Tcl Source Code

I experimented with modifying the code to suppress pullback/pushfwd when there's a circular loop of constraints involving the target state.  While this wasn't hard to code, it turns out not to work at all, because the problem case then ends up hitting the assert(NOTREACHED) in compact().  So my fear that the pullback processing is necessary was correct: not only the regex executor, but even the data representation used for CNFAs, depend on the assumption that pullback/pushfwd can eliminate *every* non-LACON constraint arc.

I don't know whether this was a fundamental oversight on Henry Spencer's part, or whether this assumption is really OK but there's some bug in the way pullback/pushfwd operate.  In any case, I'm clueless at the moment about how we might fix this.