Improve int_tac and scalar_tac

5 files changed, 87 insertions, 32 deletions

diff --git a/backends/lean/Base/Arith/Int.lean b/backends/lean/Base/Arith/Int.lean
index fa957293..3415866e 100644
--- a/backends/lean/Base/Arith/Int.lean
+++ b/backends/lean/Base/Arith/Int.lean
@@ -24,12 +24,29 @@ class PropHasImp (x : Prop) where
   concl : Prop
   prop : x → concl
 
+instance (p : Int → Prop) : HasIntProp (Subtype p) where
+  prop_ty := λ x => p x
+  prop := λ x => x.property
+
 -- This also works for `x ≠ y` because this expression reduces to `¬ x = y`
 -- and `Ne` is marked as `reducible`
 instance (x y : Int) : PropHasImp (¬ x = y) where
   concl := x < y ∨ x > y
   prop := λ (h:x ≠ y) => ne_is_lt_or_gt h
 
+-- Check if a proposition is a linear integer proposition.
+-- We notably use this to check the goals.
+class IsLinearIntProp (x : Prop) where
+
+instance (x y : Int) : IsLinearIntProp (x < y) where
+instance (x y : Int) : IsLinearIntProp (x > y) where
+instance (x y : Int) : IsLinearIntProp (x ≤ y) where
+instance (x y : Int) : IsLinearIntProp (x ≥ y) where
+instance (x y : Int) : IsLinearIntProp (x ≥ y) where
+/- It seems we don't need to do any special preprocessing when the *goal*
+   has the following shape - I guess `linarith` automatically calls `intro` -/
+instance (x y : Int) : IsLinearIntProp (¬ x = y) where
+
 open Lean Lean.Elab Lean.Meta
 
 -- Explore a term by decomposing the applications (we explore the applied
@@ -189,14 +206,27 @@ def intTacPreprocess (extraPreprocess :  Tactic.TacticM Unit) : Tactic.TacticM U
 elab "int_tac_preprocess" : tactic =>
   intTacPreprocess (do pure ())
 
-def intTac (extraPreprocess :  Tactic.TacticM Unit) : Tactic.TacticM Unit := do
+-- Check if the goal is a linear arithmetic goal
+def goalIsLinearInt : Tactic.TacticM Bool := do
+  Tactic.withMainContext do
+  let gty ← Tactic.getMainTarget
+  match ← trySynthInstance (← mkAppM ``IsLinearIntProp #[gty]) with
+  | .some _ => pure true
+  | _ => pure false
+
+def intTac (splitGoalConjs : Bool) (extraPreprocess :  Tactic.TacticM Unit) : Tactic.TacticM Unit := do
   Tactic.withMainContext do
   Tactic.focus do
+  let g ← Tactic.getMainGoal
+  trace[Arith] "Original goal: {g}"
+  -- Introduce all the universally quantified variables (includes the assumptions)
+  let (_, g) ← g.intros
+  Tactic.setGoals [g]
   -- Preprocess - wondering if we should do this before or after splitting
   -- the goal. I think before leads to a smaller proof term?
   Tactic.allGoals (intTacPreprocess extraPreprocess)
   -- Split the conjunctions in the goal
-  Tactic.allGoals (Utils.repeatTac Utils.splitConjTarget)
+  if splitGoalConjs then Tactic.allGoals (Utils.repeatTac Utils.splitConjTarget)
   -- Call linarith
   let linarith := do
     let cfg : Linarith.LinarithConfig := {
@@ -204,10 +234,25 @@ def intTac (extraPreprocess :  Tactic.TacticM Unit) : Tactic.TacticM Unit := do
       splitNe := false
     }
     Tactic.liftMetaFinishingTactic <| Linarith.linarith false [] cfg
-  Tactic.allGoals linarith
-
-elab "int_tac" : tactic =>
-  intTac (do pure ())
+  Tactic.allGoals do
+    -- We check if the goal is a linear arithmetic goal: if yes, we directly
+    -- call linarith, otherwise we first apply exfalso (we do this because
+    -- linarith is too general and sometimes fails to do this correctly).
+    if ← goalIsLinearInt then do
+      trace[Arith] "linarith goal: {← Tactic.getMainGoal}"
+      linarith
+    else do
+      let g ← Tactic.getMainGoal
+      let gs ← g.apply (Expr.const ``False.elim [.zero])
+      let goals ← Tactic.getGoals
+      Tactic.setGoals (gs ++ goals)
+      Tactic.allGoals do
+        trace[Arith] "linarith goal: {← Tactic.getMainGoal}"
+        linarith
+
+elab "int_tac" args:(" split_goal"?): tactic =>
+  let split := args.raw.getArgs.size > 0
+  intTac split (do pure ())
 
 example (x : Int) (h0: 0 ≤ x) (h1: x ≠ 0) : 0 < x := by
   int_tac_preprocess
@@ -219,10 +264,14 @@ example (x : Int) (h0: 0 ≤ x) (h1: x ≠ 0) : 0 < x := by
 
