Up to now, we've used the more manual part of Coq's tactic facilities. In this chapter, we'll learn more about some of Coq's powerful automation features: proof search via the auto tactic, automated forward reasoning via the Ltac hypothesis matching machinery, and deferred instantiation of existential variables using eapply and eauto. Using these features together with Ltac's scripting facilities will enable us to make our proofs startlingly short! Used properly, they can also make proofs more maintainable and robust to changes in underlying definitions. A deeper treatment of auto and eauto can be found in the UseAuto chapter in Programming Language Foundations.
There's another major category of automation we haven't discussed much yet, namely built-in decision procedures for specific kinds of problems: omega is one example, but there are others. This topic will be deferred for a while longer.
Our motivating example will be this proof, repeated with just a few small changes from the Imp chapter. We will simplify this proof in several stages.
First, define a little Ltac macro to compress a common pattern into a single command.
Thus far, our proof scripts mostly apply relevant hypotheses or lemmas by name, and one at a time.
The auto tactic frees us from this drudgery by searching for a sequence of applications that will prove the goal:
The auto tactic solves goals that are solvable by any combination of
Using auto is always safe
in the sense that it will never fail
and will never change the proof state: either it completely solves
the current goal, or it does nothing.
Here is a more interesting example showing auto's power:
Proof search could, in principle, take an arbitrarily long time, so there are limits to how far auto will search by default.
When searching for potential proofs of the current goal, auto considers the hypotheses in the current context together with a hint database of other lemmas and constructors. Some common lemmas about equality and logical operators are installed in this hint database by default.
We can extend the hint database just for the purposes of one
application of auto by writing auto using ...
.
Of course, in any given development there will probably be some specific constructors and lemmas that are used very often in proofs. We can add these to the global hint database by writing
Hint Resolve T.
at the top level, where T is a top-level theorem or a constructor of an inductively defined proposition (i.e., anything whose type is an implication). As a shorthand, we can write
Hint Constructors c.
to tell Coq to do a Hint Resolve for all of the constructors from the inductive definition of c.
It is also sometimes necessary to add
Hint Unfold d.
where d is a defined symbol, so that auto knows to expand uses of d, thus enabling further possibilities for applying lemmas that it knows about.
It is also possible to define specialized hint databases that can be activated only when needed. See the Coq reference manual for more.
Let's take a first pass over ceval_deterministic to simplify the proof script.
When we are using a particular tactic many times in a proof, we can use a variant of the Proof command to make that tactic into a default within the proof. Saying Proof with t (where t is an arbitrary tactic) allows us to use t1... as a shorthand for t1;t within the proof. As an illustration, here is an alternate version of the previous proof, using Proof with auto.
The proof has become simpler, but there is still an annoying amount of repetition. Let's start by tackling the contradiction cases. Each of them occurs in a situation where we have both
H1: beval st b = false
and
H2: beval st b = true
as hypotheses. The contradiction is evident, but demonstrating it is a little complicated: we have to locate the two hypotheses H1 and H2 and do a rewrite following by an inversion. We'd like to automate this process.
(In fact, Coq has a built-in tactic congruence that will do the job in this case. But we'll ignore the existence of this tactic for now, in order to demonstrate how to build forward search tactics by hand.)
As a first step, we can abstract out the piece of script in question by writing a little function in Ltac.
That was a bit better, but we really want Coq to discover the relevant hypotheses for us. We can do this by using the match goal facility of Ltac.
This match goal looks for two distinct hypotheses that have the form of equalities, with the same arbitrary expression E on the left and with conflicting boolean values on the right. If such hypotheses are found, it binds H1 and H2 to their names and applies the rwinv tactic to H1 and H2.
Adding this tactic to the ones that we invoke in each case of the induction handles all of the contradictory cases.
Let's see about the remaining cases. Each of them involves applying a conditional hypothesis to extract an equality. Currently we have phrased these as assertions, so that we have to predict what the resulting equality will be (although we can then use auto to prove it). An alternative is to pick the relevant hypotheses to use and then rewrite with them, as follows:
Now we can automate the task of finding the relevant hypotheses to rewrite with.
The pattern forall x, ?P x -> ?L = ?R matches any hypothesis of
the form for all x, some property of x implies some
equality.
The property of x is bound to the pattern variable
P, and the left- and right-hand sides of the equality are bound
to L and R. The name of this hypothesis is bound to H1.
Then the pattern ?P ?X matches any hypothesis that provides
evidence that P holds for some concrete X. If both patterns
succeed, we apply the rewrite tactic (instantiating the
quantified x with X and providing H2 as the required
evidence for P X) in all hypotheses and the goal.
One problem remains: in general, there may be several pairs of hypotheses that have the right general form, and it seems tricky to pick out the ones we actually need. A key trick is to realize that we can try them all! Here's how this works:
The big payoff in this approach is that our proof script should be more robust in the face of modest changes to our language. To test this, let's try adding a REPEAT command to the language.
REPEAT behaves like WHILE, except that the loop guard is checked after each execution of the body, with the loop repeating as long as the guard stays false. Because of this, the body will always execute at least once.
Our first attempt at the determinacy proof does not quite succeed: the E_RepeatEnd and E_RepeatLoop cases are not handled by our previous automation.
Fortunately, to fix this, we just have to swap the invocations of find_eqn and find_rwinv.
These examples just give a flavor of what hyper-automation
can achieve in Coq. The details of match goal are a bit
tricky (and debugging scripts using it is, frankly, not very
pleasant). But it is well worth adding at least simple uses to
your proofs, both to avoid tedium and to future proof
them.
To close the chapter, we'll introduce one more convenient feature of Coq: its ability to delay instantiation of quantifiers. To motivate this feature, recall this example from the Imp chapter:
In the first step of the proof, we had to explicitly provide a
longish expression to help Coq instantiate a hidden
argument to
the E_Seq constructor. This was needed because the definition
of E_Seq...
E_Seq : forall c1 c2 st st' st'', st = c1 => st' -> st' = c2 => st'' -> st = c1 ;; c2 => st''
is quantified over a variable, st', that does not appear in its
conclusion, so unifying its conclusion with the goal state doesn't
help Coq find a suitable value for this variable. If we leave
out the with, this step fails (Error: Unable to find an
instance for the variable st'
).
What's silly about this error is that the appropriate value for st' will actually become obvious in the very next step, where we apply E_Ass. If Coq could just wait until we get to this step, there would be no need to give the value explicitly. This is exactly what the eapply tactic gives us:
The eapply H tactic behaves just like apply H except that, after it finishes unifying the goal state with the conclusion of H, it does not bother to check whether all the variables that were introduced in the process have been given concrete values during unification.
If you step through the proof above, you'll see that the goal state at position 1 mentions the existential variable ?st' in both of the generated subgoals. The next step (which gets us to position 2) replaces ?st' with a concrete value. This new value contains a new existential variable ?n, which is instantiated in its turn by the following reflexivity step, position 3. When we start working on the second subgoal (position 4), we observe that the occurrence of ?st' in this subgoal has been replaced by the value that it was given during the first subgoal.
Several of the tactics that we've seen so far, including exists, constructor, and auto, have similar variants. For example, here's a proof using eauto:
The eauto tactic works just like auto, except that it uses eapply instead of apply.
Pro tip: One might think that, since eapply and eauto are more powerful than apply and auto, it would be a good idea to use them all the time. Unfortunately, they are also significantly slower -- especially eauto. Coq experts tend to use apply and auto most of the time, only switching to the e variants when the ordinary variants don't do the job.