Library mtac_tutorial

Tutorial for Mtac.

Author: Beta Ziliani
Institution: Max Planck Institute for Software Systems (MPI-SWS)

Introduction

Mtac is a typechecked language for proof automation. It consists of a monadic type M A for a type A, which is interpreted via the new operator run. The best way of understanding the type M A is as maybe A, so, for instance, a function of type M nat may return a natural number. It can also fail or loop forever, but it can never produce a value of a different type (that is, it is sound). We call functions of type M A Mtactics, to distinguish them from the usual tactics provided by Coq.
One of the key aspects of Mtac is that it subsumes Gallina, the language of Coq, and it inherits from Coq the beta delta iota zeta reduction rules. This makes programming tactics very pleasant, since developers only need to learn the new features and their semantics, since the rest is exactly the same. These new features are:
  • Exceptions,
  • Unbounded fixpoints,
  • Unification match,
  • Fresh name generation.
In this tutorial we illustrate these features, building up from simple examples. In order to execute the code in this file you will need to install our modified version of Coq. For details on how to install it, follow this link: Mtac home page

Simple examples

To begin working with the new language we need to import the M type.
Require Import mtac.

In addition, we import a couple of modules from the standard library that we are going to use in some examples.

Require Import Arith.Arith Arith.Div2.
Require Import Lists.List.
Require Import Strings.String.

Set Implicit Arguments.
Notation "x == y" := (beq_nat x y) (at level 60).

We start by showing the standard unit and bind operators, which in our language are called ret (for return) and bind. The language also defines the standard notation x <- a; b for bind. This example computes the value 1 by passing the result of computing 0 to the successor.

Definition produces_a_value :=
  x <- ret 0;
  ret (S x).

We check the type of the definition. It has type M nat.
Let's execute it using the new keyword run and print the result.

Definition the_value := run produces_a_value.

Print the_value.

The result should be the_value = 1 : nat. As you can see, run produces_a_value was replaced by the effect of computing the code in produces_a_value. Mathematically, run is a partial function from type M A to type A.

Exceptions

The monad includes exceptions, like the following silly example illustrates. Exceptions are constructed with the constructor exception. In order to make distinguishable exceptions we make them opaque, sealing the definition with the Qed word.

Definition AnException : Exception.
  exact exception.
Qed.


Definition MyException (s : string) : Exception.
  exact exception.
Qed.

Note how they are equal to exception, but we can still differentiate them.

Definition test_ex e :=
  mtry raise e
  with
  | AnException => ret ""%string
  | MyException "hello"%string => ret "world"%string
  | [s] MyException s => ret s
  end.

Definition empty_string := run test_ex AnException.
Definition world_string := run test_ex (MyException "hello"%string).
Definition other_string := run test_ex (MyException "other"%string).

Print empty_string.
Print world_string.
Print other_string.

Results should be the empty string, the string "world" and the string "other" respectively.
If an exception is not caught, then we get a meaningful error. The Fail command below will show the exception thrown by the code:

Fail Check (run (raise (MyException "This is printed out"%string))).

Unbounded fixpoints

Fixpoints in Coq should terminate to ensure soundness. Checking termination is hard, so Coq relies on a pretty restrictive syntactic condition to ensure termination. We allow non-termination in our language via an unbounded fixpoint, which we call mfix. For instance, an endless loop can be written simply as:

Definition endless_loop := mfix f (n : nat) : M False := f n.

In this definition we decided to add the type annotation M False, since otherwise it is impossible for the type inference mechanism to guess the type. It is important to note that the body of mfix should always be of type M.
Uncomment the code below and execute it: it will loop forever! You will have to interrupt the proof assistant (C-c C-c in Emacs).
Check (run (endless_loop 0)).

Endless loop... Is it still safe?

The key to understanding why it is perfectly safe to allow for such effects is to notice that run is not a function living in the kernel typechecker of Coq. That is, for t of type M A, run t constructs a witness for A only if it's safe to do so, but it itself is not a witness for A. Take as example the definitions we constructed so far: we used run but when we printed them we saw no run in their proof terms.
As an exercise, we can try to break soundness of Coq by constructing an element of type False without any further hypothesis. Take the function endless_loop above, which has type nat -> M False. To get an element of type False we have to execute it through run as in the commented code. Since it will not terminate, run (endless_loop 0) doesn't produce an offending witness.

