R devel - Jul 1999 - 0.65 HPUX/AIX update

The current situation is as follows.

AIX:

* I still need f77 for linking.  Using ld with the suggested flags gives
  a binary with `exec format error'.  The problem seems to be with an
  unreferenced __start.

* The floating point (finite|isnan) stuff is strange, but under control,
  I think.  Plain cc works, gcc has a problem and seems to need having
  prototypes turned off for IEEE fp.  (More below.)

* Apart from that, make check goes all the way through at least for me
  under both cc/f77/make and gcc/f77/make.

HPUX:

* make check fails in arith-true.R.  The problem is that

        is.infinite(.Machine$double.base ^ .Machine$double.max.exp)

gives FALSE.

Interestingly
> .Machine$double.base
[1] 2
> .Machine$double.max.exp
[1] 1024
> 2 ^ 1024
[1] 1.797693e+308
> 2 ^ 1025
[1] 1.797693e+308
> 2 ^ 2024
[1] 1.797693e+308

****

Here are some observations re IEEE fp.  I have the following small test
program:

#include <math.h>
#include <stdio.h>

main () {
    printf("1/0: %lf 0/0: %lf\n", 1./0., 0./0.);
    printf("finite(1/0): %d\n", finite(1./0.));
    printf("isnan(0/0): %d\n", isnan(0./0.));
    printf("log(0.): %lf\n", log(0.));
    return(0);
}

Under Linux, this gives:

1/0: Inf 0/0: NaN
finite(1/0): 0
isnan(0/0): -1
log(0.): -Inf

Under HPUX:

1/0: ++.000000 0/0: ?.000000
finite(1/0): 0
isnan(0/0): 1
log(0.):
-179769313486231571000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000.000000

This is also why log(0) is broken on HPUX.  See also the above why make
check fails.  However, the finite and isnan test is o.k.

Under AIX with cc:

1/0: INF 0/0: NaNQ
finite(1/0): 0
isnan(0/0): 1
log(0.): -INF

Under AIX with gcc:

1/0: INF 0/0: NaNQ
finite(1/0): 1
isnan(0/0): 1
log(0.): -INF

So there is a problem in the finite() test with gcc.  It somehow is
related to the use of ANSI C prototypes, and disappears if we define
_NO_PROTO.  I am in the process of adding this as a short term fix, but
we should think a bit more about that.

-k
-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-devel mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To:
r-devel-request@stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._

Kurt,

  thanks for the changes.  I've downloaded the snapshot on July 22.
It looks like we are in good shape now for AIX:

"Kurt" == Kurt Hornik <Kurt.Hornik@ci.tuwien.ac.at> writes:

Kurt> The current situation is as follows.
Kurt> 
Kurt> AIX:
Kurt> 
Kurt> * I still need f77 for linking.  Using ld with the suggested flags
gives
Kurt> a binary with `exec format error'.  The problem seems to be with an
Kurt> unreferenced __start.

I apologize for not seeing this earlier: "__start" is defined in
/lib/crt0.o.  So all you have to do is add /lib/crt0.o to the objects
when linking.  (You have to look carefully for that in the output from
gcc --verbose anything.o to learn what gcc is doing when linking an executable.)

Kurt> * The floating point (finite|isnan) stuff is strange, but under
control,
Kurt> I think.  Plain cc works, gcc has a problem and seems to need having
Kurt> prototypes turned off for IEEE fp.  (More below.)

Quick fix and insanity check:  Add the following lines at the end of Arith.h
  #include <fp.h>
  static int finite (double x) {
        return FINITE (x);
  }
This will generate lots of warnings.  The resulting executable
(compiled with gcc) works wrt Inf and NaN!  My current assumption is
that there is a bug in gcc that generates the wrong inline version of
finite.  Overwriting it solves the problem.  A better fix would be to
ensure that finite is never called when configure finds that it is
buggy.  (I believe that the issue whether finite () exists is a
separate one.  So there should be two flags, like HAVE_FINITE and
HAVE_FINITE_THAT_STINKS... )

Kurt> So there is a problem in the finite() test with gcc.  It somehow is
Kurt> related to the use of ANSI C prototypes, and disappears if we define
Kurt> _NO_PROTO.  I am in the process of adding this as a short term fix, but
Kurt> we should think a bit more about that.

  -tom

--
mailto:tov@ece.cmu.edu (Tom Vogels)   Tel: (412) 268-6638   FAX: -3204

-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-devel mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To:
r-devel-request@stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._