------- Comment #1 from rguenth at gcc dot gnu dot org 2007-07-04 16:11 ------- This seems to be an aliasing problem. Somehow we do not prune SMTs enough after the patch. MPT grouping doesn't trigger in.
An example function to look at is (x86_64) _ZN14MultiArgKernelI9MultiArg4I5FieldI22UniformRectilinearMeshI10MeshTraitsILi3Ed21UniformRectilinearTag12CartesianTagLi3EEEd10BrickViewUES9_S9_S9_E15EvaluateLocLoopIN6Forgas5CentYILi3EEELi3EEE3runEv if you look at the alias6 dump (the one before lim which does no longer move invariant integer loads out of the inner loop) you can see: without the patch: # VUSE <SMT.21607_161(D)> D.852018_33 = op$multiArg_m_7->a1_m.fieldEngine_m.data_m.blockControllerPtr_m.ptr_m; with the patch (only the tree-inline.c part is actually necessary): # VUSE <SMT.21592_132(D), SMT.21595_227> D.854673_30 = op$multiArg_m_19->a1_m.fieldEngine_m.data_m.blockControllerPtr_m.ptr_m; it doesn't make sense that there is a difference in pruning. Really. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32624
