https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113431
--- Comment #14 from Richard Biener <rguenth at gcc dot gnu.org> --- So the issue is that we ignore the dependence because vect_preserves_scalar_order_p - which is correct iff we'd execute the stmt within the vectorized loop body. But as we discover it invariant we hoist it, making the outcome of vect_preserves_scalar_order_p invalid. I have a fix doing t.c:5:17: note: === vect_prune_runtime_alias_test_list === t.c:5:17: note: can tell at compile time that a[0][1] and a[b.2_8][1] alias t.c:7:15: missed: not vectorized: compilation time alias: _18 = a[0][1]; a[b.2_8][1] = _20; and then falling back to SLP vectorizing the unrolled loop as desired.
