Unify GCM across chapters

Remove the _anti field.
Remove the isPinned() call.
Add a copyEmpty call.

6 files changed, 48 insertions, 45 deletions

diff --git a/chapter16/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java b/chapter16/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java
index 6a4728c3..31302ffd 100644
--- a/chapter16/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java
+++ b/chapter16/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java
@@ -69,10 +69,15 @@ public abstract class GlobalCodeMotion {
             cfg.loopDepth();    // Pre-compute loop depth
             for( Node n : cfg._inputs )
                 _schedEarly(n,visit);
-            if( cfg instanceof RegionNode )
-                for( Node phi : cfg._outputs )
-                    if( phi instanceof PhiNode )
+            // In dead infinite loops, entire code blocks may be unreachable
+            // from below.  Reach down from the CFG to their Phis so their
+            // inputs are scheduled too.
+            if( cfg instanceof RegionNode ) {
+                int len = cfg.nOuts();
+                for( int i=0; i<len; i++ )
+                    if( cfg.out(i) instanceof PhiNode phi )
                         _schedEarly(phi,visit);
+            }
         }
     }
 
@@ -90,20 +95,17 @@ public abstract class GlobalCodeMotion {
         if( n==null || visit.get(n._nid) ) return; // Been there, done that
         assert !(n instanceof CFGNode);
         visit.set(n._nid);
-        // Schedule not-pinned not-CFG inputs before self.  Since skipping
-        // Pinned, this never walks the backedge of Phis (and thus spins around
-        // a data-only loop, eventually attempting relying on some pre-visited-
-        // not-post-visited data op with no scheduled control.
+        // Schedule inputs first, except Phis: following their backedges would
+        // enter a data cycle before its control has been scheduled.
         for( Node def : n._inputs )
             if( def!=null && !(def instanceof PhiNode) )
                 _schedEarly(def,visit);
-        // If not-pinned (e.g. constants, projections, phi) and not-CFG
-        if( !n.isPinned() ) {
+        // An existing edge 0 already supplies control (or a Phi/Proj binding).
+        if( n.in(0)==null ) {
             // Schedule at deepest input
             CFGNode early = Parser.START; // Maximally early, lowest idepth
-            if( n.in(0) instanceof CFGNode cfg ) early = cfg;
             for( int i=1; i<n.nIns(); i++ )
-                if( n.in(i).cfg0().idepth() > early.idepth() )
+                if( n.in(i)!=null && n.in(i).cfg0().idepth() > early.idepth() )
                     early = n.in(i).cfg0(); // Latest/deepest input
             n.setDef(0,early);              // First place this can go
         }
@@ -113,8 +115,11 @@ public abstract class GlobalCodeMotion {
     private static void schedLate( StopNode stop) {
         CFGNode[] late = new CFGNode[Node.UID()];
         Node[] ns = new Node[Node.UID()];
+        // Record Load NIDs at all their CFG block choices, then check against
+        // Store block choices to force a Load above an anti-dependent Store.
+        int[] anti = new int[Node.UID()];
         // Breadth-first scheduling
-        breadth(stop,ns,late);
+        breadth(stop,ns,late,anti);
 
         // Copy the best placement choice into the control slot
         for( int i=0; i<late.length; i++ )
@@ -122,7 +127,7 @@ public abstract class GlobalCodeMotion {
                 ns[i].setDef(0,late[i]);
     }
 
-    private static void breadth(Node stop, Node[] ns, CFGNode[] late) {
+    private static void breadth(Node stop, Node[] ns, CFGNode[] late, int[] anti) {
         // Things on the worklist have some (but perhaps not all) uses done.
         WorkList<Node> work = new WorkList<>();
         work.push(stop);
@@ -153,22 +158,26 @@ public abstract class GlobalCodeMotion {
                             continue outer;
 
                 // All uses done, schedule
-                _doSchedLate(n,ns,late);
+                _doSchedLate(n,ns,late,anti);
             }
 
-            // Walk all inputs and put on worklist, as their last-use might now be done
-            for( Node def : n._inputs )
-                if( def!=null && late[def._nid]==null ) {
-                    work.push(def);
-                    // if the def has a load use, maybe the load can fire
-                    for( Node ld : def._outputs )
-                        if( ld instanceof LoadNode && late[ld._nid]==null )
-                            work.push(ld);
-                }
+            // A use just finished; reconsider its inputs and waiting loads,
+            // even when the shared memory input was already scheduled.
+            for( Node def : n._inputs ) {
+                if( def==null ) continue;
+                if( late[def._nid]==null ) work.push(def);
+                for( Node out : def._outputs )
+                    if( out instanceof LoadNode ld && late[ld._nid]==null )
+                        work.push(ld);
+            }
+            if( n instanceof LoopNode loop )
+                for( Node phi : loop._outputs )
+                    if( phi instanceof PhiNode && late[phi._nid]==null )
+                        work.push(phi);
         }
     }
 
