R devel - Aug 2004 - Installing R on DEC Alpha - problems with dynamic loading

I have recently installed R on DEC Alpha OSF 5.1.

The first regression test (reg-test-1.R) failed with an unresolved symbol in
lapack.so:

2246839:/data/h1/bifdev/apps/wga/R-1.9.1/bin/R.bin: /sbin/loader: Fatal
Error: call to unresolved symbol from
/data/h1/bifdev/apps/wga/R-1.9.1/modules/lapack.so (pc=0x3ffbfe22b60)

Further investigation has revealed that there is NO problem with lapack.so.
The problem seems to lie in the use of 'unloadNamespace' earlier in the
script.

On our machine the following line:

	unloadNamespace(splines)

seem to really confuse the dynamic linking/symbol resolution. I have a very
small example to demonstrate:

	data(iris)
	library(splines)
	unloadNamespace(splines)
	data(iris)

works fine, but:

	library(splines)
	unloadNamespace(splines)
	data(iris)

falls over with:

Error in dyn.load(x, as.logical(local), as.logical(now)) : 
        unable to load shared library
"/data/h1/bifdev/apps/wga/R-1.9.1/library/tools/libs/tools.so":
  dlopen: /data/h1/bifdev/apps/wga/R-1.9.1/library/stats/libs/stats.so:
symbol "bdrsplerr_" unresolved
Execution halted

The unloadNamespace has effectively unloaded a symbol from a completely
different shared object.

Can anybody shed any light on this problem?

Thanks.
-----
Mike Firth
Senior Technical Specialist
AstraZeneca Pharmaceuticals
Mereside
Alderley Park
Macclesfield
Cheshire
SK10 3LT

Tel: 01625 513919

The line in the script as distributed is

   unloadNamespace("splines")

That is not the same thing at all.  Please confirm that you actually have
a problem with that form.  (Your examples should not work, and do not work
for me.)  If you do, it is likely that your system's dlclose is
non-standard/broken.  However, unloadNamespace is not normally used so it
may be that this does not impact your users.

Please read the posting guide: this is not an appropriate question for 
R-help so I have diverted it to R-devel.


On Mon, 23 Aug 2004, Firth, Mike A (ALDP) wrote:
> I have recently installed R on DEC Alpha OSF 5.1.
> 
> The first regression test (reg-test-1.R) failed with an unresolved symbol
in
> lapack.so:
> 
> 2246839:/data/h1/bifdev/apps/wga/R-1.9.1/bin/R.bin: /sbin/loader: Fatal
> Error: call to unresolved symbol from
> /data/h1/bifdev/apps/wga/R-1.9.1/modules/lapack.so (pc=0x3ffbfe22b60)
> 
> Further investigation has revealed that there is NO problem with lapack.so.
> The problem seems to lie in the use of 'unloadNamespace' earlier in
the
> script.
> 
> On our machine the following line:
> 
> 	unloadNamespace(splines)
> 
> seem to really confuse the dynamic linking/symbol resolution. I have a very
> small example to demonstrate:
> 
> 	data(iris)
> 	library(splines)
> 	unloadNamespace(splines)
> 	data(iris)
> 
> works fine, but:
> 
> 	library(splines)
> 	unloadNamespace(splines)
> 	data(iris)
> 
> falls over with:
> 
> Error in dyn.load(x, as.logical(local), as.logical(now)) : 
>         unable to load shared library
> "/data/h1/bifdev/apps/wga/R-1.9.1/library/tools/libs/tools.so":
>   dlopen: /data/h1/bifdev/apps/wga/R-1.9.1/library/stats/libs/stats.so:
> symbol "bdrsplerr_" unresolved
> Execution halted
> 
> The unloadNamespace has effectively unloaded a symbol from a completely
> different shared object.
> 
> Can anybody shed any light on this problem?
-- 
Brian D. Ripley,                  ripley@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