https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107295
--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Ah, so that is a different issue then. And with slightly modified testcase quite old one, which regressed with: r8-727-g6b6ae9eb9c06b6911573bb9a13cf98b5a7c98b78 template <typename T> struct VecHelper { typedef T __attribute__((vector_size(sizeof(int)))) V; }; template <int, typename T> using Vec = typename VecHelper<T>::V; template <typename T> using V = Vec<4, T>; using F = V<float>; constexpr F F0 = F() + (float) 0.0L; The problem is that the NOP_EXPR around long double REAL_CST has TREE_CONSTANT flag and so the CONSTRUCTOR around it and thus we just fold the CONSTRUCTOR in: case CONSTRUCTOR: if (TREE_CONSTANT (t) && reduced_constant_expression_p (t)) { /* Don't re-process a constant CONSTRUCTOR, but do fold it to VECTOR_CST if applicable. */ verify_constructor_flags (t); if (TREE_CONSTANT (t)) return fold (t); } r = cxx_eval_bare_aggregate (ctx, t, lval, non_constant_p, overflow_p); break; but that handles the case where it has REAL_CST elts, but doesn't handle the case where it has NOP_EXPRs around them. I wonder if we just shouldn't call cxx_eval_bare_aggregate always for VECTOR_TYPE.