From: Toon Knapen (toon.knapen_at_[hidden])
Date: 2002-11-13 09:59:23
On Wednesday 13 November 2002 14:39, Fernando Cacciola wrote:
> Unfortunately, some compilers -like bcc55- provide a 'default'
> numeric_limits<> with no compile-time constants, while the specializations
> do have compile-time values. Therefore, boost's numeric_limits<> is not
> picked up for these compilers; but trying to use numeric_limits<> with user
> defined types fails to compile unless the types explicitly specialize it
> properly.
>
> I see three possible solutions here:
>
> (1) REQUIRE user defined types to *properly* specialize numeric_limits.
> (2) Enhance boost/limits.hpp so that it 'fixes' only the default
> numeric_limits<> if only that one is broken.
> (3) Define my own user specializable traits.
>
> (3) is IMO a very bad idea since we would end up with overlapping traits
> and potential ODR violations.
> (2) might be a better solution but requires further examination.
> (1) is the easiest way to go from the library developer POV, but might
> annoy users.
IIRC, the problem even arises when using std::complex< double > as argument
to std::numeric_limits<>. On gcc 3.2 for instance, the epsilon() is defined
to be (0,0) which is certainly not correct and even highly dangerous.
toon