Issue 6 November 20, 1995
______________________________________________________________________________
Core_Warrior_ is a weekly newsletter promoting the game of corewar. Emphasis
is placed on the most active hills--currently the '94 draft hill and the
beginner hill. Coverage will follow where ever the action is. If you have
no clue what I'm talking about then check out these five-star internet locals
for more information:
FAQs are available by anonymous FTP from rtfm.mit.edu as
pub/usenet/news.answers/games/corewar-faq.Z
FTP site is: ftp.csua.berkeley.edu /pub/corewar
Web pages are at:
http://www.stormking.com/~koth
http://www.ecst.csuchico.edu/~pizza/koth
http://pauillac.inria.fr/~doligez/corewar/
______________________________________________________________________________
Greetings.
The weeks keep getting shorter, but the tournament entries keep getting more
complex. Even so, action on the '94 hill has been furious. And some old
faces have returned--J. Pohjalainen, originator of the replication engine
about everyone is using now. And Schitzo aka Michael Nonemacher ( I think; I
haven't been playing all that long . . . ) whose quiz briefly ruled the hill
this week. Welcome back.
Damien Doligez seems to have a semi-regular column in the Core_Warrior_.
This issue he discusses another variation of the continuously launching imp.
Have a good Thanksgiving everyone.
______________________________________________________________________________
Tournament Time
(details at http://www.stormking.com/~koth/nsfcwt.html)
Round 6 is finally over. There was some controversy over self fights
and handshakes. At the end this didn't make much of a difference.
The round was run twice. First without self fights and then with self fights.
The results are almost the same except in one case. Karl Lewin sent in two
versions of his warrior (The Thinker). One version with handshakes and
another without handshakes. I included both of these versions in both runs.
Special thanks for Karl Lewin for Test warrior. It was a valuable tool to
find out about buggy submissions.
Test warrior was run also in the tournament, and it gained the first place
as it was supposed to. Notice that in the self fight run Test warrior didn't
win 100 % .
Since the only difference in the results of the two runs is the order of
Thinker and Winner, we decided to call it a draw between these two
submissions.
Here is the result of the no self fights run:
Elapsed time: 833 seconds (00:13:53)
Rank Name Author %W %L %T Score
___________________________________________________________________________
1 Test for Timeouts Karl Lewin 100 0 0 75000
2 Blizzard Anders Ivner 71 18 10 56123
3 Psst Robert Macrae 67 21 12 53019
4 Round 6 Kline P.Kline 64 20 16 51936
5 nt entry #6 M R Bremer 64 21 16 51529
6 Winner Steven Morrell 63 21 16 51183
7 The Thinker (w/Handshak Karl Lewin 62 23 15 50479
8 The Thinker (w/o handsh Karl Lewin 62 23 15 50479
9 Machiavelli Maurizio Vittuari 62 23 15 50356
10 Maverick Randy Graham 63 25 12 50142
11 Miss Playful Derek Ross 61 22 16 50109
12 ?{whateverCT#$~S|[]~+'i Paulsson 59 27 14 47404
13 PlinySwitch - 2990 G. Eadon 58 27 15 47140
14 Bim bum bam Beppe Bezzi 57 30 13 46149
15 WhileAway JKW 54 31 16 44161
16 Paper-Scissors-Stone #4 Nandor Sieben & Stefan 35 58 7 27887
17 Paper-Scissors-Stone #6 Nandor Sieben & Stefan 17 68 15 16168
18 Paper-Scissors-Stone #7 Nandor Sieben & Stefan 19 74 7 16097
