Library HoTT.Tactics.EvalIn

(* -*- mode: coq; mode: visual-line -*-  *)

Evaluating tactics on terms

It sometimes happens, in the course of writing a tactic, that we have some term in an Ltac variable (more precisely, we have what Ltac calls a "constr") and we would like to act on it with some tactic such as cbv or rewrite. Ordinarily, such tactics only act on the current *goal*, and generally they have a version such as rewrite ... in ... which acts on something in the current *context*, but neither of these is the same as acting on a term held in an Ltac variable.
For some tactics, such as cbv and pattern, we can write eval TAC in H, where H is the term in question; this form *returns* the modified term so we can place it in another Ltac variable. However, other tactics such as rewrite do not support this syntax. (There is a feature request for it at https://coq.inria.fr/bugs/show_bug.cgi?id=3677.)
The following tactic eval_in TAC H fills this gap, allowing us to act by rewrite on terms in Ltac variables. The argument TAC must be a tactic that takes one argument, which is an Ltac function that gets passed the name of a hypothesis to act on, such as ltac:(fun H' rewrite H in H'). (Unfortunately, however, eval_in cannot be used to exactly generalize eval pattern in H; see below.)
There is also a variant called eval_in_using, which also accepts a second user-specified tactic and uses it to solve side-conditions generated by the first one. We actually define eval_in in terms of eval_in_using by passing idtac as the second tactic.
Ltac eval_in_using tac_in using_tac H :=
  
The syntax $(...)$ allows execution of an arbitrary tactic to supply a needed term. By prefixing it with constr: which tells Ltac to expect a term, we obtain a pattern constr:($(...)$) which allows us to execute an arbitrary tactic in the situation of a fresh goal. This way we avoid modifying the existing context, and we can also get our hands on a proof term corresponding to the stateful modification. We pose H in the fresh context so we can play with it nicely, regardless of if it's a hypothesis or a term. Then we run tac_in on the hypothesis to modify it, use exact to "return" the modified hypothesis, and give a nice error message if using_tac fails to solve some side-condition.
  let ret := constr:(ltac:(
                       let H' := fresh in
                       pose H as H';
                       tac_in H';
                       [ exact H'
                       | solve [ using_tac
                               | let G := match goal with |- ?Gconstr:(G) end in
                                 repeat match goal with H : _ |- _revert H end;
                                   let G' := match goal with |- ?Gconstr:(G) end in
                                   fail 1
                                        "Cannot use" using_tac "to solve side-condition goal" G "."
                                        "Extended goal with context:" G' ].. ])) in
  
Finally, we play some games to format the return value nicely. We want to zeta-reduce the let-in generated by pose, but not any other let-ins; we do this by matching for it and doing the substitution manually. Additionally, pose/exact also results in an extra idmap; we remove this with cbv beta, which unfortunately also beta-reduces everything else. (This is why eval_in pattern H doesn't strictly generalize eval pattern in H, since the latter doesn't beta-reduce.) Perhaps we want to zeta-reduce everything, and not beta-reduce anything instead?
  let T := type of ret in
  let ret' := (lazymatch ret with
              | let x := ?x' in @?P xconstr:(P x')
               end) in
  let ret'' := (eval cbv beta in ret') in
  constr:(ret'' : T).

Ltac eval_in tac_in H := eval_in_using tac_in idtac H.

Example eval_in_example : A B : Set, A = B A B.
Proof.
  intros A B H a.
  let x := (eval_in ltac:(fun H'rewrite H in H') a) in
  pose x as b.
we get a b : B We Abort, so that we don't get an extra constant floating around.
Abort.

Rewriting with reflexivity

As an example application, we define a tactic that takes a lemma whose definition is idpath and behaves like rewrite, except that it doesn't insert any transport lemmas like Overture.internal_paths_rew_r. In other words, it does a change, but leverages the pattern-matching and substitution engine of rewrite to decide what to change into.
We use a dummy inductive type since rewrite acts on the *type* of a hypothesis rather than its body (if any).
Inductive dummy (A:Type) := adummy : dummy A.

Ltac rewrite_refl H :=
  match goal with
    | [ |- ?X ] ⇒
      let dX' := eval_in ltac:(fun H'rewrite H in H') (adummy X) in
      match type of dX' with
        | dummy ?X'change X'
      end
  end.

Here's what it would look like with ordinary rewrite:
Example rewrite_refl_example {A B : Type} (x : A) (f : A B) :
  ap f idpath = idpath :> (f x = f x).
Proof.
  rewrite ap_1.
  reflexivity.
Show Proof. ==> (fun (A B : Type) (x : A) (f : A -> B) => Overture.internal_paths_rew_r (f x = f x) (ap f 1) 1 (fun p : f x = f x => p = 1) 1 (ap_1 x f))
Abort.

And here's what we get with rewrite_refl:
Example rewrite_refl_example {A B : Type} (x : A) (f : A B) :
  ap f idpath = idpath :> (f x = f x).
Proof.
  rewrite_refl @ap_1.
  reflexivity.
Show Proof. ==> (fun (A B : Type) (x : A) (f : A -> B) => 1)
Abort.