This PR implements support for **guards** in `grind_pattern`. The new
feature provides additional control over theorem instantiation. For
example, consider the following monotonicity theorem:
```lean
opaque f : Nat → Nat
theorem fMono : x ≤ y → f x ≤ f y := ...
```
We can use `grind_pattern` to instruct `grind` to instantiate the
theorem for every pair `f x` and `f y` occurring in the goal:
```lean
grind_pattern fMono => f x, f y
```
Then we can automatically prove the following simple example using
`grind`:
```lean
/--
trace: [grind.ematch.instance] fMono: f a ≤ b → f (f a) ≤ f b
[grind.ematch.instance] fMono: f a ≤ c → f (f a) ≤ f c
[grind.ematch.instance] fMono: f a ≤ a → f (f a) ≤ f a
[grind.ematch.instance] fMono: f a ≤ f (f a) → f (f a) ≤ f (f (f a))
[grind.ematch.instance] fMono: f a ≤ f a → f (f a) ≤ f (f a)
[grind.ematch.instance] fMono: f (f a) ≤ b → f (f (f a)) ≤ f b
[grind.ematch.instance] fMono: f (f a) ≤ c → f (f (f a)) ≤ f c
[grind.ematch.instance] fMono: f (f a) ≤ a → f (f (f a)) ≤ f a
[grind.ematch.instance] fMono: f (f a) ≤ f (f a) → f (f (f a)) ≤ f (f (f a))
[grind.ematch.instance] fMono: f (f a) ≤ f a → f (f (f a)) ≤ f (f a)
[grind.ematch.instance] fMono: a ≤ b → f a ≤ f b
[grind.ematch.instance] fMono: a ≤ c → f a ≤ f c
[grind.ematch.instance] fMono: a ≤ a → f a ≤ f a
[grind.ematch.instance] fMono: a ≤ f (f a) → f a ≤ f (f (f a))
[grind.ematch.instance] fMono: a ≤ f a → f a ≤ f (f a)
[grind.ematch.instance] fMono: c ≤ b → f c ≤ f b
[grind.ematch.instance] fMono: c ≤ c → f c ≤ f c
[grind.ematch.instance] fMono: c ≤ a → f c ≤ f a
[grind.ematch.instance] fMono: c ≤ f (f a) → f c ≤ f (f (f a))
[grind.ematch.instance] fMono: c ≤ f a → f c ≤ f (f a)
[grind.ematch.instance] fMono: b ≤ b → f b ≤ f b
[grind.ematch.instance] fMono: b ≤ c → f b ≤ f c
[grind.ematch.instance] fMono: b ≤ a → f b ≤ f a
[grind.ematch.instance] fMono: b ≤ f (f a) → f b ≤ f (f (f a))
[grind.ematch.instance] fMono: b ≤ f a → f b ≤ f (f a)
-/
#guard_msgs in
example : f b = f c → a ≤ f a → f (f a) ≤ f (f (f a)) := by
set_option trace.grind.ematch.instance true in
grind
```
However, many unnecessary theorem instantiations are generated.
With the new `guard` feature, we can instruct `grind` to instantiate the
theorem **only if** `x ≤ y` is already known to be true in the current
`grind` state:
```lean
grind_pattern fMono => f x, f y where
guard x ≤ y
x =/= y
```
If we run the example again, only three instances are generated:
```lean
/--
trace: [grind.ematch.instance] fMono: a ≤ f a → f a ≤ f (f a)
[grind.ematch.instance] fMono: f a ≤ f (f a) → f (f a) ≤ f (f (f a))
[grind.ematch.instance] fMono: a ≤ f (f a) → f a ≤ f (f (f a))
-/
#guard_msgs in
example : f b = f c → a ≤ f a → f (f a) ≤ f (f (f a)) := by
set_option trace.grind.ematch.instance true in
grind
```
Note that `guard` does **not** check whether the expression is
*implied*. It only checks whether the expression is *already known* to
be true in the current `grind` state. If this fact is eventually
learned, the theorem will be instantiated.
