More Pep for Boyer-Moore Grep (part 1 of 2)

[James A. Woods]

#  The chief defect of Henry King
   Was chewing little bits of string.

	-- Hilaire Belloc, Cautionary Tales [1907]

#  Attempt the end, and never stand to doubt
   Nothing's so hard but search will find it out.

	-- Robert Herrick, Hesperides [1648]

     The world does not need another 'grep' variant.  And so, what is this
we offer?  On the surface, the exact same 'egrep' actually, but underneath,
a swift Boyer-Moore hybrid, in C, which can beat assembler versions utilizing
microcoded string search instructions.  The offering, designed in the
Kernighanian sense to utilize the existing 'egrep' when it must, also
makes use of Mr. Henry Spencer's regexp(3) functions in an unusual way.
For the edification of those without on-line access to system source code,
the vendor-supplied 'egrep' is left in a pristine state.

     With code now wending its way to mod.sources, we obtain the following
results.  Times (in seconds) are all measured on a VAX 11/750 system running
BSD 4.2 on Fujitsu Eagles, although our 'egrep' has been run on the Sun 2,
V7 Unix/PDP 11, Vaxen configured with System V, and, for added effect, the
NASA Ames Cray 2.

			200K bytes       user   sys	notes

  (new) egrep  astrian /usr/dict/words	 0.4    0.5    implementation by "jaw"
	match	  "           "		 0.5    0.5    VAX-only (Waterloo)
	bm	  "           "		 1.1    0.6    Peter Bain's version 2
  (old) egrep     "           " 	 5.6    1.7    standard	

[note:  the output here is the single word "Zoroastrian".]

     Aha, you quip -- this is all very fine for the 99 and 44/100's percent
metacharacter-free world, but what about timing for shorter strings, character
folding, as well as for the more interesting universe of extended regular 
expressions?  Samples forthwith.  (Egrep below refers to the new one, times for
the /usr/bin code being about the same as above on most any pattern.)

	egrep 	 zurich		0.4  0.5	0 words output
	egrep -i zuRich  	0.4  0.5	1 
	egrep -i zeus  		0.6  0.6	1
	egrep -i zen  		0.7  0.6	11
	bm 	 zen  		2.2  0.6	10
	egrep 	 ZZ  		0.8  0.6	0
	bm 	 ZZ  		3.0  0.7	0
	egrep -c Z  		1.5  0.6	19
	bm -c 	 Z  		5.9  0.7	19

Admittedly, most people (or programs) don't search for single characters,
where Boyer-Moore is a bit slow, but it's important for the layered regular
expression approach described herein.  We might point out from the above that
the popular "fold" option crippled by 'bm' costs little; it's only a slight
adjustment of the precomputed "delta" table as well as a single character
array reference in a secondary loop.  Why has Bain claimed complexity for this?
Also, the times show that the inner loop chosen for our code (modeled after
the original speedup done by Boyer-Moore for the PDP 10) consistently betters
the "blindingly fast" version by a factor of two to three.  The tipoff was
from previous paper studies (esp. Horspool, see header notes in code) noting
that the algorithm should, when implemented efficiently, best typical microcode.
Now it does. 

	while ( (k += delta0 ( *k )) < strend )
		;		/* over 80% of time spent here */

is the key (modulo precomputation tricks), and takes but three or four
instructions on most machines.

     Basic method for regular expressions:

	(1) isolate the longest metacharacter-free pattern string via the
	    "regmust" field provided by H. Spencer's regcomp() routine.

	    (Non-kosher, but worth not re-inventing the wheel.
	    v8 folks just might have to reverse-engineer Spencer's
	    reverse-engineering to provide equivalent functionality.
	    You see, there are many more sites running his code than v8.
	    Besides, we enjoy using regexpr technology on itself.

	(2) for "short" input, submatching lines are passed to regexec().

	(3) for "long" input, start up a standard 'egrep' process via
	    popen() or equivalent.  Why not just use regexec()?  Unfortunately
	    for our application, Spencer's otherwise admirable finite-state
	    automaton exhibits poor performance for complex expressions.
	    Setting a threshold on input length, though not perfect, helps.
	    If pipes on Unix were free, we'd use this way exclusively.
	    Until then, we buy happiness for those who might

			egrep stuff /usr/spool/news/net/unix/*

	    or on other directories full of short files.

So,
	newegrep -i 'hoe.*g' 	    words 	1.2  1.1
					 	{shoestring,Shoenberg}
	newegrep '(a|b).*zz.*[od]$' words 	1.5  1.1
					 	{blizzard,buzzword,palazzo}
	oldegrep 				6.3  1.4
but,
	{new,old}egrep -c '(first|second)'	similar times (no isolate)

Again, we stress that given the different nature of the simulations of the two
nondeterministic reg. expr. state-machines (one functionless), cases can be
"cooked" to show things in a bad light, so a hybrid is warranted.
We can generally do better incorporating the Boyer-Moore algorithm directly
into the AT&T code.  For the last example, the abstraction

	(egrep first words &; egrep second words) | sort -u | wc

ideally would work better on a parallel machine, but if you're expecting
something as amazing in this draft as, say, Morwen B. Thistletwaite's 52-move
Rubik's Cube solution, you're in the wrong place.

     About options -- system V ones are supported (-c, -l, bonus -i for BSD);
the 'egrep' here just hands off patterns to the old code for things like -n, -b,
-v, and multiple patterns.  As a bone to throw to the enemies of the cat-v
school, there is a -h (halt after printing first match), but we don't talk
about it much.  Multiple patterns can done ala 'bm' but laziness in the
presence of lack of knowledge of where 'fgrep' wins has prevailed for version 1.

     Personally I feel that adapting ("internationalizing") the 'egrep' effort
for two-byte Kanji is FAR more important than tweeking options or tradeoffs,
so for you large-alphabet Boyer-Moore algorithm specialists, send ideas
this way.
     
     Further historical/philosophical comments follow in the sequel.

     James A. Woods (ames!jaw)
     NASA Ames Research Center