Re: [Boost-users] [TR1 Review] TR1 Formal Review Beginning

From: John Maddock (john_at_[hidden])
Date: 2005-09-30 05:10:29

> In Mathematical Special Functions, why are the functions not template
> specialisations?

You would have to ask the original author of that section, but I believe it 
was C compatibility.

There is also the issue that many of these functions can only be implemented 
efficiently if you know the precission of the argument type up front, so the 
float/double/long double versions would have different implementations in 
any case (different Chebyshev polynomial approximations for example) and 
there would be no "generic" implementation readily possible.

I've also found that some well known formula can fail quite spectacularly 
once the compilers optimisations are turned on: either through double 
rounding of long double intermediates, or through non IEEE arithmetic 
optimisations.  This makes implementing many mathematical functions as 
templates error prone - because you have no control over the compilation 
environment - where as separate file compilation effectly solves that issue, 
albeit in a compiler specific manner.  See the proposed log1p implementation 
(and associated code comments) for one particularly fustrating, if trivial 
example.

I realise you're suggesting template specialisations rather than a generic 
template, but if there is no generic template provided, what advantages do 
specialisations provide over regular overloads?

John.