The fact that parsec and attoparsec exist and can be pressed into service with reasonable performance (I think?) on tasks for which regexps are suitable is probably another big part of the reason no one's done it yet. I expect much of the plumbing would wind up looking a lot like those, actually.
When I started out with Haskell, one of my early thoughts was about designing a DSL for Icon-style pattern matching; I dropped it when I realized I was reinventing (almost identically, at least for its lower level combinators) Parsec. Nothing really to be gained except from a tutelary standpoint. And the mapping from Icon patterns to regex patterns is pretty much mechanical if you phrase it so you aren't executing code in the middle.