Some timing tests...

[Anders Magnusson <ragge%ludd.ltu.se@localhost>]

Hm, I did send this to the mailing list last year, but it don't seem to 
have reached it (not in the archives...?)
Well, let's try again.  This is the third time :-)

-------- Vidarebefordrat meddelande --------
Ämne:   Some timing tests...
Datum:  Sun, 15 Apr 2018 14:36:15 +0200
Från:   Anders Magnusson <ragge%ludd.ltu.se@localhost>
Till:   port-vax List <port-vax%NetBSD.org@localhost>


Hi all,

while searching after why the vax port has gotten so slow, I noticed 
that there are a significant increase of function calls everywhere.
While it is a well-known fact that the current calling convertion on vax 
is slow, I decided to get some real numbers.

I tested on a moderately slow vax (3100/48 I think?) and timed this 
small program:

    movl $10000000,%r0
1:  sobgtr %r0,1b
    halt

Basically it just loops 10 million times.
With L1 cache turned off it took 11 seconds (roughly 1Mips).
Turning on the cache did drop to 4.5s (two MIPS)

So, adding calls/ret to the same code:

    movl $10000000,%r0
1:  calls $0,2f
    sobgtr %r0,1b
    halt
2:  .word 0
    ret

without cache (which could symbolize a cache miss) it took 83s.
Removing time for sobgtr it means a calls/ret pair takes 7.2us for a 
cache miss.
Turning on cache made it take just over 42s, which means that a hit 
takes 3.8us.

Doing the same with jsb/rsb:

    movl $10000000,%r0
1:  jsb 2f
    sobgtr %r0,1b
    halt
2:  rsb

this took 37us without cache, giving 2.6us uncached.
With cache, it only took just over 12 (!) us, giving .8us for this pair.

This is quite interesting numbers.  Changing the vax calling convention 
to use jsb/rsb + args in regs would be something quite interesting to 
do, since it seems to take 1/5 in speed compared to calls/ret.

-- Ragge