[net.arch] First PC/RT benchmark!!

bcase@amdcad.UUCP (Brian case) (01/24/86)

Ok, guys, here it is:  the first "benchmark" of the NEW, improved,
whiter, brighter, IBM PC/RT.  This is the guy with the ROMP RISC
processor in it.  We saw one at the IBM Palo Alto Engineering
Support Center.  It seems ok, but I personally hate system V UNIX.
We ran the famous while loop benchmark:

main ()
{
    register int i;
    for (i = 0; i < 1000000; ++i)
	;
}

Now, the compiler invoked by the cc comand, even with the -O
option, produced the most awful code I have ever seen.  So,we
edited the .s file and made the loop only 3 instructions:

L.15:
	c	5,0
	bltx	L.15
	sis	5,1

(Note the use of the delayed branch.)  I can't quite remember if the
increment was called sis or six, but whatever.  The compiler -O
produced a 6 instruction sequence with two branches, etc.  Yeech.

Well, the hand-optimized version runs in 1.5 secs (user).  On a
Vax 780, the cc -O version (compiles to one instruction) runs in
1.9 secs.  The RT cc -O (i.e., without hand-optimization) version
ran in 1.8 secs.  So, the RT looks to be indeed faster than a 780,
but by just how much we will not know until definative benchmarks
(like drystone (:-) are run.  Looks like a good box for the price.

We saw the low-res screen, but it was interlaced and had a very
good appearance.  We saw an automatic demo of a Mac-like editor
and the fonts and windows were very user-inviting.  I didn't like
the fact that IBM chose not to go with the tactile-feedback-type
of keyboard that they are famous for (IBM 3101, PC, 3270, etc).
We also tried another experiment:  we forked off 11 infinite loop
processes and then tried some vi and ps.  The machine still seemed
to have reasonable response to vi commands like 'j'.  They also
had a 5080 graphics substation on display, but not connected to 
the RT.  It is nice, I guess, but is HUGE compared to the RT.  I
tend to think that using an AT color graphics card and a small
monitor may be the way to go (unless the 5080 can turn somersaults
or something).

Does anyone know if the C compiler we used was indeed the one
written for the RT by Tartan Labs???

    bcase

johnl@ima.UUCP (01/25/86)

/* Written  9:19 pm  Jan 23, 1986 by bcase@amdcad in ima:net.arch */
> Does anyone know if the C compiler we used was indeed the one
> written for the RT by Tartan Labs???

No, most likely it's the one I wrote.  (Now it can be told.) The AIX 
compiler is a straightforward port of the System V PCC compiler.  PCC does 
a pretty good job of generating good code for expressions, but falls down 
when given lots of registers to use and an architecture which makes loads 
and stores very expensive relative to register operations, since it makes 
little effort to remember the contents of registers from one expression to 
the next.  This means that the compiler's code is of similar quality to 
that generated by other PCC compilers, but the ROMP chip is one where PCC's 
limitations are more evident than, say, on a Vax.  In its defense, I 
suspect that you'll find that the code looks better on real programs than 
on tiny loops.  I eagerly await some Dhrystone reports.  

John Levine, ima!johnl

PS:  No, I didn't write it all by myself.

johna@ima.UUCP (01/27/86)

The PC/RT compiler has been changed since johnl left ISC.
Dhrystone benchmarks done before I left ISC are:
	VAX 11-780	approx	1625
	PC/RT		approx	1850

These numbers are from memory, they may vary by 50 +-
		johna @ ima