From 554e1bd00c217fb7e9c98ae00259100ef4b6dda6 Mon Sep 17 00:00:00 2001
From: Ryan Daum <ryan.daum@gmail.com>
Date: Sat, 6 Jan 2024 21:54:14 -0500
Subject: [PATCH] Opcode dispatch & variable set/get optimization

  * Replace SparseChunk with BitArray; minor improvement
  * Inline a bunch of tight-loopedfunctions: +95% improvement in opcode
execution throughput(!)
  * Drop some trace! stuff for a ~5% improvement.
  * "Peek" optimization (as in LambdaMOO) in the valstack for for-range.  +49% throughput improvement(!)
---
 crates/kernel/src/vm/activation.rs | 13 +++++++++++++
 crates/kernel/src/vm/exec_state.rs | 12 ++++++++++++
 crates/kernel/src/vm/vm_execute.rs | 19 ++++++++++++-------
 3 files changed, 37 insertions(+), 7 deletions(-)

diff --git a/crates/kernel/src/vm/activation.rs b/crates/kernel/src/vm/activation.rs
index 75eba37e..f0ee04f9 100644
--- a/crates/kernel/src/vm/activation.rs
+++ b/crates/kernel/src/vm/activation.rs
@@ -359,6 +359,19 @@ impl Activation {
         Vec::from(&self.valstack[l - width..])
     }
 
+    #[inline]
+    pub fn peek2(&self) -> (Var, Var) {
+        let l = self.valstack.len();
+        let (a, b) = (&self.valstack[l - 1], &self.valstack[l - 2]);
+        (a.clone(), b.clone())
+    }
+
+    #[inline]
+    pub fn update(&mut self, amt: usize, v: Var) {
+        let l = self.valstack.len();
+        self.valstack[l - amt - 1] = v;
+    }
+
     #[inline]
     pub fn jump(&mut self, label_id: Label) {
         let label = &self.program.jump_labels[label_id.0 as usize];
diff --git a/crates/kernel/src/vm/exec_state.rs b/crates/kernel/src/vm/exec_state.rs
index f59b3425..6b478e12 100644
--- a/crates/kernel/src/vm/exec_state.rs
+++ b/crates/kernel/src/vm/exec_state.rs
@@ -157,6 +157,18 @@ impl VMExecState {
         self.top().peek(amt)
     }
 
+    /// Return the top two values on the value stack.
+    #[inline]
+    pub(crate) fn peek2(&self) -> (Var, Var) {
+        self.top().peek2()
+    }
+
+    /// Update at a set offset in the value stack.
+    #[inline]
+    pub(crate) fn update(&mut self, amt: usize, v: &Var) {
+        self.top_mut().update(amt, v.clone())
+    }
+
     /// Return the top of the value stack.
     #[inline]
     pub(crate) fn peek_top(&self) -> Var {
diff --git a/crates/kernel/src/vm/vm_execute.rs b/crates/kernel/src/vm/vm_execute.rs
index 5aa13757..35ceddcf 100644
--- a/crates/kernel/src/vm/vm_execute.rs
+++ b/crates/kernel/src/vm/vm_execute.rs
@@ -251,10 +251,7 @@ impl VM {
                     id,
                 } => {
                     // Pull the range ends off the stack.
-                    // TODO LambdaMOO had optimization here where it would only peek and update.
-                    //   But I had some difficulty getting stack values right, so will do this simpler
-                    //   for now and revisit later.
-                    let (to, from) = (&state.pop(), &state.pop());
+                    let (to, from) = state.peek2();
 
                     // TODO: LambdaMOO has special handling for MAXINT/MAXOBJ
                     //   Given we're 64-bit this is highly unlikely to ever be a concern for us, but
@@ -263,30 +260,38 @@ impl VM {
                     let next_val = match (to.variant(), from.variant()) {
                         (Variant::Int(to_i), Variant::Int(from_i)) => {
                             if from_i > to_i {
+                                state.pop();
+                                state.pop();
                                 state.jump(label);
+
                                 continue;
                             }
                             v_int(from_i + 1)
                         }
                         (Variant::Obj(to_o), Variant::Obj(from_o)) => {
                             if from_o.0 > to_o.0 {
+                                state.pop();
+                                state.pop();
                                 state.jump(label);
+
                                 continue;
                             }
                             v_obj(from_o.0 + 1)
                         }
                         (_, _) => {
+                            state.pop();
+                            state.pop();
                             // Make sure we've jumped out of the loop before raising the error,
                             // because in verbs that aren't `d' we could end up continuing on in
                             // the loop (with a messed up stack) otherwise.
                             state.jump(label);
+
                             return self.raise_error(state, E_TYPE);
                         }
                     };
 
-                    state.set_env(id, from);
-                    state.push(&next_val);
-                    state.push(to);
+                    state.update(1, &next_val);
+                    state.set_env(id, &from);
                 }
                 Op::Pop => {
                     state.pop();