This is a hard coded extra case. It is not an instance of has_coe.
Even if we change has_coe to accomodate this case, it will not be a
satisfactory solution because this coercion depends on the element and
not the type, and the element usually contains metavariables.
We should eventually write a tactic for synthesizing coercions.
I kept a few core methods (e.g., exact_core and apply_core). Reason:
if we use default parameters
meta constant exact (e : expr) (md := semireducible) : tactic unit
then, we will not be able to write
to_expr p >>= exact
The workaround is
do t <- to_expr p, exact t
or
to_expr p >>= (fun x, exact x)
One alternative is to change how we handle default parameters, and
eta-expand applications that involve default parameters.
We may also have an attribute [eta_expand]. Then
attribute [eta_expand] foo
instructs the elaborator to automatically eta-expand foo-applications.
The attribute would give users more control, and avoid potential
performance problems. Without the attribute, then for every function
application the elaborator has to check the type and decide whether it
must be eta-expanded or not.
@gebner @kha What do you think?
- `eval_expr` instantiate assigned metavariables occuring in the input
expression.
- Rename pp.instantiate_goal_mvars to pp.instantiate_mvars
- `format_expr` also instantiates assigned metavariables before pretty printing
when pp.instantiate_mvars is set to true.
eval_expr creates auxiliary definitions in the VM. These auxiliary
definitions are gone after the VM finishes.
We store vm_obj's in the attribute_manager.
Before this commit, Lean was crashing in the following scenario:
1- A new caching_user_attribute is defined, and the user data structure
contains closures.
2- The closures are created using eval_expr.
3- When reusing the cached values, the system crashes when trying
to apply a closure created using eval_expr. The closure points to
an auxiliary definition that has already been deleted.
The new test exposes the problem. This is not a hypothetical scenario,
the new test is based on the Lean - Mathematica integration being
developed by @rlewis1988.
The fix consists in making sure we do not cache anything if
the VM environment has been updated by eval_expr.
I believe this is acceptable behavior. eval_expr is a very low level
tactic, and I don't see a good motivation for invoking it when
constructing the cache.
BTW, the test can be relaxed if the vm_attr does not contain closures.
However, it doesn't seem to pay off.
Another potential fix would be to propagate the definitions created
by eval_expr to the main environment. However, I think this is not
acceptable.
We will be flooding the main environment with useless temporary definitions
created by `eval_expr`.
This commit also stores the environment at caching time, and make
sure the cache is only reused if the current environment is a descendant
of the the one at caching time. This is fixing a different potential
bug.