If you want `grind` to check whether the expression is implied, you
should use:
```lean
grind_pattern fMono => f x, f y where
check x ≤ y
x =/= y
```
Remark: we can use multiple `guard`/`check`s in a `grind_pattern`
command.
This PR adds per-module `.ilean` and `.olean` file size metrics, global
and per-module cycle counting, and adds back `lean --stat`-based
metrics. It also renames some `size/*` metrics to get rid of the name
`stdlib`.
This PR reverts https://github.com/leanprover/lean4/pull/11396, which
changed `set_library_suggestions` to create an auxiliary definition
marked with `@[library_suggestions]`, rather than storing `Syntax`
directly in the environment extension.
It wasn't tested properly.
Co-authored-by: Claude <noreply@anthropic.com>
This PR fixes a kernel type mismatch error in grind's denominator
cleanup feature. When generating proofs involving inverse numerals (like
`2⁻¹`), the proof context is compacted to only include variables
actually used. This involves renaming variable indices - e.g., if
original indices were `{0: r, 1: 2⁻¹}` and only `2⁻¹` is used, it gets
renamed to index 0.
The bug was that polynomials were correctly renamed via `varRename`, but
the variable index `x` stored in `cancelDen` constraints was passed
directly to the proof without renaming, causing a mismatch between the
polynomial's variable references and the theorem's variable argument.
Added `ringVarDecls` to track ring variable indices that need renaming,
similar to how `ringPolyDecls` tracks polynomials. The `mkRingContext`
function now also renames these variable indices.
See zulip discussion at [#nightly-testing > Mathlib status updates @
💬](https://leanprover.zulipchat.com/#narrow/channel/428973-nightly-testing/topic/Mathlib.20status.20updates/near/560575295).
🤖 Prepared with Claude Code
Co-authored-by: Claude <noreply@anthropic.com>
This PR changes `set_library_suggestions` to create an auxiliary
definition marked with `@[library_suggestions]`, rather than storing
`Syntax` directly in the environment extension. This enables better
persistence and consistency of library suggestions across modules.
The change requires a stage0 update before tests can be restored. After
CI updates stage0, a follow-up PR will restore the test cases.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
This PR fixes an issue where `grind` would fail after multiple
`norm_cast`
calls with the error "unexpected metadata found during internalization".
The `norm_cast` tactic adds mdata nodes to expressions, and when called
multiple times it creates nested mdata. The `eraseIrrelevantMData`
preprocessing function was using `.continue e` when stripping mdata,
which causes `Core.transform` to reconstruct the mdata node around the
visited children. By changing to `.visit e`, the inner expression is
passed back to `pre` for another round of processing, allowing all
nested mdata layers to be stripped.
Closes#11411🤖 Prepared with Claude Code
Co-authored-by: Claude <noreply@anthropic.com>
This PR sorts the declarations fed into ElimDeadBranches in increasing
size. This can improve performance when we are dealing with a lot of
iterations.
The motivation for this change is as follows. Currently the algorithm
for doing one step of abstract interpretation is:
```
for decl in scc do
interpDecl
if summaryChanged decl then
return true
return false
```
whenever we return true we run another step. Now suppose we are in a
situation where we have an SCC with one big decl in the front and then
`n` small ones afterwards. For each time that the small ones change
their summary, we will re-run analysis of the big one in the front.
Currently the ordering is basically at "random" based on how other
compilers inject things into the SCC. This change ensures the behavior
is consistent and at least somewhat intelligent. By putting the small
declarations first, whenever we trigger a rerun of the loop we bias
analyzing the small declarations first, thus decreasing run time.
Note that this change does not have much effect on the current pipeline
because: We usually construct the SCCs in a way such that small ones
happen to be in front anyways. However, with upcomping changes on
specialization this is about to change.
