This PR implements the main framework of the model search procedure for
the linarith component in grind. It currently handles only inequalities.
It can already solve simple goals such as
```lean
example [IntModule α] [Preorder α] [IntModule.IsOrdered α] (a b c : α)
: a < b → b < c → c < a → False := by
grind
example [IntModule α] [LinearOrder α] [IntModule.IsOrdered α] (a b c : α)
: a < b → b < c + d → a - d < c := by
grind
```
This PR implements the infrastructure for constructing proof terms in
the linarith procedure in `grind`. It also adds the `ToExpr` instances
for the reified objects.
This PR introduces an explicit `defeq` attribute to mark theorems that
can be used by `dsimp`. The benefit of an explicit attribute over the
prior logic of looking at the proof body is that we can reliably omit
theorem bodies across module boundaries. It also helps with intra-file
parallelism.
If a theorem is syntactically defined by `:= rfl`, then the attribute is
assumed and need not given explicitly. This is a purely syntactic check
and can be fooled, e.g. if in the current namespace, `rfl` is not
actually “the” `rfl` of `Eq`. In that case, some other syntax has be
used, such as `:= (rfl)`. This is also the way to go if a theorem can be
proved by `defeq`, but one does not actually want `dsimp` to use this
fact.
The `defeq` attribute will look at the *type* of the declaration, not
the body, to check if it really holds definitionally. Because of
different reduction settings, this can sometimes go wrong. Then one
should also write `:= (rfl)`, if one does not want this to be a defeq
theorem. (If one does then this is currently not possible, but it’s
probably a bad idea anyways).
The `set_option debug.tactic.simp.checkDefEqAttr true`, `dsimp` will
warn if could not apply a lemma due to a missing `defeq` attribute.
With `set_option backward.dsimp.useDefEqAttr.get false` one can revert
to the old behavior of inferring rfl-ness based on the theorem body.
Both options will go away eventually (too bad we can’t mark them as
deprecated right away, see #7969)
Meta programs that generate theorems (e.g. equational theorems) can use
`inferDefEqAttr` to set the attribute based on the theorem body of the
just created declaration.
This builds on #8501 to update Init to `@[expose]` a fair amount of
definitions that, if not exposed, would prevent some existing `:= rfl`
theorems from being `defeq` theorems. In the interest of starting
backwards compatible, I exposed these function. Hopefully many can be
un-exposed later again.
A mathlib adaption branch exists that includes both the meta programming
fixes and changes to the theorems (e.g. changing `:= by rfl` to `:=
rfl`).
With the module system there is now no special handling for `defeq`
theorem bodies, because we don’t look at the body anymore. The previous
hack is removed. The `defeq`-ness of the theorem needs to be checked in
the context of the theorem’s *type*; the error message contains a hint
if the defeq check fails because of the exported context.
This PR makes the equational theorems of non-exposed defs private. If
the author of a module chose not to expose the body of their function,
then they likely don't want that implementation to leak through
equational theorems. Helps with #8419.
There is some amount of incidential complexity due to how `private`
works in lean, by mangling the name: lots of code paths that need now do
the right thing™ about private and non-private names, including the
whole reserved name machinery.
So this includes a number of refactorings:
* The logic for calculating an equational theorem name (or similar) is
now done by a single function, `mkEqLikeNameFor`, rather than all over
the place.
* Since the name of the equational theorem now depends on the current
context (in particular whether it’s a proper module, or a non-module
file), the forward map from declaration to equational theorem doesn’t
quite work anymore. This map is deleted; the list of equational theorems
are now always found by looking for declaration of the expected names
(`alreadyGenerated). If users define such theorems themselves (and make
it past the “do not allow reserved names to be declared”) they get to
keep both pieces.
* Because this map was deleted, mathlib’s `eqns` command can no longer
easily warn if equational lemmas have already been generated too early
(adaption branch exists). But in general I think lean could provide a
more principled way of supporting custom unfold lemmas, and ideally the
whole equational theorem machinery is just using that.
* The ReservedNamePredicate is used by `resolveExact`, so we need to
make sure that it returns the right name, including privateness. It is
not ok to just reserve both the private and non-private name but then
later in the ReservedNameAction produce just one of the two.
