Koch Snowflake - codegolf

Python, 650 612 594 574 characters

n='\n'
S='_a/G\F I\n'
A=dict(zip(S,('III','   ','__/','  G','\  ','F__','   ','III','')))
B=dict(zip(S,('III','   ','\  ',' aF','/a ','  G','   ','III','')))
C=dict(zip(S,('___','aaa','/  ','GII','II\\','  F','   ','III','')))
def T(s):
 a=b=c=d=r=u''
 for k in s:
    a+=A[k];b+=B[k];c+=C[k]
    if k=='I':a=a[:-3]+('II\\'if'a '==d[1:3]else'GII'if' a'==d[:2]else 3*k)
    d=d[3:]
    if k==n:d=c.replace('____','__/F').replace('aaaa','aa  ').replace('/  a','/a  ').replace('a  F','  aF');r+=a+n+b+n+d+n;a=b=c=''
 return r
print T(T(T('__\n\G\n'))).translate({97:95,71:47,73:32,70:92})

This works by expanding the triangle by a factor of 3 each time. To do that, we need to keep track of whether each symbol is a left or right boundary (e.g. how / is expanded depends on which side of the / is the inside). We use different symbols for the two possible cases, as follows:

_: _, outside on the top
a: _, outside on the bottom
/: /, outside on the left
G: /, outside on the right
\: \, outside on the left
F: \, outside on the right
<space>: inside
I: outside

The d variable handles the special case where the expansion of an a needs to extend into the 3x3 in the next row.

MS-DOS 16 bit machine code: 199 bytes

Decode using this site, save as 'koch.com' file and execute from WinXP command prompt.

sCAAxo7ajsKLz/OquF9fulwvvUoBM9u+BADoiQDodgDocwDogADobQDoagDodwCK8TLSs0+I98cHDQrGRwIktAnNIf7GOO5+7MNWAVwBYwFsAXoBgwGJB4DDAsOIN/7D6QQA/suIF/7P6R0A/suAPyB1AogH/suIB8OBw/8AiDfpBgD+x4gX/sM4734Ciu84z30Cis/Dg8UIg8UCgf1WAXLzg+0Mw07/dgB0GV/o9v/o5v/o8P/o3f/o2v/o5//o1//o4f9Gww==

Update

Here's an easy-to-read assembler version:

  ; L-System Description
  ;
  ; Alphabet : F
  ; Constants : +, -
  ; Axiom : F++F++F
  ; Production rules: F -> F-F++F-F 
  ;
  ; Register usage:
  ;                             _        _
  ; bp = direction: 0 = ->, 1 = /|, 2 = |\, 3 = <-, 4 = |/_, 5 = _\|
  ; cl = min y, ch = max y
  ; bl = x (unsigned)
  ; bh = y (signed)
  ; si = max level

  ; clear data
  mov al,20h
  add dh,al
  mov ds,dx
  mov es,dx
  mov cx,di
  rep stosb
  mov ax,'__'
  mov dx,'/\'

  ; initialise variables
  mov bp,Direction0
  xor bx,bx
  mov si,4

  call MoveForward
  call TurnRight
  call TurnRight
  call MoveForward
  call TurnRight
  call TurnRight
  call MoveForward

  mov dh,cl
  xor dl,dl
  mov bl,79
OutputLoop:
  mov bh,dh
  mov w [bx],0a0dh
  mov b [bx+2],24h
  mov ah,9
  int 21h
  inc dh
  cmp dh,ch
  jle OutputLoop  
  ret

Direction0:
  dw MoveRight
  dw MoveUpRight
  dw MoveUpLeft
  dw MoveLeft
  dw MoveDownLeft
  dw MoveDownRight
Direction6:

MoveRight:
  mov w [bx],ax
  add bl,2
  ret

MoveUpRight:
  mov b [bx],dh
  inc bl
  jmp DecBHCheckY

MoveUpLeft:
  dec bl
  mov b [bx],dl
DecBHCheckY:  
  dec bh
  jmp CheckY

MoveLeft:
  dec bl  
  cmp b [bx],20h
  jne MoveLeftAgain
  mov [bx],al
MoveLeftAgain:
  dec bl  
  mov [bx],al
  ret

MoveDownLeft:
  add bx,255
  mov b [bx],dh
  jmp CheckY

MoveDownRight:
  inc bh
  mov b [bx],dl
  inc bl

CheckY:
  cmp bh,ch
  jle NoMaxChange
  mov ch,bh
NoMaxChange:  
  cmp bh,cl
  jge NoMinChange
  mov cl,bh
NoMinChange:  
  ret

TurnRight:
  add bp,8

TurnLeft:
  add bp,2

  cmp bp,Direction6
  jb ret
  sub bp,12
  ret

MoveForward:
  dec si
  push [bp]
  jz DontRecurse
  pop di
  call MoveForward
  call TurnLeft
  call MoveForward
  call TurnRight
  call TurnRight
  call MoveForward
  call TurnLeft
  call MoveForward
DontRecurse:
  inc si
  ret

Perl, 176 175 bytes

Posting this as a separate answer because it uses a binary source file, which is perhaps a bit cheaty. But considering that it’s still Perl source code, I think it’s remarkable that it beats the MS-DOS machine code solution!

Source as base64-encoded

JF89IsLApwag0dhnMmAmMEcGIAcGQNHYwsDRFLsQ0djCwKcGoNHYwsDRFDdbECYwcRUxe1DCwNEUuxDR2
CI7c14uXiR4PW9yZCQmOyQieCgkeD4+MykucXcoXCAvXyBfXy8gXC8gX18gX1wgLyBfXy9cX18pWyR4Jj
ddXmVnO3NeLnsyN31eJF89cmV2ZXJzZSQmO3l+L1xcflxcL347cHJpbnQkJi4kXy4kL15lZw==

Somewhat more readable

Replace all instances of /<[0-9a-f]+>/ with the relevant binary data:

# Raw data!
$_="<c2c0a706a0d1d86732602630470620070640d1d8c2c0d114bb10d1d8c2>".
   "<c0a706a0d1d8c2c0d114375b1026307115317b50c2c0d114bb10d1d8>";

# Decode left half of the snowflake (without newlines)
s^.^$x=ord$&;$"x($x>>3).qw(\ /_ __/ \/ __ _\ / __/\__)[$x&7]^eg;

# Reconstruct the right half and the newlines
s^.{27}^$_=reverse$&;y~/\\~\\/~;print$&.$_.$/^eg

In this version, the snowflake is encoded in the following way:

• The 8 bits in each byte are divided like this:

  +---+---+---+---+---+---+---+---+
  |      5 bits       |   3 bits  |
  +---+---+---+---+---+---+---+---+
            R               C
  

• R encodes a run of spaces. The longest run is 27 characters, so all runs fit into 5 bits.

• C encodes a sequence of characters which are simply looked up in the literal array. (I used to have slightly crazier encodings here where the array contained only / \ _, but the Perl code necessary to decode it was longer...)

• I am lucky that the binary data does not contain any "/' or \ that would need escaping. I didn’t plan for this. But even if it did, I probably could have just changed the order of the items in the array to fix that.

• It is amazing how simple this solution is compared to the tens of other solutions I went through before I came up with this. I experimented with many different bitwise encodings more complex than this one, and it never occurred to me that a simpler one might be worth it simply because the Perl code to decode it would be shorter. I also tried to compress repetitions in the data using variable interpolation (see the other answer), but with the newest version that doesn’t gain any characters anymore.