Finish forward-port mem ops thru ch25

21 files changed, 743 insertions, 521 deletions

diff --git a/chapter23/src/main/java/com/seaofnodes/simple/Parser.java b/chapter23/src/main/java/com/seaofnodes/simple/Parser.java
index 5b3f4d1a..2dc83905 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/Parser.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/Parser.java
@@ -138,7 +138,7 @@ public class Parser {
         _scope.define(ScopeNode.ARG0, TypeInteger.BOT, false, null, _lexer);
 
         ctrl(XCTRL);
-        _scope.mem(new MemMergeNode(false));
+        _scope.mem(con(TypeMem.BOT));
 
         // Parse the sys import
         _lexer = new Lexer(com.seaofnodes.simple.sys.SYS);
@@ -214,7 +214,7 @@ public class Parser {
         // Build a multi-exit return point for all function returns
         RegionNode r = new RegionNode( null,null,null).init();
         assert r.inProgress();
-        PhiNode rmem = new PhiNode(ScopeNode.MEM0,TypeMem.BOT,r,null).init();
+        PhiNode rmem = new BulkMemPhiNode(ScopeNode.MEM0,r,null).init();
         PhiNode rrez = new PhiNode(ScopeNode.ARG0,Type.BOTTOM,r,null).init();
         ReturnNode ret = new ReturnNode(r, rmem, rrez, rpc, fun).init();
         fun.setRet(ret);
@@ -227,11 +227,8 @@ public class Parser {
         // external calls.
         _scope.push(new Kind.Func());
         ctrl(fun);              // Scope control from function
-        // Private mem alias tracking per function
-        MemMergeNode mem = new MemMergeNode(true);
-        mem.addDef(null);       // Alias#0
-        mem.addDef(new ParmNode(ScopeNode.MEM0,1,TypeMem.BOT,fun,con(TypeMem.BOT)).peephole()); // All aliases
-        _scope.mem(mem);
+        // Unknown callers supply whole memory; field aliases are discovered lazily.
+        _scope.mem(new ParmNode(ScopeNode.MEM0,1,TypeMem.BOT,fun,con(TypeMem.BOT)).peephole());
         // All args, "as-if" called externally
         for( int i=0; i<ids.length; i++ ) {
             Type t = sig.arg(i);
@@ -246,7 +243,7 @@ public class Parser {
         // Last expression is the return except for the top-level main
         if( ctrl()._type==Type.CONTROL )
             if( fun.sig() != _code._main )
-                fun.addReturn(ctrl(), _scope.mem().merge(), last);
+                fun.addReturn(ctrl(), _scope.mem(), last);
             else {
                 fun.setDef(1,XCTRL); // Kill default main
                 _code.addAll(fun._outputs);
@@ -627,9 +624,8 @@ public class Parser {
      */
     private Node parseReturn() {
         var expr = require(parseAsgn(), ";");
-        // Need default memory, since it can be lazy, need to force
-        // a non-lazy Phi
-        _fun.addReturn(ctrl(), _scope.mem().merge(), expr);
+        // Force the lazy whole-memory Phi before closing this exit.
+        _fun.addReturn(ctrl(), _scope.mem(), expr);
         ctrl(XCTRL);            // Kill control
         return expr;
     }
@@ -1380,31 +1376,31 @@ public class Parser {
         Field[] fs = obj._fields;
         if( fs==null )
             throw error("Unknown struct type '" + obj._name + "'");
-        int len = fs.length;
-        Node[] ns = new Node[2+len];
-        ns[0] = ctrl();         // Control in slot 0
-        // Total allocated length in bytes
-        ns[1] = size;
-        // Memory aliases for every field
-        for( int i = 0; i < len; i++ )
-            if( !fs[i]._one )
-                ns[2+i] = memAlias(fs[i]._alias);
-        Node nnn = new NewNode(TypeMemPtr.make(obj), ns).peephole().keep();
-        for( int i = 0; i < len; i++ )
-            if( !fs[i]._one )
-                memAlias(fs[i]._alias, new ProjNode(nnn,i+2,memName(fs[i]._alias)).peephole());
-        Node ptr = new ProjNode(nnn.unkeep(),1,obj._name).peephole().keep();
-
-        // Initial nonzero values for every field
-        for( int i = 0; i < len; i++ ) {
-            Node val = init.get( i + idx );
+        Node prior = mem().keep();
+        MemMergeNode input = new MemMergeNode((Node)null);
+        for( Field f : fs ) if( !f._one ) input.alias(f._alias,prior);
+        // Allocation zeroes its fields; nonzero initializers become explicit Stores.
+        Node[] ns = new Node[3];
+        ns[0] = ctrl();
+        ns[1] = input.peephole();
+        ns[2] = size;
+        Node nnn = new NewNode(TypeMemPtr.make(obj),ns).peephole().keep();
+        Node ptr = new ProjNode(nnn,0,obj._name).peephole().keep();
+        Node out = new ProjNode(nnn,1,"$mem").peephole().keep();
+        MemMergeNode after = new MemMergeNode(prior);
+        for( Field f : fs ) if( !f._one ) after.alias(f._alias,out);
+        mem(after.peephole());
+        out.unkeep();
+        nnn.unkeep();
+        prior.unkeep();
+        for( int i=0; i<fs.length; i++ ) {
+            Node val = init.get(idx+i);
             if( !fs[i]._one && val._type != val._type.makeZero() ) {
-                Node mem = memAlias(fs[i]._alias);
-                Node st = new StoreNode(loc(),fs[i]._fname,fs[i]._alias,fs[i]._t,mem,ptr,off(obj,fs[i]._fname),val,true).peephole();
-                memAlias(fs[i]._alias,st);
+                Field f = fs[i];
+                Node st = new StoreNode(loc(),f._fname,f._alias,f._t,mem(),ptr,off(obj,f._fname),val,true).peephole();
+                mem(new MemMergeNode(mem(),f._alias,st).peephole());
             }
         }
-
         return ptr.unkeep();
     }
 
