From: Hubert Holin (Hubert.Holin_at_[hidden])
Date: 2005-04-19 08:37:04
Somewhere in the E.U., le 19/04/2005
In article <004e01c541d0$d6436070$cf200d52_at_fuji>,
"John Maddock" <john_at_[hidden]> wrote:
> > Sounds fine with me. In the mean time, I'll also try to deal with
> > the "long double" deficiency already identified that some platforms
> > suffer from, though that only affects the tests, not the library proper
> > (I don't know if that will make it for 1.33).
> According to http://docs.sun.com/app/docs/doc/816-5173/6mbb8adu7?a=view the
> long double math functions on solaris should be pretty accurate, so I guess
> this could be a problem with the overloads declared in <cmath>, rather than
> the platform headers?
> Just guessing, you'd need to see the preprocessed source to be sure.
Yes, it is possible. Unfortunately, I do not have access to that
platform. I found out about the problem while investigating a
regression, and asking M. Caleb Epstein who ran the tests on that
platform. Either that platform does not have tailored "long double"
versions of some special functions, or they are not visible to the C++
compiler (note that numeric_limits is indeed correct for "long double"
on that platform).
Do we have some template metamagic to see if a given function
signature is available? I had intended to test computation accuracy for
some select values and compare with epsilon for that type (could still
use it to check for QOI anyway...).
Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk