Re: Caller/Callee saved Registers

nandu@cs.clemson.edu writes:


>Henry Baker referred to Guy Steele's paper "Dream of a Lifetime:..." in
>the 1980 LISP conference. Steele's idea quickly degenerates to a callee
>saves scheme with additional overhead.
>
>Steele's idea is to save the intersection of the registers that the caller
>and callee require.


No! While several people _have_ proposed such a thing; Steele and
Sussman's paper is much more interesting. They propose an alternative
resource called "racks". Rather than a set of registers and a stack,
their machine would have a set of these racks. A rack has 4 operations:


Assign(R, V) -- make V the current value of the rack R


Fetch(R) -- return the current value of the rack R


Save(R) -- save (probably on an internal stack) the current
value of the rack R


Restore -- restore the current value of the rack R


Consider the following sequence, for an example


Assign(R, 1)
Fetch(R) -- returns 1
Fetch(R) -- returns 1
Assign(R, 2)
Fetch(R) -- returns 2
Save(R)
Assign(R, 3)
Fetch(R) -- returns 3
Restore(R)
Fetch(R) -- returns 2


So, obviously we can use a bunch of these things as registers, simply
issuing Saves and Restores around procedure calls.


Steele and Sussman described 4 simple implementations of the rack
abstraction and compared them experimentally. I won't describe them (read
the paper!), but I can assure you that the idea does _not_ "quickly
degenerate to a callee-saves scheme with additional overhead." On the
other hand, I don't expect to see the idea taking over the risc community.


The complete reference is


The Dream of a Lifetime: A Lazy Variable Extent Mechanism
Steele and Sussman
Conference Record of the 1980 LISP Conference


Preston Briggs
--

Return to the comp.compilers page.
Search the comp.compilers archives again.