sisyphus / math-nv Goto Github PK

Perl fails to assign many floating point values correctly ... and by "many", I mean millions.
(Perhaps there are some platforms where this is not the case, but I don't know of any.)

The errors are not huge (generally off-by-one or off-by-two errors), and many people are not concerned
about them.

Anyway, this module just provides a few functions for examining the values, and an nv() function that
assigns the values via the C compiler/libc.
On perl builds that have an nvtype of "double", the nv() function is essentially the same as the
POSIX::strtod() function.
Unfortunately, the POSIX module doesn't provide a strtold() function (prior to 5.21), and can therefore
not be used for the same purpose on builds of perl where the nvtype is "long double". The nv() function
does, however, also work on "long double" builds of perl.
It also works on perls (5.21.4 and later) where $config{nvtype} can be '__float128'.

Then there's always the possibility of bugs in C's strtod()/strtold() functions. My mingw.org port
of gcc-4.7.0 suffers a bug in strtod(), though it's the only compiler I've found to be buggy in this
regard. It's a bug that doesn't strike very often, but it's enough to destroy one's confidence that
strtod() is behaving correctly. (Just my luck !!)

If you want to be really paranoid about it, the surest way of assigning the correct floating point
value in perl is to have the mpfr library assign the value. That is, instead of doing:

my $nv = 1e-298;

Do something like:

use Math::MPFR qw(:mpfr);
Rmpfr_set_default_prec(Math::MPFR::_DBL_MANT_DIG());    # if nvtype is double
# alternatively:
# Rmpfr_set_default_prec(Math::MPFR::_LDBL_MANT_DIG()); # if nvtype is long double
# or
# Rmpfr_set_default_prec(Math::MPFR::_FLT128_MANT_DIG()); # if nvtype is __float128
my $nv = Rmpfr_get_NV(Math::MPFR->new('1e-298'), MPFR_RNDN);

Cheers,
Rob