https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96814
--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> --- OK, so the extra folding for PR97085 also fixed this testcase, now expanding from _1 = { 0, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1 }; _2 = .VCOND_MASK (_1, { -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1 }, { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 }); _3 = VIEW_CONVERT_EXPR<V>(_2); I still have a patch for the CTOR expansion which I'll test and propose.