R devel - Jan 2011 - Warning: you may need to use R-patched with recent R distros

Two things have emerged in testing on x86_64 Fedora 14 which mean that 
a recent R-patched is probably needed.

1) That OS uses zlib 1.2.5: that claims to be binary-compatible with 
zlib 1.2.3 but is not, as we found (painfully) on Windows.  The remedy 
was to remap _all_ the symbols in R's own copy of zlib (not just those 
zlib arranged to remap).

The symptoms were crashes using packages XML and rgoobi (both of which 
link to zlib) and incorrect results in RJaCGH (which contains a copy 
of zlib).  There may well be other problems ....

2)  Fedora 14 uses gcc 4.5.1. With CFLAGS containing the default -O2 
or higher, HAVE_C99_COMPLEX was detected as false because there is a 
(genuine) incompatibility between types Rcomplex and C99's double 
complex.  This means that R's fallback code is used, and regretably 
that contains a serious bug in an 'optimization' by a colleague, so 
z^n is incorrect for most complex z and integer n (and has been since 
2.10.0).  The remedy is to use R-patched or R-devel, or only optimize 
to -O.

We've also seen incorrect results from package mvtnorm when C 
optimization was -O3.

The upshot is that there is likely to be a 2.12.2 to fix these issues.

-- 
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

For the complex-numbers bug, do you know a reliable way (besides looking
at version numbers) to determine whether the bug is present or absent in a
given build?  

I don't know what version of gcc was used in my build nor the optimization
flags, so I did a few test exponentiations z^n and the results look okay,
but maybe I'm not tickling the right bits.


--
Ken Williams
Senior Research Scientist
Thomson Reuters
Phone: 651-848-7712
ken.williams at thomsonreuters.com
http://labs.thomsonreuters.com





On 1/31/11 1:48 PM, "Prof Brian Ripley" <ripley at
stats.ox.ac.uk> wrote:
>Two things have emerged in testing on x86_64 Fedora 14 which mean that
>a recent R-patched is probably needed.
>
>1) That OS uses zlib 1.2.5: that claims to be binary-compatible with
>zlib 1.2.3 but is not, as we found (painfully) on Windows.  The remedy
>was to remap _all_ the symbols in R's own copy of zlib (not just those
>zlib arranged to remap).
>
>The symptoms were crashes using packages XML and rgoobi (both of which
>link to zlib) and incorrect results in RJaCGH (which contains a copy
>of zlib).  There may well be other problems ....
>
>2)  Fedora 14 uses gcc 4.5.1. With CFLAGS containing the default -O2
>or higher, HAVE_C99_COMPLEX was detected as false because there is a
>(genuine) incompatibility between types Rcomplex and C99's double
>complex.  This means that R's fallback code is used, and regretably
>that contains a serious bug in an 'optimization' by a colleague, so
>z^n is incorrect for most complex z and integer n (and has been since
>2.10.0).  The remedy is to use R-patched or R-devel, or only optimize
>to -O.
>
>We've also seen incorrect results from package mvtnorm when C
>optimization was -O3.
>
>The upshot is that there is likely to be a 2.12.2 to fix these issues.
>
>-- 
>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
>
>______________________________________________
>R-devel at r-project.org mailing list
>https://stat.ethz.ch/mailman/listinfo/r-devel
>