Reconciling PowerPC simulator traces with performance monitor results

["Wealand, Barry" <barry.wealand@lmco.com>]

Hello -

I'm using the PowerPC simulator in GDB to collect instruction (-t
semantics) and load-store (-t load-store) traces for a pretty-good-sized
program (several hundred million trace lines).  I then apply various
filters to the traces to collect measurements of interest.  All of this
seems to work very well.

Then I run the exact same code on a PowerPC 750FX processor, which I set
up to collect performance monitor statistics.  For example, I can
collect "completed load and store instructions", and "number of
instructions completed from the FPU".  What puzzles me is that the
results that I collect from the simulator for these two counts in
particular are 10 to 15 percent lower than the results reported by the
performance monitor.

I'm guessing that this discrepancy might have something to do with the
details of how the performance monitor works, but the descriptions above
are typical of the level of detail provided in the 750 FX manual.  I
wonder if anyone else has grappled with this, or might have any
suggestions.

Thanks!

Barry Wealand
Lockheed Martin Space Systems Company