while/break - The pure-python FSM implementation to Rule Them All.

[Carl Cerecke]

Well, it doesn't quite rule them all, but it is fast: About three times
faster than using one function per state. Faster than using generators.
Faster than using code objects.

Some, possibly minor, problems:
1. The generated code is ugly.
2. The generated code can be quite large, depending on the shape of the
FSM (Maximum theoretical size left as an exercise for the reader ;-)
3. Not as fast as byte code hacks, or using pyrex/psyco. Peter Hansen is
right. One of those is likely a better solution if you don't need pure
python.

The example FSM has input alphabet ['a','b','c']
and looks like:
state 0:
a -> 1
b -> 2
c -> 1

state 1:
a -> 1
b -> 3
c -> 2

state 2:
a -> 1
b -> 2
c -> 3

state 3:
a -> 2
b -> 3
c -> 0

The algorithm basically transforms the FSM into a tree of while loops,
with breaks letting us go back up the tree, and a variable ("goto"!)
telling us where the chain of breaks should stop.

Note:
1. It would be more efficient with a labelled break (or a labelled
continue) such as is available in Java.
2. There are some opportunities for optimisation of the generated code.

Running this code will print out the generated FSM code for both a
while/break implementation of the FSM and a function-based
implementation. It then does some timing measurements.

Cheers,
Carl.

#!/usr/bin/env python

from StringIO import StringIO

# number of random inputs fed to the FSM (if we are using that input
generator)
r = 1000000
#r = 10

# specific string of input chars (if we are using the appropriate input
generator)
inp = ['a','a','b','a','a','c','b','c','b','c','c','c','c','c','b']

# list of states, each state is a dict describing transitions to other
states.
states = [
{'a':1,'b':2,'c':1},
{'a':1,'b':3,'c':2},
{'a':1,'b':2,'c':3},
{'a':2,'b':3,'c':0}]

ind = " "

class State(object):
def __init__(self, num):
self.num = num
self.gotos = {}

def show(self, indent = ""):
'''
Representaion of this state, and the states 'underneath' it
'''
print indent+`self.num`+':'
for i,g in self.gotos.items():
if type(g) == State:
print indent+"-"+i+"->"
g.show(indent+" ")
else:
print indent+"-"+i+"-> "+`g`

def code(self, out, lvl):
'''
Spit out code for a while/break based state machine
'''
print >>out,lvl+"while 1: # state",self.num
print >>out,lvl+ind+"#print '%d'"%self.num
print >>out,lvl+ind+"n = next()"
for i,g in self.gotos.items():
print >>out,lvl+ind+"if n == '"+i+"':"
if type(g) == State:
g.code(out,lvl+ind+ind)
else:
if g != self.num:
print >>out,lvl+ind+ind+"goto = "+`g`
print >>out,lvl+ind+ind+"break"
else:
print >>out,lvl+ind+ind+"continue"
print >>out,lvl+ind+"if n == None: return",self.num
print >>out,lvl+ind+"if goto != "+`self.num`+":"
print >>out,lvl+ind+ind+"break"

def functions(out,states, lvl = ""):
'''
Spit out code for a function-based state machine
'''
for num,transitions in enumerate(states):
print >>out,lvl+"def state_"+`num`+"():"
print >>out,lvl+ind+"#print '%d'"%num
print >>out,lvl+ind+"n = next()"
for i,g in transitions.items():
print >>out,lvl+ind+"if n == '"+i+"':"
print >>out,lvl+ind+ind+"return state_"+`g`
print >>out,lvl+ind+"if n == None: return None"


start = State(0)

# set up the hierarchy of State objects
def dfs(state, history):
for i,s in states[state.num].items():
if s in history:
state.gotos = s
else:
child = State(s)
state.gotos = child
dfs(child, history+)

dfs(start,[start.num])

#print start
#start.show()
fun = StringIO()
print >>fun,"def fsm():"
start.code(fun," ")
functions(fun,states)
def next_gen(): # for when we want specific input
for i in inp:
print i
yield i
yield None

import random
def next_genr(): # for when we want lots of input
for i in range(r):
n = random.choice(['a','b','c'])
#print n
yield n
yield None

next = next_genr().next

# show the generated code of both FSM implementations
print fun.getvalue()

exec fun.getvalue()
import time

# we want to ignore the cost of getting input, so we measure that first.
a = time.clock()
n = 1
while n:
n = next()
z = time.clock()
in_time = z-a
print "input time:",in_time


print "--"
next = next_genr().next
a = time.clock()
f = fsm()
z = time.clock()
print "while/break:",z-a-in_time

next = next_genr().next
a = time.clock()
state = state_0
while state:
state = state()
z = time.clock()
print "functions:",z-a-in_time

 

[Paul Rubin]

3. Not as fast as byte code hacks, or using pyrex/psyco. Peter Hansen
is right. One of those is likely a better solution if you don't need
pure python.

If you don't need pure python than your approach still beats
everything else. Just generate C code (or assembly code) instead of
Python.

 

[Carl Cerecke]

If you don't need pure python than your approach still beats
everything else. Just generate C code (or assembly code) instead of
Python.

If you are generating C code, then you can, instead, use the much
maligned goto to jump directly between states. All this
while/break/continue hackery is just to emulate the humble goto.

Using gotos is probably about as fast as you can get.

Cheers,
Carl.