[Bug middle-end/42777] New: [graphite] loop interchange heuristic is broken

["spop at gcc dot gnu dot org" <gcc-bugzilla at gcc dot gnu dot org>]

Starting from r155273: Fix fall outs from Fix-PR42334.

2009-12-15  Sebastian Pop  <sebastian.pop@amd.com>

       PR middle-end/42178
       PR middle-end/42334
       * graphite-interchange.c (lst_try_interchange): Do not increment the
       the OUTER index when there is no AFTER kernel.  Do not increment the
       OUTER index for after processing the AFTER kernel.
       (lst_interchange_select_inner): Call lst_try_interchange only on loops.
       (lst_interchange_select_outer): Do not pass in a pointer to the OUTER
       index.  Do not pass to lst_interchange_select_inner the OUTER index.
       (scop_do_interchange): Update use of lst_interchange_select_outer.

       * testsuite/gfortran.dg/graphite/graphite.exp
       (DEFAULT_FLAGS_GRAPHITE_IDENTITY): Remove -fdump-tree-graphite-all.
       * testsuite/gfortran.dg/graphite/interchange-1.f: Add comment.  Clean
       the graphite dump file.
       * testsuite/gfortran.dg/graphite/interchange-2.f: Same.
       * testsuite/gfortran.dg/graphite/pr42334-1.f: New.

and with r155248: Fix PR42334: correct the update of the LST on loop
interchange and distribution.

2009-12-15  Sebastian Pop  <sebastian.pop@amd.com>

       PR middle-end/42178
       PR middle-end/42334
       * graphite-interchange.c (lst_perfect_nestify): Reset to NULL the LSTs
       that are empty.
       (lst_do_interchange_1): Renamed lst_interchange_select_inner.
       (lst_try_interchange): Reimplemented.
       (lst_interchange_select_inner): Same.
       (lst_do_interchange): Renamed lst_interchange_select_outer.
       Reimplemented.
       (scop_do_interchange): Update use of lst_interchange_select_outer.

       * testsuite/g++.dg/graphite/pr42130.C: Add -fgraphite-identity.
       * testsuite/gcc.dg/graphite/block-0.c: Un-XFAILed.
       * testsuite/gcc.dg/graphite/pr42211.c: New.
       * testsuite/gfortran.dg/graphite/pr42334.f90: New.

we have at least one benchmark in {calculix,gamess,mcf} from the
SPEC2006 that performs more than 20% slower at -O2 than at -O2
-fno-{graphite,loop-interchange,...} on the graphite branch: see the
recent test on http://groups.google.com/group/gcc-graphite-test

I am suspecting the heuristic of loop interchange to be in fault here,
as the above commit was fixing the traversal of the loops where the
interchange is applied.


-- 
           Summary: [graphite] loop interchange heuristic is broken
           Product: gcc
           Version: 4.5.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
        AssignedTo: spop at gcc dot gnu dot org
        ReportedBy: spop at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42777