When using L1 smoothing, the Schwarz preconditioner creates a new, modified diagonal matrix to give to the local solver. In the current implementation, it ends up calling spgeam when creating this modified matrix. For some backends, spgeam requires the diagonal matrix to be sorted, so if it's not, the Schwarz preconditioner generation will fail "silently" (no errors or warnings, just incorrect results).
Without L1, the only sorting requirements would come from the local solvers, which will apply it if necessary (e.g., using ILU in the local solver).