I was actually building something very similar to this, though I stopped about six months ago due to time constraints. The goal was to be able to check which of a set of regular expressions matches a given string, and then invoke PCRE or similar to get the subexpression capture. The end result: Django-style regex URL routing against an arbitrary number of expressions, in the same amount of time as routing against a single expression.
Of course, I generally never use any of those features except the capture. (And greedy/nongreedy matching, but that is irrelevant if you're not doing a subexpression capture)
When dealing with very very large datasets, the ability to filter fast is important. So it becomes a 2-stage (actually n-stage...) problem: filter the raw data, and remove the parts that are known irrelevant (alternately include only the 'maybes'), then do the more intensive processing on the smaller resultant data-set. When done in parallel it looks a lot like map-reduce :)
Greedy / non-greedy is one of the main things I use in regexes (well, second to capturing). That sucks. Though if it's non-capturing, I guess it makes sense.
How many people do use regexes for boolean operations? I can only think of an instance or two where I have, aside from regular input validation.
If that is the case, wouldn't we just use more specialized functions, I forget the boost_regex function for C++, but for PHP we can do strpos() instead strstr() to see simply if a string exists? In that case, I'm guessing the benchmarks would be closer.
Well that rules out nearly every use case I have for needing a regex in the first place.