Before this commit, we "stored" macro arguments using applications.
This representation had some issues. Suppose we use [m a] to denote a macro
application. In the old representation, ([m a] b) and [m a b] would have
the same representation. Another problem is that some procedures (e.g., type inference)
would not have a clean implementation.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
The main motivation is that we will be able to move equalities between universes.
For example, suppose we have
A : (Type i)
B : (Type i)
H : @eq (Type j) A B
where j > i
We didn't find any trick for deducing (@eq (Type i) A B) from H.
Before this commit, heterogeneous equality as a constant with type
heq : {A B : (Type U)} : A -> B -> Bool
So, from H, we would only be able to deduce
(@heq (Type j) (Type j) A B)
Not being able to move the equality back to a smaller universe is
problematic in several cases. I list some instances in the end of the commit message.
With this commit, Heterogeneous equality is a special kind of expression.
It is not a constant anymore. From H, we can deduce
H1 : A == B
That is, we are essentially "erasing" the universes when we move to heterogeneous equality.
Now, since A and B have (Type i), we can deduce (@eq (Type i) A B) from H1. The proof term is
(to_eq (Type i) A B (to_heq (Type j) A B H)) : (@eq (Type i) A B)
So, it remains to explain why we need this feature.
For example, suppose we want to state the Pi extensionality axiom.
axiom hpiext {A A' : (Type U)} {B : A → (Type U)} {B' : A' → (Type U)} :
A = A' → (∀ x x', x == x' → B x == B' x') → (∀ x, B x) == (∀ x, B' x)
This axiom produces an "inflated" equality at (Type U) when we treat heterogeneous
equality as a constant. The conclusion
(∀ x, B x) == (∀ x, B' x)
is syntax sugar for
(@heq (Type U) (Type U) (∀ x : A, B x) (∀ x : A', B' x))
Even if A, A', B, B' live in a much smaller universe.
As I described above, it doesn't seem to be a way to move this equality back to a smaller universe.
So, if we wanted to keep the heterogeneous equality as a constant, it seems we would
have to support axiom schemas. That is, hpiext would be parametrized by the universes where
A, A', B and B'. Another possibility would be to have universe polymorphism like Agda.
None of the solutions seem attractive.
So, we decided to have heterogeneous equality as a special kind of expression.
And use the trick above to move equalities back to the right universe.
BTW, the parser is not creating the new heterogeneous equalities yet.
Moreover, kernel.lean still contains a constant name heq2 that is the heterogeneous
equality as a constant.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
Unification constraints of the form
ctx |- ?m[inst:i v] == T
and
ctx |- (?m a1 ... an) == T
are delayed by elaborator because the produce case-splits.
On the other hand, the step that puts terms is head-normal form is eagerly applied.
This is a bad idea for constraints like the two above. The elaborator will put T in head normal form
before executing process_meta_app and process_meta_inst. This is just wasted work, and creates
fully unfolded terms for solvers and provers.
The new test demonstrates the problem. In this test, we mark several terms as non-opaque.
Without this commit, the produced goal is a huge term.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
Projections build more general solutions. This commit also adds a test that demonstrates the issue. Before this commit, the elaborator would produce the "constant" predicate (fun x, a + b = b + a).
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
The example tests/lua/simp1.lua demonstrates the issue.
The higher-order matcher matches closed terms that are definitionally equal.
So, given a definition
definition a := 1
it will match 'a' with '1' since they are definitionally equal.
Then, if we have a theorem
theorem a_eq_1 : a = 1
as a rewrite rule, it was triggering the following infinite loop when simplifying the expression "a"
a --> 1 --> 1 --> 1 ...
The first simplification is expected. The other ones are not.
The problem is that "1" is definitionally equal to "a", and they match.
The rewrite_rule_set manager accepts the rule a --> 1 since the left-hand-side does not occur in the right-hand-side.
To avoid this loop, we test if the new expression is not equal to the previous one.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
Convertability constraints are harder to solve than equality constraints, and it seems they don't buy us anything definitions. They are just increasing the search space for the elaborator.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
Now, we are again using the following invariant for simplifier_fn::result
The type of in the equality of the result is definitionally equal to the
type of the resultant expression.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
For example, in the hpiext axiom, the resultant equality if for (Type M+1)
axiom hpiext {A A' : TypeM} {B : A -> TypeM} {B' : A' -> TypeM} :
A = A' -> (∀ x x', x == x' -> B x = B' x') -> (∀ x, B x) = (∀ x, B' x)
even if the actual arguments A, A’, B, B’ "live" in a much smaller universe (e.g., Type).
So, it would be great if we could move the resultant equality back to the right universe.
I don't see how to do it right now.
The other solution would require a major rewrite of the code base.
We would have to support universe level arguments like Agda, and write the axiom hpiext as:
axiom hpiext {l : level} {A A' : (Type l)} {B : A -> (Type l)} {B' : A' -> (Type l)} :
A = A' -> (∀ x x', x == x' -> B x = B' x') -> (∀ x, B x) = (∀ x, B' x)
This is the first instance I found where it is really handy to have this feature.
I think this would be a super clean solution, but it would require a big rewrite in the code base.
Another problem is that the actual semantics that Agda has for this kind of construction is not clear to me.
For instance, sometimes Agda reports that the type of an expression is (Set omega).
An easier to implement hack is to support "axiom templates".
We create instances of hipext "on-demand" for different universe levels.
This is essentially what Coq does, since the universe levels are implicit in Coq.
This is not as clean as the Agda approach, but it is much easier to implement.