@@ -1430,13 +1426,18 @@ public class Parser {
         return con(str);
     }
 
-    // We set up memory aliases by inserting special vars in the scope these
-    // variables are prefixed by $ so they cannot be referenced in Simple code.
-    // Using vars has the benefit that all the existing machinery of scoping
-    // and phis work as expected
-    private Node memAlias(int alias         ) { return _scope.mem(alias    ); }
-    private void memAlias(int alias, Node st) {        _scope.mem(alias, st); }
-    public static String memName(int alias) { return ("$"+alias).intern(); }
+    // Memory is one hidden SSA variable, including across branches and loops.
+    private Node mem() { return _scope.mem(); }
+    private void mem(Node n) { _scope.mem(n); }
+
+    private void store(String name, int alias, Type glb, Node ptr, Node off, Node val, boolean init, Node ctrl) {
+        Node prior = mem().keep();
+        Node st = new StoreNode(loc(),name,alias,glb,prior,ptr,off,val,init);
+        st.setDef(0,ctrl);
+        st = st.peephole();
+        mem(new MemMergeNode(prior,alias,st).peephole());
+        prior.unkeep();
+    }
 
     /**
      * Parse postfix expression; this can be a field expression, an array
@@ -1494,6 +1495,7 @@ public class Parser {
         if( tf instanceof TypeMemPtr ftmp && ftmp.isFRef() )
             tf = ftmp.makeFrom(((TypeMemPtr)(TYPES.get(ftmp._obj._name)))._obj);
 
+        expr.keep();
         // Field offset; fixed for structs, computed for arrays
         Node off = (name.equals("[]")       // If field is an array body
             // Array index math
@@ -1506,21 +1508,14 @@ public class Parser {
             Node val = parseAsgn().keep();
             Node lift = liftExpr( val, tf, f._final, false );
 
-            Node st = new StoreNode(loc(), name, f._alias, tf, memAlias(f._alias), expr, off.unkeep(), lift, false);
-            // Arrays include control, as a proxy for a safety range check.
-            // Structs don't need this; they only need a NPE check which is
-            // done via the type system.
-            if( base.isAry() )  st.setDef(0,ctrl());
-            memAlias(f._alias, st.peephole());
+            store(name,f._alias,tf,expr.unkeep(),off.unkeep(),lift,false,base.isAry() ? ctrl() : null);
             return val.unkeep();        // "obj.a = expr" returns the expression while updating memory
         }
 
