The theory underlying the Car-Parrinello extended-lagrangian approach to {\em ab initio} molecular dynamics (CPMD) is reviewed and reexamined using 'heavy' ice as a test system. It is emphasized that the adiabatic decoupling in CPMD is not a decoupling of electronic orbitals from the ions but only a decoupling of a subset of the orbital vibrational modes from the rest of the necessarily-coupled system of orbitals and ions. Recent work (J. Chem. Phys. {\bf 116}, 14 (2002)) has pointed out that, due to the orbital-ion coupling that remains once adiabatic-decoupling has been achieved, a large value of the fictitious mass \(\mu\) can lead to systematic errors in the computed forces in CPMD. These errors are further investigated in the present work with a focus on those parts of these errors that are not corrected simply by rescaling the masses of the ions. It is suggested that any comparison of the efficiencies of Born-Oppenheimer molecular dynamics (BOMD) and CPMD should be performed at a similar level of accuracy. If accuracy is judged according to the average magnitude of the systematic errors in the computed forces, the efficiency of BOMD compares more favorably to that of CPMD than previous comparisons have suggested.