[Bug rtl-optimization/37534] [4.4 Regression] IRA causes 17% degradation in 187.facerec benchmark

["vmakarov at redhat dot com" <gcc-bugzilla@gcc.gnu.org>]

------- Comment #11 from vmakarov at redhat dot com  2009-02-03 14:48 -------
I have a patch (a new spill heuristic) which makes facerec even faster with IRA
on power6.  The patch is in

http://gcc.gnu.org/ml/gcc-patches/2008-11/msg00368.html

Unfortunately the new spill heuristic results in worse SPECINT2000 scores for
x86, x86_64.  So I did not committed the patch.

Facerec has one very hot (80%) function localmove::graphRoutines.f90 and
there is only one hot loop in the function.  Although the loop is
pretty big because of inlining TopCostFct.  The loop contains a few
if-statements and several big switch-statements.  But only a small part of
the loop body is hot.  Static branch prediction makes very inaccurate results
in this case and the old RA accidentally generates a better code than IRA.

 This is classical example of heuristic approach drawback.  You never
get best code using one heuristic for all programs.  Sometimes the old
RA will generate better code especially when the static branch probability
prediction is bad.  What are goal should be is to achieve
better or not worse code in "average", i.e. for some credible benchmark like
SPECFP2000 and IRA is successful in this. 

I am still working on some better spill heuristics.  Classical graph coloring
spill heuristics are bad and IMHO they should be based on register pressure not
on number of allocnos conflicting with allocno being spilled.

I don't think the problem will be solved for gcc4.4 but it might be solved for
gcc 4.5.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37534