Thank you all for your time and answers.

I did use a state machine in an older, similar problem where I was reading the text instead of processing strings.

The strings I deal with can get to the multi-megabyte size range so reversing them is not really attractive.

I'll probably go with Roy Johnson's very neat solution unless benchmarking shows that a finite machine beats the regexp engine (which I doubt).

Again, thanks to all for an instructive thread.