fix: simp caching

closes #3943

TODO: remove `cache` field from `Simp.Result`.

src/Lean/Meta/Tactic/Simp/Main.lean

@@ -241,7 +241,17 @@ def getSimpLetCase (n : Name) (t : Expr) (b : Expr) : MetaM SimpLetCase := do
     else
       return SimpLetCase.dep
 
-def withNewLemmas {α} (xs : Array Expr) (f : SimpM α) : SimpM α := do
+/--
+We use `withNewlemmas` whenever updating the local context.
+We use `withFreshCache` because the local context affects `simp` rewrites
+even when `contextual := false`.
+For example, the `discharger` may inspect the current local context. The default
+discharger does that when applying equational theorems, and the user may
+use `(discharger := assumption)` or `(discharger := omega)`.
+If the `wishFreshCache` introduces performance issues, we can design a better solution
+for the default discharger which is used most of the time.
+-/
+def withNewLemmas {α} (xs : Array Expr) (f : SimpM α) : SimpM α := withFreshCache do
   if (← getConfig).contextual then
     let mut s ← getSimpTheorems
     let mut updated := false
@@ -250,7 +260,7 @@ def withNewLemmas {α} (xs : Array Expr) (f : SimpM α) : SimpM α := do
         s ← s.addTheorem (.fvar x.fvarId!) x
         updated := true
     if updated then
-      withSimpTheorems s f
+      withTheReader Context (fun ctx => { ctx with simpTheorems := s }) f
     else
       f
   else
@@ -304,30 +314,27 @@ def simpArrow (e : Expr) : SimpM Result := do
   trace[Debug.Meta.Tactic.simp] "arrow [{(← getConfig).contextual}] {p} [{← isProp p}] -> {q} [{← isProp q}]"
   if (← pure (← getConfig).contextual <&&> isProp p <&&> isProp q) then
     trace[Debug.Meta.Tactic.simp] "ctx arrow {rp.expr} -> {q}"
-    withLocalDeclD e.bindingName! rp.expr fun h => do
-      let s ← getSimpTheorems
-      let s ← s.addTheorem (.fvar h.fvarId!) h
-      withSimpTheorems s do
-        let rq ← simp q
-        match rq.proof? with
-        | none    => mkImpCongr e rp rq
-        | some hq =>
-          let hq ← mkLambdaFVars #[h] hq
-          /-
-            We use the default reducibility setting at `mkImpDepCongrCtx` and `mkImpCongrCtx` because they use the theorems
-            ```lean
-            @implies_dep_congr_ctx : ∀ {p₁ p₂ q₁ : Prop}, p₁ = p₂ → ∀ {q₂ : p₂ → Prop}, (∀ (h : p₂), q₁ = q₂ h) → (p₁ → q₁) = ∀ (h : p₂), q₂ h
-            @implies_congr_ctx : ∀ {p₁ p₂ q₁ q₂ : Prop}, p₁ = p₂ → (p₂ → q₁ = q₂) → (p₁ → q₁) = (p₂ → q₂)
-            ```
-            And the proofs may be from `rfl` theorems which are now omitted. Moreover, we cannot establish that the two
-            terms are definitionally equal using `withReducible`.
-            TODO (better solution): provide the problematic implicit arguments explicitly. It is more efficient and avoids this
-            problem.
-            -/
-          if rq.expr.containsFVar h.fvarId! then
-            return { expr := (← mkForallFVars #[h] rq.expr), proof? := (← withDefault <| mkImpDepCongrCtx (← rp.getProof) hq) }
-          else
-            return { expr := e.updateForallE! rp.expr rq.expr, proof? := (← withDefault <| mkImpCongrCtx (← rp.getProof) hq) }
+    withLocalDeclD e.bindingName! rp.expr fun h => withNewLemmas #[h] do
+      let rq ← simp q
+      match rq.proof? with
+      | none    => mkImpCongr e rp rq
+      | some hq =>
+        let hq ← mkLambdaFVars #[h] hq
+        /-
+          We use the default reducibility setting at `mkImpDepCongrCtx` and `mkImpCongrCtx` because they use the theorems
+          ```lean
+          @implies_dep_congr_ctx : ∀ {p₁ p₂ q₁ : Prop}, p₁ = p₂ → ∀ {q₂ : p₂ → Prop}, (∀ (h : p₂), q₁ = q₂ h) → (p₁ → q₁) = ∀ (h : p₂), q₂ h
+          @implies_congr_ctx : ∀ {p₁ p₂ q₁ q₂ : Prop}, p₁ = p₂ → (p₂ → q₁ = q₂) → (p₁ → q₁) = (p₂ → q₂)
+          ```
+          And the proofs may be from `rfl` theorems which are now omitted. Moreover, we cannot establish that the two
+          terms are definitionally equal using `withReducible`.
+          TODO (better solution): provide the problematic implicit arguments explicitly. It is more efficient and avoids this
+          problem.
+          -/
+        if rq.expr.containsFVar h.fvarId! then
+          return { expr := (← mkForallFVars #[h] rq.expr), proof? := (← withDefault <| mkImpDepCongrCtx (← rp.getProof) hq) }
+        else
+          return { expr := e.updateForallE! rp.expr rq.expr, proof? := (← withDefault <| mkImpCongrCtx (← rp.getProof) hq) }
   else
     mkImpCongr e rp (← simp q)
 
