This PR adds a number of simple comparison lemmas to the top/bottom
element for BitVec. Then they are applied to teach bv_normalize that
`(a<1) = (a==0)` and to remove an intermediate proof that is no longer
necessary along the way.
This PR fixes a significant auto-completion performance regression that
was introduced in #5666, i.e. v4.14.0.
#5666 introduced tactic docstrings, which were attempted to be collected
for every single completion item. This is slow for hundreds of thousands
of completion items. To fix this, this PR moves the docstring
computation into the completion item resolution, which is only called
when users select a specific completion item in the UI.
A downside of this approach is that we currently can't test completion
item resolution, so we lose a few tests that cover docstrings in
completions in this PR.
This PR adds infrastructure for the `grind?` tactic. It also adds the
new modifier `usr` which allows users to write `grind only [usr
thmName]` to instruct `grind` to only use theorem `thmName`, but using
the patterns specified with the command `grind_pattern`.
This PR adds support for closing goals using `match`-expression
conditions that are known to be true in the `grind` tactic state.
`grind` can now solve goals such as:
```lean
def f : List Nat → List Nat → Nat
| _, 1 :: _ :: _ => 1
| _, _ :: _ => 2
| _, _ => 0
example : z = a :: as → y = z → f x y > 0
```
Without `grind`, we would use the `split` tactic. The first two goals,
corresponding to the first two alternatives, are closed using `simp`,
and the the third using the `match`-expression condition produced by
`split`. The proof would proceed as follows.
```lean
example : z = a :: as → y = z → f x y > 0 := by
intros
unfold f
split
next => simp
next => simp
next h =>
/-
...
_ : z = a :: as
_ : y = z
...
h : ∀ (head : Nat) (tail : List Nat), y = head :: tail → False
|- 0 > 0
-/
subst_vars
/-
...
h : ∀ (head : Nat) (tail : List Nat), a :: as = head :: tail → False
|- 0 > 0
-/
have : False := h a as rfl
contradiction
```
Here is the same proof using `grind`.
```lean
example : z = a :: as → y = z → f x y > 0 := by
grind [f.eq_def]
```
This PR implements the `zetaUnused` simp and reduction option (added in
#6754).
True by default, and implied by `zeta`, this can be turned off to make
simp even more careful about preserving the expression structure,
including unused let and have expressions.
Breaking change: The `split` tactic no longer removes unused let and
have expressions as a side-effect, in rare cases this may break proofs.
`dsimp only` can be used to remove unused have and let expressions.
This PR fixes the support for case splitting on data in the `grind`
tactic. The following example works now:
```lean
inductive C where
| a | b | c
def f : C → Nat
| .a => 2
| .b => 3
| .c => 4
example : f x > 1 := by
grind [
f, -- instructs `grind` to use `f`-equation theorems,
C -- instructs `grind` to case-split on free variables of type `C`
]
```
This PR fixes a bug in the internalization of offset terms in the
`grind` tactic. For example, `grind` was failing to solve the following
example because of this bug.
```lean
example (f : Nat → Nat) : f (a + 1) = 1 → a = 0 → f 1 = 1 := by
grind
```
This PR fixes a `partial_fixpoint` error message to suggest the option
`trace.Elab.Tactic.monotonicity` rather than the nonexistent
`trace.Elab.Tactic.partial_monotonicity`.
This PR fixes a few bugs in the `grind` tactic: missing issues, bad
error messages, incorrect threshold in the canonicalizer, and bug in the
ground pattern internalizer.
This PR ensures that conditional equation theorems for function
definitions are handled correctly in `grind`. We use the same
infrastructure built for `match`-expression equations. Recall that in
both cases, these theorems are conditional when there are overlapping
patterns.
Avoids build time overhead until the option is proven to speed up
average projects. Adds Init.Prelude (many tiny declarations, "worst
case") and Init.List.Sublist (many nontrivial theorems, "best case")
under -DElab.async=true as new benchmarks for tracking.
This PR adds support for `bv_decide` to automatically split up
non-recursive structures that contain information about supported types.
It can be controlled using the new `structures` field in the `bv_decide`
config.
This PR ensures that the branches of an `if-then-else` term are
internalized only after establishing the truth value of the condition.
This change makes its behavior consistent with the `match`-expression
and dependent `if-then-else` behavior in `grind`.
