+The MD5 hashing is not satisfactory. It is computed for a file only if
+the said file has to be read fully for a comparison (i.e. two files
+match and we have to read them completely).
+
+Hence, in practice lot of partial MD5s are computed, which costs a lot
+of cpu and is useless. This often hurts more than it helps. The only
+case when it should really be useful is when you have plenty of
+different files of same size, and lot of similar ones, which does not
+happen often.
+
+Forcing the files to be read fully so that the MD5s are properly
+computed is not okay neither, since it would fully read certain files,
+even if we will never need their MD5s.
+
+Anyway, it has to be compiled in with 'make WITH_MD5=yes', and even in
+that case it will be off by default
+