This issue is similar to a bug where `isDefEqOffset` was exposing
`Nat.add` when processing `HAdd.hAdd`.
Fixes#561
The example at issue #561 is now working, but we may have other places
where raw literals are being accidentally exposed.
This makes sure we can properly quote e.g. `deriving` clauses and avoids
a suspicious `eraseMacroScopes` call (though not at `Elab.Syntax`, since
categories do not have to be declaration names)
Try to improve the performance issue described at #587.
The issue is that Mathlib contains thousands of theorems where the
associated key for the discrimination tree is just
`Key.other`. The indexing is not effective for them. This happens because
1- Lambda expressions are indexed using `Key.other`. The
discrimination tree mainly focus on the first-order structure.
2- It unfolds reducible constants when inserting and retrieving
entries. The motivation is that users expect simp theorems to fire
modulo reducible constants.
Then, we have many theorems such as
```lean
map ?g ∘ map ?f = map (?g ∘ ?f)
```
when we expand the function composition on the left-hand side, we get
```lean
fun (x : List ?α) => map ?g (map ?f x)
```
Which is indexed as `Key.other`.
We should not avoid the `Array`s in the discrimination tree nodes
If the index is working effectively, these arrays are all very small.
In this commit, we try to address the problem by using a different
approach. When processing the root of a pattern, we interrupt
reduction as soon as the we hit something that would be indexed
as `Key.other`. Note that, in Lean 3, the root of a pattern also
receives special treatment.
New test demonstrates how to use them.
The user-defined extensions cannot be used in the same file where they
were declared because the `initialize` commands are only executed when
we import the modules containing them.
TODO: user-defined attributes.
Also consistently use binders as reference position for an elided binder type.
Before, type errors were always given extent 1, the length of the
synthetic `_` token.