A super dirty trick is to include some instances of hpiext for commonly used universes
(e.g., Type and (Type 1)).
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
The optimization was incorrect if the term indirectly contained a metavariable.
It could happen if the term contained a free variable that was assigned in the context to a term containing a metavariable.
This commit also adds a new test that exposes the problem.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
Before this commit, the elaborator was solving constraints of the form
ctx |- (?m x) == (f x)
as
?m <- (fun x : A, f x) where A is the domain of f.
In our kernel, the terms f and (fun x, f x) are not definitionally equal.
So, the solution above is not the only one. Another possible solution is
?m <- f
Depending of the circumstances we want ?m <- (fun x : A, f x) OR ?m <- f.
For example, when Lean is elaborating the eta-theorem in kernel.lean, the first solution should be used:
?m <- (fun x : A, f x)
When we are elaborating the axiom_of_choice theorem, we need to use the second one:
?m <- f
Of course, we can always provide the parameters explicitly and bypass the elaborator.
However, this goes against the idea that the elaborator can do mechanical steps for us.
This commit addresses this issue by creating a case-split
?m <- (fun x : A, f x)
OR
?m <- f
Another solution is to implement eta-expanded normal forms in the Kernel.
With this change, we were able to cleanup the following "hacks" in kernel.lean:
@eps_ax A (nonempty_ex_intro H) P w Hw
@axiom_of_choice A B P H
where we had to explicitly provided the implicit arguments
This commit also improves the imitation step for Pi-terms that are actually arrows.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
It is not incorrect to use size, but it can easily overflow due to sharing.
The following script demonstrates the problem:
local f = Const("f")
local a = Const("a")
function mk_shared(d)
if d == 0 then
return a
else
local c = mk_shared(d-1)
return f(c, c)
end
end
print(mk_shared(33):size())
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
It now can handle (?m t) where t is not a locally bound variable, but ?m and all free variables in t are assigned.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
The idea is to support conditional equations where the left-hand-side does not contain all theorem arguments, but the missing arguments can be inferred using type inference.
For example, we will be able to have the eta theorem as rewrite rule:
theorem eta {A : TypeU} {B : A → TypeU} (f : ∀ x : A, B x) : (λ x : A, f x) = f
:= funext (λ x : A, refl (f x))
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
The elaborator was failing in the following higher-order constraint
ctx |- (?M a) = (?M b)
This constraint has solution, but the missing condition was making the elaborator to reduce this problem to
ctx |- a = b
That does not have a solution.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
The method is_proposition was using an optimization that became incorrect after we identified Pi and forall.
It was assuming that any Pi expression is not a proposition.
This is not true anymore. Now, (Pi x : A, B) is a proposition if B is a proposition.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
The universe constraint manager is more flexible now.
We don't need to start with a huge universe U >= 512.
We can start small, and increase it on demand.
If module mod1 needs it, it can always add
universe U >= 3
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
The elaborator was not handling correctly constraints of the form
ctx |- ?m << (Pi x : A, B)
and
ctx |- (Pi x : A, B) << ?m
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
This commit affects different modules.
I used the following approach:
1- I store the metavariable environment at unification_failure_justifications. The idea is to capture the set of instantiated metavariables at the time of failure.
2- I added a remove_detail function. It removes propagation steps from the justification tree object. I also remove the backtracking search space associated with higher-order unificiation. I keep only the search related to case-splits due to coercions and overloads.
3- I use the metavariable environment captured at step 1 when pretty printing the justification of an elaborator_exception.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
This modification improves the effectiveness of the process_metavar_inst procedure in the Lean elaborator.
For example, suppose we have the constraint
ctx |- ?M1[inst:0 ?M2] == a
If ?M1 and ?M2 are unassigned, then we have to consider the two possible solutions:
?M1 == a
or
?M1 == #0 and ?M2 == a
On the other hand, if ?M2 is assigned to b, then we can ignore the second case.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
There is a lot to be done. We should do the same for Nat, Int and Real.
We also should cleanup the file builtin.cpp and builtin.h.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
This commit also adds several new theorems that are useful for implementing the simplifier.
TODO: perhaps we should remove the declarations at basic_thms.h?
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
The idea is to allow users to define their own commands using Lua.
The builtin command Find is now written in Lua.
This commit also fixes a bug in the get_formatter() Lua API.
It also adds String arguments to macros.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
After this commit, in the type checker, when checking convertability, we first compute a normal form without expanding opaque terms.
If the terms are convertible, then we are done, and saved a lot of time by not expanding unnecessary definitions.
If they are not, instead of throwing an error, we try again expanding the opaque terms.
This seems to be the best of both worlds.
The opaque flag is a hint for the type checker, but it would never prevent us from type checking a valid term.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
The elaborator produces better proof terms. This is particularly important when we have to prove the remaining holes using tactics.
For example, in one of the tests, the elaborator was producing the sub-expression
(λ x : N, if ((λ x::1 : N, if (P a x x::1) ⊥ ⊤) == (λ x : N, ⊤)) ⊥ ⊤)
After, this commit it produces
(λ x : N, ¬ ∀ x::1 : N, ¬ P a x x::1)
The expressions above are definitionally equal, but the second is easier to work with.
Question: do we really need hidden definitions?
Perhaps, we can use only the opaque flag.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
The elaborator was failing in the following scenario:
- Failing constraint of the form
ctx |- ?m1 =:= ?m2
where
?m2 is assigned to ?m1,
and ?m1 is unassigned.
has_metavar(?m2, ?m1) returns true, and a cycle is incorrectly reported.
Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