-        // Keep expr for possible store update
-        expr.keep();
         // Load field
         Node load = f._one && INITS.containsKey( base._name )
             ? INITS.get(base._name).in(base.find(name))
-            : new LoadNode(loc(),name, f._alias, tf, memAlias(f._alias), expr, off);
+            : new LoadNode(loc(),name, f._alias, tf, mem(), expr, off);
         // Arrays include control, as a proxy for a safety range check
         // Structs don't need this; they only need a NPE check which is
         // done via the type system.
@@ -1530,13 +1525,8 @@ public class Parser {
         // Check for assign-update, "ptr.fld += expr" or "ary[idx]++"
         char ch = _lexer.matchOperAssign();
         if( ch!=0 ) {
-            Node op = opAssign(ch,load, tf );
-            Node st = new StoreNode(loc(), name, f._alias, tf, memAlias(f._alias), expr.unkeep(), off.unkeep(), op, false);
-            // Arrays include control, as a proxy for a safety range check.
-            // Structs don't need this; they only need a NPE check which is
-            // done via the type system.
-            if( base.isAry() )  st.setDef(0,ctrl());
-            memAlias(f._alias, peep(st));
+            Node op = opAssign(ch,load,tf);
+            store(name,f._alias,tf,expr.unkeep(),off.unkeep(),op,false,base.isAry() ? ctrl() : null);
             load = postfix(ch) ? load.unkeep() : op;
             // And use the original loaded value as the result
             return load;
@@ -1647,7 +1637,7 @@ public class Parser {
         }
         // Control & memory after parsing args
         args.set(0,ctrl().keep());
-        args.set(1,_scope.mem().merge().keep());
+        args.set(1,_scope.mem().keep());
         args.push(fcn);        // Function pointer
         // Unkeep them all
         for( Node arg : args )
@@ -1668,11 +1658,8 @@ public class Parser {
         call.peephole();        // Rerun peeps after CallEnd, allows early inlining
         // Control from CallEnd
         ctrl(new CProjNode(cend,0,ScopeNode.CTRL).peephole());
-        // Memory from CallEnd
-        MemMergeNode mem = new MemMergeNode(true);
-        mem.addDef(null);       // Alias#0
-        mem.addDef(new ProjNode(cend,1,ScopeNode.MEM0).peephole());
-        _scope.mem(mem);
+        // A call may update any alias, so its memory result is the new whole state.
+        _scope.mem(new ProjNode(cend,1,ScopeNode.MEM0).peephole());
         // Call result
         return new ProjNode(cend,2,"#2").peephole();
     }
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/codegen/GlobalCodeMotion.java b/chapter23/src/main/java/com/seaofnodes/simple/codegen/GlobalCodeMotion.java
index b8a8d2b3..1d802cae 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/codegen/GlobalCodeMotion.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/codegen/GlobalCodeMotion.java
@@ -50,7 +50,7 @@ public abstract class GlobalCodeMotion {
         var cons = new ArrayList<Node>();
         for( Node con : start._outputs )
             if( con!=null && !(con instanceof CFGNode) &&
-                (con.isConst() || con instanceof MachNode mach && mach.outregmap()!=null) ) {
+                (con.isConst() || (con instanceof MachNode mach && mach.outregmap()!=null)) ) {
                 globals.put(con,true);
                 cons.add(con.keep()); // Preserve original inputs while rewiring users.
             }
@@ -172,16 +172,8 @@ public abstract class GlobalCodeMotion {
 
                 // Loads need their memory inputs' uses also done
                 if( n instanceof MemOpNode ld && ld._isLoad )
-                    for( Node memuse : ld.mem()._outputs )
-                        if( late[memuse._nid]==null &&
-                            // New makes new memory, never crushes load memory
-                            !(memuse instanceof NewNode) &&
-                            // Load-use directly defines memory
-                            (memuse._type instanceof TypeMem ||
-                             // Load-use directly defines memory
-                             memuse instanceof CallNode ||
-                             // Load-use indirectly defines memory
-                             (memuse._type instanceof TypeTuple tt && tt._types[ld._alias] instanceof TypeMem)) )
+                    for( Node memuse : ld.antiDeps() )
+                        if( late[memuse._nid]==null )
                             continue outer;
 
