support of 128 bit long real on linux for ia32

[Winfrid Tschiedel <winfrid dot tschiedel at hpc dot fujitsu-siemens dot com>]

Hello,

I have a small question concerning the support of
long double on linux ia32 with gcc v3.x

I am currently using gcc-3.0.1 on redhat 7.2, kernel 2.4.7-10smp

I tried this small program :
#include <stdio.h>

/*
  icl -DTYP=float precision.c
  icl -DTYP="long double" -Qlong_double -Qpc80 precision.c
*/

int same(TYP* t1, TYP* t2) { return *t1 == *t2; }
int same2(TYP* t1, TYP* t2) { return *t1 == *t2 / *t1; }

main()
{
   int mbits = 0, ebits = 1;
   TYP t = 1.0, delta = 0.5, tt = t + delta;
   while ( !same( &t, &tt)) {
      delta *= 0.5;
      tt = t + delta;
      mbits++;
   }
   printf( "mantissa precision: 1 + %3d = %3d bits\n", mbits, mbits+1);

   t = 2.0; tt = t * t;
   while ( same2( &t, &tt)) {
      t *= t;
      tt = t * t;
      ebits++;
   }
   printf( "exponent precision: 1 + %3d = %3d bits\n", ebits, ebits+1);

   printf( "-----------------------------------------------\n"
           "total significant bits:       %3d "
           "(out of %3d)\n", 2 + mbits + ebits, sizeof( TYP) * 8);
}

and got the following results :
gcc -DTYP="long double" -m128bit-long-double precision.c -o prec_128
./prec_128
mantissa precision: 1 +  63 =  64 bits
exponent precision: 1 +  14 =  15 bits
-----------------------------------------------
total significant bits:        79 (out of 128)


gcc -DTYP="long double" -m96bit-long-double precision.c -o prec_96
./prec_96
mantissa precision: 1 +  63 =  64 bits
exponent precision: 1 +  14 =  15 bits
-----------------------------------------------
total significant bits:        79 (out of  96)

Now my question, is it true, that the precision for both compilations
is the same, except that for -m128bit-long-double the result of
sizeof(long double) is 16, while in the other case sizeof(long double) is 12?

Do plan to enhance your compiler in order to get 
"long double" in C compatible with real(16) in fortran ?

      program prec
      real (kind=kind(1.e0_16)) , parameter :: one = 1.0
      write ( unit=*,fmt=* ) huge(one)
      write ( unit=*,fmt=* ) tiny(one)
      write ( unit=*,fmt=* ) maxexponent(one)
      write ( unit=*,fmt=* ) minexponent(one)
      write ( unit=*,fmt=* ) precision(one)
      write ( unit=*,fmt=* ) epsilon(one)
      write ( unit=*,fmt="(z32)" ) one+epsilon(one)
      end
ifc precision.F
ifc: NOTE: The evaluation period for this product ends on 15-may-2002 UTC.
   program PREC

11 Lines Compiled
winfrid@dino csh 23: ./a.out
   1.189731495357231765085759326628007Q+4932
   3.362103143112093506262677817321753Q-4932
       16384
      -16381
          33
   1.925929944387235853055977942584928Q-0034
3FFF0000000000000000000000000001

Many thanks in advance for your information.

Take care,

Winfrid


-- 
_______________________________________________________________________________

Fujitsu Siemens Computers
OEA VC HPC                   Email: winfrid.tschiedel@hpc.fujitsu-siemens.com
Otto-Hahn-Ring 6             Tel. : ++49-89-636-45652
81739 Muenchen               Fax  : ++49-89-636-44088 

MS Exchange : winfrid.tschiedel@fujitsu-siemens.com