This PR implements the following `grind_pattern` constraints:
```lean
grind_pattern fax => f x where
depth x < 2
grind_pattern fax => f x where
is_ground x
grind_pattern fax => f x where
size x < 5
grind_pattern fax => f x where
gen < 2
grind_pattern fax => f x where
max_insts < 4
grind_pattern gax => g as where
as =?= _ :: _
```
This PR fixes the compilation of structure projections with unboxed
arguments marked `extern`, adding missing `dec` instructions. It led to
leaking single allocations when such functions were used as closures or
in the interpreter.
This is the minimal working fix; `extern` should not replicate parts of
the compilation pipeline, which will be possible via #10291.
This PR is a followup of #11381 and enforces the invariants on ordering
of closed terms and constants required by the EmitC pass properly by
toposorting before saving the declarations into the Environment.
This PR fixes a bug where the closed term extraction does not respect
the implicit invariant of the
c emitter to have closed term decls first, other decls second, within an
SCC. This bug has not yet
been triggered in the wild but was unearthed during work on upcoming
modifications of the
specializer.
This PR makes the library suggestions extension state available when
importing from `module` files.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude <noreply@anthropic.com>
This PR adds support for cleaning up denominators in `grind linarith`
when the type is a `Field`.
Examples:
```lean
open Std Lean.Grind
section
variable {α : Type} [Field α] [LE α] [LT α] [LawfulOrderLT α] [IsLinearOrder α] [OrderedRing α]
example (a b : α) (h : a < b / 2) : 2 * a < b := by grind
example (a b : α) (_ : 0 ≤ a) (h : a ≤ b) : a / 7 ≤ b / 2 := by grind
example (a b : α) (_ : b < 0) (h : a < b) : (3/2) * a < (5/4) * b := by grind
example (a b : α) (h : a = b * (3⁻¹)^2) : 9 * a ≤ b := by grind
example (a b : α) (h : a / 2 ≠ b / 9) : 9 * a < 2 * b ∨ 9 * a > 2 * b := by grind
example (a b : α) (h : a < b / (2^2 - 3/2 + -1 + 1/2)) : 2 * a < b := by grind
end
example (a b : Rat) (h : a < b / 2) : a + a < b := by grind
example (a b : Rat) (h : a < b / 2) : a + a ≤ b := by grind
example (a b : Rat) (h : a ≠ b * (3⁻¹)^2) : 9 * a < b ∨ 9 * a > b := by grind
example (a b : Rat) (h : a / 2 ≠ b / 9) : 9 * a < 2 * b ∨ 9 * a > 2 * b := by grind
```
This PR makes the `Std.Time.Format` import in
`Lean.Elab.Tactic.Grind.Annotated` private rather than public,
preventing the entire `Std.Time` infrastructure (including timezone
databases) from being re-exported through `import Lean`.
The `grindAnnotatedExt` extension is kept private, with a new public
accessor function `isGrindAnnotatedModule` exposed for use by
`LibrarySuggestions.Basic`.
This should address the +2.5% instruction increase on `import Lean`
observed after merging #11332.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude <noreply@anthropic.com>
This PR enables parallelism in `try?`. Currently, we replace the
`attempt_all` stages (there are two, one for builtin tactics including
`grind` and `simp_all`, and a second one for all user extensions) with
parallel versions. We do not (yet?) change the behaviour of `first`
based stages.
This PR implements a helper simproc for `grind`. It is part of the
infrastructure used to cleanup denominators in `grind linarith`.
---------
Co-authored-by: Kim Morrison <kim@tqft.net>
This PR adds a focused error explanation aimed at the case where someone
tries to use Natural-Numbers-Game-style `induction` proofs directly in
Lean, where such proofs are not syntactically valid.
## Discussion
The natural numbers game uses a syntax that overlaps with Lean's
`induction` syntax despite having more structural similarity to
`induction'`. This means that fully correct proofs in the natural
numbers game, like this...