* We create `foo.def_eq` eagerly for well-founded recursion. This is
needed because we need feed in the proof of the rewriting done by
`wf_preprocess`. But if `foo.def_eq` is private in a module, then a
non-module importing it will still expect a non-private `foo.def_eq` to
exist. To patch that, we install a `copyPrivateUnfoldTheorem :
GetUnfoldEqnFn` that declares a theorem aliasing the private one. Seems
to work.
This PR removes the `NatCast (Fin n)` global instance (both the direct
instance, and the indirect one via `Lean.Grind.Semiring`), as that
instance causes causes `x < n` (for `x : Fin k`, `n : Nat`) to be
elaborated as `x < ↑n` rather than `↑x < n`, which is undesirable. Note
however that in Mathlib this happens anyway!
This PR fixes a bug in the equality-resolution procedure used by
`grind`.
The procedure now performs a topological sort so that every simplified
theorem declaration is emitted **before** any place where it is
referenced.
Previously, applying equality resolution to
```lean
h : ∀ x, p x a → ∀ y, p y b → x ≠ y
```
in the example
```lean
example
(p : Nat → Nat → Prop)
(a b c : Nat)
(h : ∀ x, p x a → ∀ y, p y b → x ≠ y)
(h₁ : p c a)
(h₂ : p c b) :
False := by
grind
```
caused `grind` to produce the incorrect term
```lean
p ?y a → ∀ y, p y b → False
```
The patch eliminates this error, and the following correct simplified
theorem is generated
```lean
∀ y, p y a → p y b → False
```
This PR sets `ring := true` by default in `grind`. It also fixes a bug
in the reification procedure, and improves the term internalization in
the ring and cutsat modules.
This PR improves the rendering of hints in error messages by
consistently indenting diffs and splitting large diffs less granularly;
it also improves the ergonomics of `Lean.MessageData.hint`. Note that
the changes to the signature of `Lean.MessageData.hint` are breaking.
This PR depends on #8457.
This PR simplifies the interface between the `grind` core and the cutsat
procedure. Before this PR, core would try to minimize the number of
numeric literals that have to be internalized in cutsat. This
optimization was buggy (see `grind_cutsat_zero.lean` test), and produced
counterintuitive counterexamples.
This PR fixes the hash function used to implement congruence closure in
`grind`. The hash of an `Expr` must not depend on whether the expression
has been internalized or not.
This PR fixes two inappropriate uses of `whnfD` in `grind`. They were
potential performance foot guns, and were producing unexpected errors
since `whnfD` is not consistently used (and it should not be) in all
modules.
This PR implements `match`-expressions in `grind` using `match`
congruence equations. The goal is to minimize the number of `cast`
operations that need to be inserted, and avoid `cast` over functions.
The new approach support `match`-expressions of the form `match h : ...
with ...`.
This PR changes the behavior of `pp.showLetValues` to use a hoverable
`⋯` to hide let values. This is now false by default, and there is a new
option `pp.showLetValues.threshold` for allowing small expressions to be
shown anyway. For tactic metavariables, there is an additional option
`pp.showLetValues.tactic.threshold`, which by default is set to the
maximal value, since in tactic states local values are usually
significant.
This PR upstreams and extends the Mathlib `clear_value` tactic. Given a
local definition `x : T := v`, the tactic `clear_value x` replaces it
with a hypothesis `x : T`, or throws an error if the goal does not
depend on the value `v`. The syntax `clear_value x with h` creates a
hypothesis `h : x = v` before clearing the value of `x`. Furthermore,
`clear_value *` clears all values that can be cleared, or throws an
error if none can be cleared.
This PR adds `seal` commands at `grind_ite.lean` to workaround expensive
definitionally equality tests in the canonicalizer. The new module
system will automatically hide definitions such as `HashMap.insert` and
`TreeMap.insert` which are being unfolded by the canonicalizer in this
test.
This PR also adds a `profileItM` for tracking the time spent in the
`grind` canonicalizer.
This PR implements non-chronological backtracking for the `grind`
tactic. This feature ensures that `grind` does not need to process
irrelevant branches after performing a case-split that is not relevant.
It is not just about performance, but also the size of the final proof
term. The new test demonstrates this feature in practice.
