[TUHS] Short history of 'grep'

[Doug McIlroy]

> I'm still trying to get my around about how a program such as "egrep"
> which handles complex patterns can be faster than one that doesn't

> Is there a simple explanation, involving small words?

First, the big-word explanation. Grep is implemented as a nondeterministic
finite automaton, egrep as a deterministic finite automaton. Theory folk 
abbreviate these terms to "NFA" and "DFA", but these are probably not
what you had in mind as "small words".

The way grep works is to keep a record of possible partial parsings
as it scans the subject text; it doesn't know which (if any) will
ultimately be the right one, hence "nondeterministic". For example, 
suppose grep seeks pattern '^a.*bbbc' in text 'a.*bbbbbbc'. When
it has read past 3 b's, the possible partial parses are 'a.*', 'a.*b',
'a.*bb' and 'a.*bbb'. If it then reads another b, it splits the
first partial parse into 'a.*' and 'a.*b', updates the next two
to 'a.*bb' and 'a.*bbb', and kills off the fourth. If instead it
reads a c, recognition is successful; if anything else, all partials 
are killed and recognition fails.

Egrep, by preprocessing the expression, produces separate code for
each of several possible states: "no b's", "1 b", 2 b's" and "3 bs".
When it's in state "1 b", for example, it switches on the next 
character into "2 b's" or fails, depending on whether the 
character is b or not--very fast. Grep, on the other hand, has to  
update all the live parses.

So if egrep is so fast, why do we have grep? One reason is that
grep only needs to keep a list of possible progress points
in the expression. This list can't be longer than the expression.
In egrep, though, each state is a subset of progress points.
The number of possible subsets is exponential in the length
of the expression, so the recognition machine that egrep constructs
before attempting the parse can explode--perhaps overflowing
memory, or perhaps taking so long in construction that grep
can finish its slow parse before egrep even begins its fast parse.

To revert to the words of theory folks, grep takes time O(m*n) and
space O(m) worst case, while egrep takes time and space O(2^m+n).
(2^m is an overestimate, but you get the idea.)

That's the short story. In real life egrep overcomes the exponential
by lazily constructing the machine--not generating a state until it
is encountered in the parse, so no more than n states get constructed.
It's a complex program, though, for the already fancy preprocessing 
must be interleaved with the parsing.

Egrep is a tour de force of classical computer science, and pays
off big on scanning big files. Still, there's a lot to say for 
the simple elegance of grep (and the theoretical simplicity
of nondeterministic automata). On small jobs it can often win.
And it is guaranteed to work in O(m) memory, while egrep may need
O(n). 

-------------------------------------------------

Ken Thompson invented the grep algorithm and published it in CACM.
A pointy-headed reviewer for Computing Reviews scoffed: why would
anybody want to use the method when a DFA could do the recognition
in time O(n)? Of course the reviewer overlooked the potentially
exponential costs of constructing a DFA.

Some years later Al Aho wrote the more complicated egrep in the 
expectation that bad exponential cases wouldn't happen in everyday life. 
But one did. This job took 30 seconds' preprocessing to prepare
for a fraction of a second's parsing. Chagrined, Al conceived
the lazy-evaluation trick to overcome the exponential and
achieved O(n) run time, albeit with a big linear constant.

In regard to the "short history of grep", I have always thought
my request that Ken liberate regular expressions from ed caused
him to write grep. I asked him one afternoon, but I can't remember
whether I asked in person or by email. Anyway, next morning I
got an email message directing me to grep. If Ken took it
from his hip pocket, I was unaware. I'll have to ask him.

Doug