This is an optimization for speed. Many subroutines need to use only a few registers for temporary values. They can use the T registers without the added expense of saving and restoring them. A subroutine that needs more registers or that needs to have values saved across a subroutine call may use some S registers, but must pay for them by saving and restoring them.
These rules are somewhat complicated. Here is a picture. It shows the four sections of subroutine linkage. The basic tasks of each section are:
Subroutine Call: Push all T registers that contain
values that will be needed after the call.
Put arguments in A registers.
jal to the subroutine.
Prolog: If this subroutine calls other subroutines,
push $ra. Push all S registers that the subroutine alters.
Body: Normal code, except that it must follow these conventions if it calls a subroutine. T registers and A registers can be used freely, as can any S registers that were saved in the prolog.
Epilog: Put return values in V registers. Pop any S registers. Pop$raif it was pushed in the prolog.jr $raback to the caller.
Regaining Control: Pop any T registers that were previously pushed.
Is there any limit in these rules about how many levels deep subroutine calls may be?