This PR fixes two bugs in `grind`.
1. Model-based theory combination was creating type incorrect terms.
2. `Nat.cast` vs `NatCast.natCast` issue during normalization.
This PR fixes a regression where elaboration of a previous document
version is not cancelled on changes to the document.
Done by removing the default from `SnapshotTask.cancelTk?` and
consistently passing the current thread's token for synchronous
elaboration steps.
This PR eliminates another source of facts of the form `-1 *
NatCast.natCast x <= 0` for each `x : Nat` in the local context. These
facts are now stored internally in the cutsat state.
cc @kim-em
This PR adjusts the `TryThis` widget to also work in widget messages
rather than only as a panel widget. It also adds additional
documentation explaining why this change was needed.
This PR improves the normalization of `Bool` terms in `grind`. Recall
that `grind` currently does not case split on Boolean terms to reduce
the size of the search space.
This PR updates `rw?`, `show_term`, and other tactic-suggesting tactics
to suggest `expose_names` when necessary and validate tactics prior to
suggesting them, as `exact?` already did, and it also ensures all such
tactics produce hover info in the messages showing tactic suggestions.
This introduces a breaking change in the `TryThis` API: the `type?`
parameter of `addRewriteSuggestion` is now an `LOption`, not an
`Option`, to obviate the need for a hack we previously used to indicate
that a rewrite closed the goal.
Closes#7350
This PR fixes an issue in the cutsat counterexamples. It removes the
optimization (`Cutsat.State.terms`) that was used to avoid the new
theorem `eq_def`. In the two new tests, prior to this PR, `cutsat`
produced a bogus counterexample with `b := 2`.
This PR prevents redundant invocations to `markAsCutsatTerm` which would
trigger equalities of the form `x = x` being propagated. This redundancy
only affected performance and "polluted" trace messages with redundant
information.
This PR improves support for `Nat` in the `cutsat` procedure used in
`grind`:
- `cutsat` no longer *pollutes* the local context with facts of the form
`-1 * NatCast.natCast x <= 0` for each `x : Nat`. These facts are now
stored internally in the `cutsat` state.
- A single context is now used for all `Nat` terms.
The PR also introduces a mapping mechanism for all "foreign" types that
can be converted to `Int`. Currently, only `Nat` is supported, but
additional types will be added in the future.
This PR adds a new propagation rule for `Bool` disequalities to `grind`.
It now propagates `x = true` (`x = false`) from the disequality `x =
false` (`x = true`). It ensures we don't have to perform case analysis
on `x` to learn this fact. See tests.
This PR adds missing propagation rules for `LawfulBEq A` to `grind`.
They are needed in a context where the instance `DecidableEq A` is not
available. See new test.
This PR improves how `grind` normalizes dependent implications during
introduction.
Previously, `grind` would introduce a hypothesis `h : p` for a goal of
the form `.. ⊢ (h : p) → q h`, and then normalize and assert a
non-dependent copy of `p`. As a result, the local context would contain
both `h : p` and a separate `h' : p'`, where `p'` is the normal form of
`p`. Moreover, `q` would still depend on the original `h`.
After this commit, `grind` avoids creating a copy. The context will now
contain only `h : p'`, and the new goal becomes `.. ⊢ q (he.mpr_prop
h)`, where `he` is a proof of `p = p'`.
This PR adds a feature to `structure`/`class` where binders without
types on a field definition are interpreted as overriding the type's
parameters binder kinds in that field's projection function. The rules
are (1) only a prefix of the binders are interpreted this way, (2)
multi-identifier binders are allowed but they must all be for
parameters, (3) only parameters that appear in the declaration itself
(not from `variables`) can be overridden and (4) the updates will be
applied after parameter binder kind inference is done. Binder updates
are not allowed in default value redefinitions. Example application: In
the following, `(R p)` causes the `R` and `p` parameters to be explicit,
where normally they would be implicit.
```
class CharP (R : Type u) [AddMonoidWithOne R] (p : Nat) : Prop where
cast_eq_zero_iff (R p) : ∀ x : Nat, (x : R) = 0 ↔ p ∣ x
#guard_msgs in #check CharP.cast_eq_zero_iff
/-
info: CharP.cast_eq_zero_iff.{u} (R : Type u) {inst✝ : AddMonoidWithOne R} (p : Nat) [self : CharP R p] (x : Nat) :
↑x = 0 ↔ p ∣ x
-/
```
The rationale for (3) is that there are cases where a module starts with
a large `variables` list and a field only incidentally uses the binder.
Without the restriction, the field ends up depending on that variable,
counterintuitively causing it to be introduced as an additional
parameter for the type. Instead, there is an explicit error. The easy
fix is to add `: _`, which is the bare minimum to make the binder have a
type.
We should consider warning when binders shadow parameters.
