* The original code performs ( n ) normalizations
 * and ( n * ( n - 1 ) / 2 ) matches, which hide
 * the same number of normalizations.  The new code
 * performs the same number of normalizations ( n )
 * and ( n * ( n - 1 ) / 2 ) mem compares, far less
 * expensive than an entire match, if a match is
 * equivalent to a normalization and a mem compare ...
 *
 * This is far more memory expensive than the previous,
 * but it can heavily improve performances when big
 * chunks of data are added (typical example is a group
 * with thousands of DN-syntax members; on my system:
 * for members of 5-RDN DNs,

 members         orig            bvmatch (dirty) new
 1000            0m38.456s       0m0.553s        0m0.608s
 2000            2m33.341s       0m0.851s        0m1.003s

 * Moreover, 100 groups with 10000 members each were
 * added in 37m27.933s (an analogous LDIF file was
 * loaded into Active Directory in 38m28.682s, BTW).
 *
 * Maybe we could switch to the new algorithm when
 * the number of values overcomes a given threshold?
 */