Re: [boost] [lexical_cast][math] can't unconditionally call math::changesign with long double

Subject: Re: [boost] [lexical_cast][math] can't unconditionally call math::changesign with long double
From: Antony Polukhin (antoshkka_at_[hidden])
Date: 2013-02-24 07:41:25

2013/2/24 Paul A. Bristow <pbristow_at_[hidden]>:
>> On Sunday 24 February 2013 09:51:27 Antony Polukhin wrote:
>> > 2013/2/24 Jeroen Habraken <vexocide_at_[hidden]>:
>> > > On 23 February 2013 18:53, Eric Niebler <eniebler_at_[hidden]> wrote:
>> > >> lexical_cast is using math::changesign with long double regardless
>> > >> of whether BOOST_MATH_NO_LONG_DOUBLE_MATH_FUNCTIONS is defined or
>> > >> not. This is causing Boost.Test build problems (see here
>> > >> http://tinyurl.com/af5453h ).
>> > >>
>> > >> I'm pretty sure this is a bug in lexical_cast, so I've opened a ticket:
>> > >> https://svn.boost.org/trac/boost/ticket/8162
>> >
>> > So, portable way of changing sign is just to multiply on -1 ?
>>
>> You could toggle the most significant bit, if FP numbers have IEEE 754 format (which should be
> common
>> nowdays). Unlike multiplication, this would keep the precision and is faster.
>
> I think this is what changesign does?  So it works for NaN where all comparisons and tests fail.  If
> changesign is implemented, it should not rely on IEEE 754 format.
>
> As John says, portability is all good fun ;-)

OK, I'll just disable optimized conversions to doubles when there is
no support for long doubles in Boost.Math. This may take a few weeks.

-- 
Best regards,
Antony Polukhin