Implement the pad cache

Frames are now kept on a number of auxilliary doubly-linked lists.  In
each list, frames can be reused after they return, because the lists
are kept monotonic by call depth.  When the call stack needs to change
shape away from a stack, for closures and gather/take, a new list is
created and the relevant frames are relinked.  +30% speedup on the call
benchmark.

lib/Cursor.cs

@@ -281,6 +281,7 @@ public sealed class RxFrame {
         if (return_one) {
             return Kernel.Take(th, Kernel.NewROScalar(m));
         } else {
+            th.MarkSharedChain();
             return_one = true;
             VarDeque ks = new VarDeque();
             ks.Push(Kernel.NewROScalar(m));

lib/Kernel.cs

@@ -215,12 +215,15 @@ public class SubInfo {
     // We need hashy frames available to properly handle BEGIN; for the time
     // being, all frames will be hashy for simplicity
     public class Frame: IP6 {
-        public readonly Frame caller;
-        public readonly Frame outer;
-        public readonly SubInfo info;
+        public Frame caller;
+        public Frame outer;
+        public SubInfo info;
+        // a doubly-linked list of frames being used by a given coroutine
+        public Frame reusable_child;
+        public Frame reuser;
         public object resultSlot = null;
         public int ip = 0;
-        public readonly DynBlockDelegate code;
+        public DynBlockDelegate code;
         public Dictionary<string, object> lex;
         // statistically, most subs have between 1 and 4 anonymous lexicals
         public object lex0;
@@ -234,6 +237,10 @@ public class Frame: IP6 {
         public Variable[] pos;
         public Dictionary<string, Variable> named;
 
+        // after MakeSub, GatherHelper
+        public const int SHARED = 1;
+        public int flags;
+
         public Frame(Frame caller_, Frame outer_,
                 SubInfo info_) {
             caller = caller_;
@@ -242,6 +249,29 @@ public class Frame: IP6 {
             info = info_;
         }
 
+        public Frame() {}
+
+        public Frame MakeChild(Frame outer, SubInfo info) {
+            if (reusable_child == null) {
+                reusable_child = new Frame();
+                reusable_child.reuser = this;
+            }
+            reusable_child.ip = 0;
+            reusable_child.resultSlot = null;
+            reusable_child.lexn = null;
+            reusable_child.lex = null;
+            reusable_child.lex0 = null;
+            reusable_child.lex1 = null;
+            reusable_child.lex2 = null;
+            reusable_child.lex3 = null;
+            reusable_child.caller = this;
+            reusable_child.outer = outer;
+            reusable_child.info = info;
+            reusable_child.code = info.code;
+            reusable_child.rx = null;
+            return reusable_child;
+        }
+
         public Frame Continue() {
             return code(this);
         }
@@ -256,6 +286,21 @@ public class Frame: IP6 {
             }
         }
 
+        public void MarkShared() {
+            if (0 == (flags & SHARED)) {
+                flags |= SHARED;
+                if (reuser != null) reuser.reusable_child = reusable_child;
+                if (reusable_child != null) reusable_child.reuser = reuser;
+                reuser = reusable_child = null;
+            }
+        }
+
+        // when control might re-enter a function
+        public void MarkSharedChain() {
+            for (Frame x = this; x != null; x = x.caller)
+                x.MarkShared();
+        }
+
         public override DynMetaObject GetMO() { return Kernel.CallFrameMO; }
 
         public int ExecutingLine() {
@@ -617,9 +662,9 @@ public class Kernel {
 
         public static Frame GatherHelper(Frame th, IP6 sub) {
             DynObject dyo = (DynObject) sub;
-            Frame n = new Frame(th,
-                                (Frame) dyo.slots[0],
-                                (SubInfo) dyo.slots[1]);
+            Frame n = th.MakeChild((Frame) dyo.slots[0],
+                    (SubInfo) dyo.slots[1]);
+            n.MarkSharedChain();
             th.resultSlot = n;
             return th;
         }
@@ -630,7 +675,7 @@ public class Kernel {
             Frame outer = (Frame) dyo.slots[0];
             SubInfo info = (SubInfo) dyo.slots[1];
 
-            Frame n = new Frame(caller, outer, info);
+            Frame n = caller.MakeChild(outer, info);
             n.pos = pos;
             n.named = named;
 
@@ -678,6 +723,7 @@ public class Kernel {
         public static IP6 MakeSub(SubInfo info, Frame outer) {
             DynObject n = new DynObject(info.mo ?? SubMO);
             n.slots[0] = outer;
+            if (outer != null) outer.MarkShared();
             n.slots[1] = info;
             return n;
         }
@@ -744,7 +790,7 @@ public class Kernel {
 
             th.resultSlot = rhs;
 
-            n = new Frame(th, null, BindSI);
+            n = th.MakeChild(null, BindSI);
             n.pos = new Variable[1] { rhs };
             return n;
         }
@@ -783,7 +829,7 @@ public class Kernel {
                 return th;
             }
 
-            Frame n = new Frame(th, null, AssignSI);
+            Frame n = th.MakeChild(null, AssignSI);
             n.pos = new Variable[2] { lhs, rhs };
             return n;
         }
@@ -886,8 +932,11 @@ public class Kernel {
         }
 
         public static Frame StartP6Thread(Frame th, IP6 sub) {
+            th.MarkSharedChain();
             Thread thr = new Thread(delegate () {
-                    Frame current = sub.Invoke(th, new Variable[0], null);
+                    Frame current = new Frame();
+                    current = sub.Invoke(current, new Variable[0], null);
+                    current.caller = current.caller.caller;
                     RunCore(current, th);
                 });
             thr.Start();
@@ -1021,7 +1070,7 @@ public class Kernel {
                     Environment.Exit(1);
                 }
                 in_unhandled = true;
-                Frame r = new Frame(th, null, UnhandledSI);
+                Frame r = th.MakeChild(null, UnhandledSI);
                 r.lex0 = mp;
                 return r;
             } else {

perf/callmark.pl

@@ -29,4 +29,4 @@
 sub x29 () { x28; x28 } # 536870911
 sub x30 () { x29; x29 } # 1073741823
 
-x25;
+x26;