Add a note about the inability to model FP -> int conversions which

perform rounding other than truncation in the IR. Common C code for this
turns into really an LLVM intrinsic call that blocks a lot of further
optimizations.

llvm-svn: 123135