There are many types of automata, but if your description of "automata" includes phrases like "when the VM forked" I'm guessing that you're doing something pret

There are many types of automata, but if your description of "automata" includes phrases like "when the VM forked" I'm guessing that you're doing something pretty different to what we're doing. :-) In our normal execution model, we also can't have a stack. Hyperscan is not allowed to go get more memory at run-time nor is it allowed to grab more storage proportional to the size of the input. When we had a capturing prototype, we broke a lot of these rules (you needed either O(N) or O(lg(N)) side storage). We never had something that entirely fit our normal strictures and still did capturing; we couldn't do it in streaming mode either. We had something that could exactly simulate what you'd get with libpcre in terms of capturing and quantifiers as long as you could live with block mode and some limits on scan size dictated by our side storage use. The main problem? It was complex, and we didn't have a customer for it (this was back when we were a startup).