                 // All uses done, schedule
@@ -193,9 +185,7 @@ public abstract class GlobalCodeMotion {
             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 MemOpNode ld && ld._isLoad && late[ld._nid]==null )
-                        work.push(ld);
+                wakeLoads(def,late,work,new BitSet());
             }
             if( n instanceof LoopNode loop )
                 for( Node phi : loop._outputs )
@@ -255,23 +245,32 @@ public abstract class GlobalCodeMotion {
             best instanceof IfNode;
     }
 
+    // An allocation may consume a load's memory through a partial aggregate.
+    private static void wakeLoads(Node def, CFGNode[] late, WorkList<Node> work, BitSet visit) {
+        if( visit.get(def._nid) ) return;
+        visit.set(def._nid);
+        for( Node out : def._outputs )
+            if( out instanceof MemOpNode ld && ld._isLoad && late[ld._nid]==null ) work.push(ld);
+        if( def instanceof MemMergeNode )
+            for( int i=1; i<def.nIns(); i++ )
+                if( def.in(i)!=null ) wakeLoads(def.in(i),late,work,visit);
+    }
+
     private static CFGNode find_anti_dep(CFGNode lca, MemOpNode 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() )
             anti[cfg._nid] = load._nid;
         // Walk load->mem uses, looking for Stores causing an anti-dep
-        for( Node mem : load.mem()._outputs ) {
+        for( Node mem : load.antiDeps() ) {
             switch( mem ) {
             case MemOpNode st:
-                if( !st._isLoad && load._alias == st._alias ) {
-                    assert late[mem._nid] != null;
-                    lca = anti_dep(load,late[mem._nid],lca,st,anti);
-                }
-                break; // Loads do not cause anti-deps on other loads
-            case CallNode call:
-                assert late[call._nid]!=null;
-                lca = anti_dep(load,late[call._nid],lca,call,anti);
+                assert late[st._nid]!=null;
+                lca = anti_dep(load,late[st._nid],lca,st,anti);
+                break;
+            case CallNode st:
+                assert late[st._nid]!=null;
+                lca = anti_dep(load,late[st._nid],lca,st,anti);
                 break;
             case PhiNode phi:
                 // Repeat anti-dep for matching Phi inputs.
@@ -280,10 +279,10 @@ public abstract class GlobalCodeMotion {
                     if( phi.in(i)==load.mem() )
                         lca = anti_dep(load,phi.region().cfg(i),lca,null,anti);
                 break;
-            case NewNode st: break;
-            case ReturnNode ret: break; // Load must already be ahead of Return
-            case MemMergeNode ret: break; // Mem uses now on ScopeMin
-            case NeverNode never: break;
+            case NewNode st:
+                assert late[st._nid]!=null;
+                lca = anti_dep(load,late[st._nid],lca,st,anti);
+                break;
             default: throw Utils.TODO();
             }
         }
@@ -295,7 +294,7 @@ public abstract class GlobalCodeMotion {
         // Stores are already placed.  Constrain the load at that block only.
         if( anti[stblk._nid]==load._nid ) {
             lca = stblk.domLCA(lca,null);
-            if( lca==stblk && st!=null && st._inputs.find(load) == -1 )
+            if( lca==stblk && st!=null && !(st instanceof CallNode) && st._inputs.find(load) == -1 )
                 st.addDef(load); // Same-block load must precede the store.
         }
         return lca;
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/BulkMemPhiNode.java b/chapter23/src/main/java/com/seaofnodes/simple/node/BulkMemPhiNode.java
new file mode 100644
index 00000000..7d4f1fb8
--- /dev/null
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/BulkMemPhiNode.java
@@ -0,0 +1,240 @@
+package com.seaofnodes.simple.node;
+
+import com.seaofnodes.simple.codegen.CodeGen;
+import com.seaofnodes.simple.type.Type;
+import com.seaofnodes.simple.type.TypeMem;
+import java.util.BitSet;
+
+/**
+ * A Phi for bulk memory.  It covers all aliases except {@link #_aliases};
+ * those aliases must be represented by parallel precise MemPhis at the same
+ * program slice.
+ */
+public class BulkMemPhiNode extends PhiNode {
+
+    private static final BitSet EMPTY = new BitSet();
+
+    // Excluded Aliases; empty means NO exclusions.
+    // Each split clones the exclusions; published sets are never mutated.
+    public final BitSet _aliases;
+
