I wonder about /others/bar|baz|others/ if there are a lot of alternations, the probability of common prefixes decreases with the number of alternations or does it try to find common prefixes for close neighbours ?

In the last 2-3 months on p5p was mentioned a few times the non-linearity of the regexp engine. The problem wrt to the internal utf8 representation is that at a given byte position you cannot really say what is the closest "character" position without knowing a previous correct one (the start for example is assumed ok). Is that the only cause of non-linearity? could you use markers to always keep correct "last" char. positions (cases with lots of backtracking could benefit of this, no? or is all that taken care of already in some smart way)

Actually the "keeping all markers" trick reminds me for some reason of packrat parsing. Have you looked at packrat parsing in the context of regex? (wikipedia has links on the subject, and to the original haskell implementation of the algorithm) the algorithm seems to limit worse time behaviour of pathological NFA (non-posix) regexes to something roughly linear in the regex size

thanks --stephan

update: corrected typo if to is