19 Paper-Scissors-Stone #3 Nandor Sieben & Stefan 17 74 9 15006
20 Paranoid Calvin Loh 16 73 10 14829
21 Paper-Scissors-Stone #1 Nandor Sieben & Stefan 17 76 7 14507
22 Paper-Scissors-Stone #8 Nandor Sieben & Stefan 15 74 11 14084
23 Paper-Scissors-Stone #5 Nandor Sieben & Stefan 13 70 18 13834
24 Paper-Scissors-Stone #9 Nandor Sieben & Stefan 15 76 9 13589
25 Paper-Scissors-Stone #2 Nandor Sieben & Stefan 15 76 9 13386
26 Paper-Scissors-Stone #1 Nandor Sieben & Stefan 1 94 4 2096
These is the result of the self fight run:
Elapsed time: 942 seconds (00:15:42)
Rank Name Author %W %L %T Score
___________________________________________________________________________
1 Test for Timeouts Karl Lewin 93 0 7 77000
2 Blizzard Anders Ivner 66 17 17 58123
3 Psst Robert Macrae 65 23 11 56019
4 Round 6 Kline P.Kline 59 19 22 53936
5 nt entry #6 M R Bremer 59 19 22 53529
6 The Thinker (w/Handshak Karl Lewin 61 25 14 53477
7 Winner Steven Morrell 58 20 22 53183
8 The Thinker (w/o handsh Karl Lewin 58 21 21 52479
9 Machiavelli Maurizio Vittuari 57 21 22 52356
10 Maverick Randy Graham 58 23 19 52142
11 Miss Playful Derek Ross 57 21 22 52109
12 ?{whateverCT#$~S|[]~+'i Paulsson 54 25 20 49404
13 PlinySwitch - 2990 G. Eadon 54 25 21 49140
14 Bim bum bam Beppe Bezzi 53 27 20 48149
15 WhileAway JKW 50 28 22 46161
16 Paper-Scissors-Stone #4 Nandor Sieben & Stefan 32 54 14 29887
17 Paper-Scissors-Stone #6 Nandor Sieben & Stefan 15 63 21 18168
18 Paper-Scissors-Stone #7 Nandor Sieben & Stefan 18 68 14 18097
19 Paper-Scissors-Stone #3 Nandor Sieben & Stefan 16 68 16 17006
20 Paranoid Calvin Loh 15 68 17 16829
21 Paper-Scissors-Stone #1 Nandor Sieben & Stefan 16 70 14 16507
22 Paper-Scissors-Stone #8 Nandor Sieben & Stefan 14 68 18 16084
23 Paper-Scissors-Stone #5 Nandor Sieben & Stefan 12 65 24 15834
24 Paper-Scissors-Stone #9 Nandor Sieben & Stefan 14 70 16 15589
25 Paper-Scissors-Stone #2 Nandor Sieben & Stefan 14 71 16 15386
26 Paper-Scissors-Stone #1 Nandor Sieben & Stefan 1 87 12 4096
This is the rank after round 6:
Name 1 2 3 4 5 6 total
Steven Morrell 5 10 9 13 14 10 61
P.Kline 7.5 9 7 11 11 12 57.5
Paulsson 7.5 11 11 9 2 5 45.5
Anders Ivner 5.5 8 4 0 10 14 41.5
Beppe Bezzi 7 7 13 2 8 3 40
M R Bremer 7 4 3.6 5 7 11 37.6
Maurizio Vittuari 6.5 5 6 3 9 8 37.5
John K. Wilkinson 4 6 12 0 13 2 37
Robert Macrae 0 0 0 12 12 13 37
Karl Lewin 0 0 10 4 6 10 30
Randy Graham 0 0 8 10 4 7 29
Derek Ross 3.5 3 3.3 7 3 6 25.8
G. Eadon 1.5 2 5 6 1 4 19.5
Kurt Franke 0 0 0 8 0 0 8
Michael Constant 0 0 0 0 5 0 5
Anders Scholl 0 1 2 1 0 0 4
John Lewis 0 0 3 0 0 0 3
Calvin Loh 0 0 1 0 0 1 2
Detailed results and the submissions will be sent in a separate file. It wasn't
clear for a while if handshake is possible. So looking at the submissions
will be very interesting. We let Corewarrior discuss this topic in depth.
It is clear from game theory that assuming an optimal opponent the best
strategy is to play randomly. If the opponent does not play using optimal
strategy then we can do much better. Of course trying to outguess the opponent
makes our strategy different from the optimal strategy.
There was some concern about mts. The Borland C compiled version of mts has
difficulty handling big scores. This didn't caused problem in this round
since the djgpp compiled version was used which can handle huge scores.
Thanks for playing,
Nandor & Stefan.
