Remark: the attribute actions used by the frontend are all in IO.
These actions access attributes by name, and need access to the IO.ref
that contains all registered attributes in the system.
@kha Trying again :)
I started using the prefix `f` for the `Fin` version (e.g., `fswap` and
`fswapAt`). So, it seemed natural to use the prefix `f` for the `Fin`
versions of `get` and `set`.
BTW, is it too crazy to use `a[i]` (without spaces) as notation for
`a.get i`? The constraint is to make sure `f a [i]` is not ambiguous.
That is, `f a[i]` is `f (a.get i)` and `f a [i]` is `f a (i::nil)`.
This is doable with the new parser.
Then, we could have `a[i]` as notation for `a.fget i` if `i` is a `Fin`,
and `a.get i` if `i` is `Nat`.
Similarly, `a[i := v]` would be notation for `a.fset i v` if `i` is
`Fin`, and `a.set i` if `i` is `Nat`.
It would also be awesome to have
```
let a[i] := v in
e
```
as notation for
```
let a := a[i := v] in
```
This commit adds the auxiliary function `expand` to break a nasty
interaction between code specialization, erasure and memory reuse.
Suppose we have
```lean
let new_array := array.update array i v in
have pr : <some property that mentions array>, from <proof>,
f new_array pr
```
Suppose `f` is not marked with `[specialize]`. Then, `pr` is erased and we get:
```lean
let new_array := array.update array i v in
f new_array box(0)
```
If `RC(array) == 1`, then the update performs a destructive update, and
we are happy.
Now, assume that `f` is marked with `[specialize]`.
Moreover, function specialization occurs before erasure since we want to
be able to use the to-be-implemented user-defined rewriting rules after specialization.
When we specialize `f` we compute a closure of its dependencies, and one of them is array because of `pr`.
Thus, we get
```lean
let new_array := array.update array i v in
f_spec new_array array
```
after specialization and erasure, but now, we don't perfom the destructive update.
BTW, I assumed we should be able to reduce the arity of `f_spec` after
erasure since the parameter `array` be dead after it. However, I found the following TODO:
https://github.com/leanprover/lean4/blob/master/src/library/compiler/reduce_arity.cpp#L50
The array read and write operations are now called:
- "Comfortable" version (with runtime bound checks):
`Array.get` and `Array.set` like OCaml.
It is also consistent with `Ref.get` and `Ref.put`,
and `get` and `set` for `MonadState`.
- `Fin` version (without runtime bound checks):
`Array.index` and `Array.update` like in F*.
- `USize` version (without runtime bound checks and unboxing):
`Array.idx` and `Array.updt`.
cc @kha
Without these annotations, Lean will timeout when trying to synthesize
the type class instance `decidable_eq uint32`. The type class resolution
problem will produce the unification problem:
```
decidable (@eq uint32 a b) =?= decidable (@eq usize ?x ?y)
```
which Lean tries to solve by assigning `?x := a`.
During the assignment, the types of `?x` and `a` are unified with "full
force". Thus, we get the constraint
```
usize_sz =?= uint32_sz
```
which will take forever to be solved when peforming the computation in
unary arithmetic.
Remark: this commit also makes sure that `type_context` will not unfold
irreducible definitions when trying to unify/match the types.
The new test `type_class_performance1.lean` exposes the problem fixed
by this commit.
Most efficient hash functions use uint32/uint64 and produce values
that do not fit in out small nat representation. Thus, GMP big numbers
would have to be created.
The array dimension is now stored inside the array.
The main motivation is that it reflects the actual runtime implementation.
We need to store the array size to be able to GC it.
So, it feels silly to have the array size stored in each array object,
but we cannot use this information.
For example, in the `hashmap` we implemented the bucket array using
`array`, and we store the `size` of the array.
Same for the Lean3 `buffer` object.
The `buffer` object doesn't even need to exist.
The actual `array` implementation is the `buffer`
We use the same approach used to define rbtree:
1- Structure with minimal number of invariants, AND
2- well_formed inductive predicate
We can use the well_formed predicate to prove auxiliary invariants later.
Example: the keys stored in every bucket have the correct hash code.
This implementation does not depend on the tactic framework,
and it is not a mess like the one in mathlib.
