Agner`s CPU blog

The memory map is so small that it can be contained on the chip. There are separate memory maps for system code. In benign cases where a process uses only a small part of the total memory map area on the chip, you may swap memory maps when switching processes.

All code is addressed relative to the instruction pointer and all read/write memory is addressed relative to the data pointer and the stack pointer. This means that everything is addressed relative to just three registers which can quickly be changed. There are separate registers that are only accessible to system code. These can be useful in a device driver. The device driver needs access to the registers of the application program because these registers (actually, only half of them) can be used for parameter transfer to the device driver. The extra registers can be used by the device driver for temporary storage of the application's registers.

I interpret your Setup # 1 as something like register banks that can be swapped. This may be added as an optional feature on ForwardCom, but I don't know if the costs in terms of central silicon space can justify it.

Your setup # 2 is reading and writing all registers to some memory area or cache. I have avoided an instruction for "read all registers" or "write all registers" because I want to avoid complex micro-coded instructions and because these instructions are very slow on x86 (I don't know about other architectures). Maybe a special on-chip cache for the sole purpose of saving registers on context switch would be useful.