-    private static void _doSchedLate(Node n, Node[] ns, CFGNode[] late) {
+    private static void _doSchedLate(Node n, Node[] ns, CFGNode[] late, int[] anti) {
         // Walk uses, gathering the LCA (Least Common Ancestor) of uses
         CFGNode early = n.in(0) instanceof CFGNode cfg ? cfg : n.in(0).cfg0();
         assert early != null;
@@ -179,7 +188,7 @@ public abstract class GlobalCodeMotion {
 
         // Loads may need anti-dependencies, raising their LCA
         if( n instanceof LoadNode load )
-            lca = find_anti_dep(lca,load,early,late);
+            lca = find_anti_dep(lca,load,early,late,anti);
 
         // Walk up from the LCA to the early, looking for best place.  This is
         // the lowest execution frequency, approximated by least loop depth and
@@ -202,8 +211,7 @@ public abstract class GlobalCodeMotion {
         CFGNode found=null;
         for( int i=1; i<phi.nIns(); i++ )
             if( phi.in(i)==n )
-                if( found==null ) found = phi.region().cfg(i);
-                else Utils.TODO(); // Can be more than once
+                found = phi.region().cfg(i).domLCA(found,null); // Can be more than one matching input.
         assert found!=null;
         return found;
     }
@@ -211,32 +219,34 @@ public abstract class GlobalCodeMotion {
 
     // Least loop depth first, then largest idepth
     private static boolean better( CFGNode lca, CFGNode best ) {
-        return lca._loopDepth < best._loopDepth ||
-                (lca.idepth() > best.idepth() || best instanceof IfNode);
+        return lca.loopDepth() < best.loopDepth() ||
+            lca instanceof NeverNode ||
+            lca.idepth() > best.idepth() ||
+            best instanceof IfNode;
     }
 
-    private static CFGNode find_anti_dep(CFGNode lca, LoadNode load, CFGNode early, CFGNode[] late) {
+    private static CFGNode find_anti_dep(CFGNode lca, LoadNode load, CFGNode early, CFGNode[] late, int[] anti) {
         // We could skip final-field loads here.
         // Walk LCA->early, flagging Load's block location choices
         for( CFGNode cfg=lca; early!=null && cfg!=early.idom(); cfg = cfg.idom() )
-            cfg._anti = load._nid;
+            anti[cfg._nid] = load._nid;
         // Walk load->mem uses, looking for Stores causing an anti-dep
         for( Node mem : load.mem()._outputs ) {
             switch( mem ) {
             case StoreNode st:
                 assert late[st._nid]!=null;
-                lca = anti_dep(load,late[st._nid],st.cfg0(),lca,st);
+                lca = anti_dep(load,late[st._nid],st.cfg0(),lca,st,anti);
                 break;
             case NewNode st:
                 assert late[st._nid]!=null;
-                lca = anti_dep(load,late[st._nid],st.cfg0(),lca,st);
+                lca = anti_dep(load,late[st._nid],st.cfg0(),lca,st,anti);
                 break;
             case PhiNode phi:
                 // Repeat anti-dep for matching Phi inputs.
                 // No anti-dep edges but may raise the LCA.
                 for( int i=1; i<phi.nIns(); i++ )
                     if( phi.in(i)==load.mem() )
-                        lca = anti_dep(load,phi.region().cfg(i),load.mem().cfg0(),lca,null);
+                        lca = anti_dep(load,phi.region().cfg(i),load.mem().cfg0(),lca,null,anti);
                 break;
             case LoadNode ld: break; // Loads do not cause anti-deps on other loads
             case ReturnNode ret: break; // Load must already be ahead of Return
@@ -249,11 +259,12 @@ public abstract class GlobalCodeMotion {
     }
 