This feature is particularly important for recursive functions defined
by well-founded recursion and `if-then-else`. Without lazy
`if-then-else` branch internalization, the equation theorem for the
recursive function would unfold until reaching the generation depth
threshold, and before performing any case analysis. See new tests for an
example.
This PR adds support for case splitting on `match`-expressions with
overlapping patterns to the `grind` tactic. `grind` can now solve
examples such as:
```
inductive S where
| mk1 (n : Nat)
| mk2 (n : Nat) (s : S)
| mk3 (n : Bool)
| mk4 (s1 s2 : S)
def g (x y : S) :=
match x, y with
| .mk1 a, _ => a + 2
| _, .mk2 1 (.mk4 _ _) => 3
| .mk3 _, .mk4 _ _ => 4
| _, _ => 5
example : g a b > 1 := by
grind [g.eq_def]
```
This PR adds better support for overlapping `match` patterns in `grind`.
`grind` can now solve examples such as
```lean
inductive S where
| mk1 (n : Nat)
| mk2 (n : Nat) (s : S)
| mk3 (n : Bool)
| mk4 (s1 s2 : S)
def f (x y : S) :=
match x, y with
| .mk1 _, _ => 2
| _, .mk2 1 (.mk4 _ _) => 3
| .mk3 _, _ => 4
| _, _ => 5
example : b = .mk2 y1 y2 → y1 = 2 → a = .mk4 y3 y4 → f a b = 5 := by
unfold f
grind (splits := 0)
```
---------
Co-authored-by: Leonardo de Moura <leodemoura@amazon.com>
This PR adds `foo.fun_cases`, an automatically generated theorem that
splits the goal according to the branching structure of `foo`, much like
the Functional Induction Principle, but for all functions (not just
recursive ones), and without providing inductive hypotheses.
The design isn't quite final yet as to which function parameters should
become targets of the motive, and which parameters of the theorem, but
the current version is already proven to be useful, so start with this
and iterate later.
This PR adds the ability to define possibly non-terminating functions
and still be able to reason about them equationally, as long as they are
tail-recursive or monadic.
Typical uses of this feature are
```lean4
def ack : (n m : Nat) → Option Nat
| 0, y => some (y+1)
| x+1, 0 => ack x 1
| x+1, y+1 => do ack x (← ack (x+1) y)
partial_fixpiont
def whileSome (f : α → Option α) (x : α) : α :=
match f x with
| none => x
| some x' => whileSome f x'
partial_fixpiont
def computeLfp {α : Type u} [DecidableEq α] (f : α → α) (x : α) : α :=
let next := f x
if x ≠ next then
computeLfp f next
else
x
partial_fixpiont
noncomputable def geom : Distr Nat := do
let head ← coin
if head then
return 0
else
let n ← geom
return (n + 1)
partial_fixpiont
```
This PR contains
* The necessary fragment of domain theory, up to (a variant of)
Knaster–Tarski theorem (merged as
https://github.com/leanprover/lean4/pull/6477)
* A tactic to solve monotonicity goals compositionally (a bit like
mathlib’s `fun_prop`) (merged as
https://github.com/leanprover/lean4/pull/6506)
* An attribute to extend that tactic (merged as
https://github.com/leanprover/lean4/pull/6506)
* A “derecursifier” that uses that machinery to define recursive
function, including support for dependent functions and mutual
recursion.
* Fixed-point induction principles (technical, tedious to use)
* For `Option`-valued functions: Partial correctness induction theorems
that hide all the domain theory
This is heavily inspired by [Isabelle’s `partial_function`
command](https://isabelle.in.tum.de/doc/codegen.pdf).
This PR changes the identifier parser to allow for the ⱼ unicode
character which was forgotten as it lives by itself in a codeblock with
coptic characters.
This PR deprecates `List.iota`, which we make no essential use of. `iota
n` can be replaced with `(range' 1 n).reverse`. The verification lemmas
for `range'` already have better coverage than those for `iota`.
Any downstream projects using it (I am not aware of any) are encouraged
to adopt it.
This PR modifies LCNF.toMonoType to use a more refined type erasure
scheme, which distinguishes between irrelevant/erased information
(represented by lcErased) and erased type dependencies (represented by
lcAny). This corresponds to the irrelevant/object distinction in the old
code generator.
This PR fixes a bug in the equational theorem generator for
`match`-expressions. See new test for an example.
