Re: Changes to stable-5-0 branch

[Hans-Bernhard B. <HBB...@t-...>]

Am 15.02.2016 um 23:31 schrieb sfeam:

> I am having trouble to understand how in our particular case C++ can
> fail to find the C language definition of isnan() from <math.h>
>
> and <math.h> as I understand it is required to define isnan() by C99.

It would be.  _But_ we're not running a C99 compiler here; we're running 
a C++ compiler.  Now a C++11 compiler would be required (C++1x 
26.8p3,p4) to effectively pull in the C99 version of <math.h>, including 
isnan(), except that in C++ it's not a macro: it's a set of three 
overloaded functions.  C++ really doesn't want macros for this sort of 
thing.

A compiler that's not running in C++11 mode (or equivalent) is not only 
not required to do that: it's even kind-of forbidden.  Earlier editions 
of the C++ standard explicitly listed which functions <math.h> shall 
bring into a C++ program's global name space (C++03 26.5p1,p2). isnan() 
was not among them.  To overcome this limit might take an equivalent of 
what the AC_USE_SYSTEM_EXTENSIONS method of autoconf does for us about 
the difference between "strict" and "extended" C, but for C++.

> So why does the C++ compiler not pick up the C99 macro even if
> it wouldn't otherwise define it on its own?

Because the header itself hides it from C++'s view, probably by querying 
(__STDC_VERSION__ >= 199901L) and/or (__cplusplus >= 201103L)

> If it's only this one place, may the simplest work-around is:
> -                               if (isnan(*image))
> +                               if (*image != *image)   // this test works even if isnan() is missing

That might well be the case.