@Kha @dselsam:
This hack was preventing us from making `Expr` a "real" Lean type.
This was bad for a few reasons:
- It was hard to extend/modify `Expr` in Lean since we would also have
to modify the C++ code that creates the `Expr` objects with the hidden
fields.
- `Expr.lam` and `Expr.forallE` were not following the Lean layout
standard where we sort fields by size. @Kha: recall we used that to
avoid a UB. The issue with `Expr.lam` and `Expr.forallE` is that they
have a "visible" field (`BinderInfo`), which is smaller than
hidden fields such as hash code.
- `Expr.fvar` had only one field at `Expr.lean,` but four behind the
scenes.
I added a new constructor `Local` that is only accessible from C++.
It is only used in legacy code we inherited from Lean2.
We will eventually delete it.
This refactoring was quite painful since many parts of the codebase
were mixing the new `Expr.fvar` with the old `Expr.local`.
I doubt I would be able to do it without the new staging framework
@Kha built.
BTW, some of the patches are horrible. I didn't care much since we
are going to deleted the super ugly files. That being said,
you should expect new weird bevaior due to `Expr.fvar` vs `Expr.local`.
Next step: use the new `ExprCachedData` to make all `Expr` hidden visibles
accessible from Lean.
checkpoint
This is a temporary hack. After we eliminate the old elaborator,
we will delete the C++ class `local_context`.
In Lean4, we will not have two different kinds of local context:
`local_ctx` and `local_context`.
The datastructures at `local_context` used to manage used user_names
introduce a lot of overhead. They do guarantee that `get_unused_name` is
`O(log(n))`, but they slowdown much more common operations such as:
local declaration creation/deletion. We create/delete local declarations
much more often than we use `get_unused_name`.
The corelib build time is now 34.18 secs on my desktop. It was 39.5 secs.
We will re-implement the type class resolution algorithm, and the new
implementation will not rely on a persistent cache. We will improve
performance by:
1) Using better indexing data-structures.
2) Using a local cache during the search.
memory_pool object introduces memory contention and unnecessary
complexity. Moreover, it actually reduces performance when we compile
Lean using JEMALLOC.
Here are the numbers for corelib
jemalloc with memory_pool: 13.83 secs
jemalloc without memory_pool: 13.60 secs
This is one of the performance problems at issue #1646.
The method `local_context::erase_user_name(local_decl const & d)` was
inefficient when there are many locals with the same user facing name.
For size 7 in the example described at issue #1646, the average size of the
declaration list was 400. Here are the runtimes for size 7
Before: 19.021 secs
After: 16.433 secs
There are more performance issues.
Fixes#1363
After error recovery has been implemented in the elaborator, a few
assumptions made in the type context are not valid anymore since we may
be recovering from errors, and the local and metavariable contexts may
be invalid.
I used the approach used in the class environment.
- find* methods return optional<...>
- get* methods throw exception for unknown elements
Remarks:
I preserved code patterns such as
optional<local_decl> d = lctx.find_local_decl(...)
lean_assert(d)
and did not convert them into
local_decl d = lctx.get_local_decl(...)
Reason: the intention is clear that the local must be defined there.
If it is not we should analyze the problem and decide whether we should
throw an exception or not.
However, I converted code patterns such as
local_decl d = *lctx.find_local_decl(...)
into
local_decl d = lctx.get_local_decl(...)
Disclaimer: this change fixes issue #1363, but it may obfuscate other bugs.
