Re: [hatari-devel] Suspicious instruction & data cache hit/miss accounting

[ Thread Index | Date Index | More Archives ]


On 02/06/2018 07:20 PM, Nicolas Pomarède wrote:
Le 06/02/2018 à 18:09, Eero Tamminen a écrit :
I added that statistic, but now I started to wonder about terminology,
what would be best understood by the users of the profiler.

Is "prefetch" correct name for moving data from i-cache to CAHR
register, or do people normally interpret "prefetch" to mean reading
instructions from system RAM to i-cache? >
I.e. maybe the zero i-cache hit/miss case should be named as:
   "Cache holding register (CAHR) already refilled from i-cache"

instead of my current "Already prefetched" name?

In fact, I don't know if all users of the profiler are aware of these low details of the CPU inner work, so finding a brief and correct wording might be difficult.

Most users who look for optimisations should try to increase hit values and decrease miss values, ignoring this "prefetched" case (as it's basically a 50% chance).

It's only visible with the cache statistics histograms command,
not in disassembly, so I think it gets pretty much ignored. :-)

Still the motorola docs tells about CAHR in "11.2.2" :

"This long word is obtained from the instruction cache o the external bus in response to a prefetch request from the microsequencer."

So I think that "already prefetched" is correct.
Moving from ram to i-cache should be simply called "caching" :)


	- Eero

Mail converted by MHonArc 2.6.19+