```lean4
import Mathlib
theorem zero_mul (m : ℕ) : 0 * m = 0 := by
induction m with n n_ih
rw [mul_zero]
rfl
rw [mul_succ]
rw [add_zero]
rw [n_ih]
rfl
```
...have completely baffling error messages from a newcomers'
perspective:
```
notNaturalNumbersGame.lean:3:20: error: unknown tactic
notNaturalNumbersGame.lean:3:2: error: Alternative `zero` has not been provided
notNaturalNumbersGame.lean:3:2: error: Alternative `succ` has not been provided
```
(the Mathlib import here only provides the `ℕ` syntax here; equivalently
`ℕ` could be renamed to `Nat` and the import could be removed, [like
this](https://live.lean-lang.org/#codez=C4Cwpg9gTmC2AEAvMUIH1YFcA28AUCAXPAHICGwAlPMQAzwBU8CAvPPYWwEYCeAUPHgBLAHYATTAGNgQiCObwA7kNDx5ItEJAD4URfADaWbGmSoAujqgAzbFf1GcaAM5TJlwXsNkxY0yggPXQcNLSCbbCA))
There are many problems with this proof from the perspective of "stock"
Lean, but the error messages in the `induction` case are particularly
unfriendly and provide no guidance from a NNG learner's perspective.
This PR provides more information about what is wrong:
```
notNaturalNumbersGame.lean:3:20: error: unknown tactic
notNaturalNumbersGame.lean:3:14: error(lean.inductionWithNoAlts): Invalid syntax for induction tactic: The `with` keyword must followed by a tactic or by an alternative (e.g. `| zero =>`), but here it is followed by the identifier `n`.
```
The error explanation it links to explicitly flags the transition of
NNG-style proofs to Lean as the likely culprit, and gives an example of
an effective translation.
This PR adds a new [radar]-based [temci]-less bench suite that replaces
the `stdlib` benchmarks from the old suite and also measures per-module
instruction counts. All other benchmarks from the old suite are
unaffected.
The readme at `tests/bench-radar/README.md` explains in more detail how
the bench suite is structured and how it works. The readmes in the
benchmark subdirectories explain what each benchmark does and which
metrics it collects.
All metrics except `stdlib//max dynamic symbols` were ported to the new
suite, though most have been renamed.
[radar]: https://github.com/leanprover/radar
[temci]: https://github.com/parttimenerd/temci
This PR changes the interface of the `ForIn`, `ForIn'`, and `ForM`
typeclasses to not take a `Monad m` parameter. This is a breaking change
for most downstream `instance`s, which will will now need to assume
`[Monad m]`.
The rationale is that if the provider of an instance requires `m` to be
a Monad, they should assume this up front. This makes it possible for
the instanve to assume `LawfulMonad m` or some other stronger
requirement, and also to provided a concrete instance for a particular
`m` without assuming a non-canonical `Monad` structure on it.
Zulip: [#lean4 > Monad assumptions in fields of other typeclasses @
💬](https://leanprover.zulipchat.com/#narrow/channel/270676-lean4/topic/Monad.20assumptions.20in.20fields.20of.20other.20typeclasses/near/537102158)
This PR activates the `grind_annotated` command in
`Init.Data.List.Lemmas` by removing the TODO comment and uncommenting
the command.
This PR depends on #11346 (implement `grind_annotated` command) and
should be merged after that PR (and after CI has done an
`update-stage0`.
This PR enables the syntax `use [ns Foo]` and `instantiate only [ns
Foo]` inside a `grind` tactic block, and has the effect of activating
all grind patterns scoped to that namespace. We can use this to
implement specialized tactics using `grind`, but only controlled subsets
of theorems.
---------
Co-authored-by: Claude <noreply@anthropic.com>
This PR upstreams the `with_weak_namespace` command from Mathlib:
`with_weak_namespace <id> <cmd>` changes the current namespace to `<id>`
for the duration of executing command `<cmd>`, without causing scoped
things to go out of scope. This is in preparation for upstreaming the
`scoped[Foo.Bar]` syntax from Mathlib, which will be useful now that we
are adding `grind` annotations in scopes.
This PR adds a `grind_annotated "YYYY-MM-DD"` command that marks files
as manually annotated for grind.
