[Rd] O2 optimization produces wrong code (PR#5315)

[jean.coursol at math.u-psud.fr]

Full_Name: jean coursol
Version: 1.7.1, 1.8.0
OS: linux & Windows-XP
Submission from: (NULL) (129.175.52.7)


Binary MS-Windows akima module from CRAN (1.8.0 version) produces wrong results
with some data.

Installing akima source in linux, with same data: 
-with gcc-2.95.3 -O2 : give correct results (under R 1.7.1);
-with gcc-3.2.3  -O2 : give  wrong results (under R-1.7.1 and R-1.8.0);
-with gcc-3.2.3 (no optimization) give correct results (under R-1.7.1 and
R-1.8.0);

Installing akima module made by RCrossBuild (with gcc-2.95.3, R-1.7.1) in
Windows-XP 
and R-1.8.0 gives correct results (it is possible to use cross-compiled akima
module 
using R-1.7.1, because there is no call from akima to R...).

It would be better to use RCrossBuild and R-1.8.0, but actually, RCrossBuild is
not running with gcc-3.2.3 and R-1.8.0...