Constructing Collatz sequences

To show the use of this unbounded fixpoint we define a function computing the Collatz sequence, which cannot be defined in vanilla Coq since its termination is a conjecture.


Definition collatz :=
  mfix f n :=
    let rest :=
      if n == 1 then
        ret nil
      else if is_even n then
        f (div2 n)
      else
        f (3 * n + 1)
    in
      s <- rest;
      ret (n :: s).

We try it with the value 6.
Definition the_sequence_6 := (run (collatz 6)).

Print the_sequence_6.
Result: (6 :: 3 :: 10 :: 5 :: 16 :: 8 :: 4 :: 2 :: 1 :: nil) : list nat

Unification match

Mtac provides a powerful new construct: the unification match. Unlike the native Coq pattern matching, the unification match let us specify any term as a pattern, even patterns containing variables bound in the context.
For instance, the code below shows a function that searches for an element in a list.

Definition NotFound : Exception.
  exact exception.
Qed.

Definition inlist A (x : A) :=
  mfix f (s : list A) :=
    mmatch s as s' return M (In x s') with
    | [s'] (x :: s') => ret (in_eq _ _)
    | [y s'] (y :: s') =>
      r <- f s';
      ret (in_cons y _ _ r)
    | _ => raise NotFound
    end.

Check inlist.

We also depart from the standard notation for patterns: since they may now refer to variables in the context, we need to specify a list of pattern variables, like [s'] in the first pattern. All the variables not included in this list should be bound by the context, like x in the same pattern, which is bound to the argument of the definition. That is, this pattern matches a list containing the element x in the head.
So far we have constructed the proof terms directly, without using the interactive mode of Coq. We can use any standard tactic (apply, refine, exact, set, ...) with run, although not all of them are suitable if we want to avoid writing inferable arguments. For instance, if we have to prove a goal of the form In x s for some list s and some element x, then we would like to use run (inlist _ _), that is, without specifying the arguments. This will help us build more robust proof scripts, since tomorrow we may replace x by some other element in the list and still get a valid proof script. In order to avoid writing the arguments, we use the tactic refine, which provides run with the goal to prove, unlike apply, as the following examples show:

Example x_in_zyx (x y z : nat) : In x (z :: y :: x :: nil).
Proof.
  refine (run (inlist _ _)).
Qed.

Example y_in_zyx (x y z : nat) : In y (z :: y :: x :: nil).
Proof.
  Fail apply (run (inlist _ (z :: y :: x :: nil))).
Abort.

Fail above shows that indeed it has failed to apply the Mtactic.
Of course, we can always provide the proof term directly instead of going into interactive mode. In this case we don't need to explicitly provide the arguments.

Example z_in_xyz (x y z : nat) : In z (x :: y :: z :: nil)
  := run (inlist _ _).

Ssreflect[5] users can use the apply: tactic, which is better than refine. An alternative is to use eval, which is similar to run, except that it performs the execution of the Mtactic after the type inference mechanism of Coq has done its job:

Example y_in_zyx (x y z : nat) : In y (z :: y :: x :: nil).
Proof.
   apply (eval (inlist _ _)).
Qed.

Interaction with Program

When writing tactics, we can use Program to avoid having to write the proof terms ourselves. As an example, we will extend our inlist function to handle list concatenation in order to handle more cases and get shorter proof terms. By using Program, Coq will ask us to provide (interactively) the proof terms for the cases where there is a hole (_) and it cannot guess what to fill in that hole.

Program
Definition inlist' A (x : A) :=
  mfix f (s : list A) : M (In x s) :=
    mmatch s as s' return M (In x s') with
    | [l r] l ++ r =>
      mtry
        il <- f l;
        ret _
      with _ =>
        ir <- f r;
        ret _
      end
    | [s'] (x :: s') => ret (in_eq _ _)
    | [y s'] (y :: s') =>
      r <- f s';
      ret (in_cons y _ _ r)
    | _ => raise NotFound
    end.
Next Obligation.
intros; apply in_or_app; left; assumption.
Qed.
Next Obligation.
intros; apply in_or_app; right; assumption.
Qed.