@@ -389,7 +396,7 @@ def simpLet (e : Expr) : SimpM Result := do
     | SimpLetCase.dep => return { expr := (← dsimp e) }
     | SimpLetCase.nondep =>
       let rv ← simp v
-      withLocalDeclD n t fun x => do
+      withLocalDeclD n t fun x => withNewLemmas #[x] do
         let bx := b.instantiate1 x
         let rbx ← simp bx
         let hb? ← match rbx.proof? with
@@ -402,7 +409,7 @@ def simpLet (e : Expr) : SimpM Result := do
         | _,      some h => return { expr := e', proof? := some (← mkLetCongr (← rv.getProof) h) }
     | SimpLetCase.nondepDepVar =>
       let v' ← dsimp v
-      withLocalDeclD n t fun x => do
+      withLocalDeclD n t fun x => withNewLemmas #[x] do
         let bx := b.instantiate1 x
         let rbx ← simp bx
         let e' := mkLet n t v' (← rbx.expr.abstractM #[x])

src/Lean/Meta/Tactic/Simp/Types.lean

@@ -21,7 +21,11 @@ structure Result where
   /-- A proof that `$e = $expr`, where the simplified expression is on the RHS.
   If `none`, the proof is assumed to be `refl`. -/
   proof?         : Option Expr := none
-  /-- If `cache := true` the result is cached. -/
+  /--
+  If `cache := true` the result is cached.
+  Warning: we will remove this field in the future. It is currently used by
+  `arith := true`, but we can now refactor the code to avoid the hack.
+  -/
   cache          : Bool := true
   deriving Inhabited
 
@@ -284,9 +288,6 @@ Save current cache, reset it, execute `x`, and then restore original cache.
   modify fun s => { s with cache := {} }
   try x finally modify fun s => { s with cache := cacheSaved }
 
-@[inline] def withSimpTheorems (s : SimpTheoremsArray) (x : SimpM α) : SimpM α := do
-  withFreshCache <| withTheReader Context (fun ctx => { ctx with simpTheorems := s }) x
-
 @[inline] def withDischarger (discharge? : Expr → SimpM (Option Expr)) (x : SimpM α) : SimpM α :=
   withFreshCache <| withReader (fun r => { MethodsRef.toMethods r with discharge? }.toMethodsRef) x
 

tests/lean/run/3943.lean

@@ -0,0 +1,38 @@
+example (f : Nat → Nat) : (if f x = 0 then f x else f x + 1) + f x = y := by
+  simp (config := { contextual := true })
+  guard_target =ₛ (if f x = 0 then 0 else f x + 1) + f x = y
+  sorry
+
+example (f : Nat → Nat) : f x = 0 → f x + 1 = y := by
+  simp (config := { contextual := true })
+  guard_target =ₛ f x = 0 → 1 = y
+  sorry
+
+example (f : Nat → Nat) : let _  : f x = 0 := sorry; f x + 1 = y := by
+  simp (config := { contextual := true, zeta := false })
+  guard_target =ₛ let _  : f x = 0 := sorry; 1 = y
+  sorry
+
+def overlap : Nat → Nat
+  | 0 => 0
+  | 1 => 1
+  | n+1 => overlap n
+
+example : (if (n = 0 → False) then overlap (n+1) else overlap (n+1)) = overlap n  := by
+  simp only [overlap]
+  guard_target =ₛ (if (n = 0 → False) then overlap n else overlap (n+1)) = overlap n
+  sorry
+
+opaque p : Nat → Bool
+opaque g : Nat → Nat
+@[simp] theorem g_eq (h : p x) : g x = x := sorry
+
+example : (if p x then g x else g x + 1) + g x = y := by
+  simp (discharger := assumption)
+  guard_target =ₛ (if p x then x else g x + 1) + g x = y
+  sorry
+
+example : (let _  : p x := sorry; g x + 1 = y) ↔ g x = y := by
+  simp (config := { zeta := false }) (discharger := assumption)
+  guard_target =ₛ (let _  : p x := sorry; x + 1 = y) ↔ g x = y
+  sorry