I have been looking into making as.double() and friends do less copying (and they now do in R-devel). The problem is that people have been writing code which assumes copying is guaranteed, so we decided not to put any changes into 2.5.0. All but one of the cases I have detected (RandomFields, cluster::diana and stats::smooth, but there may be others that I have not) are due to the use of DUP=FALSE. Now, there is a pretty strong warning against this on the help page, but people still do it and some of you are relying on as.xxxx making a new object which the compiled code then alters. I think the message is clear: do not use DUP=FALSE unless you can prove that alteration of arguments in your C/Fortran code never has a side-effect. Watch out for what that side-effect might be: in one case it is to change the parsed code in the loaded namespace which affects subsequent uses of the function but the first example works correctly (and few test suites try functions repeatedly).
Brian D. Ripley, ripley at stats.ox.ac.uk Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/ University of Oxford, Tel: +44 1865 272861 (self) 1 South Parks Road, +44 1865 272866 (PA) Oxford OX1 3TG, UK Fax: +44 1865 272595