If the list is a concatenation of two lists l and r, we first try to search for the element on l and, if it fails, on r. Notice that the pattern is not a constructor, but the application of the function ++ to two lists. As mentioned before, we can use any Coq term as a pattern! It is important to make this case the first case of the match, as the unification of the scrutinee with the pattern takes into account beta delta iota zeta reductions. That is, if the concatenation case were put third in the match, then the list (x :: nil) ++ (z :: nil) will be matched against the pattern (x :: s'), by reducing it to (x :: z :: nil).
One problem with Program is that it generates big proof terms. Let's look at the proof terms generated in the obligations and plug those terms into the holes.

Print inlist'_obligation_1.
Print inlist'_obligation_2.

The important bits are in_or_app l r x (or_introl H) and in_or_app l r x (or_intror H). We write our function again filling in the holes with these two terms.

Definition inlist'' A (x : A) :=
  mfix f (s : list A) :=
    mmatch s as s' return M (In x s') with
    | [l r] l ++ r =>
      mtry
        il <- f l;
        ret (in_or_app l r x (or_introl il))
      with _ =>
        ir <- f r;
        ret (in_or_app l r x (or_intror ir))
      end
    | [s'] (x :: s') => ret (in_eq _ _)
    | [y s'] (y :: s') =>
      r <- f s';
      ret (in_cons y _ _ r)
    | _ => raise NotFound
    end.

Let's prove an example using the three functions just created to compare the proof terms they generate.

Example ex_inlist (x y z : nat) : In x ((y :: z :: nil)++(x :: z :: nil)).
Proof.
  refine (run (inlist _ _)).
Qed.

Example ex_inlist' (x y z : nat) : In x ((y :: z :: nil)++(x :: z :: nil)).
Proof.
  refine (run (inlist' _ _)).
Qed.

Example ex_inlist'' (x y z : nat) : In x ((y :: z :: nil)++(x :: z :: nil)).
Proof.
  refine (run (inlist'' _ _)).
Qed.

Print ex_inlist.
Print ex_inlist'.
Print ex_inlist''.

Inspect the result. The last example has the shortest proof term.

A simple tautology prover

We show by example some useful constructs for dealing with Higher Order Abstract Syntax (HOAS). As the driving example we will write a rudimentary tautology prover similar to that found in VeriML [1] and CPDT [2]. Compared to VeriML, our approach has the benefit that it doesn't require any special context treatment, since for us a context is nothing more than a Coq list. And unlike in the Ltac version presented in [2], we have meaningful types to prevent ourselves from shooting ourselves in the foot.

Warming the engine: a simple propositional prover

We start with a very simple propositional prover. It considers only three cases:
  • The proposition is True. In this case, it returns the trivial proof I.
  • The proposition is a conjunction of p1 and p2. In this case, it proves both propositions and returns the introduction form of the conjunction.
  • The proposition is a disjunction of p1 and p2. In this case, it tries to prove the proposition p1, and if it fails it tries to prove the proposition p2. The corresponding introduction form of the disjunction is returned.
  • In any other case, it raises an exception, since no proof could be found.

Definition simpl_prop_auto :=
  mfix f (p : Prop) : M p :=
    mmatch p with
    | True => ret I
    | [ p1 p2 ] p1 /\ p2 =>
         r1 <- f p1 ;
         r2 <- f p2 ;
         ret (conj r1 r2)
    | [p1 p2] p1 \/ p2 =>
         mtry
           r1 <- f p1 ;
           ret (or_introl r1)
         with _ =>
           r2 <- f p2 ;
           ret (or_intror r2)
         end
    | _ => raise NotFound
    end.

Given this definition we can easily discharge the following example.
Example ex1 : True /\ (False \/ True).
Proof.
  refine (run (simpl_prop_auto _)).
Qed.

Print ex1.

The proof term is exactly what we would have written by hand:
ex1 = conj I (or_intror I)

Adding a context

Our previous function is very limited since it cannot prove tautologies as simple as P -> P. To handle implications we need a list of hypotheses where we can search for a proof of the atom we are considering. We create a record type containing a proposition and a witness for the proposition.

