doc: tc triggers nested tc, potentially with tmp metavar leak

This commit is contained in:
Daniel Selsam
2019-11-04 16:34:52 -08:00
committed by Leonardo de Moura
parent cf381f68b7
commit 7cddeaa0d3
4 changed files with 149 additions and 0 deletions

View File

@@ -2411,9 +2411,14 @@ expr type_context_old::complete_instance(expr const & e) {
expr const & m = new_arg;
expr m_type = instantiate_mvars(infer(m));
if (!has_expr_metavar(m_type) && is_class(m_type)) {
lean_trace(name({"type_context", "complete_instance"}), tout() << "about to synth: " << m_type << "\n";);
if (mk_nested_instance(m, m_type)) {
new_arg = instantiate_mvars(new_arg);
}
} else if (is_class(m_type)) {
if (is_class(m_type)) {
lean_trace(name({"type_context", "complete_instance"}), tout() << "would have synthed: " << m_type << "\n";);
}
}
} else {
new_arg = complete_instance(new_arg);
@@ -3943,6 +3948,7 @@ mk_pp_ctx(type_context_old const & ctx) {
void initialize_type_context() {
register_trace_class("class_instances");
register_trace_class(name({"type_context", "complete_instance"}));
register_trace_class(name({"type_context", "is_def_eq"}));
register_trace_class(name({"type_context", "mvar_deps"}));
register_trace_class(name({"type_context", "is_def_eq_detail"}));

View File

@@ -0,0 +1,44 @@
/-
Here is an artificial example where temporary metavariables created by typeclass resolution
would otherwise leak into nested typeclass resolution unless we catch it.
-/
class Foo (α : Type) : Type := (x : Unit)
class Bar (α : Type) : Type := (x : Unit)
class HasParam (α : Type) [Bar α] : Type := (x : Unit)
instance FooToBar (α : Type) [Foo α] : Bar α := {x:=()}
-- Note: there is an implicit `FooToBar` inside the second argument of the return type.
instance HasParamInst (α : Type) [h : Foo α] : HasParam α := {x:=()}
class Top : Type := (x : Unit)
-- Note: `AllFoo` is a weird, universal instance.
instance AllFoo (α : Type) : Foo α := {x:=()}
/-
When the subgoal `[@HasParam α (@FooToBar α (AllFoo α))]` is triggerred, `α` is not known.
This happens for two reasons:
1. The return type `Top` does not include `α`.
2. The universal `AllFoo α` instance obviates the need to take a building-block class as argument.
Thus we would be tempted to call nested typeclass resolution on `Foo <tmp-mvar>`.
-/
instance Bad (α : Type) [HasParam α] : Top := Top.mk ()
def foo [Top] : Unit := ()
set_option pp.all true
set_option trace.class_instances true
set_option trace.type_context.complete_instance true
#check @foo _
/-
[class_instances] class-instance resolution trace
[class_instances] (0) ?x_0 : Top := @Bad ?x_1 ?x_2
[class_instances] (1) ?x_2 : @HasParam ?x_1 (@FooToBar ?x_1 (AllFoo ?x_1)) := @HasParamInst ?x_3 ?x_4
[type_context.complete_instance] would have synthed: Foo ?x_3
-/

View File

@@ -0,0 +1,66 @@
/-
Here is an example where temporary metavariables from typeclass resolution
could spill into nested typeclass resolution, for which the outer TC succeeds
iff:
1. the inner TC is still called, even with the leaked tmp mvars
2. the inner TC is allowed to assign the leaked tmp mvars
This example will *NOT* work in Lean4.
Although it would be easy to allow inner TC to set the outer TC tmp mvars,
the issue is that the solution to the inner TC problem may not be unique,
and it would be extremely difficult to allow backtracking from the outer TC
through to the inner TC.
In Lean4, inner TC can be called with leaked temporary mvars from the outer TC,
but they are treated as opaque, just as regular mvars are treated in the outer TC.
So, this example will fail.
-/
class Foo (α : Type) : Type := (x : Unit)
class Zoo (α : Type) : Type := (x : Unit)
class Bar (α : Type) : Type := (x : Unit)
class HasParam (α : Type) [Bar α] : Type := (x : Unit)
instance FooToBar (α : Type) [f : Foo α] : Bar α :=
match f.x with
| () => {x:=()}
instance ZooToBar (α : Type) [z : Zoo α] : Bar α :=
match z.x with
| () => {x:=()}
instance HasParamInst (α : Type) [h : Foo α] : HasParam α := {x:=()}
class Top : Type := (x : Unit)
instance FooInt : Foo Int := {x:=()}
instance AllZoo (α : Type) : Zoo α := {x:=()}
instance Bad (α : Type) [HasParam α] : Top := Top.mk ()
set_option pp.all true
set_option trace.class_instances true
set_option trace.type_context.complete_instance true
def foo [Top] : Unit := ()
#check @foo _
/-
[class_instances] class-instance resolution trace
[class_instances] (0) ?x_0 : Top := @Bad ?x_1 ?x_2
[class_instances] (1) ?x_2 : @HasParam ?x_1 (@ZooToBar ?x_1 (AllZoo ?x_1)) := @HasParamInst ?x_3 ?x_4
[type_context.complete_instance] would have synthed: Foo ?x_3
[type_context.complete_instance] would have synthed: Foo ?x_3
failed is_def_eq
-/
def couldWork : Top :=
@Bad Int (@HasParamInst Int FooInt)
#print couldWork
/-
def couldWork : Top :=
@Bad Int (@HasParamInst Int FooInt)
-/

View File

@@ -0,0 +1,33 @@
/-
Here is an example where typeclass resolution triggers nested typeclass resolution.
-/
class Ring (α : Type) : Type := (x : Unit)
class Double (α : Type) : Type := (x : Unit)
class Foo (α : Type) [Double α] : Type := (x : Unit)
instance RingToDouble (α : Type) [Ring α] : Double α := Double.mk α ()
-- Note: The return type is really `@Foo α (@RingToDouble α s)`.
instance RingFoo (α : Type) [s : Ring α] : Foo α := Foo.mk α ()
instance RingInt : Ring Int := Ring.mk Int ()
instance DoubleInt : Double Int := Double.mk Int ()
def foo [@Foo Int DoubleInt] : Unit := ()
set_option pp.all true
set_option trace.class_instances true
set_option trace.type_context.complete_instance true
#check @foo _
/-
[class_instances] class-instance resolution trace
[class_instances] (0) ?x_0 : @Foo Int DoubleInt := @RingFoo ?x_1 ?x_2
[type_context.complete_instance] about to synth: Ring Int
[class_instances] class-instance resolution trace
[class_instances] (0) ?x_3 : Ring Int := RingInt
[class_instances] (1) ?x_2 : Ring Int := RingInt
@foo (@RingFoo Int RingInt) : Unit
-/