+    public BulkMemPhiNode(String label, Node... inputs) {
+        this(label,EMPTY,inputs);
+    }
+
+    private BulkMemPhiNode(String label, BitSet aliases, Node... inputs) {
+        super(label,TypeMem.BOT,inputs);
+        _aliases = aliases;
+    }
+
+    public boolean isSplit(int alias) { return _aliases.get(alias); }
+
+    public BulkMemPhiNode(BulkMemPhiNode phi) { super(phi); _aliases = phi._aliases; }
+
+    @Override public String label() { return "BulkPhi"+_aliases; }
+
+    @Override
+    public Type compute() {
+        Type t = super.compute();
+        return t instanceof TypeMem mem ? mem :
+            t==Type.BOTTOM ? TypeMem.BOT :
+            TypeMem.TOP;
+    }
+
+    @Override
+    public Node idealize() {
+        Node progress = super.idealize();
+        if( progress != null ) return progress;
+        if( ((RegionNode)region()).inProgress() || region().nIns()<=1 )
+            return null;        // Input is in-progress
+
+        // "Peek through" a MemMerge that covers this alias set on its default
+        for( int i=1; i<nIns(); i++ )
+            if( in(i) instanceof MemMergeNode mmm && canPeek(mmm) ) {
+                setDef(i,CodeGen.CODE.add(mmm.in(1)));
+                return this;
+            }
+
+        // If any input or output uses or defines a specific alias we cover,
+        // make a private MemPhi for it, and add it to the excluded list.
+        for( int i=1; i<nIns(); i++ ) {
+            int alias = inputAlias(in(i));
+            if( alias!=0 )
+                return slice(alias);
+        }
+
+        for( int i=0; i<_outputs.size(); i++ ) {
+            int alias = outputAlias(_outputs.get(i));
+            if( alias!=0 )
+                return slice(alias);
+        }
+
+        return null;
+    }
+
+    static public boolean checkMem(Node n, int alias) {
+        for( Node m : n._inputs )
+            // Bulk memory supplies this alias
+            if( m instanceof BulkMemPhiNode bulk ) {
+                if( bulk._aliases.get(alias) )
+                    return false;
+            } else if( m instanceof MemMergeNode mmm &&
+                       mmm.in(1) instanceof BulkMemPhiNode bulk &&
+                       bulk._aliases.get(alias) &&
+                       (alias >= mmm.nIns() || mmm.in(alias)==null) )
+                return false;
+        return true;
+    }
+
+
+    // True if the MemMerge has no precise slice still covered by this Phi.
+    private boolean canPeek(MemMergeNode mmm) {
+        return mmm.in(1)!=null && missingAlias(mmm)==0;
+    }
+
+    // Return an alias required by an input but still covered by this Phi, or 0.
+    private int inputAlias(Node n) {
+        return switch(n) {
+        case ProjNode proj -> 0;
+        case ParmNode parm -> 0;
+        case ConstantNode con -> 0;
+        case MemMergeNode mmm -> missingAlias(mmm);
+        case MemOpNode mem -> unsplit(mem._alias);
+        case BulkMemPhiNode bulk -> missingAlias(bulk);
+        default -> throw new AssertionError("Unexpected memory input "+n);
+        };
+    }
+
+    // Return an alias required by a user but still covered by this Phi, or 0.
+    private int outputAlias(Node use) {
+        // This decision depends backwards on an immediate user's alias/set.
+        // Normal graph-neighbor enqueueing flows from defs to uses, so retain
+        // an explicit forward dependency to revisit this Phi when the user
+        // sharpens.
+        addDepForwards(use);
+        return switch(use) {
+        case ScopeNode scope -> 0;
+        case MemMergeNode mmm -> missingAlias(mmm);
+        case BulkMemPhiNode bulk -> missingAlias(bulk);
+        case MemOpNode mem -> unsplit(mem._alias);
+        case MemPhiNode phi -> unsplit(phi._alias);
+        default -> {
+            assert use instanceof ReturnNode || use instanceof CallNode || use instanceof ParmNode
+                : "Unexpected bulk-memory user "+use;
+            yield 0;
+        }
+        };
+    }
+
+    // Alias zero/one is bulk, and an already excluded precise alias cannot
+    // trigger another split of this Phi.
+    private int unsplit(int alias) {
+        if( alias <= 1 ) return 0;
+        assert !_aliases.get(alias);
+        return alias;
+    }
+
+    // First alias excluded by the neighboring bulk Phi but not by this one.
+    private int missingAlias(BulkMemPhiNode bulk) {
+        for( int alias = bulk._aliases.nextSetBit(0); alias >= 0; alias = bulk._aliases.nextSetBit(alias+1) )
