Actually probably we could remove that from every function except from kernel entry point. Currently biggest problems that this causes is that we kernel entry point is not allowed to inline the safepointer version of kernel, which causes pretty big overhead on small kernels and also nvptx compiler cannot allocate registers to handle smart pointer value arguments.
Actually probably we could remove that from every function except from kernel entry point. Currently biggest problems that this causes is that we kernel entry point is not allowed to inline the safepointer version of kernel, which causes pretty big overhead on small kernels and also nvptx compiler cannot allocate registers to handle smart pointer value arguments.