Updated Windows toolchain

Hello,

On Mon, Jan 9, 2012 at 12:28 AM, Prof Brian Ripley
<ripley at stats.ox.ac.uk>  wrote:

CRAN Windows binary packages built for R-devel are now online, and Uwe's
winbuilder has gained the ability to check source packages under R-devel.

Windows check results are available from
http://cran.r-project.org/bin/windows/contrib/checkSummaryWin.html
and in due course from the main CRAN check page.

There have been a few updates to the toolchain:

(i) It is now based on a beta of gcc 4.6.3, and so reports almost the same
compilation warnings/errors as the CRAN check machines.

Is the binary R-devel provided by CRAN built against this toolchain?
If it is, should I expect "R --arch x64 CMD config CC" to report
"gcc"? It still reports "x86_64-w64-mingw32-gcc" (R-devel 58077).

Also, I have the latest Rtools installed (VERSION.txt reads "Rtools
version 2.15.0.1911") and it is first on my PATH.

I expect "gcc --version" to report 4.6.3 but it still says 4.5.0.

"x86_64-w64-mingw32-gcc --version" reports 4.5.2.
"which gcc" reports "/cygdrive/c/Rtools215/MinGW/bin/gcc".

"which x86_64-w64-mingw32-gcc" reports
"/cygdrive/c/Rtools215/MinGW64/bin/x86_64-w64-mingw32-gcc".

(ii) There are various bug-fixes to the toolchain: notably x^n and exp(x)
use gradual underflow to denormal numbers rather than abrubtly underflowing
to zero.

(iii) This is a 'multilib' toolchain: the compiler is named 'gcc.exe' for
both architectures, selected by flag -m32 (the default) and -m64.

Looks like to check gcc versions I should (instead of what I do above)
do simply "gcc --version". I imagine that "gcc -m32 --version" would
report the same thing as "gcc -m64 --version".

Thanks,
Dan

On 29/11/2011 07:56, Prof Brian Ripley wrote:

An updated toolchain is now being used for Windows' builds of R-devel:
details are in the R-admin manual and at
http://www.murdoch-sutherland.com/Rtools/ and
http://www.stats.ox.ac.uk/pub/Rtools/

Both 32- and 64-bit parts of the toolchain use v2.0.1 of the Mingw-w64
project's runtime and a beta of gcc 4.5.4: the Mingw.org project's
builds are no longer used. This should mean that code which compiles for
64-bit Windows also compiles for 32-bit Windows, and v.v. unless code
makes (incorrect but common) assumptions that pointers fit into longs.

A very few packages will need modifications because they contain
declarations which clash with the headers in this toolchain: where we
are aware of problems the maintainers have been informed.

At DLL level different Windows' toolchains should be compatible: at C
level they mostly are but at C++ level they are pretty much incompatible
(so that for example GDAL has to be re-compiled for every toolchain: and
Rcpp users need to be careful to use only one toolchain for Rcpp and
their packages). All the external software previously made available
(and more) is made available at http://www.stats.ox.ac.uk/pub/Rtools .

The toolchain has support for OpenMP and pthreads: however OpenMP
support is not enabled by default in R (it is too slow to be much use).
If you do make use of it in your packages, be aware that you will need
to ship the appropriate pthreads DLL(s).

It is expected that there will be several further minor updates prior to
the release of 2.15.0 in ca 4 months, but this step is the major one.

--
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