[Fortran] PR 34482: RFC and patch: BOZ to real/complex conversion for some systems

[Jerry DeLisle]

Tobias Burnus wrote:
> Jerry DeLisle wrote:
>> Testing on ppc64 tonight after a rebuild.  When I updated I see we
>> have been promoted to mpfr-2.3.0 which I now have to install locally
>> since the distro does not have it out yet.
> Well, MPFR 2.3.0 is the new recommended version, but it should still be
> buildable with 2.2.1:
> 
> + dnl MPFR 2.2.1 is acceptable, but MPFR 2.3.0 is better.
> + #if MPFR_VERSION < MPFR_VERSION_NUM(2,2,1)
>  choke me
> 
> Thus it really should work with 2.2.1, though MPFR 2.3.0 is nicer as it
> supports, e.g., GAMMA and LOG_GAMMA simplification.
> 
> 
>> Jerry DeLisle wrote:
>>> Tobias Burnus wrote:
>>>> Jerry, can you check that the patch does the right thing on such
>>>> platforms and  real(z'00000000FFFFFFFF',kind=8) ?
>> Hmm, looking at the logs now I have 12575 failures.  I will revert the
>> patch and see if I can get a working system here.
> The patch itself should work on PPC; at least Dominique reported no
> problems and John wrote in the PR: "Tested second version posted and it
> fixes the fail on hppa-unknown-linux-gnu."
> 
> Tobias
> 
OK, I got ppc64 working again.

The patch regression tests OK.  However, with the following test program.

   print *, real(z'FFFFFFFF', 4)
   print *, transfer(4294967295_8, 0.0_4)
   print '(z12)', transfer(real(z'FFFFFFFF',4), 1_4)
   print *, transfer(2143289344_8, 0.0_4)
   print '(z12)', transfer(real(z'7FC00000',4), 1_4)
   print *, real(z'7FC00000',4)
   print *, real(z'00000000ffffffff',8)
   print *, real(z'FFFFFFFF00000000',8)
   print *, real(z'3333')
   END

On x86-64 I get:

$ gfc -fno-range-check nantest.f90
$ ./a.out
              NaN
              NaN
     7FC00000
              NaN
     7FC00000
              NaN
   2.12199579096527232E-314
                        NaN

and on ppc64-linux:

$ gfc -fno-range-check nantest.f90
$ ./a.out
              NaN
    0.0000000
     7FC00000
    0.0000000
     7FC00000
              NaN
   2.12199579096527232E-314
                        NaN
   1.83668190E-41

Is there something wrong with 'transfer' ?

Jerry