RTEMS CPU Architecture Supplement
The SPARC architecture includes the concept of register windows. An overly simplistic way to think of these windows is to imagine them as being an infinite supply of "fresh" register sets available for each subroutine to use. In reality, they are much more complicated.
The save instruction is used to obtain a new register window. This instruction decrements the current window pointer, thus providing a new set of registers for use. This register set includes eight fresh local registers for use exclusively by this subroutine. When done with a register set, the restore instruction increments the current window pointer and the previous register set is once again available.
The two primary issues complicating the use of register windows are that (1) the set of register windows is finite, and (2) some registers are shared between adjacent registers windows.
Because the set of register windows is finite, it is possible to execute enough save instructions without corresponding restore's to consume all of the register windows. This is easily accomplished in a high level language because each subroutine typically performs a save instruction upon entry. Thus having a subroutine call depth greater than the number of register windows will result in a window overflow condition. The window overflow condition generates a trap which must be handled in software. The window overflow trap handler is responsible for saving the contents of the oldest register window on the program stack.
Similarly, the subroutines will eventually complete and begin to perform restore's. If the restore results in the need for a register window which has previously been written to memory as part of an overflow, then a window underflow condition results. Just like the window overflow, the window underflow condition must be handled in software by a trap handler. The window underflow trap handler is responsible for reloading the contents of the register window requested by the restore instruction from the program stack.
The Window Invalid Mask (wim) and the Current Window Pointer (cwp) field in the psr are used in conjunction to manage the finite set of register windows and detect the window overflow and underflow conditions. The cwp contains the index of the register window currently in use. The save instruction decrements the cwp modulo the number of register windows. Similarly, the restore instruction increments the cwp modulo the number of register windows. Each bit in the wim represents represents whether a register window contains valid information. The value of 0 indicates the register window is valid and 1 indicates it is invalid. When a save instruction causes the cwp to point to a register window which is marked as invalid, a window overflow condition results. Conversely, the restore instruction may result in a window underflow condition.
Other than the assumption that a register window is always available for trap (i.e. interrupt) handlers, the SPARC architecture places no limits on the number of register windows simultaneously marked as invalid (i.e. number of bits set in the wim). However, RTEMS assumes that only one register window is marked invalid at a time (i.e. only one bit set in the wim). This makes the maximum possible number of register windows available to the user while still meeting the requirement that window overflow and underflow conditions can be detected.
The window overflow and window underflow trap handlers are a critical part of the run-time environment for a SPARC application. The SPARC architectural specification allows for the number of register windows to be any power of two less than or equal to 32. The most common choice for SPARC implementations appears to be 8 register windows. This results in the cwp ranging in value from 0 to 7 on most implementations.
The second complicating factor is the sharing of registers between adjacent register windows. While each register window has its own set of local registers, the input and output registers are shared between adjacent windows. The output registers for register window N are the same as the input registers for register window ((N - 1) modulo RW) where RW is the number of register windows. An alternative way to think of this is to remember how parameters are passed to a subroutine on the SPARC. The caller loads values into what are its output registers. Then after the callee executes a save instruction, those parameters are available in its input registers. This is a very efficient way to pass parameters as no data is actually moved by the save or restore instructions.
RTEMS CPU Architecture Supplement
Copyright © 1988-2004 OAR Corporation