optimizing code using declarations

["Roland Zito-Wolf" <rjz@chaos.cs.brandeis.edu>]

I am trying to speed up a loop in a critical section. The loop does 
nothing but integer math, array and structure references, and 
calls to bit-and and bit-or. My question is:

1. Given how liberal the Commonlisp spec is about what declarations 
can and cant be ignored; and

2. Given how hard it is to tell when one has managed to declare every
instance/reference to a given variable/function (eg, which declaration scoping
rule does macl 1.3 and 2.0 use? do i need to put (the X) formas around 
all intermediate results? Do i have to declare built-in functions? ); and

3. Given  that (to my knowledge) the compiler gives no messages about what has
or has not been optimized and/or coded inline, or any compilation statistics such 
as size of the resulting code, subprimitive calls, etc,
or as to which decls have been  ignored or utilized; and

4. Given that, in all fairness to MACL, optimizing LISP code is in general
a fairly intricate and painful process:

Then:  how does one go about
optimizing a piece of code other than iterating thorgh hundreds of trials of
(loop while (not (exasperated-p)) 
      do (tweak-declarations) 
      do (compile)
      do (disassemble-and-guess-what-code-is-doing))?

thanks,
rjz

(PS: anything along the line of #3 would i think be of great general utility.)

(PPS: How useful is it to move oft-referenced quantities into local variables,
and if so, how many such efficient local variables can one have?)