[Bug tree-optimization/68379] New: [6 Regression] BB vectorization: definition in block 13 follows the use for SSA_NAME

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68379

            Bug ID: 68379
           Summary: [6 Regression] BB vectorization: definition in block
                    13 follows the use for SSA_NAME
           Product: gcc
           Version: 6.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rguenth at gcc dot gnu.org
  Target Milestone: ---
            Target: aarch64

! { dg-do compile }
! { dg-additional-options "-Ofast" }
! { dg-additional-options "-mavx" { target x86_64-*-* i?86-*-* } }

      SUBROUTINE PASSB4 (IDO,L1,CC,CH,WA1,WA2,WA3)
      IMPLICIT REAL(4) (A-H, O-Z)
      DIMENSION       CC(IDO,4,L1)           ,CH(IDO,L1,4)           ,&
                     WA1(*)     ,WA2(*)     ,WA3(*)
  102 DO 104 K=1,L1
         DO 103 I=2,IDO,2
            TI1 = CC(I,1,K)-CC(I,3,K)
            TI2 = CC(I,1,K)+CC(I,3,K)
            TI3 = CC(I,2,K)+CC(I,4,K)
            TR2 = CC(I-1,1,K)+CC(I-1,3,K)
            TI4 = CC(I-1,2,K)-CC(I-1,4,K)
            TR3 = CC(I-1,2,K)+CC(I-1,4,K)
            CH(I-1,K,1) = TR2+TR3
            CH(I,K,1) = TI2+TI3
            CI4 = TI1-TI4
            CH(I-1,K,4) = CI4
            CH(I,K,4) = CI4
  103    CONTINUE
  104 CONTINUE
      RETURN
      END

on aarch64-linux ICEs with

vect-9.f90:5:0:

       SUBROUTINE PASSB4 (IDO,L1,CC,CH,WA1,WA2,WA3)


Error: definition in block 11 follows the use
for SSA_NAME: _54 in statement:
vectp.20_198 = &*cc_36(D)[_54];

Breakpoint 2, internal_error (gmsgid=0x1718ff4 "verify_ssa failed")

which is because when we vectorize

vect-9.f90:10:0: note: node
vect-9.f90:10:0: note:  stmt 0 _37 = *cc_36(D)[_35];
vect-9.f90:10:0: note:  stmt 1 _37 = *cc_36(D)[_35];

the interleaving with _55 = *cc_36(D)[_54]; and a permutation of { 1 1 } will
cause us to base the reference off a pointer to a ref that does not
participate in the SLP node and thus is not considered when we look for
the insertion place with vect_find_last_scalar_stmt_in_slp.