Simple, faster parser for "value terms" in Isabelle.
The Isabelle parser does some fancy mixfix stuff, but as a result it's quite complicated. For some tasks, such as parsing a deep embedding of a moderate-sized program, using the full parser can be overkill. Then, after we've parsed the term, we need to do full type inference. These two together can be incredibly slow – a list with a thousand elements can take 7 seconds to parse and type check.
In cases where we know ahead of time that our definition will only contain a subset of syntactic constructs, we don't need the full power of the Isabelle parser. Furthermore, if there are no binders, and the only functions are relatively simple data constructors, typechecking is much simpler. With these specialisations, the same list with a thousand elements takes around 0.8s to parse and type check - about ten times faster than Syntax.read_term. (This is, of course, still appallingly slow.)
Example usage
value_definition a_big_list :: "int list" where
"[1, 2, 3, 4, 5]"
Example times
(* 1.0s *)
value_definition tuples_val :: "(int \<times> int) list" where
"[(1, 1), (2, 2), (3, 3), (4, 4), (5, 5), <snip...> (995, 995), (996, 996), (997, 997), (998, 998), (999, 999), (1000, 1000)]"
(* 10s *)
definition tuples :: "(int \<times> int) list" where
"tuples = [(1, 1), (2, 2), (3, 3), (4, 4), (5, 5), <snip...> (995, 995), (996, 996), (997, 997), (998, 998), (999, 999), (1000, 1000)]"
For this example, tuples_val spends about ninety percent of its time inserting the definition into the local theory.
I'm not sure whether this can be improved.
See ValueParser_Test.thy for details. Not all examples have such a dramatic improvement; on more
realistic examples value_definition seems to be roughly twice the speed of definition.