Yes, you're right. I was fixating on the wrong interpretation of your post. We do have exactly the problem you describe (startup time) and the cross-over analys

Yes, you're right. I was fixating on the wrong interpretation of your post. We do have exactly the problem you describe (startup time) and the cross-over analysis with RE2 ("How much data do I have to scan to catch a simpler compiler") would be interesting against libpcre or another backtracker. Generally - if you can amortize the startup - we find that it's not that hard to outperform back-trackers. Some of this is chucking effort at the problem (writing SIMD implementations of some parts of the search problem) and could be replicated on the backtracking side of the fence. I do think it's simpler to make an NFA or DFA go fast, however, as the problem is very regular and predictable. The analysis of how you would do more than trivial SIMD optimizations on a true back-tracker break my brain. :-)