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
250 KiB
250 KiB