support cycle
exactness.

You should use this commit or later for comparisons:
https://hg.tuxfamily.org/mercurialroot/hatari/hatari/rev/c8fb3b2e7852

See the release notes for full list of impacted programs:
https://hg.tuxfamily.org/mercurialroot/hatari/hatari/raw-file/tip/doc/release-notes.txt


> There's certainly a regression in the current version and MMU.


	- Eero

> Regards
> Laurent
>
>
>
> Le 30/07/2017 à 13:19, Nicolas Pomarède a écrit :
> > Le 30/07/2017 à 12:53, Laurent Sallafranque a écrit :
> > > Hi,
> > >
> > > A few tests done with the new core :
> > >
> > > - rainbow 2 (use the MMU) : OK
> > > - X_tasie demo (runs with MMU) : KO (probably because of CPU timings, 
> > > as it's locked between DSP and CPU exchanges)
> > > - _ demo : it seems to have regressed (both 68030 and non 68030 cpu). 
> > > Before, it was running at least the 3 first screens before freezing, 
> > > but now it freezes at start.
> > >
> > > - a few more demo tested, that are working, (they probably don't use 
> > > the MMU). But with the 2 cycles instructions timings, they look 
> > > strange, or music is full of glitches, ...
> > >
> > > I'll do some more tests.
> > >
> > > The best now for hatari would be to have 68030 + MMU + cycle exact 
> > > (nevermind if the MMU instructions are not cycle exact, there's never 
> > > involved in time critic loops).
> > >
> > > That's the Falcon standard CPU.
> > >
> > > Thanks for the new CPU Tony and Nicolas.
> >
> > hi
> >
> > thanks for the tests ; for those that don't work, did they work with 
> > Hatari 2.0 and the same settings ? It's important to know if 
> > regression is due to latest MMU changes or if it was already wrong 
> > before.
> >
> > Nicolas