+            if( !_aliases.get(alias) )
+                return alias;
+        return 0;
+    }
+
+    // Every explicit slot requests that alias, including a partial allocation
+    // input whose slot still points to this bulk Phi.
+    private int missingAlias(MemMergeNode mmm) {
+        for( int alias=2; alias<mmm.nIns(); alias++ )
+            if( mmm.in(alias)!=null &&
+                mmm.alias(alias)!=mmm.in(1) &&
+                !_aliases.get(alias) )
+                return alias;
+        return 0;
+    }
+
+    // Slice out given alias
+    private MemMergeNode slice( int alias ) {
+        assert !_aliases.get(alias);
+        // Inputs are the same; MemPhi will sharpen his own inputs.  If a
+        // parallel precise Phi already exists, reuse it instead of building a
+        // duplicate for the same Region/alias pair.
+        Node mem = precisePhi(alias);
+
+        BitSet aliases = ((BitSet)_aliases.clone());
+        aliases.set(alias);
+        BulkMemPhiNode bphi = new BulkMemPhiNode(_label,aliases);
+        for( int i=0; i<nIns(); i++ )
+            bphi.addDef(in(i));
+        // Installing bphi creates a new bulk user with a different exclusion
+        // set.  Queue the original memory inputs before bphi itself can peep
+        // away and hide those neighbor relationships.
+        for( int i=1; i<bphi.nIns(); i++ )
+            CodeGen.CODE.add(bphi.in(i));
+        // Do not peephole bphi here, as BulkMemPhi can recursively start a
+        // second bulk rewrite before this one has finished.  Let the worklist
+        // discover any further splits
+        bphi.setType(TypeMem.BOT);
+        Node bulk = bphi;
+        CodeGen.CODE.add(bulk);
+        CodeGen.CODE.add(mem);
+        return aggregate(bulk,alias,mem);
+    }
+
+    // Build aggregate memory over a bulk Phi, preserving every precise slice
+    // which the bulk has already excluded.  The named alias is newly supplied
+    // (or replaces its old slice).
+    MemMergeNode aggregate(Node bulk, int alias, Node precise) {
+        MemMergeNode mmm = new MemMergeNode(bulk);
+        for( int old = _aliases.nextSetBit(0); old >= 0; old = _aliases.nextSetBit(old+1) )
+            { assert old != alias; mmm.alias(old,precisePhi(old));}
+        mmm.alias(alias,precise);
+        assert checkMerge(mmm);
+        return mmm;
+    }
+
+    // Find or make the precise Phi parallel to this bulk Phi at the same merge.
+    private Node precisePhi(int alias) {
+        MemPhiNode mphi = _findPhi(alias);
+        if( mphi!=null ) return mphi;
+
+        mphi = new MemPhiNode("$"+alias,alias);
+        mphi.addDef(region());
+        // Due to cycles, must set before calling peephole
+        mphi.setType(TypeMem.make(alias,Type.BOTTOM));
+        for( int i=1; i<nIns(); i++ )
+            mphi.addDef(CodeGen.CODE.add(preciseInput(in(i),alias)));
+        return CodeGen.CODE.add(mphi);
+    }
+
+    MemPhiNode _findPhi(int alias) {
+        for( Node use : region()._outputs )
+            if( use instanceof MemPhiNode mphi && mphi._alias==alias )
+                return mphi;
+        return null;
+    }
+
+    // Select an alias already split out of a predecessor's bulk memory.
+    private Node preciseInput(Node mem, int alias) {
+        if( mem instanceof MemMergeNode mmm )
+            return mmm.alias(alias);
+        if( mem instanceof BulkMemPhiNode bulk && bulk.isSplit(alias) )
+            return bulk.precisePhi(alias);
+        return mem;
+    }
+
+    // Every alias excluded by a bulk default must be supplied explicitly.
+    static boolean checkMerge(MemMergeNode mmm) {
+        if( !(mmm.in(1) instanceof BulkMemPhiNode bulk) ) return true;
+        for( int alias = bulk._aliases.nextSetBit(0); alias >= 0; alias = bulk._aliases.nextSetBit(alias+1) )
+            if( alias >= mmm.nIns() || mmm.in(alias)==null )
+                return false;
+        return true;
+    }
+
+    @Override public boolean eq(Node n) {
+        return _aliases.equals(((BulkMemPhiNode)n)._aliases) && super.eq(n);
+    }
+
+    @Override int hash() { return _aliases.hashCode(); }
+}
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/LoadNode.java b/chapter23/src/main/java/com/seaofnodes/simple/node/LoadNode.java
index e752c9ce..8b526be2 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/node/LoadNode.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/LoadNode.java
@@ -1,7 +1,7 @@
 package com.seaofnodes.simple.node;
 
