Re: Progressively slow pattern match

[Alan Mackenzie <acm@muc.de>]

'n Abend, Ralf!

On Wed, 17 May 2006, Ralf Angeli wrote:

>In AUCTeX there is a regexp used with `looking-at' where pattern
>matching seems to progressively get slower the longer a part of the
>(possible) match gets.  I reduced the regexp to a bare minimum for
>testing and the code now looks something like this:

>(looking-at "\\(%+\\)*foo")

>The problem occurs if this is used against a line with only %
>characters in it.  The more of these characters there are the slower
>it gets.  I checked the time one call of `looking-at' takes with
>(abs (- (float-time) (progn (looking-at "\\(%+\\)*foo") (float-time))))
>and got the following results (in seconds):

>%%%%%%%%%%                 0.0006
>%%%%%%%%%%%%%%%            0.0154
>%%%%%%%%%%%%%%%%%%%%       0.5132
>%%%%%%%%%%%%%%%%%%%%%%%%%  7.8058

>The regexp is used with `looking-at' for checking if there are LaTeX
>macros which have to be treated specially during paragraph movement.
>As paragraph movement is used quite extensively when a region is to be
>filled, users might get the notion that they are experiencing a hang
>if they have such line for visually separating parts in the file.

>Is this a deficiency in Emacs?  Is there a way matching can be sped up
>with this or maybe another, equivalent regexp?

It's a bad regexp.  It's got a sort of ambiguity, in that you've got both
a "+" and a "*" on the "%" you want to match.  The number of ways %+* can
divide a row of n %s is the number of ordered partions of n[*].  The
Emacs regexp engine tries out ALL of these, I think.  Each time it comes
to the non-% at the end, it goes back to try a different subdivision, to
see if that gives a longer match.  Evidently, each % you add doubles the
time the regexp engine takes, approximately (though I suspect that for
your last timining, the jump was only 4%, not 5%).

And yes, I've done this too.  :-(  The solution is to write either %+ or
%*, but not both together.

[*] As in, 4 = 3 + 1
             = 2 + 1 + 1
             = 1 + 2 + 1
             = 1 + 3
             = 2 + 2 
             = 1 + 1 + 2
             = 1 + 1 + 1 + 1



>Ralf

-- 
Alan.