     //
-    private static CFGNode anti_dep( LoadNode load, CFGNode stblk, CFGNode defblk, CFGNode lca, Node st ) {
-        // Walk store blocks "reach" from its scheduled location to its earliest
+    private static CFGNode anti_dep( LoadNode load, CFGNode stblk, CFGNode defblk, CFGNode lca, Node st, int[] anti ) {
+        // Preserve the full store range for the earlier evaluator scheduler.
+        // It places nodes independently of GCM and may hoist this store.
         for( ; stblk != defblk; stblk = stblk.idom() ) {
             // Store and Load overlap, need anti-dependence
-            if( !(stblk instanceof IfNode) && stblk._anti==load._nid ) {
+            if( !(stblk instanceof IfNode) && anti[stblk._nid]==load._nid ) {
                 lca = stblk.domLCA(lca,null); // Raise Loads LCA
                 if( lca == stblk && st != null && st._inputs.find(load) == -1 ) // And if something moved,
                     st.addDef(load);   // Add anti-dep as well
diff --git a/chapter16/src/main/java/com/seaofnodes/simple/node/CFGNode.java b/chapter16/src/main/java/com/seaofnodes/simple/node/CFGNode.java
index 704dea8d..06596ce0 100644
--- a/chapter16/src/main/java/com/seaofnodes/simple/node/CFGNode.java
+++ b/chapter16/src/main/java/com/seaofnodes/simple/node/CFGNode.java
@@ -75,9 +75,6 @@ public abstract class CFGNode extends Node {
     public int _loopDepth;
     public int loopDepth() { return _loopDepth==0 ? (_loopDepth = cfg(0).loopDepth()) : _loopDepth; }
 
-    // Anti-dependence field support
-    public int _anti;           // Per-CFG field to help find anti-deps
-
     // ------------------------------------------------------------------------
     // Support routines for Global Code Motion
 
diff --git a/chapter16/src/main/java/com/seaofnodes/simple/node/CastNode.java b/chapter16/src/main/java/com/seaofnodes/simple/node/CastNode.java
index 462d6cba..38d6a126 100644
--- a/chapter16/src/main/java/com/seaofnodes/simple/node/CastNode.java
+++ b/chapter16/src/main/java/com/seaofnodes/simple/node/CastNode.java
@@ -21,7 +21,6 @@ public class CastNode extends Node {
     @Override
     public String uniqueName() { return "Cast_" + _nid; }
 
-    @Override public boolean isPinned() { return true; }
 
     @Override
     StringBuilder _print1(StringBuilder sb, BitSet visited) {
diff --git a/chapter16/src/main/java/com/seaofnodes/simple/node/Node.java b/chapter16/src/main/java/com/seaofnodes/simple/node/Node.java
index e1e6283a..daf5994c 100644
--- a/chapter16/src/main/java/com/seaofnodes/simple/node/Node.java
+++ b/chapter16/src/main/java/com/seaofnodes/simple/node/Node.java
@@ -542,8 +542,6 @@ public abstract class Node implements OutNode {
     /** Return block start from a isCFG() */
     public Node getBlockStart() { return null; }
 
-    /** Pinned in the schedule; these are data nodes whose input#0 is not allowed to change */
-    public boolean isPinned() { return false; }
 
     // Semantic change to the graph (so NOT a peephole), used by the Parser.
     // If any input is a float, flip to a float-flavored opcode and widen any
diff --git a/chapter16/src/main/java/com/seaofnodes/simple/node/PhiNode.java b/chapter16/src/main/java/com/seaofnodes/simple/node/PhiNode.java
index f0140ce6..6cd9efb0 100644
--- a/chapter16/src/main/java/com/seaofnodes/simple/node/PhiNode.java
+++ b/chapter16/src/main/java/com/seaofnodes/simple/node/PhiNode.java
@@ -38,7 +38,6 @@ public class PhiNode extends Node {
     public CFGNode region() { return (CFGNode)in(0); }
     @Override public boolean isMultiTail() { return true; }
     @Override public boolean isMem() { return _declaredType instanceof TypeMem; }
-    @Override public boolean isPinned() { return true; }
 
     @Override
     public Type compute() {
diff --git a/chapter16/src/main/java/com/seaofnodes/simple/node/ProjNode.java b/chapter16/src/main/java/com/seaofnodes/simple/node/ProjNode.java
index 73c1fa15..cd57b85c 100644
--- a/chapter16/src/main/java/com/seaofnodes/simple/node/ProjNode.java
+++ b/chapter16/src/main/java/com/seaofnodes/simple/node/ProjNode.java
@@ -29,7 +29,6 @@ public class ProjNode extends Node {
 
     @Override public boolean isMultiTail() { return in(0).isMultiHead(); }
     @Override public boolean isMem() { return _type instanceof TypeMem; }
-    @Override public boolean isPinned() { return true; }
 
     @Override
     public Type compute() {