Re: Major issues with g++ 6 and gnulib

[John W. Eaton]

On 02/20/2016 10:34 AM, Orion Poplawski wrote:

> Some comments from fedora-devel and a patch -
> https://lists.fedoraproject.org/archives/list/address@hidden/message/UY3VX3W7XEXYTUKHG5BALU4ACUD7ZLGE/

I don't think we want to use something like "#include "libgnu/math.h" 
and I would really prefer to avoid including <math.h> directly in C++ files.

As I see it, the goal of gnulib is to allow us to write software that 
assumes a POSIX environment that works, and to have broken or missing 
functions replaced without having to do anything special.  So we should 
be able to include standard C++ and POSIX headers unconditionally and 
expect the functions they provide to work.

I think gnulib meets this goal for C programs fairly well.

For C++ there are some problems.

The assumption that <cmath> will include a <math.h> in a way that would 
allow us to wrap it with the gnulib <math.h> file was always 
questionable.  I think we went with that option because gnulib was (and 
is) primarily aimed at C language programs and making it work for C++ 
was a secondary goal.  But it seemed to work for all platforms of 
interest and it was easier than any other option.  But now if this 
assumption isn't valid we'll have to find some other solution.

With C++, many (but not all) of the gnulib replacement functions are in 
a separate namespace.  I suppose we could add a using statement so that 
functions could be called without the namespace tag, but that's still 
not quite the same as being able to call them with the "std::" namespace 
prefix.

Because of the assumption that standard C++ headers like <cmath> include 
the corresponding C headers like <math.h>, we may end up with symbols 
that we don't need or want.  For example, if you use gnulib and include 
<cmath> a signbit function will be defined outside of the std namespace. 
 That's not what should happen if you just include <cmath>.

As much as possible, I'd like to use only C++ headers and std:: 
namespace tags.  But I'd still like to get replacements for functions 
that are missing or broken.

Can we do something similar to what gnulib was attempting, but directly 
wrapping <cmath> instead of wrapping <math.h> and assuming that <cmath> 
would include our wrapper?

jwe