As noted by Tobias in the PR, the loop_vec_info constructor wasn't initializing the new partial_load_store_bias field.
Tested on aarch64-linux-gnu and pushed as obvious. Richard gcc/ PR middle-end/104026 * tree-vect-loop.c (_loop_vec_info::_loop_vec_info): Initialize partial_load_store_bias. --- gcc/tree-vect-loop.c | 1 + 1 file changed, 1 insertion(+) diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c index 1dbdb9e5fa0..f1410b03fb3 100644 --- a/gcc/tree-vect-loop.c +++ b/gcc/tree-vect-loop.c @@ -846,6 +846,7 @@ _loop_vec_info::_loop_vec_info (class loop *loop_in, vec_info_shared *shared) can_use_partial_vectors_p (param_vect_partial_vector_usage != 0), using_partial_vectors_p (false), epil_using_partial_vectors_p (false), + partial_load_store_bias (0), peeling_for_gaps (false), peeling_for_niter (false), no_data_dependencies (false), -- 2.25.1