______________________________________________________________________________
Current Status of the Internet Pizza Server ICWS '94 Draft Hill:
Hill Specs:
coresize: 8000
max. processes: 8000
duration: after 80,000 cycles, a tie is declared.
max. entry length: 100
minimum distance: 100
rounds fought: 250
instruction set: ICWS '94 Draft
Last challenge: Sun Nov 19 10:25:06 PST 1995
# %W/ %L/ %T Name Author Score Age
1 42/ 25/ 33 test L 7e Beppe Bezzi 159 2
2 41/ 25/ 34 La Bomba Beppe Bezzi 157 46
3 47/ 39/ 14 quiz Schitzo 155 51
4 42/ 34/ 24 Derision M R Bremer 149 70
5 34/ 31/ 35 Torch t18 P.Kline 136 430
6 28/ 20/ 52 Fly Fishing w/Plastic Wor Karl Lewin 136 28
7 37/ 39/ 24 myVamp v3.7 Paulsson 134 398
8 40/ 46/ 13 Leprechaun on speed Anders Ivner 134 226
9 35/ 37/ 28 endpoint . M R Bremer 134 75
10 34/ 35/ 31 Phq Maurizio Vittuari 132 532
11 38/ 43/ 20 Porch Swing + Randy Graham 132 131
12 33/ 35/ 31 Son & Father Maurizio Vittuari 131 81
13 32/ 33/ 35 Father & Son Maurizio Vittuari 131 80
14 35/ 40/ 25 Armory - A5 Wilkinson 130 569
15 38/ 48/ 14 Scissors v.a1 John K. Wilkinson 129 22
16 25/ 22/ 54 juliet and paper M R Bremer, B. Bezzi 128 47
17 27/ 28/ 45 Jack in the box Beppe Bezzi 125 418
18 20/ 18/ 63 TimeScape (1.7) J. Pohjalainen 121 3
19 15/ 10/ 75 Evolve 6.3 John Wilkinson 121 11
20 31/ 45/ 23 myTry Paulsson 117 1
Bezzi continues to dominate the hill with La Bomba. His test L 7e is no
doubt an improvement that no one wants to see ( but him ). quiz is the only
serious contender for the throne at this point. Forty-seven percent wins
is quite impressive. Karl Lewin makes the jump from the b hill with his
Fly Fishing w/ Plastic Worms. One of the more original names I've seen.
Jack in the box and Armory have been floating in the bottom half lately.
It's too soon to tell if their effectiveness is gone.
______________________________________________________________________________
94 - What's New
1 42/ 25/ 33 test L 7e Beppe Bezzi 159 2
3 47/ 39/ 14 quiz Schitzo 155 51
6 28/ 20/ 52 Fly Fishing w/Plastic Wor Karl Lewin 136 28
15 38/ 48/ 14 Scissors v.a1 John K. Wilkinson 129 22
16 25/ 22/ 54 juliet and paper M R Bremer, B. Bezzi 128 47
19 15/ 10/ 75 Evolve 6.3 John Wilkinson 121 11
20 31/ 45/ 23 myTry Paulsson 117 1
______________________________________________________________________________
94 - What's No More
21 38/ 48/ 14 Anti Die-Hard Bevo (3c) John Wilkinson 128 176
21 27/ 28/ 45 .Brain Vamp. B.Bezzi, M.Paulsson 125 74
21 37/ 48/ 15 Leprechaun deluxe Anders Ivner 126 284
21 28/ 34/ 37 Thieves Like Us M R Bremer 122 45
21 31/ 47/ 22 Frontwards Steven Morrell 116 323
21 15/ 7/ 78 Chugging Along Karl Lewin 122 84
21 24/ 22/ 54 Impfinity v3e7 Planar 125 70
Carnage on the '94 hill! A combined age of 1,056 was kicked off the hill
this week. Hopefully the seven new warriors are worthy successors.
______________________________________________________________________________
Current Status of the Internet Pizza Server Beginner's Hill:
Hill Specs:
coresize: 8000
max. processes: 8000
duration: after 80,000 cycles, a tie is declared.
max. entry length: 100
minimum distance: 100
maximum age: At age 100, warriors are retired.
