https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63593
--- Comment #3 from Marek Polacek <mpolacek at gcc dot gnu.org> --- So it seems like tree_transform_and_unroll_loop creates a wrong PHI: _101 = PHI <pretmp_103(8), ivtmp_47(4)> where the type of _101 and pretmp_103 is int, and ivtmp_47 is unsigned. How that ivtmp_47 gets there is a mystery to me, but it happens during duplicate_block.