Re: [PATCH] ftoastr: new module, for lossless conversion of floats to short strings

[Bruno Haible]

Hi Paul,

Paul Eggert wrote:
> +2010-11-17  Paul Eggert  <address@hidden>
> +
> +     ftoastr: new module, for lossless conversion of floats to short strings
> +     * lib/ftoastr.h, lib/ftoastr.c, lib/dtoastr.c, lib/ldtoastr.c:
> +     * modules/ftoastr: New files.

Three comments on this:

* The code uses the snprintf() function but the module does not depend on the
  'snprintf' or 'snprintf-posix' module. Do you intend to ignore these
  portability problems?

  This function is missing on some platforms:
  IRIX 5.3, OSF/1 4.0, Solaris 2.5.1.

  This function overwrites memory even when a size argument <= 1 is passed on 
some
  platforms:
  Linux libc5, BeOS, OSF/1 5.1.

  printf of @samp{long double} numbers is unsupported on some platforms:
  mingw, BeOS.

  printf @code{"%g"} of Infinity and NaN yields an incorrect result on some 
platforms:
  AIX 5.2, OSF/1 5.1, Solaris 10, mingw.

  This function behaves incorrectly when a @samp{-} flag and a negative width
  are specified together, on some platforms:
  HP-UX 10.20.

  This function does not support precisions larger than 512 or 1024 in integer,
  floating-point and pointer output on some platforms:
  mingw, BeOS.

  This function can crash in out-of-memory conditions on some platforms:
  MacOS X 10.3, FreeBSD 6.0, NetBSD 3.0.

  This function does not truncate the result as specified in C99 on some 
platforms:
  mingw.

  This function does not return a byte count as specified in C99 on some 
platforms:
  HP-UX 11, IRIX 6.5, OSF/1 5.1, mingw.

* This module uses the function strtof(), strtod(), strtold(), which also have
  known portability problems. See
    doc/posix-functions/strtof.texi
    doc/posix-functions/strtod.texi
    doc/posix-functions/strtold.texi
  In particular, when the gnulib strtod() function is in use, which has known
  rounding errors, this will lead to wrong results of the function dtoastr(),
  no?

* The URL to the paper that you gave is a closed-content one: it requires
  payment. Fortunately, the author of the paper makes it also available online:
  <http://florian.loitsch.com/tmp/article.pdf>

  His algorithm depends on the existence of a precomputed table of powers of 10
  whose size depends on the exponent range. But the 'long double' exponent
  range on IA-64 is -16382..16383 (see the IA-64 Architecture Software
  Developer's Manual, section 5.1.1), and I can hardly imagine a reasonable
  programmer will spend 32766 * 16 bytes = 512 KB on _just_ a table of powers
  of 10, even in a shared library. Even for 'double', people may prefer to
  allocate a couple of more bits during the precision than to spend
  1022 * 8 bytes = 8 KB on a table.

Bruno