Fix typos in #if and nonstandard implicit cast.

[lyuma]

Regarding the changes in math.h, it appears that _MSC_VER was typoed in preprocessor statements in two different ways: once as __MSV_VER and again as __MSC_VER (note the double underscore) which means they will be treated as 0. Combined with < or <=, it means that the check will be true on all compilers building for windows.

The typos in math.h (and missing defined(_MSC_VER)) caused the following errors when cross-compiling with llvm-mingw on a Linux system:

In file included from kernels/common/accelset.cpp:5:
In file included from kernels/common/scene.h:16:
In file included from kernels/common/scene_subdiv_mesh.h:11:
In file included from kernels/common/../subdiv/patch.h:6:
In file included from kernels/common/../subdiv/catmullclark_patch.h:6:
kernels/common/../subdiv/catmullclark_ring.h:365:20: error: call to 'isinf' is ambiguous
      if (unlikely(std::isinf(vertex_crease_weight)))
                   ^~~~~~~~~~
kernels/common/../../common/tasking/../sys/platform.h:152:48: note: expanded from macro 'unlikely'
#define unlikely(expr) __builtin_expect((bool)(expr),false)
                                               ^~~~
/opt/llvm-mingw/x86_64-w64-mingw32/include/c++/v1/math.h:491:1: note: candidate function
isinf(float __lcpp_x) _NOEXCEPT { return __libcpp_isinf(__lcpp_x); }
^
kernels/common/../../common/algorithms/../math/math.h:19:22: note: candidate function
  __forceinline bool isinf ( const float x ) { return _finite(x) == 0; }

The second instance guards what appear to be fallback implementations of roundf and nextafter.

One concern I have is these Visual Studio 2013(?) implementations of nextafter(float, float) and roundf(float) are actually wildly incorrect, and specifically roundf was likely being used on all windows builds:

• nextafter(x, -1.0) was unused except in tutorial code, but if it were used, as in tutorials/common/texture/texture2d.cpp, it would cause the texture dimensions to be multiplied by 0.9
• roundf(float x) seems mostly correct for float values which are representable as a 32-bit integer, but for those which are not, will cause some form of overflow. This seems like potential for bugs.

This PR will inadvertently fix these issues, by fixing the condition for the "fallback" implementations to be used. Hopefully nothing depended on the overflow mechanic of the old roundf implementation...

---

Regarding the other change, the missing (void*) cast in common/sys/library.cpp is apparently a non-standard Microsoft extension, allowed only in the Microsoft-compatible setting of clang, and not by default. Using the same command line as on Windows, this implicit cast from FARPROC to void* causes the following error.

common/sys/library.cpp:30:12: error: cannot initialize return object of type 'void *' with an rvalue of type 'FARPROC' (aka 'long long (*)()')
    return GetProcAddress(HMODULE(lib),sym.c_str());
           ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1 error generated.

By adding -fms-compatibility to the command line on a clang build, it is possible to reduce this to a warning instead of an error:

warning: implicit conversion between pointer-to-function and pointer-to-object is a Microsoft extension [-Wmicrosoft-cast]

However, it would be preferable to solve this by performing an explicit cast to void*.

Collectively, these two fixes allow embree to be cross compiled using llvm-mingw on Linux without any non-standard options.

[svenwoop]

Just merged this into our internal devel branch. Fix will be in the next release.