```lean
-- In the following test, the first 8 case-splits are irrelevant,
-- and non-choronological backtracking is used to avoid searching
-- (2^8 - 1) irrelevant branches
/--
trace:
[grind.split] p8 ∨ q8, generation: 0
[grind.split] p7 ∨ q7, generation: 0
[grind.split] p6 ∨ q6, generation: 0
[grind.split] p5 ∨ q5, generation: 0
[grind.split] p4 ∨ q4, generation: 0
[grind.split] p3 ∨ q3, generation: 0
[grind.split] p2 ∨ q2, generation: 0
[grind.split] p1 ∨ q1, generation: 0
[grind.split] ¬p ∨ ¬q, generation: 0
-/
#guard_msgs (trace) in
set_option trace.grind.split true in
theorem ex
: p ∨ q →
¬ p ∨ q →
p ∨ ¬ q →
¬ p ∨ ¬ q →
p1 ∨ q1 →
p2 ∨ q2 →
p3 ∨ q3 →
p4 ∨ q4 →
p5 ∨ q5 →
p6 ∨ q6 →
p7 ∨ q7 →
p8 ∨ q8 →
False := by
grind (splits := 10)
```
This PR fixes `split` in the presence of metavariables in the target.
The fix consists of replacing an internal use of `apply` for
instantiating match splitters by a new, simpler variant `applyN`. This
new `applyN` is not prone to #8436, which is the ultimate cause for
`split` failing on targets containing metavariables.
---------
Co-authored-by: Sebastian Graf <sg@lean-fro.org>
Co-authored-by: Joachim Breitner <mail@joachim-breitner.de>
This PR reworks the `simp` set around the `Id` monad, to not elide or
unfold `pure` and `Id.run`
In particular, it stops encoding the "defeq abuse" of `Id X = X` in the
statements of theorems, instead using `Id.run` and `pure` to pass back
and forth between these two spellings. Often when writing these with
`pure`, they generalize to other lawful monads; though such changes were
split off to other PRs.
This fixes the problem with the current simp set where `Id.run (pure x)`
is simplified to `Id.run x`, instead of the desirable `x`.
This is particularly bad because the` x` is sometimes inferred with type
`Id X` instead of `X`, which prevents other `simp` lemmas about `X` from
firing.
Making `Id` reducible instead is not an option, as then the `Monad`
instances would have nothing to key on.
---------
Co-authored-by: Sebastian Graf <sg@lean-fro.org>
Co-authored-by: Kim Morrison <kim@tqft.net>
Co-authored-by: Paul Reichert <6992158+datokrat@users.noreply.github.com>
This PR introduces a `noConfusionType` construction that’s sub-quadratic
in size, and reduces faster.
The previous `noConfusion` construction with two nested `match`
statements is quadratic in size and reduction behavior. Using some
helper definitions, a linear size construction is possible.
With this, processing the RISC-V-AST definition from
https://github.com/opencompl/sail-riscv-lean takes 6s instead of 60s.
The previous construction is still used when processing the early
prelude, and can be enabled elsewhere using `set_option
backwards.linearNoConfusionType false`.
This PR fixes a bug where the unknown identifier code actions wouldn't
work correctly for some unknown identifier error spans and adjusts
several unknown identifier spans to actually end on the identifier in
question.
The following additional adjustments are made:
- The fallback mechanism of the unknown identifier code actions is
removed, since it could produce severely incorrect suggestions for
unknown identifier errors on fields.
- A performance bug when using the code action to import all unknown
identifiers is fixed.
- A bug that occurs when the elaborator produces multiple overlapping
completion infos is fixed.
- A bug in the snapshot selection that could cause it to wait for
snapshots in snapshots with non-canonical syntax is fixed.
- Some invariants of the snapshot tree are documented.
- The snapshot tree formatting is adjusted to display the final info
tree again.
This PR adds the attribute `[grind?]`. It is like `[grind]` but displays
inferred E-matching patterns. It is a more convinient than writing.
Thanks @kim-em for suggesting this feature.
```lean
set_option trace.grind.ematch.pattern true
```
This PR also improves some tests, and adds helper function
`ENode.isRoot`.