Record dyn := Dyn { prop : Prop ; elem : prop }.

We will need to search a list of dyns to find a witness for some proposition. The search function below is similar to the inlist above, but keying on the prop projector of the record. We have to prepend Program because it calls a more agressive typechecker, otherwise it fails to notice that the element in the body of the first case should return a P.

Program
Definition search (P : Prop) :=
  mfix f (s:list dyn) : M P :=
    mmatch s with
    | [(x:P) s'] (Dyn x) :: s' => ret x
    | [d s'] d :: s' => f s'
    | _ => raise NotFound
    end.

The proposition in the Dyn constructor is implicit, since it can be inferred from the element, so we write Dyn x instead of Dyn A x.
The tautology prover takes a context c (e.g., a list of dyns) and a proposition. The first three cases are the same as before.

Program
Definition prop_auto' :=
  mfix f (c : list dyn) (p : Prop) : M p :=
    mmatch p as p' return M p' with
    | True => ret I
    | [ p1 p2 ] p1 /\ p2 =>
         r1 <- f c p1 ;
         r2 <- f c p2 ;
         ret (conj r1 r2)
    | [p1 p2] p1 \/ p2 =>
         mtry
           r1 <- f c p1 ;
           ret (or_introl r1)
         with _ =>
           r2 <- f c p2 ;
           ret (or_intror r2)
         end
    | [(p1 p2 : Prop)] p1 -> p2 =>
          nu (x:p1),
          r <- f (Dyn x :: c) p2;
          abs x r
    | [p':Prop] p' => search p' c
    end.

Let's look at the new case for handling the implication. We need to return an element of type M (p1 -> p2), that is, maybe a function from p1 to p2. Of course, we cannot simply write
ret (fun x:p1 => f (Dyn x :: c) p2)
since this code has type M (p1 -> M p2) which is not what we want. Instead, we use two new operators: nu and abs. The first one is analogous to the nu operator in [3] and [4].
nu has type forall A B, (A -> M B) -> M B where A and B are left implicit. The effect of computing nu (fun x=>b), where b : T B, is the result of executing b, replacing any occurrence of x with a fresh parameter a. If the execution results in a term ret t for some t with a not appearing free in it, then the value ret t is used as result for nu (fun x => b). Otherwise, a failure is raised. Intuitively, the idea is that it is safe to execute the body of a function as long as it doesn't get stuck (i.e., it shouldn't inspect its argument), and the returning value doesn't return the argument (i.e., it shouldn't violate the context).
abs abstracts over parameters created by nu. It has type forall A P (x : A), P x -> M (forall x, P x) where A and P are left implicit. If a is a parameter created by nu and t is a term with a appearing free in it, then abs a t is replaced by ret (fun x=>r), where r is t with a replaced by x. That is, a is abstracted from t.
Coming back to the implication case, we use nu to create a parameter x as a witness for p1. Then we add it to the list of hypothesis to prove p2 and get the result r, which may refer to x. Therefore, we use abs x r to abstract x from the result. We encourage the reader to check that the type of the whole expression returned in the implication case has type M (p1 -> p2).
Finally, we changed the last case of the algorithm: instead of throwing an error, now we search for a witness for the proposition in the list using the search function defined before.
We create a definition to avoid passing the empty list
Definition prop_auto P :=
  prop_auto' nil P.

We can now easily prove this tautology.
Example ex_with_implication (p q : Prop) : p -> q -> p /\ q.
Proof.
  refine (run (prop_auto _)).
Qed.

Again, the proof term generated is exactly what we would expect for such a proof.

Print ex_with_implication.

Result: ex_with_implication = fun (p q : Prop) (H : p) (H0 : q) => conj H H0

Getting first order

We can generalize our algorithm very easily to deal with forall and exists. Below is the code, where the first four cases and the last one are the same as before.