rounds fought: 250
instruction set: ICWS '94 Draft
Last challenge: Sun Nov 19 15:33:24 PST 1995
# %W/ %L/ %T Name Author Score Age
1 44/ 37/ 19 Fire Master v1 JS Pulido 150 8
2 37/ 26/ 37 teamwork? Kurt Franke 149 4
3 40/ 31/ 29 Lurker 1.1 Kurt Franke 148 2
4 39/ 40/ 21 Look J. Doster 138 31
5 28/ 20/ 52 Schizo J. Doster 136 34
6 25/ 16/ 59 Impfinity v3c11 Planar 133 46
7 24/ 21/ 55 Paper8-II G. Eadon 127 18
8 32/ 39/ 29 Test-Fc G. Eadon 125 89
9 31/ 38/ 31 Hint Test v4 M R Bremer 123 35
10 24/ 26/ 49 Loh_tst_1.3 Calvin Loh 123 7
11 28/ 37/ 34 0.5 jmz scanner Kurt Franke 120 3
12 33/ 48/ 19 Recon V7.2 A. Nevermind 118 5
13 17/ 19/ 64 Sheet 1.0 J. Doster 115 44
14 13/ 10/ 77 Impfinity v1 Planar 115 78
15 12/ 10/ 79 RingWorm_v2.5 Calvin Loh 113 17
16 15/ 19/ 66 RingWorm_v2.4 Calvin Loh 112 32
17 28/ 46/ 25 Paper Shredder 2.1 Kurt Franke 110 1
18 10/ 5/ 5 teamwork? Kurt Franke 33 15
19 9/ 6/ 4 Lurker 1.1 Kurt Franke 30 64
20 8/ 7/ 4 Searching Kurt Franke 29 70
juliet storm was retired from the hill this week, meaning the hill has aged
100 warriors since the introduction of the Core_Warrior_. Franke managed to
kill all his warriors somehow. Some new strategy? Franke, Doster, and Eadon
have all been doing well on the hill, but none are savvy enough as Pulido to
gain the top spot. I predict Kurt will be the next author to gain entry to
the '94 hill. He better prove me right ( or somebody prove me wrong ).
______________________________________________________________________________
The Hint
Hey kids. Look what we have in store for you today: the next installment
of the Core_Warrior_ hint exploring standard warrior strategies. Scanners
and replicators have been thouroughly ( hopefully ) discussed. Next in line
are the dumb brutes--bombers. Bombers have been well treated in Morrell's
tutorial 'chapter 2' accessible from the pizza or stormking pages at the
berkeley ftp site, so we'll concentrate on two veterans--Blue Funk 3 and
juliet storm along with the comments of their authors. These two warriors
are typical of the 'classic' stone. Another group of bombers such as Torch,
HeremScimitar, and Tornado represent another type that will be discussed
next week. Stay tuned. ( For a detailed explanation on how stones work, see
Morrell's book, chapter 2 ). Thanks to Morrell for his contribution on Blue
Funk.
______________________________________________________________________________
BLUE FUNK 3 by Steven Morrell
The two things a stone tries hardest to optimize are size and speed. A good
metric to start out with is what I call the "stone index:" size/speed. Blue
Funk 3 has a stone index of 5 instructions / approximately 33% c = 15. Iron
Gate 1.5 has a stone index of 13 instructions/ 2/3 c = 19.5. Hence, if all
else is equal, Blue Funk 3 will win more often.
How much more often? If warrior A has a stone index of m and warrior B
has a stone index of n, then warrior A will win n/(m+n) of the time. So
Blue Funk 3 should win 56.5% of the time. (Iron Gate is very resistant to
stones.)
All else, of course, is not equal. The imp-ring slows Blue Funk 3 down
dramatically and lends no aid to the warrior on the current hill, with every
anti-imp strategy working against Blue Funk 3's imp. But, color is on the
side of Blue Funk 3: While the scanner focuses upon the decoy, the bomber
can color even more core with it's post-inc and DJN-stream. These aren't
the main weapon, but they are free attacks, so quite useful. Against most
scanners, a DJN.f stream can be remarkably lethal (as all the Genetic
Algorithm folks find out).
This is all academic, as Blue Funk 3 doesn't stand a chance on the modern
hill. Frontwards, for instance, also has a stone index of 15 during the
scanning phase, kills this kind of imp-ring, is very durable in the sense
that 30% of its instructions are not used after the initial scan, and kills
a better grade of paper than Blue Funk 3 could hope to.
