The new `partial def`s allow us to define `fix` in Lean, but the Lean
implementation is not as efficient as the native one. The native one
in C++ use weak pointers to prevent a closure allocation at every
recursive invocation.
This commit also fixes the `fixCore` helper functions that were broken
after we switched to camelCase.
We have updated the test `fix1.lean` to demonstrate the native
implementation is faster. Here are the numbers on my desktop.
```
./run.sh fix1.lean 24
721420279
Time for 'native fix': 816ms
721420279
Time for 'fix in lean': 1.34s
```
After we erase types and proofs, `Decidable.toBool` can be replaced with
the identity function since `Decidable A` and `Bool` have the same
runtime representation. By eagerly expanding `toBool`, we introduce
unnecessary `cases` expressions.
`elim_jp1_fn` was incorrectly expanding join points that were used more
than once. The issue is that the `foreach` combinator "may" skip nodes
that have already been visited.
We can now write
```
@[init] def initFn : IO Unit := ...
```
instead of
```
def initFn : IO Unit := ...
@[init initFn] constant execInitFn : Unit := ()
```
In the Haskell proposal for top level mutable state
https://wiki.haskell.org/Top_level_mutable_state, they describe the
following problems with using the `IO` monad during initialization.
"A more serious problem is that there is nothing to prevent arbitrary
observable IO actions from appearing to the right of the arrow. If we
perform all actions before executing main, then import becomes a
side-effectful operation, rather than simply a way of bringing names
into scope; furthermore we must specify the order in which actions from
different modules are executed, which would appear to be difficult in
general. If we execute actions on demand (as the unsafePerformIO hack
does) then we are building an unsafe syntactic construct into the
language."
I believe this is not applicable to us. First, our imports are already
side-effectful since we update attributes and the order we import
modules already matters. Second, we have already a well-defined order
in which we import modules. Finally, all global constants are already
being initialized eagerly.
Their ACIO proposal (`init_io` in our implementation) is too restrictive
for what we want to do. For example, to implement an environment
extension mechanism like we have discussed, we would also need `io.ref.write` and
`io.ref.read`. I imagine, we would have a global table, and `register`
would update this table. These extra actions do not satisfy the ACIO restrictions
described in the Haskell proposal. From their document:
"AC stands for Affine Central.
An IO action u is affine if its effect is not indirectly observable, hence need not be performed if the result is unneeded. That is, if u >> v === v for all actions v.
It is central if its effect commutes with every other IO action. That is, if do { x <- u; y <- v; w } === do { y <- v; x <- u; w } for all actions v and w."
It feels like we would have to keep fighting with the ACIO
restrictions. As I said above, our initialization order is well
defined. So, we must document the `[init]` feature and tell users they
should be aware that the `import` is important for initialization
purposes, and that their initialization actions should be
affine central whenever possible.
@kha I found yet another bug in the specializer code :(
The bug is related to the previous bug fix where we try avoid
duplication of work by lambda abstracting let-variables.
We knew this could introduce type errors, but I thought it would only
happen in very complicated programs that make a heavy use of dependent
types. Actually, this is not the case. I just found an instance when
I was playing with the new parser.
