https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99721
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
So we end up scheduling
t.f90:1:16: note: node 0x35514b8 (max_nunits=1, refcnt=1)
t.f90:1:16: note: op: VEC_PERM_EXPR
t.f90:1:16: note: { }
t.f90:1:16: note: lane permutation { 0[0] 1[1] }
t.f90:1:16: note: children 0x3550bb0 0x3550d48
t.f90:1:16: note: node (constant) 0x3550bb0 (max_nunits=1, refcnt=1)
t.f90:1:16: note: { -4.1e+1, -4.1e+1 }
t.f90:1:16: note: node (constant) 0x3550d48 (max_nunits=1, refcnt=1)
t.f90:1:16: note: { 4.3e+1, 4.3e+1 }
we cannot compute an insertion place for the permute. Ideally we'd have
constant folded this permute node, it gets built by vect_build_combine_node,
but for now avoid the situation during SLP analysis.
