> It does not need to be a function of the data because it is never
> again checked against the data -- it is simply used as a unique identifier.

But it should be able to be regenerated from that data. In other words,
for any given file, there should be only *one* unique identifier.

> > A random number generator destroys the important property that
> > identical source run through identical toolchain bits (which should
> > include prebinding) produce identical files.  Violating this property is
> > a major PITA, especially for testing...
> 
> You do have a point there. OTOH, timestamps etc. in the ELF files
> already violate this, as you noted.

No.  I did not note that, because it's not true.  There are no
timestamps in our ELF files, except where someone took explicit action
to put them there (eg, by actually generating source on the fly that
contains a timestamp, like in kernels). 

For example, I just compiled bin/cat from the 1.6 sources. I download
bin/cat from ftp.netbsd.org and it is identical to mine. There are a
zillion small ways in which this is useful. Don't screw it up.