Agner`s CPU blog

You have a stack in memory. The contents of said stack area in memory is normally undefined (!), except instructions can read / write the top, say, 32 elements. Or whatever the window size is. Call / return adjust the stack - although I am not sure how much the stack should be adjusted. There are advantages and disadvantages to fixed and variable-sized adjustments, as well as the question of if returns must be matched with calls. There may also be instructions to force an explicit and up-to-date read / write of any element of the stack.

To the processor, treat it as a ring-buffer (or potentially tagged-memory, to allow for easier context switching) cache of the top part of the stack. It speculatively saves anything "below" the current window to the stack to make room for calls, and speculatively loads from the stack to keep the buffer full on returns. As usual here, there is the question of how speculative it should be.

Some architectures have a separate instruction cache - this one has a separate stack cache.