[Bug tree-optimization/56933] New: [4.9 Regression] Vectorizer missing read-write dependency for interleaved accesses

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56933

             Bug #: 56933
           Summary: [4.9 Regression] Vectorizer missing read-write
                    dependency for interleaved accesses
    Classification: Unclassified
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: wschmidt@gcc.gnu.org
                CC: bergner@gcc.gnu.org, dje@gcc.gnu.org,
                    rguenth@gcc.gnu.org
              Host: powerpc*-*-*
            Target: powerpc*-*-*
             Build: powerpc*-*-*


Created attachment 29861
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=29861
Vectorization details dump for the test case

vect_analyze_group_access() in tree-vect-data-refs.c contains a test for
load-store dependencies:

              if (GROUP_READ_WRITE_DEPENDENCE (vinfo_for_stmt (next))
                  || GROUP_READ_WRITE_DEPENDENCE (vinfo_for_stmt (prev)))

Currently this always returns false because this field has not yet been set in
the vinfo.  This began with r196872, where the code to analyze accesses was
moved ahead of the code to analyze dependences.

I put together a test demonstrating that it's possible for us to generate
incorrect code as a result:

       subroutine test(a,b,c,d,e,f)
       integer k
       real*4, intent(out) :: a(1000)
       real*4, intent(out) :: b(1000)
       real*4, intent(in) :: c(1000)
       real*4, intent(inout) :: d(2000)
       real*4, intent(out) :: e(1000)
       real*4, intent(out) :: f(1000)

       do k = 1,1000
          a(k) = 3.0 * d(2*k)
          e(k) = 3.3 * d(2*k+1)
          d(2*k) = 2.0 * c(k)
          d(2*k+1) = 2.3 * c(k)
          b(k) = d(2*k) - 5.5;
          f(k) = d(2*k+1) + 5.5;
       enddo

       return
       end

I'm attaching a detailed dump of the vectorization pass that shows that the
values of d(2*k) and d(2*k+1) used to compute b(k) and f(k) are the ones loaded
prior to the stores to those locations.

To reproduce on powerpc64-unknown-linux-gnu:

$ gfortran -O3 -ffast-math -mcpu=power7 -fno-vect-cost-model interl-lsl-2.f