[Bug tree-optimization/53346] [4.6/4.7/4.8 Regression] Bad vectorization in the proc cptrf2 of rnflow.f90

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53346

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
         AssignedTo|unassigned at gcc dot       |rguenth at gcc dot gnu.org
                   |gnu.org                     |

--- Comment #5 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-05-18 11:02:53 UTC ---
Yeah, this is sort-of related to what is observed in PR53355.  I suppose at
runtime nxtr is comparatively small.

Reduced testcase:

      subroutine cptrf2 (nxtr, ixtrt)
      integer, dimension (1:nxtr), intent (out) :: ixtrt
      ixtrt = 0
      end subroutine

we peel the loop to possibly align the stores, and we peel the loop
to possibly take care of a remaining store at the end of the array.
And of course we compute that we need at least 6 scalar iterations
to make executing the vectorized loop profitable.

And apart from all that we should have recognized the loop as memset.

Mine.