https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116415
--- Comment #8 from Kewen Lin <linkw at gcc dot gnu.org> --- Some more information: bisection showed it started to fail from r12-4240-g2b8453c401b699 which enabled vectorization at -O2. But by further checking, I confirmed that commit just exposed this latent issue, if we escalated the optimization to -O3 which enables vectorization by default, it fails at least from AT14.0 (GCC-10). So this is a latent bug.
