Re: [hatari-devel] Movem and cycles

[Laurent Sallafranque <laurent.sallafranque@xxxxxxx>]

Hi Toni,

Yes, I know div/mul algo are not known on 68020/68030.
I was just thinking about a general improvement of the cycles counting 
with memory access.

I think that memory access are not  taken into account in 68030 mode (or 
wrongly taken into acount).

I mean actually, hatari reads a movem.l (a6)+,d0-a4 in 16 cycles instead 
of about 120 cycles.
I would love to see my falcon reading all this memory in 16 cycles.
I think the memory access is not taken into account for MOVEM.

I know there's still a lot of work to do to reach a better accurate 
68030, but here, I was speaking of trying to reach a closer level of 
accuracy (we start from far with movem ;)

Regards
Laurent






Le 04/02/2017 à 17:41, Toni Wilen a écrit :
> > Hi
> >
> > for now, only 68030 in CE will count cycles for instructions with non
> > fixed cycles count.
> >
> > Toni, do you think it would be possible to extend gencpu and to define
> > "using_simple_cycles = 1" for cpu id 20 and 22, maybe 32 too ? (as it's
> > the case for id 12 and 68000 in prefetch mode).
> >
> > It could give some extra accuracy when CE mode is not needed for movem,
> > bit shift, muls/divs
> >
> > Just adding "using_simple_cycles = 1;" is not enough at the moment
> > because in that case gencpu checks for "using_prefetch" and not for
> > "using_prefetch_020", but maybe adding this would be not to invasive ?
> I don't think it helps. MOVEM depends on memory access speed, it is not
> CPU core's problem to time it ("memory cycle exact mode" should do that,
> which only counts external memory access cycles, CPU internal cycles are
> instant), shift cycle usage is static in 68020+ (it has barrel shifter).
> And finally, no one knows how mul or div work in 68020+.