profiling problems

[Mark Levison]

  I have just finished reading an article called "Inaccuracies in Program
Profilers" by C. Ponder and R. J. Fateman from Software Practice and Experience
May 1988. In this paper several problems with the traditional Unix profiler
gprof are discussed. One of the most important ones is that profiler calculates
the amount of time spent in a routine for each parent by taking the total time
spent in the routine and dividing by the total number of calls to that routine.
The problem is that this is an overall average and if the routine has two
different parents with rather different execution times for each then this
average time will be very misleading. This leads to my question is there any
profiling system out there that will do the same type of job as gprof but
witout this and the other problems the authors mention? Failing that is some
clever way of getting the compiler (cc on a Sun 3/80 at SunOs 4.0.3 or gcc on
the same system) to insert calls to special routine at the start and end of
each call?  I would prefer if such a system were available free because I
do not have any budget to buy any solutions.

Thanks in advance,
Mark Levison
Research Assistant
levisonm at qucis.queensu.ca