LoopUnrollAndJam miscompiles loop with vector load/store

[mveriksson]

I think LoopUnrollAndJam does something wrong when it checks the dependency between vector loads and stores.

Reproducer:
opt input.ll -S -passes=loop-unroll-and-jam -enable-unroll-and-jam -allow-unroll-and-jam -unroll-and-jam-count=2
input.ll:

define void @badjam(ptr %A) {
entry:
  br label %outer

outer:
  %i = phi i16 [ 0, %entry ], [ %add.i, %outer2 ]
  %add.i = add nuw nsw i16 %i, 1
  br label %inner

inner:
  %index = phi i16 [ 0, %outer ], [ %index.next, %inner ]
  %offset.idx = or disjoint i16 %index, 1
  %0 = getelementptr inbounds [10 x [10 x i16]], ptr %A, i16 0, i16 %i, i16 %offset.idx
  %wide.load = load <2 x i16>, ptr %0, align 1
  %1 = getelementptr inbounds [10 x [10 x i16]], ptr %A, i16 0, i16 %add.i, i16 %index
  store <2 x i16> %wide.load, ptr %1, align 1
  %index.next = add nuw i16 %index, 2
  %2 = icmp eq i16 %index.next, 8
  br i1 %2, label %outer2, label %inner

outer2:
  %exitcond.not = icmp eq i16 %add.i, 9
  br i1 %exitcond.not, label %exit, label %outer

exit:
  ret void
}

Originally from C:

#pragma unroll_and_jam(2)
  for (i = 0; i < 9; i++) { // unroll-and-jam miscompile
#pragma clang loop unroll(disable) vectorize_width(2)
    for (j = 1; j < 9; j++) {
      A[i + 1][j - 1] = A[i][j];
    }
  }

When this is unroll-and-jammed, results in row i=2 are copied from row i=1 too early.

godbolt link

[davemgreen]

This is an example without vectors but with array indices working on a smaller type than the loads/stores: https://godbolt.org/z/ec8vreo66

[sjoerdmeijer]

CC: @sebpop , @sjoerdmeijer