[Bug libfortran/48488] New: Wrong default format for real numbers

[Bug libfortran/48488] New: Wrong default format for real numbers

[thenlich at users dot sourceforge.net]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48488

           Summary: Wrong default format for real numbers
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: libfortran
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: thenlich@users.sourceforge.net


In write.c the intended default format for real numbers is documented as:

/* Output a real number with default format.
   This is 1PG14.7E2 for REAL(4), 1PG23.15E3 for REAL(8),
   1PG28.19E4 for REAL(10) and 1PG43.34E4 for REAL(16).  */
// FX -- FIXME: should we change the default format for __float128-real(16)?

This is reasonable, since it reflects the rounded-down number of decimal
significant digits for each format: 7, 15, 19, 34. Thus any number with less
decimal digits than the maximum precision always retains its original decimal
value which is a useful feature.

But it is implemented incorrectly as the following example shows:

print "(1PG14.7E2)", .3_4 ! 0.3000000
print *,             .3_4 ! 0.30000001 expected 0.3000000

print "(1PG23.15E3)", .3_8 ! 0.300000000000000
print *,              .3_8 ! 0.29999999999999999 expected 0.300000000000000

print "(1PG28.19E4)", .3_10 ! 0.3000000000000000000
print *,              .3_10 ! 0.30000000000000000001 expected (s.a.)

print "(1PG43.34E4)", .3_16 ! 0.300000000000000000000000000000000
print *,              .3_16 ! 0.29999999999999999999999999999999999 exp. (s.a)

[Bug libfortran/48488] Wrong default format for real numbers

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48488

--- Comment #1 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2011-04-07 09:08:01 UTC ---
> In write.c the intended default format for real numbers is documented as:
>
> /* Output a real number with default format.
>    This is 1PG14.7E2 for REAL(4), 1PG23.15E3 for REAL(8),
>    1PG28.19E4 for REAL(10) and 1PG43.34E4 for REAL(16).  */
> // FX -- FIXME: should we change the default format for __float128-real(16)?
>
> This is reasonable, since it reflects the rounded-down number of decimal
> significant digits for each format: 7, 15, 19, 34. Thus any number with less
> decimal digits than the maximum precision always retains its original decimal
> value which is a useful feature.

I think the error is in the documentation: the actual formats are 1PG16.8E2 for
REAL(4), 1PG26.17E3 for REAL(8), 1PG30.20E4 for REAL(10) and 1PG45.35E4 for
REAL(16) as shown by the following modified test

print "(1PG16.8E2)", .3_4 
print *,             .3_4 

print "(1PG26.17E3)", .3_8 
print *,              .3_8 

print "(1PG30.20E4)", .3_10 
print *,              .3_10
print *, x10

print "(1PG45.35E4)", .3_16
print *,              .3_16
end

that gives

  0.30000001    
  0.30000001    
  0.29999999999999999     
  0.29999999999999999     
  0.30000000000000000001      
  0.30000000000000000001      
  0.29999999999999999999999999999999999      
  0.29999999999999999999999999999999999      

The values 8, 17, 20, and 35 (?see FIXME) are chosen such that reading the
default output will always returns the original value up to the last bit. There
is a test in the test suite checking that, but AFAIK not for all supported
reals, in particular I don't think this has been tested for REAL(16).

[Bug libfortran/48488] Wrong default format for real numbers

[thenlich at users dot sourceforge.net]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48488

--- Comment #2 from Thomas Henlich <thenlich at users dot sourceforge.net> 2011-04-07 11:22:49 UTC ---
Ok, now I found that this was changed in
http://gcc.gnu.org/viewcvs?view=revision&revision=128967

The testcase says:

> ! This tests that the default formats for formatted I/O of reals are
> ! wide enough and have enough precision, by checking that values can
> ! be written and read back.

However, the selected width in the current version (8, 17, 20, 35) is not big
enough to guarantee that. See IEEE 754/2008:

For the purposes of discussing the limits on correctly rounded conversion,
define the following quantities:
...
―   for binary32, Pmin(binary32) = 9
―   for binary64, Pmin(binary64) = 17
―   for binary128, Pmin(binary128) = 36
―   for all other binary formats bf, Pmin(bf) = 1 + ceiling(p×log10(2)), where
p is the number of significant bits in bf
...
Conversions from a supported binary format bf to an external character sequence
and back again results in a copy of the original number so long as there are at
least Pmin(bf) significant digits specified and the rounding-direction
attributes in effect during the two conversions are round to nearest
rounding-direction attributes.

I see two possibilities:

A. We aim to make the default format useful for human readers, so that e.g. 0.3
is always represented as 0.3000...
In that case we must choose P(bf) = floor(p×log10(2))

P = {7, 15, 19, 34}

B. We aim to make the default format useful for round-trip conversion, so that
a binary value is always converted to the same binary value.
In that case we must choose P(bf) = 1 + ceiling(p×log10(2))

P = {9, 17, 21, 36}

I personally prefer A.

Currently the values are sort of in the middle (except for real64).

[Bug libfortran/48488] Wrong default format for real numbers

[jvdelisle at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48488

Jerry DeLisle <jvdelisle at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2011.04.08 23:54:07
         AssignedTo|unassigned at gcc dot       |jvdelisle at gcc dot
                   |gnu.org                     |gnu.org
     Ever Confirmed|0                           |1

--- Comment #3 from Jerry DeLisle <jvdelisle at gcc dot gnu.org> 2011-04-08 23:54:07 UTC ---
We have tried for b) up to now.  A patch is simple enough.

