(2025-05-19) How I ported Bulls and Cows to HP 12C and why it is important
--------------------------------------------------------------------------
Yes, I promised to return to the topic of tiny VMs in the next post but I
still think this story needs to be told sooner, and you can, by the way, 
consider this one a direct precursor to the next post about tiny VMs. 
Because it is about a game that's my all-time favorite when it comes to 
numeric-only I/O, and can (arguably) be viewed as the ultimate benchmark 
that sets apart computers from calculators. Yes, Bulls and Cows return, now 
on the programmable financial HP 12C I told you about in the previous post.

If you just want a short form listing, the setup and the gameplay rules, this
game has already been published in my HP 12C software collection text 
document ([1]) under the name "SlowMoo12". But I assume everyone who's been 
reading my phlog for some time already knows how to play Bulls and Cows. 
Today, I want to discuss all the technical challenges that I had to overcome 
in order to fit it into 99 program steps of a not very advanced instruction 
set, with just the seven main registers and the five financial registers 
left at my disposal. At some moments, the mission felt impossible but I 
still managed to accomplish it. Let's talk about how and then get to why.

But first, let's break down the design of the game itself on the high level.
Any Bulls and Cows session with a computer consists of two parts: one where 
the computer generates a random number and one where the player tries to 
guess the generated number. What do we know about the first part? Well, we 
know two requirements: the generated number must contain four decimal digits 
(and can start with 0) and all digits must be unique. So, we need to 
generate a digit, check if it hasn't already been used and append (or 
prepend, doesn't matter) it to the number if it hasn't been used yet. How do 
we check for uniqueness? This is where the first challange comes.

In the mu808 and n808 versions of Bulls and Cows, I had a pool of digits from
0 to 9 in separate memory cells. Then, I generated a random position from 0 
to 9 and extracted the digit from the corresponding cell. If the cell 
already contained 10, I retried with a different position. If it didn't, I 
took the digit from there and put 10 instead. This way, the result was 
guaranteed to have four unique digits. However, this approach won't work on 
the 12C at all: we don't have indirect addressing and we don't even have 
enough cells to start with. We do, however, have a ten-digit (visible) 
register size, so we can save the digit pool within a single register just 
by writing 1234567890 in there.

Once we have this value in a register, we need to generate a random position
between 0 and its length. How do we know its length? We know that it will be 
10 for the first iteration, 9 for the second iteration, 8 for the third 
iteration, and 7 for the fourth (and last) iteration. Since our iteration 
counter I starts with 4 and then decrements (with the condition to get out 
of the loop when it gets to 0), we know that the pool number length is just 
I + 6.

So, we've generated the position P as a random integer number in the [0; I+6)
interval, now what? Now we need to do two things: extract the digit at that 
position and shrink the pool number to not include that digit anymore. Digit 
extraction is seemingly easy: calculate the factor F in the form of 10^P, 
divide the pool number N by F, take the integer part of the result and the 
digit will be that integer part modulo 10. However, another hurdle arose: 
the 12C does not have any modulo operation, only the operations to take the 
integer (INTG) and fractional (FRAC) parts. So, instead of going all the way 
to obtaining the modulo, we divide the integer part by 10 and store the 
digit as the fractional part of that as it is (from 0 to 0.9). Then, we 
update the target number T (initially set to 0) by shifting it 1 position to 
the right and adding this decimal fraction representing the new digit. So, 
with the pool number N and position P:

F = 10 ^ P
DP = INTG(N/F)
D = FRAC(DP/10)
T = T/10 + D

That's why, by the way, we naturally obtain the guess numbers between 0 and
1. Now, we need to shrink the pool number N. Remember that 
we have the factor F stored and also the integer part (DP) we got from
dividing N by that factor. So, essentially, we need to divide DP further by 
10, take an integer part from that, add the fractional part of N/F and then 
multiply the result by the factor. Using only the above variables, that 
comes down to:

