This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: Merging Graphite to trunk
On Thu, Jul 30, 2009 at 08:24, Jack Howarth<howarth@bromo.med.uc.edu> wrote:
> Sebastian,
> ÂHas anyone benchmarked the current graphite branch vs
> gcc 4.4.1 with the Polyhedron 2005 benchmark suite using
> the various graphite loop optimizations? If so are we
> finally seeing significant improvements in the runtime
> benchmarks for those tests?
The automatic tester running on the GCC farm is continuously testing
the polyhedron benchmarks for compile, runtime correctness, but not
for speed. Extra benchmarking and testing is always welcome.
In particular, I do not know yet if we managed to solve "[Graphite]
70% slower using -floop* than without graphite (gas_dyn of
Polyhedron)": http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38846
My understanding of this problem is that the Graphite generated code
is not recognized by the pattern matching of the vectorizer. We did
solved some of the problems that we identified, like the part 0006:
"Leave the loop_latch basic block empty." that would have caused the
vectorizer to ignore all the loops generated by Graphite. Another
problem that is still not fixed is the bb and edges frequencies and
probabilities that are for the moment not set up correctly, making all
the loops generated by Graphite look like "cold loops", and the cost
model of the vectorizer just discards those loops.
Sebastian