https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71702
--- Comment #3 from Alexander Monakov <amonakov at gcc dot gnu.org> --- On 6/trunk, this issue is fixed or made latent by r230667 that added + STRIP_NOPS (t1); + STRIP_NOPS (t2); to tree-vect-data-refs.c:compare_tree (patch submission here: https://gcc.gnu.org/ml/gcc-patches/2015-11/msg02444.html ). Should this change be backported to 4.9/5 now that an ICE is demonstrated? On trunk, backing out that change and adding a qsort comparator checking to vec.h causes ICEs in: gcc.c-torture/execute/pr20527-1.c gcc.dg/torture/pr52028.c gcc.dg/torture/pr69174.c gcc.dg/vect/fast-math-pr43074.c gcc.dg/vect/pr60841.c gcc.dg/vect/pr66677.c (but seemingly no ICEs without that backout on check-c, so — no apparent issue in other comparators). Thus, qsort comparator checking would uncover this issue much earlier. Would the corresponding patch be welcome? (under CHECKING_P, check in O(n^2) time that all pairs in the vector are properly ordered after qsort, limiting n to 100 to avoid quadraticness hit)