N = N + (INTG(DP/10) - DP) * F

It might also be helpful to cache the DP/10 value since we're calculating
both integer and fractional parts from it. To sum it up, the pseudocode (a 
bit BASIC-like but in the terms close to the 12C's instruction set) for 
generating all four digits in the result T looks like this:

N = 1234567890
T = 0
I = 4
:GENLOOP
F = 10 ^ INTG(RAND() * (I+6))
DP = INTG(N/F)
DQ = DP / 10
T = T/10 + FRAC(DQ)
N = N + (INTG(DQ) - DP) * F
I = I - 1
IF I = 0 THEN GOTO :MAIN ELSE GOTO :GENLOOP
:MAIN
...

OK, we've got the resulting unique 4-digit number in R (which, along with the
PRNG implementation itself, took 46 steps of the real 12C program, and 
that's considering 1234567890 and 10 being set in the registers outside the 
program as a part of the first-run game setup). But that just was the first 
part of the game. The second part is the main loop where we prompt the 
player for the number to guess and then score the input by matching all 
digits. This seems to be straightforward: just set up two loops, increase 
the number of bulls if the digits and the indices match, increase the number 
of cows if just the digits match and the indices is different... but to do 
this in a size-efficient way is quite a challenge of its own. In order to 
actually compare the numbers digit by digit, we shift each of them by 
multiplying by 10 and taking the integer part, then, again, using the 
fractional trick on their integer parts difference divided by 10. We don't 
need the actual digit values to be whole, just need to get 0.0 difference to 
spot a match. If a match has been spotted, we increase the score by 1, but 
also increase it by another 9 if the indices are equal too. Finally, we 
restore the guess number after we've fully gone through the inner loop and 
the target number after we've fully gone through the outer loop by dividing 
them by 10000 (another constant that had to be placed into a register during 
the first-run setup).

As a "last-minute change", I decided to actually move the guess value
adjustment to the point _before_ the inner loop starts, so that the player 
actually can enter just four digits without the decimal point. This wasn't 
present in the first version of the game, by the way.

In the pseudocode, the main routine looks like this. Let's call the outer and
inner loop counters I and J, the target number T and the guess input G:

:MAIN
G = INPUT()
I = 4
SCORE = 0
:OUTERLOOP
G = G / 10000
T = T * 10
TI = INTG(T)
J = 4
:INNERLOOP
G = G * 10
C = FRAC((INTG(G) - TI)/10)
IF C = 0 THEN GOTO :IDXCOMP ELSE GOTO :INNEREND
:IDXCOMP
D = I - J
IF D = 0 THEN SCORE = SCORE + 10 ELSE SCORE = SCORE + 1
:INNEREND
J = J - 1
IF J = 0 THEN GOTO :OUTEREND ELSE GOTO :INNERLOOP
:OUTEREND
I = I - 1
IF I = 0 THEN GOTO :SCORESHOW ELSE GOTO :OUTERLOOP
:SCORESHOW
T = T / 10000
OUTPUT(SCORE)
GOTO :MAIN

As such, the entire Bulls and Cows game can effectively be described with the
following algorithm (that should work on any platform supporting 10 digits 
before and 4 after the decimal point):

N = 1234567890
T = 0
I = 4
:GENLOOP
F = 10 ^ INTG(RAND() * (I+6))
DP = INTG(N/F)
DQ = DP / 10
T = T/10 + FRAC(DQ)
N = N + (INTG(DQ) - DP) * F
I = I - 1
IF I = 0 THEN GOTO :MAIN ELSE GOTO :GENLOOP
:MAIN
G = INPUT()
I = 4
SCORE = 0
:OUTERLOOP
G = G / 10000
T = T * 10
TI = INTG(T)
J = 4
:INNERLOOP
G = G * 10
C = FRAC((INTG(G) - TI)/10)
IF C = 0 THEN GOTO :IDXCOMP ELSE GOTO :INNEREND
:IDXCOMP
D = I - J
IF D = 0 THEN SCORE = SCORE + 10 ELSE SCORE = SCORE + 1
:INNEREND
J = J - 1
IF J = 0 THEN GOTO :OUTEREND ELSE GOTO :INNERLOOP
:OUTEREND
I = I - 1
IF I = 0 THEN GOTO :SCORESHOW ELSE GOTO :OUTERLOOP
:SCORESHOW
T = T / 10000
OUTPUT(SCORE)
GOTO :MAIN

This, however, is just pseudocode. Implementing it in the actual 12C
instruction set required quite a lot of non-trivial thinking.

