[hatari-devel] Re: Profiler CPU cycles values

[Eero Tamminen <oak@xxxxxxxxxxxxxx>]

Hi,

On tiistai 12 helmikuu 2013, Douglas Little wrote:
> On 12 February 2013 10:38, Douglas Little <doug694@xxxxxxxxxxxxxx> wrote:
> > ...one small update - while the profiler crash is now gone, the
> > instruction cycle tally information is still peppered with -1 values,
> > so this will need diagnosed too at some point... (4294967295 ==
> > 0xffffffff) 
....
> > $023eb2 :             dbra      d0,$23eac                  0.00% (287,
> > 4294967295, 0)
> > $023eb6 :             move.w    #$1554,d0                  0.00% (41,
> > 4294967295, 0)
> > $023eba :             clr.w     d1                         0.00% (41,

All collected values are are unsigned and the collection is stopped
at maximum unsigned value to prevent unnoticed value wrapping.


> ...one more bit of supporting information - although I don't know if it
> is relevant to this. I notice every time I use the profiler I see this
> 'Counters OVERFLOW' warning. This happens regardless of how long the
> profiling session is (a few seconds is enough).
....
> - used cycles:
>   3650737951796 (99.65% of all)
> - instruction cache misses:
>   513112 (99.96% of all)
> *- Counters OVERFLOW!*

Cycles counter is always larger than instruction or cache miss count.
When statistics notices largest cycles count value to be largest unsigned
value i.e. statistics collection for some instruction to have reached
its max value, it will note the user about this.

I'll look into that after I've gotten the callgraph stuff to working
state and we otherwise look more into cycles counts.  For now cycle
counts are best ignored for CPU.


	- Eero