On 10/18/2011 02:51 PM, Michael Goffioul wrote:
Hi,With a current version of gnulib and MSVC10, I get a compilation error in octave due to signbit being #define'd by gnulib's math.h module. The configure script ends up with GNULIB_SIGNBIT being 1 and REPLACE_SIGNBIT being 1. This causes signbit being #define'd as # define signbit(x) \ (sizeof (x) == sizeof (long double) ? gl_signbitl (x) : \ sizeof (x) == sizeof (double) ? gl_signbitd (x) : \ gl_signbitf (x)) The problem is that one octave file defines a class with a method called signbit. This leads to a syntax error in the compiler. Now, when looking signbit documentation, it looks like it should be a macro. So is octave wrong when using signbit as a class method name?
If octave were written in C, then I would say you are correct, and that octave is in error; inclusion of <math.h> renders a program non-compliant for using 'signbit' in any context where the macro expansion would be undesirable. But this is because signbit _must_ be a macro in C (since C has no method overloading).
But since octave is written in C++, we could avoid the macros and instead have three overloaded functions named signbit which operate on the correct types, so as not to pollute the namespace with a macro. POSIX does not speak to C++ compliance, so using overloads instead of a macro for C++ does not cause any non-compliance issues. We've done method overloads in the past for functions like strstr which must be type-unsafe in C, but which can usefully have two overloads for proper type safety in C++.
-- Eric Blake ebl...@redhat.com +1-801-349-2682 Libvirt virtualization library http://libvirt.org