[Bug libfortran/48488] Wrong default format for real numbers

[jvdelisle at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48488

--- Comment #4 from Jerry DeLisle <jvdelisle at gcc dot gnu.org> 2011-04-29 14:56:06 UTC ---
Author: jvdelisle
Date: Fri Apr 29 14:56:02 2011
New Revision: 173166

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=173166
Log:
2011-04-29  Jerry DeLisle  <jvdelisle@gcc.gnu.org>
        Janne Blomqvist  <jb@gcc.gnu.org>

    PR libgfortran/48488
    PR libgfortran/48602
    PR libgfortran/48615
    PR libgfortran/48684
    PR libgfortran/48787
    * io/write.c (write_d, write_e, write_f, write_en,
    write_es): Add precision compemsation parameter to call.
    (set_fnode_default): Adjust default widths to assure
    round trip on write and read. (write_real): Adjust call to write_float.
    (write_real_g0): Calculate compensation for extra precision and adjust
    call to write_float. 
    * io/write_float.def (output_float_FMT_G_): Use volatile rather than
    asm volatile to avoid optimization issue. Correctly calculate the
    number of blanks (nb) to be appended and simplify calculation logic.
    (write_float): Increase MIN_FIELD_WIDTH by one to accomodate the new
    default widths. Eliminate the code that attempted to reduce the
    the precision used in later sprintf functions.  Add call parameter to
    compensate for extra precision.

Modified:
    trunk/libgfortran/ChangeLog
    trunk/libgfortran/io/write.c
    trunk/libgfortran/io/write_float.def

[Bug libfortran/48488] Wrong default format for real numbers

[jvdelisle at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48488

--- Comment #5 from Jerry DeLisle <jvdelisle at gcc dot gnu.org> 2011-04-29 15:09:02 UTC ---
Author: jvdelisle
Date: Fri Apr 29 15:08:57 2011
New Revision: 173168

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=173168
Log:
2011-04-29  Jerry DeLisle  <jvdelisle@gcc.gnu.org>

    PR libgfortran/48488
    PR libgfortran/48602
    PR libgfortran/48615
    PR libgfortran/48684
    PR libgfortran/48787
    * gfortran.dg/fmt_g.f: Adjust test.
    * gfortran.dg/fmt_g0_1.f08: Adjust test.
    * gfortran.dg/round_3.f08: New test.
    * gfortran.dg/namelist_print_1.f: Adjust test.
    * gfortran.dg/char4_iunit_1.f03: Adjust test.
    * gfortran.dg/f2003_io_5.f03: Adjust test.
    * gfortran.dg/coarray_15.f90: Adjust test.
    * gfortran.dg/namelist_65.f90: Adjust test.
    * gfortran.dg/fmt_cache_1.f: Adjust test.
    * gfortran.dg/char4_iunit_2.f03: Adjust test.
    * gfortran.dg/real_const_3.f90: Adjust test.

Added:
    trunk/gcc/testsuite/gfortran.dg/round_3.f08
Modified:
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/testsuite/gfortran.dg/char4_iunit_1.f03
    trunk/gcc/testsuite/gfortran.dg/char4_iunit_2.f03
    trunk/gcc/testsuite/gfortran.dg/coarray_15.f90
    trunk/gcc/testsuite/gfortran.dg/f2003_io_5.f03
    trunk/gcc/testsuite/gfortran.dg/fmt_cache_1.f
    trunk/gcc/testsuite/gfortran.dg/fmt_g.f
    trunk/gcc/testsuite/gfortran.dg/fmt_g0_1.f08
    trunk/gcc/testsuite/gfortran.dg/namelist_65.f90
    trunk/gcc/testsuite/gfortran.dg/namelist_print_1.f
    trunk/gcc/testsuite/gfortran.dg/real_const_3.f90

[Bug libfortran/48488] Wrong default format for real numbers

[jb at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48488

--- Comment #6 from Janne Blomqvist <jb at gcc dot gnu.org> 2011-04-29 15:58:23 UTC ---
Author: jb
Date: Fri Apr 29 15:58:20 2011
New Revision: 173169

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=173169
Log:
PR 48488 Fix comments

Modified:
    trunk/libgfortran/ChangeLog
    trunk/libgfortran/io/write.c

[Bug libfortran/48488] Wrong default format for real numbers

[jb at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48488

--- Comment #8 from Janne Blomqvist <jb at gcc dot gnu.org> 2011-04-29 16:06:55 UTC ---
Author: jb
Date: Fri Apr 29 16:06:48 2011
New Revision: 173171

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=173171
Log:
PR 48488 Typo

Modified:
    trunk/libgfortran/io/write.c

[Bug libfortran/48488] Wrong default format for real numbers

[jb at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48488

--- Comment #7 from Janne Blomqvist <jb at gcc dot gnu.org> 2011-04-29 16:03:57 UTC ---
Author: jb
Date: Fri Apr 29 16:03:54 2011
New Revision: 173170

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=173170
Log:
PR 48488 Fix comments

Modified:
    trunk/libgfortran/io/write.c

[Bug libfortran/48488] Wrong default format for real numbers

[thenlich at users dot sourceforge.net]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48488

Thomas Henlich <thenlich at users dot sourceforge.net> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED

--- Comment #9 from Thomas Henlich <thenlich at users dot sourceforge.net> 2011-05-02 11:24:13 UTC ---
Ok, the precision is now consistently the smallest value that ensures
round-trip conversion for all values.

It may not always look nice, but it fulfills the Fortran standard ("reasonable
processor-dependent values of w, d, and e are used") with "reasonable" being
defined by Pmin(bf) in IEEE 754/2008.

It's the best we can do with this, so I'm closing this one.