https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104214
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> --- So it looks like we registered q_4(D) != 2147483647 && p_3(D) != -2147483648 as assumptions as we set no_overflow to false before using (delta + step - 1) / step for niter in number_of_iterations_lt. So for the fixed cases it seems that some earlier transforms use the adjusted IVs under wrong assumptions. A way out would be to register the no overflow assumptions earlier. Oh, and for pointers we kept the no overflow guarantee, making the relational compare transform always valid - possibly because of reasoning that relational compares in C invoke undefined behavior when they compare two distinct objects and object addresses not wrapping around (ignoring maybe the special case of one-after-the-object being wrapped). So the testcase which uses pointers can possibly be fixed by re-instantiating that. As said, further improvements might involve doing sth like registering assumptions earlier and to make that possible for LE also, turn LE to LT earlier as well.
