https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106662
--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> --- As I wrote on omp-lang, I believe this is UB. simd doesn't accept firstprivate clause, but does accept lastprivate. So, each SIMD lane's private j starts uninitialized (or for C++ default constructed but for int that is the same thing) and the value from the SIMD lane corresponding to last iteration is copied over to the private copy of j in the worksharing-loop. With simd, one has basically these data-sharing possibilities: 1) no data sharing clause, then it works kind like "shared", all SIMD lanes read the same variable, they'd better not to store it as that would be a data race 2) private - each SIMD lane has its private copy of the var, uninitialized/default-constructed at the start, destroyed at the end 3) lastprivate - likewise, but before it is destroyed, the SIMD lane corresponding to last iteration copies to original var 4) linear - must be integral, each SIMD lane gets the private var initialized with the original list value + step * SIMD_lane_# and the last iteration effect like lastprivate