When LibrarySuggestions is called with `caller := "grind"` (as happens
with `grind +suggestions`), theorems from grind-annotated files are
filtered out from premise selection. The date argument validates using
Std.Time and is informational only for now, but could be used later to
detect files that need re-review.
There's no need for the library suggestions tools to suggest `grind`
theorems from files that have already been carefully annotated by hand.
This PR adds infrastructure for parallel execution across Lean's tactic
monads.
- Add IO.waitAny' to Init/System/IO.lean for waiting on task completion
- Add `Lean.Elab.Task` with `asTask` utilities for `CoreM`, `MetaM`,
`TermElabM`, `TacticM`
- Add `Lean.Elab.Parallel` with parallel execution strategies:
* `par`/`par'` - collect results in original order
* `parIter`/`parIterGreedy` - iterate over results (original or
completion order) (also variants with a cancellation token)
* `parFirst` - return first successful result
This does *not* attempt to be a monad-polymorphic framework for
parallelism. It's intentionally hard-coded to the Lean tactic monads
which I need to work with. If there's desire to make this polymorphic,
hopefully that can be done separately.
This PR renames `String.bytes` to `String.toByteArray`.
This is for two reasons: first, `toByteArray` is a better name, and
second, we have something else that wants to use the name `bytes`,
namely the function that returns in iterator over the string's bytes.
This PR renames `String.ValidPos` to `String.Pos`, `String.endValidPos`
to `String.endPos` and `String.startValidPos` to `String.startPos`.
Accordingly, the deprecations of `String.Pos` to `String.Pos.Raw` and
`String.endPos` to `String.rawEndPos` are removed early, after an
abbreviated deprecation cycle of two releases.
This PR fixes freeing memory accidentally retained for each document
version in the language server on certain elaboration workloads. The
issue must have existed since 4.18.0.
This PR adds an explicit normalization layer for ring constraints in the
`grind linarith` module. For example, it will be used to clean up
denominators when the ring is a field.
This PR renames the `cutsat` tactic to `lia` for better alignment with
standard terminology in the theorem proving community.
`cutsat` still works but now emits a deprecation warning and suggests
using `lia` instead via "Try this:". Both tactics have identical
behavior.
Co-authored-by: Claude <noreply@anthropic.com>
This PR ensures that users can provide `grind` proof parameters whose
types are not `forall`-quantified. Examples:
```lean
opaque f : Nat → Nat
axiom le_f (a : Nat) : a ≤ f a
example (a : Nat) : a ≤ f a := by
grind [le_f a]
example (a b : α) (h : ∀ x y : α, x = y) : a = b := by
grind [h a b]
```
This PR introduces a new `grind` option, `funCC` (enabled by default),
which extends congruence closure to *function-valued* equalities. When
`funCC` is enabled, `grind` tracks equalities of **partially applied
functions**, allowing reasoning steps such as:
```lean
a : Nat → Nat
f : (Nat → Nat) → (Nat → Nat)
h : f a = a
⊢ (f a) m = a m
g : Nat → Nat
f : Nat → Nat → Nat
h : f a = g
⊢ f a b = g b
```
Given an application `f a₁ a₂ … aₙ`, when `funCC := true` and function
equality is enabled for `f`, `grind` generates and tracks equalities for
all partial applications:
* `f a₁`
* `f a₁ a₂`
* …
* `f a₁ a₂ … aₙ`
This allows equalities such as `f a₁ = g` to propagate through further
applications.
**When is function equality enabled for a symbol?**
Function equality is enabled for `f` in the following cases:
1. `f` is **not a constant** (e.g., a lambda, a local function, or a
function parameter).