Closes#3574
[Zulip
discussion](https://leanprover.zulipchat.com/#narrow/channel/270676-lean4/topic/RFC.3A.20adjust.20argument.20explicitness.20on.20typeclass.20projections/near/508584627)
Mathlib fixes:
https://github.com/leanprover-community/mathlib4/pull/23469
This PR changes how `{...}`/`where` notation ("structure instance
notation") elaborates. The notation now tries to simulate a flat
representation as much as possible, without exposing the details of
subobjects. Features:
- When fields are elaborated, their expected types now have a couple
reductions applied. For all projections and constructors associated to
the structure and its parents, projections of constructors are reduced
and constructors of projections are eta reduced, and also implementation
detail local variables are zeta reduced in propositions (so tactic
proofs should never see them anymore). Furthermore, field values are
beta reduced automatically in successive field types. The example in
[mathlib4#12129](https://github.com/leanprover-community/mathlib4/issues/12129#issuecomment-2056134533)
now shows a goal of `0 = 0` rather than `{ toFun := fun x => x }.toFun 0
= 0`.
- All parents can now be used as field names, not just the subobject
parents. These are like additional sources but with three constraints:
every field of the value must be used, the fields must not overlap with
other provided fields, and every field of the specified parent must be
provided for. Similar to sources, the values are hoisted to `let`s if
they are not already variables, to avoid multiple evaluation. They are
implementation detail local variables, so they get unfolded for
successive fields.
- All class parents are now used to fill in missing fields, not just the
subobject parents. Closes#6046. Rules: (1) only those parents whose
fields are a subset of the remaining fields are considered, (2) parents
are considered only before any fields are elaborated, and (3) only those
parents whose type can be computed are considered (this can happen if a
parent depends on another parent, which is possible since #7302).
- Default values and autoparams now respect the resolution order
completely: each field has at most one default value definition that can
provide for it. The algorithm that tries to unstick default values by
walking up the subobject hierarchy has been removed. If there are
applications of default value priorities, we might consider it in a
future release.
- The resulting constructors are now fully packed. This is implemented
by doing structure eta reduction of the elaborated expressions.
- "Magic field definitions" (as reported [on
Zulip](https://leanprover.zulipchat.com/#narrow/channel/113489-new-members/topic/Where.20is.20sSup.20defined.20on.20submodules.3F/near/499578795))
have been eliminated. This was where fields were being solved for by
unification, tricking the default value system into thinking they had
actually been provided. Now the default value system keeps track of
which fields it has actually solved for, and which fields the user did
not provide. Explicit structure fields (the default kind) without any
explicit value definition will result in an error. If it was solved for
by unification, the error message will include the inferred value, like
"field 'f' must be explicitly provided, its synthesized value is v"
- When the notation is used in patterns, it now no longer inserts fields
using class parents, and it no longer applies autoparams or default
values. The motivation is that one expects patterns to match only the
given fields. This is still imperfect, since fields might be solved for
indirectly.
- Elaboration now attempts error recovery. Extraneous fields log errors
and are ignored, missing fields are filled with `sorry`.
This is a breaking change, but generally the mitigation is to remove
`dsimp only` from the beginnings of proofs. Sometimes "magic fields"
need to be provided — four possible mitigations are (1) to provide the
field, (2) to provide `_` for the value of the field, (3) to add `..` to
the structure instance notation, (4) or decide to modify the `structure`
command to make the field implicit. Lastly, sometimes parent instances
don't apply when they should. This could be because some of the provided
fields overlap with the class, or it could be that the parent depends on
some of the fields for synthesis — and as parents are only considered
before any fields are elaborated, such parents might not be possible to
use — we will look into refining this further.
There is also a change to elaboration: now the `afterTypeChecking`
attributes are run with all `structure` data set up (e.g. the list of
parents, along with all parent projections in the environment). This is
necessary since attributes like `@[ext]` use structure instance
notation, and the notation needs all this data to be set up now.
This PR fixes the `markNestedProofs` procedure used in `grind`. It was
missing the case where the type of a nested proof may contain other
nested proofs.
This PR ensures that `grind` does not use `mkEqMP`. It often triggered
type errors because `grind` uses the `[reducible]` transparency setting
by default. Increasing the transparency setting to default was another
possible, but less efficient fix.
This PR implements basic model-based theory combination in `grind`.
`grind` can now solve examples such as
```lean
example (f : Int → Int) (x : Int)
: 0 ≤ x → x ≠ 0 → x ≤ 1 → f x = 2 → f 1 = 2 := by
grind
```
This PR improves the counterexamples produced by the cutsat procedure,
and adds proper support for `Nat`. Before this PR, the assignment for an
natural variable `x` would be represented as `NatCast.natCast x`.
This PR fixes the support for nonlinear `Nat` terms in cutsat. For
example, cutsat was failing in the following example
```lean
example (i j k l : Nat) : i / j + k + l - k = i / j + l := by grind
```
because we were not adding the fact that `i / j` is non negative when we
inject the `Nat` expression into `Int`.
This PR ensures that we use the same ordering to normalize linear `Int`
terms and relations. This change affects `simp +arith` and `grind`
normalizer.
This consistency is important in the cutsat procedure. We want to avoid
a situation where the cutsat state contains both "atoms":
- `「(NatCast.natCast x + NatCast.natCast y) % 8」`
- `「(NatCast.natCast y + NatCast.natCast x) % 8」`
This was happening because we were using different orderings for
(nested) terms and relations (`=`, `<=`).