 -- Checking that things append correctly when there are several disjunctions
 example (x y : Int) (h0: 0 ≤ x) (h1: x ≠ 0) (h2 : 0 ≤ y) (h3 : y ≠ 0) : 0 < x ∧ 0 < y := by
-  int_tac
+  int_tac split_goal
 
 -- Checking that things append correctly when there are several disjunctions
 example (x y : Int) (h0: 0 ≤ x) (h1: x ≠ 0) (h2 : 0 ≤ y) (h3 : y ≠ 0) : 0 < x ∧ 0 < y ∧ x + y ≥ 2 := by
+  int_tac split_goal
+
+-- Checking that we can prove exfalso
+example (a : Prop) (x : Int) (h0: 0 < x) (h1: x < 0) : a := by
   int_tac
 
 end Arith
diff --git a/backends/lean/Base/Arith/Scalar.lean b/backends/lean/Base/Arith/Scalar.lean
index f8903ecf..a56ea08b 100644
--- a/backends/lean/Base/Arith/Scalar.lean
+++ b/backends/lean/Base/Arith/Scalar.lean
@@ -28,11 +28,11 @@ elab "scalar_tac_preprocess" : tactic =>
   intTacPreprocess scalarTacExtraPreprocess
 
 -- A tactic to solve linear arithmetic goals in the presence of scalars
-def scalarTac : Tactic.TacticM Unit := do
-  intTac scalarTacExtraPreprocess
+def scalarTac (splitGoalConjs : Bool) : Tactic.TacticM Unit := do
+  intTac splitGoalConjs scalarTacExtraPreprocess
 
 elab "scalar_tac" : tactic =>
-  scalarTac
+  scalarTac false
 
 instance (ty : ScalarTy) : HasIntProp (Scalar ty) where
   -- prop_ty is inferred
diff --git a/backends/lean/Base/IList/IList.lean b/backends/lean/Base/IList/IList.lean
index 1773e593..2443b1a6 100644
--- a/backends/lean/Base/IList/IList.lean
+++ b/backends/lean/Base/IList/IList.lean
@@ -46,21 +46,18 @@ theorem indexOpt_bounds (ls : List α) (i : Int) :
   ls.indexOpt i = none ↔ i < 0 ∨ ls.len ≤ i :=
   match ls with
   | [] =>
-    have : ¬ (i < 0) → 0 ≤ i := by intro; linarith -- TODO: simplify (we could boost int_tac)
+    have : ¬ (i < 0) → 0 ≤ i := by int_tac
     by simp; tauto
   | _ :: tl =>
     have := indexOpt_bounds tl (i - 1)
     if h: i = 0 then
       by
         simp [*];
-        -- TODO: int_tac/scalar_tac should also explore the goal!
-        have := tl.len_pos
-        linarith
+        int_tac
     else by
       simp [*]
       constructor <;> intros <;>
-      -- TODO: tactic to split all disjunctions
-      rename_i hor <;> cases hor <;>
+      casesm* _ ∨ _ <;> -- splits all the disjunctions
       first | left; int_tac | right; int_tac
 
 theorem indexOpt_eq_index [Inhabited α] (ls : List α) (i : Int) :
@@ -126,7 +123,6 @@ theorem length_update (ls : List α) (i : Int) (x : α) : (ls.update i x).length
 theorem len_update (ls : List α) (i : Int) (x : α) : (ls.update i x).len = ls.len := by
   simp [len_eq_length]
 
-
 theorem left_length_eq_append_eq (l1 l2 l1' l2' : List α) (heq : l1.length = l1'.length) :
   l1 ++ l2 = l1' ++ l2' ↔ l1 = l1' ∧ l2 = l2' := by
   revert l1'
@@ -203,7 +199,7 @@ theorem index_eq
     (l.update i x).index i = x
   :=
   fun _ _ => match l with
-  | [] => by simp at *; exfalso; scalar_tac -- TODO: exfalso needed. Son FIXME
+  | [] => by simp at *; scalar_tac
   | hd :: tl =>
     if h: i = 0 then
       by
diff --git a/backends/lean/Base/Primitives/Vec.lean b/backends/lean/Base/Primitives/Vec.lean
index be3a0e5b..35092c29 100644
--- a/backends/lean/Base/Primitives/Vec.lean
+++ b/backends/lean/Base/Primitives/Vec.lean
@@ -16,20 +16,19 @@ open Result Error
 -- VECTORS --
 -------------
 