2. `f` is a **structure field projection**, provided the structure is
**not a `class`**.
3. `f` is a constant marked with `@[grind funCC]`
Users can also enable function equality for specific constants in a
single call using:
```lean
grind [funCC f, funCC g]
```
**Examples:**
```lean
example (m : Nat) (a : Nat → Nat) (f : (Nat → Nat) → (Nat → Nat)) (h : f a = a) :
f a m = a m := by
grind
example (m : Nat) (a : Nat → Nat) (f : (Nat → Nat) → (Nat → Nat)) (h : f a = a) :
f a m = a m := by
fail_if_success grind -funCC -- fails if `funCC` is disabled
grind
```
```lean
example (a b : Nat) (g : Nat → Nat) (f : Nat → Nat → Nat) (h : f a = g) :
f a b = g b := by
grind
example (a b : Nat) (g : Nat → Nat) (f : Nat → Nat → Nat) (h : f a = g) :
f a b = g b := by
fail_if_success grind -funCC
grind
```
**Enabling per-symbol with parameters or attributes**
```lean
opaque f : Nat → Nat → Nat
opaque g : Nat → Nat
example (a b c : Nat) : f a = g → b = c → f a b = g c := by
grind [funCC f, funCC g]
attribute [grind funCC] f g
example (a b c : Nat) : f a = g → b = c → f a b = g c := by
grind
```
This feature substantially improves `grind`’s support for higher-order
and partially-applied function equalities, while preserving
compatibility with first-order SMT behavior when `funCC` is disabled.
Closes#11309
This PR significantly changes the signature of the `ToIterator` type
class. The obtained iterators' state is no longer dependently typed and
is an `outParam` instead of being bundled inside the class. Among other
benefits, `simp` can now rewrite inside of `Slice.toList` and
`Slice.toArray`. The downside is that we lose flexibility. For example,
the former combinator-based implementation of `Subarray`'s iterators is
no longer feasible because the states are dependently typed. Therefore,
this PR provides a hand-written iterator for `Subarray`, which does not
require a dependently typed state and is faster than the previous one.
Converting a family of dependently typed iterators into a simply typed
one using a `Sigma`-state iterator generates forbiddingly bad code, so
that we do provide such a combinator. This PR adds a benchmark for this
problem.
This PR improves the support for `Fin n` in `grind` when `n` is not a
numeral.
- `toInt (0 : Fin n) = 0` in `grind lia`.
- `Fin.mk`-applications are treated as interpreted terms in `grind lia`.
- `Fin.val` applications are suppressed from `grind lia`
counterexamples.
This PR fixes a breakage in Lake's TOML test caused by String API
changes. It also removes a JSON parser workaround that has since been
fixed, and it more generally polishes up the code.
This PR fixes an issue affecting `grind -revert`. In this mode, assigned
metavariables in hypotheses were not being instantiated. This issue was
affecting two files in Mathlib.
This PR fixes a local declaration internalization in `grind` that was
exposed when using `grind -revert`. This bug was affecting a `grind`
proof in Mathlib.
This PR improves the error message encountered in the case of a type
class instance resolution failure, and adds an error explanation that
discusses the common new-user case of binary operation overloading and
points to the `trace.Meta.synthInstance` option for advanced debugging.
## Example
```lean4
def f (x : String) := x + x
```
Before:
```
failed to synthesize
HAdd String String ?m.5
Hint: Additional diagnostic information may be available using the `set_option diagnostics true` command.
```
After:
```
failed to synthesize instance of type class
HAdd String String ?m.5
Hint: Type class instance resolution failures can be inspected with the `set_option trace.Meta.synthInstance true` command.
Error code: lean.failedToSynthesizeTypeclassInstance
[View explanation](https://lean-lang.org/doc/reference/latest/find/?domain=Manual.errorExplanation&name=lean.failedToSynthesizeTypeclassInstance)
```
The error message is changed in three important ways:
* Explains *what* failed to synthesize, using the "type class"
terminology that's more likely to be recognized than the "instance"
terminology
* Points to the `trace.Meta.synthInstance` option which is otherwise
nearly undiscoverable but is quite powerful (see also
leanprover/reference-manual#663 which is adding commentary on this
option)
* Gives an error explanation link (which won't actually work until the
next release after this is merged) which prioritizes the common-case
explanation of using the wrong binary operation