Also, there is a "mesh" issue when stones fight stones. Stones have an
annoying habit of bombing themselves if things are not calibrated very well,
so they don't generally use a very small mod (say, 1 to 3) for their step
size. I was happy to get Blue Funk 3 down to mod 4, but this means that
small programs, like P.Kline's deadly core-clear (3 executing instructions),
can slip through the cracks. That's 25% non-contested wins!
The best programs for both mesh and stone index that I know of are
W.Sheppard's RedRain (mod 2, stone index 8) and No Ties Allowed (mod 2,
stone index 9). Unfortunately, neither of them have a core-clear or imp
gate, so they don't do too well on the hills anymore.
Now for the details for Blue Funk 3: Ideally, a stone will bomb everywhere
except for the stone itself. That is how P.Kline's Emerald (stone
component) is supposed to work, but it has a very subtle bug in it. Blue
Funk 3 uses the same general mechanism without the bugs.
After it is booted, the stone looks like this:
emerald SPL #-step,<step
stone MOV >-step,step+1 ;hit
ADD emerald,stone
cnt DJN.f stone,<emerald-50
DAT 0,0
cc DAT #-7 ;hit
The bombing instruction (at "stone") will post-decrement the B-fields of the
instructions labeled ;hit. When "stone" gets hit (after a complete mod-4
pass), the destination pointer slips to a new mod-4 pattern. By the time
this slips the fourth time, every location in core has been bombed.
The instruction that hits "stone" looks like "MOV >0,n", which doesn't do
anything except make the bombing pattern slip. The instruction that hits
"cc" looks like "MOV >4,n" when cc contains n-4, so it is equally harmless
and self-adjusting. (For instance, the first time cc is used, stone is
MOV >4,-3, and cc is DAT #0,#-7. Hence stone-3 is copied neatly to
itself, and cc becomes DAT #0,#-6 for the next time around.) The result
of all of this foolishness is that the stone doesn't bomb itself until it
has bombed all of the rest of core 1.5 times!
;redcode-94 verbose
;name Blue Funk 3
;author Steven Morrell
;strategy Fixed another in-memory/in-register bug
;macro
org boot
step equ 3044
for 78
dat <imp,<2667
rof
emerald SPL #-step,<step
stone MOV >-step,step+1
ADD emerald,stone
cnt DJN.f stone,<emerald-50
cc dat #-7
boot mov cc,out-200+5
out spl @0,out-200
mov emerald,>out
mov emerald+1,>out
mov emerald+2,>out
mov emerald+3,>out
spl i
spl i31
i12 spl imp2
imp1 jmp >0,imp
i31 spl imp1
imp3 jmp >0,imp+5334
i spl i12
spl imp3
imp2 jmp >0,imp+2667
imp mov.i #3044,2667
______________________________________________________________________________
JULIET STORM by M R Bremer
I started coding juliet storm when B Panama X and Blue Funk 3 by Steven
Morrell were both doing _very_ well on the hill. B Panama was a combination
stone/paper and Blue Funk was a bomber/imp. Since B Panama was on top of
the hill, I decided to follow the leader. juliet storm was originally
intended to be part of my own stone/paper. That's why there are two splits
in the stone--to keep up with the replicator. When I challenged the hill
with my 'incomplete' warrior, it scored surprisingly well. Amazingly well.
So I became obsessed with making it the best bomber on the hill. Out of that
effort came juliet.
The bombing step size was chosen to find size 5 and size 10 opponents rapidly
and still be small enough for a decent imp gate. I found out that the double
split consumed so many processes, that stun attacks on my imps had little
effect. This greatly improved my score against Agony II. They also helped
juliet to 'jump' into papers when overwritten. The extra 'gate' line was used
in case the coreclear move was decremented. juliet continues to bomb as she
clears, but I don't really think this matters much. The clear ends in a
hyper-perfect gate ( I think ). The imp was taken from a posting by T. Hsu on
vector launced imps. I wasn't quite savvy enough to write my own at the time.
juliet ruled the '94 draft hill for a few weeks. It's a _HUGE_ rush when
you finally get your first warrior to be King of the Hill. At least I thought
so. The introduction of p-space killed her dead, dead, dead. A stand alone
bomber has little chance of success on the current hill.
BTW: the name: 'juliet storm' is a reference to Romeo and Juliet. Theirs
was near perfect love. But in all my relationships, I always manage to
really anger my girlfriend from time to time--hence the 'storm'.