Signed-off-by: Leonardo de Moura <leodemoura@amazon.com>
Co-authored-by: Leonardo de Moura <leodemoura@amazon.com>
This PR introduces a new feature that allows users to specify which
inductive datatypes the `grind` tactic should perform case splits on.
The configuration option `splitIndPred` is now set to `false` by
default. The attribute `[grind cases]` is used to mark inductive
datatypes and predicates that `grind` may case split on during the
search. Additionally, the attribute `[grind cases eager]` can be used to
mark datatypes and predicates for case splitting both during
pre-processing and the search.
Users can also write `grind [HasType]` or `grind [cases HasType]` to
instruct `grind` to perform case splitting on the inductive predicate
`HasType` in a specific instance. Similarly, `grind [-Or]` can be used
to instruct `grind` not to case split on disjunctions.
Co-authored-by: Leonardo de Moura <leodemoura@amazon.com>
This PR adds the attributes `[grind cases]` and `[grind cases eager]`
for controlling case splitting in `grind`. They will replace the
`[grind_cases]` and the configuration option `splitIndPred`.
After update stage0, we will push the second part of this PR.
This PR adds support for equality backward reasoning to `grind`. We can
illustrate the new feature with the following example. Suppose we have a
theorem:
```lean
theorem inv_eq {a b : α} (w : a * b = 1) : inv a = b
```
and we want to instantiate the theorem whenever we are tying to prove
`inv t = s` for some terms `t` and `s`
The attribute `[grind ←]` is not applicable in this case because, by
default, `=` is not eligible for E-matching. The new attribute `[grind
←=]` instructs `grind` to use the equality and consider disequalities in
the `grind` proof state as candidates for E-matching.
This PR adds support for beta reduction in the `grind` tactic. `grind`
can now solve goals such as
```lean
example (f : Nat → Nat) : f = (fun x : Nat => x + 5) → f 2 > 5 := by
grind
```
This PR removes the `[grind_norm]` attribute. The normalization theorems
used by `grind` are now fixed and cannot be modified by users. We use
normalization theorems to ensure the built-in procedures receive term
wish expected "shapes". We use it for types that have built-in support
in grind. Users could misuse this feature as a simplification rule. For
example, consider the following example:
```lean
def replicate : (n : Nat) → (a : α) → List α
| 0, _ => []
| n+1, a => a :: replicate n a
-- I want `grind` to instantiate the equations theorems for me.
attribute [grind] replicate
-- I want it to use the equation theorems as simplication rules too.
attribute [grind_norm] replicate
/--
info: [grind.assert] n = 0
[grind.assert] ¬replicate n xs = []
[grind.ematch.instance] replicate.eq_1: replicate 0 xs = []
[grind.assert] True
-/
set_option trace.grind.ematch.instance true in
set_option trace.grind.assert true in
example (xs : List α) : n = 0 → replicate n xs = [] := by
grind -- fails :(
```
In this example, `grind` starts by asserting the two propositions as
expected: `n = 0`, and `¬replicate n xs = []`. The normalizer cannot
reduce `replicate n xs` as expected.
Then, the E-matching module finds the instance `replicate 0 xs = []` for
the equation theorem `replicate.eq_1` also as expected. But, then the
normalizer kicks in and reduces the new instance to `True`. By removing
`[grind_norm]` we elimninate this kind of misuse. Users that want to
preprocess a formula before invoking `grind` should use `simp` instead.
This PR splits the environment used by the kernel from that used by the
elaborator, providing the foundation for tracking of asynchronously
elaborated declarations, which will exist as a concept only in the
latter.
Minor changes:
* kernel diagnostics are moved from an environment extension to a direct
environment as they are the only extension used directly by the kernel
* `initQuot` is moved from an environment header field to a direct
environment as it is the only header field used by the kernel; this also
makes the remaining header immutable after import
This PR adds support for extensionality theorems (using the `[ext]`
attribute) to the `grind` tactic. Users can disable this functionality
using `grind -ext` . Below are examples that demonstrate problems now
solvable by `grind`.
```lean
open List in
example : (replicate n a).map f = replicate n (f a) := by
grind only [Option.map_some', Option.map_none', getElem?_map, getElem?_replicate]
```
```lean
@[ext] structure S where
a : Nat
b : Bool
example (x y : S) : x.a = y.a → y.b = x.b → x = y := by
grind
```
