[R] noncommutative addition: NA+NaN != NaN+NA

[Prof Brian Ripley]

On 7 Sep 2004, Peter Dalgaard wrote:

> Robin Hankin <rksh at soc.soton.ac.uk> writes:
> 
> > Dear Professor Ripley
> > 
> > thank you for your reply.
> > 
> > >
> > >>  >  NaN +NA
> > >>  [1] NaN
> > >>  >  NA + NaN
> > >>  [1] NA
> > >>
> > >>  I thought "+" was commutative by definition.   What's going on?
> > >
> > >It is clearly not under your compiler/OS.  We could add a configure test
> > >for broken systems and fix it in arithmetic.c but it hardly seems
> > >worthwhile.
> 
> Didn't we do this discussion before? 

Certainly over NA vs 0+NA, as in this excerpt from reg-tests-1.R

## matching NAs on Solaris (MM 2002-08-02)
# x <- as.double(NA)
# identical(x + 0, x)
# stopifnot(match(x + 0, x, 0) == 1)
## match failed on Solaris with some compiler settings
## NA+0 is not guaranteed to be NA: could be NaN


> AFAIR, the thing is that IEEE
> specifies that NaN + whatever == whatever + NaN == NaN, but NaN is
> only specified a bit pattern in the first couple of bytes. R uses a
> special value in the lower bytes (1954 -- BTW, when *is* Ross'
> birthday?) to signal the NA, but we can't really expect that chip
> makers do what we hope they'd do with that part of the value. I think
> we resolved that specific checking for this issue would be too much of
> a performance killer, especially since R generally treats NaN as NA
> anyway. 

-- 
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