This is the mail archive of the
mailing list for the GCC project.
Some (small) c++ compilation profiling data (oprofile)
- From: Will Cohen <wcohen at redhat dot com>
- To: gcc at gcc dot gnu dot org, kgardas at objectsecurity dot com
- Date: Fri, 17 May 2002 11:35:34 -0400
- Subject: Some (small) c++ compilation profiling data (oprofile)
- Organization: Red Hat, Inc.
Profiling and characterization of gcc by oprofile would be very useful.
The cvs version of oprofile has a "--reverse" option for oprofpp, which
lists things from largest to smallest. This will put the largest samples
right after the information about what was being sampled, so the most
interesting information is together. It would be useful to include that
header information about what is being sampled and what the rate is.
One of the strengths of oprofile is it can measure things other than
time, e.g. cache misses and branch mispredictions. Having the header
information with the measurements will avoid confusion.
What environment (hardware and software) did you use to take these
measurements? Ideally, someone else should be able to reproduce and
verify the result. GCC is a moving target, the cvs tag used to check
out the code should be included. Is the bootstrapped compiler being used
for the tests? If just the stage1 compiler is being used, the compiler
used to build the GCC being profiled can influence the measurements.
You might want to get copies of the Intel software optimization manuals
to get an idea of what other metrics to measure.
Pentium III software optimization manual:
Right now oprofile only provide RTC sampling for Pentium 4, but here is
link to the web page for the Pentium 4 software optimization manual:
Will Cohen, GCC Engineer