computing long double NaN on Irix 6.5

[Eric Blake]

I've done some investigation into failures of test-frexpl, test-isnanl, and 
test-vasnprintf-posix on Irix 6.5.

The Irix 6.5 cc, on 0.0L/0.0L, produces:

(gdb) p sizeof x
$1 = 16
(gdb) x/4x &x
0x7ffb7ea0:     0x7ff7ffff      0xffffffff      0x00000000      0x00000000
(gdb) p x
$2 = nan(0x7ffffffffffff)

but gcc 3.4.3 on the same platform produces:

(gdb) x/4x &x
0x7ffb7e80:     0x7ff00000      0x00000000      0x00000000      0x00000000
(gdb) p x
$1 = 0

The following program discovers a way to coerce long double NaN even with gcc:

-------
#include <stdio.h>
#include <math.h>
int main()
{
  double w = 0.0 / 0.0;
  long double ww = w;
  long double x = 0.0L / 0.0L;
  long double zero = 0.0L;
  long double y = zero / zero;
  long double inf = 1.0L / 0.0L;
  long double z = inf - inf;
  printf("%d %d %d %d %d\n", isnan (w), isnan (ww), 
         isnan (x), isnan (y), isnan (z));
  return 0;
}
------
$ cc -o foo foo.c
$ ./foo
1 1 1 1 1
$ gcc -o foo foo.c
$ ./foo
1 1 0 0 0

Although both compilers made ww a NaN, gcc still doesn't come up with bit-
identical results.  Oh well.

cc:
(gdb) x/4x &ww
0x7ffb7e40:     0x7ff7ffff      0xffffffff      0x00000000      0x00000000

gcc:
(gdb) x/4x &ww
0x7ffb7e30:     0x7fffffff      0xfffffffc      0x00000000      0x00000000
(gdb) p ww
$2 = nan(0xffffffffffffc)


So, I am working on a patch that looks for all instances of 0.0L / 0.0L in the 
testsuite, and replacing them with a call to NaNl() from tests/nan.h; at the 
same time, beefing up that macro to use type conversion from double to long 
double on Irix as the only known way to guarantee a NaN.

-- 
Eric Blake