Nothing matches the fury of a woman.
;redcode-94
;name juliet storm
;author M R Bremer
;strategy bomber--yippee
ptr EQU -1333
gate dat <-445, <-446
s spl #445, <-445
spl #0, <-446
mov {445-1, -445+2
add -3, -1
djn.f -2, <-2667-500
mov 33, <-20
go dat #0, #ptr
start mov {-1, <-1
mov {-2, <-2
mov {-3, <-3
mov {-4, <-4
mov {-5, <-5
mov {-6, <-6
mov gate, ptr+24
mov gate, ptr+24
spl @go, <-4000
jmp boot, <-4013
for 73
dat #0, #0
rof
imp_sz equ 2667
boot spl 1 ,#0
spl 1 ,#0
spl <0 ,#vector+1
djn.a @vector,#0
imp mov.i #0,imp_sz
jmp imp+imp_sz*7,imp+imp_sz*6
jmp imp+imp_sz*5,imp+imp_sz*4
jmp imp+imp_sz*3,imp+imp_sz*2
vector jmp imp+imp_sz ,imp
end start
______________________________________________________________________________
You may have noticed the new Hint Warrior on the beginner hill. It is doing
quite well.
4 38/ 34/ 28 Hint Test v4 M R Bremer 142 1
The old Hint Warrior suffered from scanners. By pspacing a bomber element
as a backup, Hint Warrior vastly improves its score. Losses to papers have
gone up due to the fact that when a replicator beats the scanner part of
Mutagen, the bomber is almost sure to tie or lose the next round. By adding
an imp, I should be able to reduce the number of losses at least. Another
improvement may be to boot of a decoy. Something to think about for later
weeks.
;redcode-b
;name Hint Test v4
;author M R Bremer
;strategy Original code based on Scott Manley's Mutagen
;strategy Once through scan --> spl spl dat dat coreclear
;strategy Core_Warrior_ #2: compressed code
;strategy Core_Warrior_ #4: improved scanner
;strategy Core_Warrior_ #6: pspaced with small bomber
;kill Hint
;assert CORESIZE==8000
org start
dist equ 98
scan equ dist*2
begin SPL b1+2
jmp a
for 20
dat 0, 0
rof
a add d,c
c cmp a+2*dist,a+dist
slt.a #dist+ptr2-a, c
djn.f a,<7000
mov s, *c
x mov m, @c
jmn a,a
s spl #-dist+1, <0
jmp b1
d dat <scan,<scan
m mov.i dist, }dist
dat 0, 0
dat 0, 0
dat 0, 0
dat 0, 0
dat 0, 0
b3 dat <2667, <2667*2 ;a suggestion by Paulsson to kill imps
b2 spl #b3-ptr2, a-2-6
b1 spl #b2-ptr2, a-2-5
mov *ptr2, <ptr2
mov *ptr2, <ptr2
mov *ptr2, <ptr2
djn.f -3, <6000
ptr2 dat b1, -100+a
for 21
dat 0, 0
rof
start ldp #0, #0
ldp #1, #0
cmp.b -2, -1
jmp bomber, {4000
stp #1, #1
jmp begin, {4000
bomber stp #0, #1
jmp juliet, {3000
for 19
dat 0, 0
rof
juliet spl #445, <-445
spl #0, <-446
mov {445-1, -445+2
add -3, -1
djn.f -2, <-2667-500
mov.i 1, <-6
______________________________________________________________________________
Planar's Corner:
Continuous spiral launching (continued)
Erratum
-------
My previous article about imp spirals contains a few small typos and
a big one: the line that reads
[imp+1 imp+step+2 imp+2*step+1 imp+2 imp+step+2 imp+2*step+2 ...]
should be
[imp+1 imp+step+1 imp+2*step+1 imp+2 imp+step+2 imp+2*step+2 ...]
I hope it didn't confuse anybody. Also, I have renamed the labels
between the explanation and the code of Impfinity v1: Z becomes a,
A becomes b, etc.