Let's first perform some basic clearing:

01 CLx     ; get 0
02 STO 4   ; the target number T will be in R4
03 RCL PV  ; recall the 1234567890 constant
04 STO 2   ; copy into the operating register R2 as N
05 4       ; push 4 as the iteration counter I
06 STO 0   ; save it into R0

Let's now see how the RAND() function is implemented. We'll use the 7-step
generator described in my previous post (using Rn as the preset seed value 
from 0 to 1):

07 RCL n   ; generation loop start, recall Rn
08 FRAC    ; and take its fractional part
09 3       ; push 3
10 +       ; add 3
11 e^x     ; take its natural exponent
12 12x     ; multiply by 12 and rewrite Rn
13 FRAC    ; return its fractional part for further usage

Then we convert the value returned by the generator into our factor value
(assuming the iteration counter I already is in R0 and the constant 10 is in 
R5):

14 6       ; push 6
15 RCL 0   ; push the iteration counter
16 +       ; add it
17 x       ; multiply by the generator output
18 INTG    ; get the integer index of the target digit
19 RCL 5   ; push 10
20 x<>y    ; prepare for 10 ^ index
21 y^x     ; calculate it
22 STO 1   ; store the factor F into R1

Yes, it took 16 steps just to calculate the F = 10 ^ INTG(RAND() * (I+6))
line from the algorithm. In 15C with its built-in PRNG, it would take 10 at 
most. But we're prepared to carry that burden, so let's move on. Here, we 
just calculate DP = INTG(N/F):

23 RCL 2   ; recall N
24 x<>y    ; swap the args
25 /       ; divide N by F (already in the stack)
26 INTG    ; get DP

That's 26 steps in total already. Now, the problem is that we don't have
enough registers to store both DP and DQ in, so we resort to some stack 
manipulation to calculate these lines:

DQ = DP / 10
T = T/10 + FRAC(DQ)
N = N + (INTG(DQ) - DP) * F

In the actual code (using (X Y Z T) stack diagrams to make the picture
clearer):

27 ENTER   ; push DP
28 ENTER   ; (DP DP - -)
29 RCL 5   ; (10 DP DP -)
30 STO / 4 ; T = T/10
31 /       ; (DQ DP - -)
32 FRAC    ; (FRAC(DQ) DP - -)
33 STO + 4 ; T = T + FRAC(DQ)
34 LSTx    ; (DQ FRAC(DQ) DP -)
35 -       ; (-INTG(DQ) DP - -)
36 +       ; (DP-INTG(DQ) - - -)
37 CHS     ; (INTG(DQ)-DP - - -)
38 RCL 1   ; recall the factor F
39 x       ; F * (INTG(DQ) - DP)
40 STO + 2 ; N = N + F * (INTG(DQ) - DP)

14 more steps. But it's time to finish the loop.

