Ramana Radhakrishnan wrote:
Hi Jeffrey,
I'm seeing a few performance regressions similar to
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32306 and
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33315 in a port where I'm
working off the 4.3 branch. These regressions are caused by the
decision to stop iterating DOM on identifying jump threading
oppurtunities and splitting up VRP into a separate pass based on your
patch at http://gcc.gnu.org/ml/gcc-patches/2006-02/msg00586.html and
as described by Richard here
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32306#c7.
I'm going to try a quick experiment by re-running DOM if jump
threading is turned on in the 4.3 branch. I know there are a number of
changes I need to do and I'm looking into them. I wanted to check back
if you still had any record of the small number of performance
regressions that you mention in the posting of the afore mentioned
patch.
I didn't keep a record of things we missed due to removal of the
iteration step. My experience was that they were relatively rare and
the cost to pick them up was fairly high, so the decision was made to
remove the iteration step. So while there are some codes where we miss
threading opportunities, it's my belief that we're on the right side of
the cost/benefit curve.
I've postulated in the past that I believe the iteration step could be
faster by moving to a worklist based algorithm for DOM -- the theory
being that the cascading optimization opportunities exposed by jump
threading are actually rather limited both in terms of paths that need
revisiting and in how newly exposed equivalences can be utilized. I
was never able to convince myself that the scheme would work though and
thus I never wrote code to try it.
There was also a master's thesis which touched on these issues and
approached path specific optimizations with a PRE-like framework. I
could probably dig this up if you wanted to look at it -- the overall
approach seemed based on fairly solid fundamentals as opposed to the
ad-hoc threading+DOM stuff we're using.
I would welcome investigations into solving some of these problems and I
would certainly encourage you to submit testcases for the testsuite
which test for the missed optimizations.
Jeff
