This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: RFD: profile standard deviation of # of loop iterations?
- From: Joern Rennecke <renneckej at superh dot com>
- To: dalej at apple dot com (Dale Johannesen)
- Cc: gcc at gcc dot gnu dot org, dalej at apple dot com (Dale Johannesen), joern dot rennecke at superh dot com (Joern Rennecke)
- Date: Tue, 16 Mar 2004 21:19:31 +0000 (GMT)
- Subject: Re: RFD: profile standard deviation of # of loop iterations?
> What would you do with the standard deviation after you got it?
> I don't see how that would let you decide unrolling any better than the
> average.
It gives us a measure of how close the individual numbers of loop iterations
are clustered around the average.
E.g. some 95% of all loop iteration counts are expected to lie within +-
two times the standard deviation from the average.
IF we decide that the loop us usually executed with the average iteration
count, we want to unroll by that number, by a number the average can be
divided by with no or small rest. If it's closely spread around
the average, we want to keep a small rest.