41 1       ; push 1
42 STO - 0 ; I = I - 1
43 RCL 0   ; recall the I value from R0
44 x=0     ; if reached 0
45 GTO 47  ; go further
46 GTO 07  ; otherwise repeat the loop

Now, after the long journey of 46 steps, we finally have our desired unique
four digits abcd in the 0.abcd format in R4. We're ready to prompt the 
player to enter the number to guess, right? Well, almost. We need to clear 
the score first, which is gonna be stored in R3. We can make use of the fact 
the X register now contains 0, but at the end of the loop it will contain 
the score itself. Then, we input the guess G and store it into R2:

47 STO - 3 ; UI loop start, clear the score register R3
48 R/S     ; prompt for the guess number G
49 STO 2   ; store it into R2

Then, let's store 4 into the outer loop counter I in R0:

50 4       ; push 4
51 STO 0   ; store it in R0 (outer loop counter)

Start the outer loop by setting up the same for the inner loop counter in R1
and adjusting the guess number G to move it into the correct decimal range 
(either after the initial input, or after previous multiplications):

52 4       ; outer loop start, push 4
53 STO 1   ; store it in R1 (inner loop counter)
54 EEX     ; push 10000
55 4
56 STO / 2 ; restore the guess number G value

Then, perform the T = T * 10 and TI = INTG(T) steps (storing the TI value in
R6):

57 RCL 5   ; recall 10
58 STO x 4 ; multiply T (in R4) by 10 in place
59 RCL 4   ; recall T
60 INTG    ; get the integer part of the new T value (TI)
61 STO 6   ; store the TI value in R6

Now, we can start the inner loop. First we do the same as for the outer loop
but with a twist:

62 RCL 5   ; inner loop start, recall 10
63 STO x 2 ; multiply G (in R2) by 10 in place
64 RCL 2   ; recall G
65 INTG    ; get the integer part of the new G value
66 RCL 6   ; recall the TI value

The twist is, we don't need to store the INTG(G) value anywhere as we already
have it in the stack register Y, and the previously stored TI value is now 
in the register X, so we can just continue with our algorithm by computing C 
= FRAC((INTG(G) - TI)/10):

67 -       ; INTG(G) - TI
68 RCL 5   ; push 10
69 /       ; (INTG(G) - TI)/10
70 FRAC    ; C = FRAC((INTG(G) - TI)/10)

Again, the C value is ephemeral because we only need it to compare to 0,
which we do right now:

71 x=0     ; compare C to zero
72 GTO 74  ; proceed to compare indices if it is zero (the guess and target
digits match)
73 GTO 83  ; go to the inner loop finalization if it isn't

Now, we perform the index comparison and scoring logic in case a match is
found, and some advanced stack manipulation also is necessary here to save 
crucial step count:

74 RCL 5   ; push 10 beforehands (10 - - -)
75 1       ; push 1 beforehands (1 10 - -)
76 RCL 0   ; recall the outer index I (I 1 10 -)
77 RCL 1   ; recall the inner index J (J I 1 10)
78 -       ; compare them (I-J 1 10 10)
79 x=0     ; if they are equal
80 ROLL    ; (1 10 10 0)
81 ROLL    ; the stack top will be 10 if they are equal, 1 if not
82 STO + 3 ; add to the score in R3

The trickery here is with the correct stack setup. When the x=0 condition is
run, the stack is (I-J 1 10 10). If I-J is NOT 0, only one roll is 
performed, which makes the state (1 10 10 I-J) and returns 1 at the top. But 
if I-J actually is equal to 0, then the second roll will bring the stack to 
the (10 10 0 1) state, with 10 at the top. This top value is what finally 
gets added to the score variable in R3.

83 1       ; end of the inner loop, push 1
84 STO - 1 ; decrement the inner index
85 RCL 1   ; recall the inner index
86 x=0     ; if 0, then
87 GTO 89  ; go to the outer loop end
88 GTO 62  ; else repeat the inner loop

Now, we can finalize the outer loop too, not forgetting to reset the G value
to its initial state by multiplying by 10000:

89 1       ; outer loop end, push 1
90 STO - 0 ; decrement the outer index
91 RCL 0   ; recall the outer index
92 x=0     ; if 0, then
93 GTO 95  ; go to show the score
94 GTO 52  ; else repeat the outer loop

We've lastly come to the final part, where we prepare for the next guess
iteration by restoring the T value and showing the score before looping back 
to the new prompt:

