https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115556
--- Comment #21 from Andrew Pinski <pinskia at gcc dot gnu.org> --- The problem has only to do with POLY_INT_CST constants I think; I don't think vector constants could be prop'ed over a function not in SVE mode. If so we could just reject POLY_INT_CST constants always for at least a workaround for older branches. For the trunk we could add a new target hook that asks if the constant is ok for the current function. The default hook returns true always. For aarch64 and riscv, it returns false if the constant is a POLY_INT_CST is not allowed. Note I don't know how to easily to create a POLY_INT_CST for riscv using their intrinsics. But I could learn by the time I implmenet the hook ...
