| Re: [hatari-devel] Instruction cache misses and CPU cycles |
[ Thread Index |
Date Index
| More lists.tuxfamily.org/hatari-devel Archives
]
Hi,
On lauantai 16 helmikuu 2013, Douglas Little wrote:
> On 15 February 2013 21:45, Eero Tamminen <oak@xxxxxxxxxxxxxx> wrote:
> > Does this look even remotely right:
> > ---------
>>
>> > d 0xe2f044
>>
>> $e2f044 : 3f00 move.w d0,-(sp) 0.31% (43303, 173268,
>> $e2f046 : 4eb9 00e2 fe7e jsr $e2fe7e 0.31% (43303, 433172,
>> $e2f04c : 548f addq.l #2,sp 0.31% (43303, 34,
>> $e2f04e : 4a80 tst.l d0 0.31% (43303, 86692,
>> $e2f050 : 674a beq.s $e2f09c 0.31% (43303, 86666,
---------
>> ?
>
> There can't be fewer total cycles than visits so something is wrong
> there. However it's hard to tell from one instance if it's some sort of
> off-by-one problem or more complicated. Would need to watch what happens
> at lots of sites, and possibly on the first few visits to each affected
> site.
Old UAE core shows sane cycle values for same place (nearly same use-case):
--------------------------
> d 0xe2f044
$e2f044 : 3f00 move.w d0,-(sp) 0.31% (49077, 196364, 0)
$e2f046 : 4eb9 00e2 fe7e jsr $e2fe7e 0.31% (49077, 491054, 0)
$e2f04c : 548f addq.l #2,sp 0.31% (49077, 196372, 0)
$e2f04e : 4a80 tst.l d0 0.31% (49077, 98184, 0)
$e2f050 : 674a beq.s $e2f09c 0.31% (49077, 196420, 0)
--------------------------
And I've never seen it to complain about zero cycles, like happens
with WinUAE CPU core.
Cycles might "bleed" to next instruction from STOP instruction though:
--------------------------
WARNING: cycles 832 > 512 at 0xe015c6
....
> d 0xe015c0
$e015c0 : 4e75 rts
$e015c2 : 4e72 2300 stop #$2300 0.18% (49741, 1448002278, 0)
$e015c6 : 46c0 move d0,sr 0.18% (49742, 299398, 0)
$e015c8 : 4e75 rts 0.18% (49742, 398110, 0)
--------------------------
When looking at newcpu.c, do_specialties() can return from it
after STOP instruction without calling debugger. That probably
needs to be fixed both in old and new CPU core.
- Eero