95 EEX     ; push 10000
96 4
97 STO / 4 ; restore the target number T
98 RCL 3   ; show the score
99 GTO 47  ; go to the new input prompt

This is it. This is how it has been done, from the idea to the pseudocode to
the exact 99 steps taken by the program. I could make it a bit less but that 
would make the game require more setup than just f REG [seed] n 1234567890 
PV 10 STO 5 and still wouldn't make the program take less than 92 steps in 
order to free an extra register. Anyway, I'm pretty satisfied with the 
result, considering it has been achieved on a mere financial calculator with 
no internal PRNG or advanced programming features like indirect addressing, 
subroutines or loop helpers, with very few conditional instructions and 
limited storage arithmetic capabilities. And this, by the way, answers the 
"why" question: because I wanted to prove the point that a normal, classic 
HP 12C is also capable of running such a game, and that fact, despite all 
its limitations, puts it into a class much above "normal" scientific or 
financial calculators. Again, even though it has nowhere near the same 
programming capabilities as the MK-52/61, 15C or even 11C, it still is 
lightyears ahead of what e.g. Casio fx-3400P or Citizen SRP-145 programmable 
calculators could offer. As I already said, if it can run Bulls and Cows 
which can be programmed on the device itself, it can be considered a pocket 
computer. A very limited one, but a computer nonetheless. You won't get this 
game on a Casio fx-3400P, no matter how hard you try. Because it's just a 
totally different class of devices in a similar form factor.

And now, just for fun, here's a port of the very same algorithm (with slight
input/output decorations) to the TI-74 BASIC:

10 RANDOMIZE:N=1234567890:T=0
20 FOR I=10 TO 7 STEP -1
30 F=10^INT(RND*I)
40 DP=INT(N/F)
50 DQ=DP/10
60 IDQ=INT(DQ)
70 T=T/10+DQ-IDQ 
80 N=N+(IDQ-DP)*F
90 NEXT I
100 INPUT "Guess:";G
110 IF G<=0 THEN 270
120 BS,CS=0
130 FOR I=1 TO 4
140 G=G/10000
150 T=T*10
160 TI=INT(T)
170 FOR J=1 TO 4
180 G=G*10
190 C=(INT(G)-TI)/10
200 IF INT(C)<>C THEN 220
210 IF I=J THEN BS=BS+1 ELSE CS=CS+1
220 NEXT J
230 NEXT I
240 T=T/10000
250 PRINT USING "Bulls: #, cows: #",BS,CS:PAUSE
260 GOTO 100
270 PRINT USING "The number was ####, bye!",T*10000:PAUSE
280 END

The gameplay is the same, except you can also type 0 (or any negative number)
to give up, get the final number and exit the game. The game takes up just 
447 bytes out of the total 7710 in the internal TI-74S memory, but it just 
shows how drastically a slight increase of computing power simplified 
programming capabilities. The TI-74 is just as far ahead of the 12C as the 
12C is ahead of the fx-3400P. And the HP 41C, I guess, was somewhere in 
between the two, but closer to the TI-74, especially when it came to 
peripheral support. The latter, by the way, had been mostly ditched in the 
42S (except a proprietary infrared port), which many already considered a 
decline in the series. However, it ran on just three LR44 batteries instead 
of pretty non-standard N-size cells, while supporting the same programming 
language. As we can see, 1980s and 1990s were the time when the line between 
(programmable) calculators and computers started to blur as much as it 
could, but again, HP Voyagers, in my opinion, were among the first to blur 
it. And the fact that we can play Bulls and Cows even on the 12C just proves 
it one more time.

Since then, I have ported some other games to the 12C, like a dice simulator,
a Yahtzee box simulator, a customizable variant of Nim, even the entire 
Mugwump/Hunt the Hurkle/Snark trilogy and a 3D variant of Hurkle called 
Depth Charge. But this particular game... It, so to speak, in a way, still 
makes me horny.

--- Luxferre ---

[1]: gopher://hoi.st:70/0/docs/own/hp-12c-soft.txt