https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115669
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> --- I can reproduce on aarch64 with -O2 -march=armv8.3-a+sve -fno-vect-cost-model -fwrapv. The reduction is <bb 3> [local count: 955630224]: # b_lsm.9_11 = PHI <_9(7), pretmp_42(6)> _4 = *_3; _5 = (signed int) _4; b.2_7 = (signed int) b_lsm.9_11; _15 = b.2_7 + -1; _8 = _15 - _5; _9 = (unsigned int) _8; with -fwrapv and # b_lsm.9_11 = PHI <_7(7), pretmp_40(6)> _4 = *_3; _5 = (unsigned int) _4; _15 = b_lsm.9_11 + 4294967295; _7 = _15 - _5; without. So it seems the support for singed/unsigned converted reductions is somehow broken. At least for VLA vectors. The epilog does # vect__9.22_59 = PHI <vect__9.22_58(3)> _60 = (vector([4,4]) signed int) vect__9.22_59; _61 = .REDUC_PLUS (_60); I don't see where it goes wrong yet.