-def Vec (α : Type u) := { l : List α // List.length l ≤ Usize.max }
+def Vec (α : Type u) := { l : List α // l.length ≤ Usize.max }
 
 -- TODO: do we really need it? It should be with Subtype by default
-instance Vec.cast (a : Type): Coe (Vec a) (List a)  where coe := λ v => v.val
+instance Vec.cast (a : Type u): Coe (Vec a) (List a)  where coe := λ v => v.val
 
-instance (a : Type) : Arith.HasIntProp (Vec a) where
-  prop_ty := λ v => v.val.length ≤ Scalar.max ScalarTy.Usize
-  prop := λ ⟨ _, l ⟩ => l
+instance (a : Type u) : Arith.HasIntProp (Vec a) where
+  prop_ty := λ v => v.val.len ≤ Scalar.max ScalarTy.Usize
+  prop := λ ⟨ _, l ⟩ => by simp[Scalar.max, List.len_eq_length, *]
 
-example {a: Type} (v : Vec a) : v.val.length ≤ Scalar.max ScalarTy.Usize := by
-  intro_has_int_prop_instances
-  simp_all [Scalar.max, Scalar.min]
+@[simp]
+abbrev Vec.length {α : Type u} (v : Vec α) : Int := v.val.len
 
-example {a: Type} (v : Vec a) : v.val.length ≤ Scalar.max ScalarTy.Usize := by
+example {a: Type u} (v : Vec a) : v.length ≤ Scalar.max ScalarTy.Usize := by
   scalar_tac
 
 def Vec.new (α : Type u): Vec α := ⟨ [], by apply Scalar.cMax_suffices .Usize; simp ⟩
@@ -38,9 +37,6 @@ def Vec.len (α : Type u) (v : Vec α) : Usize :=
   let ⟨ v, l ⟩ := v
   Usize.ofIntCore (List.length v) (by simp [Scalar.min, Usize.min]) l
 
-@[simp]
-abbrev Vec.length {α : Type u} (v : Vec α) : Int := v.val.len
-
 -- This shouldn't be used
 def Vec.push_fwd (α : Type u) (_ : Vec α) (_ : α) : Unit := ()
 
@@ -115,11 +111,14 @@ theorem Vec.index_mut_spec {α : Type u} [Inhabited α] (v: Vec α) (i: Usize) :
   have h := List.indexOpt_eq_index v.val i.val (by scalar_tac) (by simp [*])
   simp only [*]
 
+instance {α : Type u} (p : Vec α → Prop) : Arith.HasIntProp (Subtype p) where
+  prop_ty := λ x => p x
+  prop := λ x => x.property
+
 def Vec.index_mut_back (α : Type u) (v: Vec α) (i: Usize) (x: α) : Result (Vec α) :=
   match v.val.indexOpt i.val with
   | none => fail .arrayOutOfBounds
   | some _ =>
-    -- TODO: int_tac: introduce the refinements in the context?
     .ret ⟨ v.val.update i.val x, by have := v.property; simp [*] ⟩
 
 @[pspec]
diff --git a/backends/lean/Base/Progress/Progress.lean b/backends/lean/Base/Progress/Progress.lean
index c0ddc63d..a281f1d2 100644
--- a/backends/lean/Base/Progress/Progress.lean
+++ b/backends/lean/Base/Progress/Progress.lean
@@ -307,7 +307,18 @@ def evalProgress (args : TSyntax `Progress.progressArgs) : TacticM Unit := do
     let args := (args.get! 2).getArgs
     (args.get! 3).getArgs.size > 0
   trace[Progress] "Split post: {splitPost}"
-  progressAsmsOrLookupTheorem keep withArg ids splitPost (firstTac [assumptionTac, Arith.scalarTac])
+  /- For scalarTac we have a fast track: if the goal is not a linear
+     arithmetic goal, we skip (note that otherwise, scalarTac would try
+     to prove a contradiction) -/
+  let scalarTac : TacticM Unit := do
+    if ← Arith.goalIsLinearInt then
+      -- Also: we don't try to split the goal if it is a conjunction
+      -- (it shouldn't be)
+      Arith.scalarTac false
+    else
+      throwError "Not a linear arithmetic goal"
+  progressAsmsOrLookupTheorem keep withArg ids splitPost (
+    firstTac [assumptionTac, scalarTac])
 
 elab "progress" args:progressArgs : tactic =>
   evalProgress args