[R] R 1.9.0 on AIX, 64-bit

[Tim Hoar]

Here is a followup to Andy's post -- I have a working 64-bit R on AIX ...

My institution is running AIX 5.1.

Several things are needed to build R 1.9.0 and use dynamic loading of .so
objects.

*) setenv OBJECT_MODE 64       or the equivalent.

*) There is a quirk (bug?) in the AIX compilers in 64bit mode that requires
   changing configure:line 36555      to initialize a variable to zero.
   Line 36555 should be:  time_t ct=0;

   without the change -- configure hits a test for POSIX times and hangs ...

   This can also be overcome by using -qinitauto=00, but it is
   a performance hit ... so it is probably undesirable in the long run.

   Note that in 32bit mode, this is not needed.

*) The dynamic linking options specified in the R Installation and
   Administration Guide: Section B.7.8 did not work for me --
   the ones that did work are:

OBJECT_MODE=64
LDFLAGS='-brtl'
CFLAGS='-O -qstrict'
FFLAGS='-O -qstrict'
CXXFLAGS='-O -qstrict'
CC=/usr/bin/xlc_r
F77=/usr/bin/xlf_r
CXX=/usr/bin/xlC_r

   furthermore, I was able to configure with the X11 libraries (and did not
   have to supply the "--without-blas" option):

   ./configure --prefix='blah_blah_blah'

   it is nice to be able to plot!
   Our IBM rep indicated that with LDFLAGS set as above, "configure" is
   generally smart enough to figure out everything else -- kudos to the R
   configure team. I have not tried a higher level of optimization.

*) Our system does not have a 32,64-bit /lib/crt0.o  -- ours is only 32bit.
   Our 64bit one is /lib/crt0_64.o   so after configure generates a
   Makeconf file, it must be edited to use /lib/crt0_64.o  instead of crt0.o

*) (credit to Andy Pierce)

> Date: Mon, 10 May 2004 13:31:22 -0400
> From: Andy Pierce <apierce at stny.rr.com>
> To: Tim Hoar <thoar at ucar.edu>
> Subject: Re: [R] R 1.9.0 on AIX, 64-bit
>
> I worked the problem down to where there seems to be something wrong
> with mktime in the way that R uses it.... I modified src/main/datetime.c
> to add the lines for _AIX which you see below...  oddly, if you
> extract out the mktime code from here and stuff it into a standalone
> routine, it seems to work fine. It's all in the handling of dates from
> before the epoch (1970).
>
> static Rboolean have_broken_mktime(void)
> {
> #ifdef _AIX
>      return TRUE;
> #else
> #ifdef Win32
>
>      return TRUE;
>
> (the rest of the existing function....
>
> #endif
> }

*) Our system is configured such that it is guaranteed to fail the
   "running tests of Internet and socket function" tests, so I have
   not run any of the tests after that -- yet.

*) (again - credit to Andy Pierce)
   When I tried to build the SparseM package -- there was an unresolved
   external -- "etime"

   Adding a file "etime.c" to the SparseM/src  with the following:

#include <sys/time.h>
#include <sys/resource.h>
float etime(tt)
float tt[2];
{
   int who;
   struct rusage used;
   who = 0;
   getrusage(who,&used);
/*   tt[0] = used.ru_exutime.tv_sec+((used.ru_exutime.tv_usec)/1000000.); */
   tt[0] = used.ru_utime.tv_sec+((used.ru_utime.tv_usec)/1000000.);
   tt[1] = used.ru_stime.tv_sec+((used.ru_stime.tv_usec)/1000000.);
   return(tt[0]+tt[1]);
}

I did not find any problems in the config.log or the build output that
indicates etime is a problem anywhere else.

-=-=-=-=-=-=-=-=-=-=-

After all the above -- I can dyn.load(), generate figures, etc.
Hope this helps someone ...

Tim

<Andy's post from Mon, 03 May 2004 08:27:47 -0400  deleted>


## Tim Hoar, Associate Scientist              email: thoar at ucar.edu     ##
## Geophysical Statistics Project             phone: 303-497-1708       ##
## National Center for Atmospheric Research   FAX  : 303-497-1333       ##
## Boulder, CO  80307                    http://www.cgd.ucar.edu/~thoar ##