Program
Definition tauto' :=
  mfix f (c : list dyn) (p : Prop) : M p :=
    mmatch p as p' return M p' with
    | True => ret I
    | [p1 p2] p1 /\ p2 =>
         r1 <- f c p1 ;
         r2 <- f c p2 ;
         ret (conj r1 r2)
    | [p1 p2] p1 \/ p2 =>
         mtry
           r1 <- f c p1 ;
           ret (or_introl r1)
         with _ =>
           r2 <- f c p2 ;
           ret (or_intror r2)
         end
    | [(p1 p2 : Prop)] p1 -> p2 =>
          nu (x:p1),
          r <- f (Dyn x :: c) p2;
          abs x r
    | [A (q:A -> Prop)] (forall x:A, q x) =>
          nu (x:A),
          r <- f c (q x);
          abs x r
    | [A (q:A -> Prop)] (exists x:A, q x) =>
          X <- evar A;
          r <- f c (q X) ;
          b <- is_evar X;
          if b then
            raise NotFound
          else
            ret (ex_intro q X r)
    | [p':Prop] p' => search p' c
    end.

The forall case is similar to the implication case from before but taking into account the following:
  • The type of x is any type A, not just Prop.
  • The possible dependency of x in q, the body of the forall. This dependency is marked by making q a function from A to Prop. The unification algorithm used to unify the pattern with the proposition p will take care of instantiating q with a function taking an element of type A and returning the body of the forall.
  • The context is not extended.
For the existential case, we create a fresh meta-variable X via the command evar, which takes a type (in this case A) and returns a new meta-variable of that type. Then, we call the function recursively with the body q of the existential, replacing the argument x with X. Hopefully, the result will instantiate X and we return this as the witness for the existential. If not, that is, if X is still an uninstantiated meta-variable, then we raise an error.
As before, we create a definition to avoid passing the empty list:

Definition tauto P :=
  tauto' nil P.

Here is an example to test tauto:
Example ex_first_order (p q : nat -> Prop) :
  forall x, p x -> q x -> exists y, p y /\ q y.
Proof.
  refine (run (tauto _)).
Qed.

If we cannot instantiate an existential, then an error is thrown.
Example ex_fail (p q : nat -> Prop) :
  exists y, p y /\ q y.
Proof.
  Fail refine (run (tauto _)).
Abort.

Actually, we can omit the check for the existential and let the user come up with the witness by itself.

Delayed execution via eval

We mentioned brefly that with eval we can delay the execution of the Mtactic in order to get arguments from the goal. However, one must use it with care, as the proof term generated is bigger than with run:

Print y_in_zyx.

Note how the procedure executed inlist ... is included in the proof term.
The function eval is particularly useful when rewriting procedures returning equalities. Here is an example using boolean equality of natural numbers. Notice how we use the [H] notation after the right arrow in the pattern. The name H will be instantiated with a proof of equality of the scrutinee with the pattern.

Program
Definition eq_nats :=
  mfix f (x : nat) (y : nat) : M (x == y = true) :=
    mmatch (x, y) with
    | (x, x) => [H] ret _
    | [x1 x2] (x1 + x2, x2 + x1) => [H]
      ret _
    end.
Next Obligation.
  intros _ _ x y H.
  inversion H.
  symmetry.
  apply beq_nat_refl.
Qed.
Next Obligation.
  intros _ _ x y x1 x2 H.
  inversion H.
  rewrite beq_nat_true_iff.
  apply plus_comm.
Qed.

Example plus_S n m : n + m == m + n = true /\ m == m = true /\ n == n = true.
Proof.
  rewrite !(eval (eq_nats _ _)).
  auto.
Qed.

References

[1] VeriML: Typed Computation of Logical Terms inside a Language with Effects. Antonis Stampoulis and Zhong Shao. In Proc. 2010 ACM SIGPLAN International Conference on Functional Programming (ICFP'10).
[2] http://adam.chlipala.net/cpdt/
[3] Aleksandar Nanevski. Meta-programming with names and necessity. In Proceedings of the seventh ACM SIGPLAN international conference on Functional programming, ICFP'02, pages 206-217, New York, NY, USA, 2002. ACM.
[4] Carsten Schuermann, Adam Poswolsky, and Jeffrey Sarnat. The nabla-calculus. functional programming with higher-order encodings. In Proceedings of the 7th international conference on Typed Lambda Calculi and Applications, TLCA'05, pages 339-353, Berlin, Heidelberg, 2005. Springer-Verlag.
[5] http://www.msr-inria.inria.fr/Projects/math-components