[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Common Lisp for SUNS
- To: franz!ucbarpa!red.ipsa.dnd.ca!bill (Bill Pase)
- Subject: Re: Common Lisp for SUNS
- From: franz!frisky!jkf@ucbarpa.Berkeley.EDU (John Foderaro)
- Date: Tue, 29 Nov 88 13:40:53 PST
- Cc: franz!sail.stanford.edu!common-lisp
- In-reply-to: Your message of Tue, 29 Nov 88 11:14:29 EST. <8811291614.AA01479@red.ipsa.dnd.ca>
Regarding benchmarking Lisps.
1. Declarations are important for speed, but you'll find that some
Lisps will require more specific declarations than others.
For example Sun's Lisp only supports one floating point type so declaring
something to be a 'float' is sufficient for the compiler to open code
the operation. Allegro supports two types of floats so you must declare
either 'single-float' or 'double-float.'
2. Different systems do different amounts of optimizations based on
the settings of the optimization parameters speed, safety and size.
Probably the only comparable values are for speed=3, safety=0, size=0
where each system does maximum optimization. While benchmarks
run at this value are interesting you probably aren't going to
do development with these settings since you aren't going to get
good diagnostics.
Regarding the Lisp on the Sun4.
Allegro is highly optimized for the Sun4 but you have to
add some declarations and specify that speed is important.
The comment was made that a C based Lisp on the Sun4 would
be very fast because of sun's optimizing C compiler.
While it is true that by using C you can take advantage of that
compiler, it is also true that the use of C constrains your
whole system design so you can't use the machine optimally.
I say this from many years experience programming in C parts of the
kernel of Franz Lisp. To cite two examples for the Sun4:
One of the nicest features of the sun4 (sparc) architecture
is the support the the Lisp fixnum data type (there are 'tagged'
versions of the add, subtract and compare instructions).
This allows the compiler to open code +,-,<,<=.. and so on
and if the operands happen to be fixnums (they very often are)
then the operation and fixnum tag check can go on in parallel.
This makes for fast computation in the typical case
and it is completely safe (if the tags aren't fixnum you find out
about it and perform the appropriate operation for the
actual data types).
A C based lisp can't make use of these tagged instructions.
The Sun4 has 7 global registers. Allegro stores nil and a
pointer to our global function table in one of them, and
uses another for tracking asychronous interrupts. Another
is used to pass argument/result counts. C doesn't permit
global register declarations so you can't take advantage
of this nice feature of the Sun4 in a C-based Lisp.
[The use of machine registers to hold certain values was
so important in Franz Lisp that for certain regular
architectures (e.g. vax) an incredible hack was written
by Keith Sklower to modify the assembler language coming
out of the C compiler to change certain memory references
to register references and to fix the register save masks.]
-- John Foderaro
Franz Inc.
jkf%franz.uucp@berkeley.edu