Attached patch fixes PR31202 by correcting the code generated for  
calls to the intrinsic NINT (and ANINT). Instead of adding 0.5 and  
rounding by casting into an integer, we simply call the lround (or  
llround) functions. A fallback implementation of these functions is  
provided for libgfortran.

I've bootstrapped and regtested on x86_64-linux, but I'd be glad if  
someone with a platform without full C99 support, and especiall  
lround, could test it.  I once had access to a sparc-solaris2.9,  
which fulfilled these criteria, but not more.

OK for mainline?



:ADDPATCH 31202: