Parallelizing profile driven optimization

[Michael Veksler <mveksler@xxxxxxxxxxxxxxxxxxxxxxx>]

I want to speed up the following process through parallelization. Is it 
safe to
run different instances of the training set in parallel? Will the 
profiling information
of one run overwrite, rather than accumulate, of a parallel run? Are 
there race
conditions?

I am using profiling driven optimization using the following steps:

  1. Run g++ 4.4.3 with  -fprofile-generate=$HOME/project
  2. Run the resulting executable with a training set
  3. Rerun g++ with exactly the same parameters but now instead of
     -fprofile-generate, I pass g++   -fprofile-use=$HOME/project


I can parallelize step 1 with "make -j2", similarly for step 3. This 
does not help me much
because step 2 takes most of the time.

My question, can I safely run several runs of step 2 and reach exactly 
the same
profiling information, without the fear of profile information loss or 
mutation?

Thanks
   Michael