Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!gem.mps.ohio-state.edu!rpi!turing.cs.rpi.edu!harrisr
From: harrisr@turing.cs.rpi.edu (Richard Harris)
Newsgroups: comp.lang.lisp
Subject: Re: CLOS speed
Message-ID: <1989Oct27.224758.13427@rpi.edu>
Date: 27 Oct 89 22:47:58 GMT
References: <18769@pasteur.Berkeley.EDU> <31037@news.Think.COM> <MUJICA.89Oct25161848@ra.cs.ucla.edu> <31074@news.Think.COM>
Reply-To: harrisr@turing.cs.rpi.edu (Richard Harris)
Organization: RPI CS Dept.
Lines: 52

In article <31074@news.Think.COM> barmar@kulla (Barry Margolin) writes:
>In article <MUJICA.89Oct25161848@ra.cs.ucla.edu> mujica@ra.cs.ucla.edu (S. Mujica) writes:
>>What about setting the value of a slot?  I have noticed that running
>>PCL on Lucid 3.0,  it takes about 30 times longer to set the value of
>>a slot than to set the value of local variable.
>
>Accessing a local variable is probably the fastest single operation in any
>Lisp implementation.  CLOS slots should be compared with DEFSTRUCT slots,
>not with local variables.
>
>I'm sure it will still be much slower in PCL, but that's because it does no
>optimization.  PCL's simplistic implementation of SLOT-VALUE makes a number
>of function calls; (SLOT-VALUE object 'slot) calls (SLOT-VALUE-USING-CLASS
>(CLASS-OF object) object 'slot), which has to do some table lookup to
>figure out where that slot is in that class.
This is wrong.  Calls to SLOT-VALUE within the body of a DEFMETHOD 
are optimized.  Look at EXPAND-DEFMETHOD-INTERNAL (in boot.lisp).

>
>In a high performance implementation, with type declarations and block
>compilation, the offset can be determined at compile time, and most of the
>above can be compiled away or inlined.
The offset can be determined at compile time if:
 1) the class of the object is known at compile time, and
 2) the metaclass and superclass chains of the object's class
    can also be determined at compile time, and
 3) the class structure uses single inheritance, and
 4) the instance variables of the class and its superclasses
    will not change.
PCL generates (reasonably) good code when the class of the object 
is known at compile time, but PCL can not assume that conditions 
(3) or (4) are true (when the object's metaclass is standard-class).

There may be better slot accessing techniques that could be used
for CLOS:  for instance, when the Flavors system can determine
the class of an object, it arranges to compute at load time
an offset into a vector (which is stored in each instance) of offsets 
into the instance.  This is probably the fastest method of
slot access that CLOS could use (2 vector accesses per slot-value form),
but it has has a significant drawback: it is hard (or impossible) 
to do load-time evalutation of constants (in code created by macros) 
in some implementations of CL, because CL does not support load-time 
constants very well (#, is too limited).
> ...

>
>Barry Margolin, Thinking Machines Corp.
>
>barmar@think.com
>{uunet,harvard}!think!barmar

Rick Harris