https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71270
--- Comment #9 from ktkachov at gcc dot gnu.org ---
(In reply to ktkachov from comment #8)
> After playing around with this I believe the problem is in the
> *neon_mov<mode> pattern in the arm backend. The way it constructs the
> const_vector of {1, 0, 1, ... } is probably wrong for big-endian.
I should say the problem is not the RTL const_vector, but rather the vmov
immediate that neon_immediate_valid_for_move generates
