Because in CUDA types do not have associated address spaces,
globals are declared in their "native" address space, and accessed
by bitcasting the pointer to address space 0.  This relies on address
space 0 being a unified address space.

llvm-svn: 157167