http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48616
--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-04-15 19:08:50 UTC --- Created attachment 24003 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24003 gcc46-pr48616.patch This untested patch seems to work for me, though the generated code is in some cases still less than ideal, because it doesn't figure it can just use a vector computed in the earlier bb. It would be nice if it could figure that out or some later pass be able to clean that up. Additionally, it doesn't seem to SLP optimize fn3's first bb, where it very well could use a scalar shift. The testcase will fail without the patch just when testing with -mxop in flags, we don't have unfortunately yet xop and xop_runtime effective targets in testsuite/lib/