Vector launching
----------------
Myer told me he wanted to see a continuous vector-launching imp spiral
(quite a cumbersome phrase, isn't it ?). We'll start with a normal
vector launch:
A spl 1
B spl 1
C jmp @V, }0
V dat 0, imp
W dat 0, imp+step
X dat 0, imp+2*step
Y dat 0, imp+1
imp mov imp, imp+step
Now this is what happens when we execute this code. I'm still using
my notation for the process queue. To execute one step, I remove the
first element of the list, and push back the new PC value(s) at the end
of the list. To execute one round, I execute one step repeatedly, or
I can equivalently replace each element with the new value(s) it pushes
back:
[A]
[B B]
[C C C C]
Now we have four processes ready to execute C. The first one will jump
to imp and increment the A-field of C, so C is now:
jmp @W, }0
The next process will jump to imp+step and increment the A-field, and
so on. On the next round we have:
[imp imp+step imp+2*step imp+1]
(recall that imp+1 is the same as imp+3*step).
How do we turn this into a continuous-launch ? The first step is to
add an SPL 0 at the top of the program (as instruction Z). The SPL
will produce a steady stream of processes to its left (in the process
queue):
[Z]
[A Z]
[B B A Z]
[C C C C B B A Z]
[imp imp+step imp+2*step imp+1 C C C C B B A Z]
But now, as the processes at C jump to the imp, we have the same
problem as last time: the new processes are at the head of the imp
spiral (the tail is the process that executes first, it is located
at the smallest address, and the head executes last, at the highest
address). Since the spiral advances by one core element at each
round, we should launch our new processes two elements further than
the previous ones, so we should add 2 to each vector element at
each round. And our pointer (the A-value of C) now points beyond
the vector, so we should reset it to C. Lastly, we have four
processes, so we get out of sync: the vector values should change
not by 2 but by step+2. This will never work.
The last problem is easy to fix: launch only three processes or
get rid of the fourth by making it jump to a DAT statement.
We still have to reset our pointer at each round. We'll do it
between Z and A:
Z spl 0
R mov.a #V-C, C
A spl 1
V-C is the address of V relative to C. When we put it into C's
A-field it becomes the address of V (here, this is 1).
Now we have to add 2 to each vector element at each round. I don't
know how to do it. We don't have enough instructions to do it by
postincrement. So we'll launch the processes at the tail instead of
launching them at the head. This means two things: first, we must
find a Z that will add processes to its right (in the process queue)
instead of the left. Second, we can launch without any increment
because the spiral advances by 1 and we add a ring at the exact
place that it left: we can launch processes at the same place while
the spiral advances.
This still won't work because I cannot find a Z that adds processes
to its right. With an SPL, the next instruction is pushed, then
the target, so we would need the SPL to be the next instruction after
itself, and the target to be R. Instead, I'll use a two-instruction
loop:
P spl R
Q jmp P
This will launch processes to its right, as follows:
[P]
[Q R]
[P A]
[Q R B B]
[P A C C C C]
But it launches only one process every two rounds. This means that
the spiral will advance by two locations while I add one ring, so I
will have to increment my vector elements after all. This is how
we'll do it: turn C into a direct jump instead of indirect and the
vector elements into JMPs instead of DATs. And use their B-fields
to increment the A-fields. This is the whole launcher:
P spl R
Q jmp P
R mov.a #V-C, C
A spl 2
B spl 1
C jmp V, }0
V jmp imp, }0
W jmp imp+step, }0
X jmp imp+2*step, }0
imp mov 0, step
The binary launcher I described last time added one ring every round,
this one adds one every two rounds. But we can double the vector and
launch two rings at once every two rounds. Then we don't need the
increments. Or three rings. Then we need to decrement the JMP values
by 1 at each round. The vector launcher is smaller than the binary
launcher, but you can't always produce the right number of processes,
so you may have to add some DAT statements to the vector to get rid of
the extra processes. Finally, tail-launching is more easily upset by
the opponent than head-launching because the newly launched process is
not refreshed right before its execution: if the tail ever gets
clobbered, the launcher will work in vain from then on.
This is quite long already, so I'll leave continuous SPL/ADD launching
and Impfinity v3 for next week (hoping it will not get thrown off the
hill in the meantime). Maybe I went into too much detail here.
<Damien.Doligez@inria.fr>
______________________________________________________________________________
Questions? Concerns? Comments? Complaints? Mail them to:
Beppe Bezzi <bezzi@iol.it> or Myer R Bremer <bremermr@ecn.purdue.edu>
|
