[Boost-testing] Re: [testing] [math] [gcc] Problems with long double math functions

From: Hubert Holin (Hubert.Holin_at_[hidden])
Date: 2005-06-09 04:01:54

Somewhere in the E.U., le 09/06/2005

   Bonjour

In article <989aceac0506070535775585f3_at_[hidden]>,
 Caleb Epstein <caleb.epstein_at_[hidden]> wrote:

> These test failures all seem to stem from problems with std::exp (long
> double) and perhaps some other math functions using long double on
> gcc.  Does this smell like a gcc bug, or is there something wrong with
> the tests?
> 
> gcc-4_0-darwin:
> http://tinyurl.com/8tts6
> http://tinyurl.com/atzbx
> 
> gcc-windows:
> http://tinyurl.com/dwmdk
> http://tinyurl.com/9to33
> 
> gcc-3_4_3-sunos:
> http://tinyurl.com/8zvpz
> http://tinyurl.com/bx6mt

      For gcc-4_0-darwin and gcc-3_4_3-sunos at least, I strongly 
suspect a libstdc++ bug (or severe QOI issue), w.r.t. "long double" on 
64 bits architectures. Jonathan Wakely is investigating that possibility 
(I sent him a much shorter test case demonstrating the problem). If 
necessary, I will turn off the tests for "long double" before the 1.33 
release.

      For gcc-windows things are not as clear, as apparently this is a 
32 bits architecture and not a 64 bits one. There is the possibility 
that "long double" may actually be (slightly) bigger than "double" on 
that platform (I seem to remember things along that line when I worked 
on NT a great while back), and that we are falling into the same kind of 
trouble.

   Merci

Hubert