Issues with C++ overloaded math functions in math.h in the wrong scope

[Martin Storsjö]

Hi,

When building a test package shared in 
https://savannah.gnu.org/bugs/?57406 
(https://haible.de/bruno/gnu/testdir-posix-mingw.tar.gz) with my 
mingw/clang/libc++ based toolchain, I ran into some errors in the gltests 
subdirectory.

One of the errors looks like this:

In file included from test-arpa_inet-c++.cc:22:
In file included from ../gllib/arpa/inet.h:34:
In file included from ../gllib/sys/socket.h:651:
In file included from /toolchain/x86_64-w64-mingw32/include/winsock2.h:23:
In file included from /toolchain/x86_64-w64-mingw32/include/windows.h:69:
In file included from /toolchain/x86_64-w64-mingw32/include/windef.h:8:
In file included from /toolchain/x86_64-w64-mingw32/include/minwindef.h:167:
In file included from /toolchain/x86_64-w64-mingw32/include/winnt.h:1554:
In file included from /toolchain/lib/clang/10.0.0/include/x86intrin.h:15:
In file included from /toolchain/lib/clang/10.0.0/include/immintrin.h:18:
In file included from /toolchain/lib/clang/10.0.0/include/xmmintrin.h:27:
In file included from /toolchain/lib/clang/10.0.0/include/mm_malloc.h:13:
In file included from ../gllib/stdlib.h:36:
In file included from /toolchain/x86_64-w64-mingw32/include/c++/v1/stdlib.h:100:
../gllib/math.h:2740:46: error: conflicting types for 'rpl_isfinite'
_GL_MATH_CXX_REAL_FLOATING_DECL_2 (isfinite, rpl_isfinite, bool)
                                             ^
../gllib/math.h:2740:46: note: previous definition is here
../gllib/math.h:2740:46: error: conflicting types for 'rpl_isfinite'
_GL_MATH_CXX_REAL_FLOATING_DECL_2 (isfinite, rpl_isfinite, bool)

For various reasons, gnulib decides to replace the isfinite function, and 
then the gnulib math.h file provides C++ overloaded functions for 
different data types.

In this particular case, this math.h header ends up included in a context 
wrapped in 'extern "C" {}' surrounding it, which breaks the C++ function 
overloading.

The series of headers leading up to this is a bit long...

- The source file includes some windows header

- windef.h includes winnt.h within an extern "C" block. windef.h also
  includes minwindef.h, which also includes winnt.h within an extern "C"
  block. Not sure if this really is the right thing to do or not.

- The mingw-w64 winnt.h has one extern "C" block covering the whole file,
  one over a section of the file, and one block specifically around an
  include of x86intrin.h. Not sure if the former two are right, but
  x86intrin.h and the files that it includes at least seem like they
  really need extern "C" around them.

- *intrinh end up including mm_malloc.h (this also goes for gcc), and
  mm_malloc.h includes stdlib.h (also goes for gcc)

- libc++'s stdlib.h (which is included before the normal stdlib.h)
  includes math.h since https://reviews.llvm.org/D60097 (this was also the
  root cause for an issue discussed earlier), and math.h (which first hits
  gnulib's math.h) now tries to declare the overloaded set of
  float/double/long double rpl_isfinite functions, which fails since
  we're in an extern "C" scope.

I'm not really sure which bit in this chain that actually can be 
considered wrong, but I wanted to post the issue I'm seeing at least...

libc++'s math.h should in principle also have the same issue (if just 
including windows.h in a C++ file), but it seems to be using some other 
mechanism for its function overloads, which makes it not break apart here.

// Martin