-import com.seaofnodes.simple.Parser;
 import com.seaofnodes.simple.codegen.CodeGen;
+import com.seaofnodes.simple.Parser;
 import com.seaofnodes.simple.type.*;
 import com.seaofnodes.simple.util.Utils;
 import java.util.BitSet;
@@ -24,6 +24,25 @@ public class LoadNode extends MemOpNode {
         super(loc, name, alias, true, glb, mem, ptr, off);
     }
 
+    @Override boolean canDrop(MemOpNode other, Node dep) {
+        return super.canDrop(other,dep) && !clobbered(dep);
+    }
+
+    // Check only immediate memory users. Writers clobber memory; Phis and
+    // aggregates might lead to a clobber, so stop rather than search further.
+    private boolean clobbered(Node dep) {
+        Node mem = mem();
+        dep.addDepForwards(mem);
+        for( Node use : mem._outputs ) {
+            if( use==null ) continue;
+            dep.addDepForwards(use);
+            if( use instanceof StoreNode || use instanceof PhiNode ||
+                use instanceof MemMergeNode ||
+                use instanceof NewNode || use instanceof CallNode ) return true;
+        }
+        return false;
+    }
+
     // Debugger label
     @Override public String  label() { return "ld_"+mlabel(); }
     @Override
@@ -42,7 +61,7 @@ public class LoadNode extends MemOpNode {
             t = ary.elem();     // TODO: if offset is known, can peek the constant
         }
         // Lift from declared type and memory input
-        t = t.join(_declaredType).join(mem._t);
+        t = t.join(_declaredType).join(MemMergeNode.contents(mem(),_alias,this));
         if( _declaredType.isFinal() )
             t = t.makeRO(); // Deep final applied
         return t;
@@ -50,35 +69,19 @@ public class LoadNode extends MemOpNode {
 
     @Override
     public Node idealize() {
-        Node ptr = ptr();
-        Node mem = mem();
-
-        if( mem instanceof CastNode cast ) {
+        if( mem() instanceof CastNode cast ) {
             setDef(1,cast.in(1));
             return this;
         }
-
-        // Simple Load-after-Store on same address.
-        if( mem instanceof StoreNode st &&
-            ptr == st.ptr() && off() == st.off() ) { // Must check same object
-            assert _name.equals(st._name); // Equiv class aliasing is perfect
-            return extend(st.val());
-        }
-
-        // Simple load-after-MemMerge to a known alias can bypass.  Happens when inlining.
-        if( mem instanceof MemMergeNode mem2 ) {
-            Node memA = mem2.alias(_alias);
-            for( Node ld : memA._outputs )
-                if( ld instanceof LoadNode )
-                    CodeGen.CODE.add(ld);
-            setDef(1,memA);
+        if( mem() instanceof MemMergeNode merge ) {
+            setDef(1,CodeGen.CODE.add(merge.alias(_alias)));
             return this;
         }
+        Node ptr = ptr();
+        Node mem = mem();
 
-        // Simple Load-after-New on same address.
-        if( mem instanceof ProjNode p && p.in(0) instanceof NewNode nnn &&
-            ptr == nnn.proj(1) ) // Must check same object
-            return zero(nnn);   // Load zero from new
+        Node win = find(mem,ptr,null,null);
+        if( win!=null ) return folded(win);
 
         // Uplift control to a prior dominating load.
         for( Node memuse : mem._outputs )
@@ -87,50 +90,7 @@ public class LoadNode extends MemOpNode {
                 cfg0()!=null && cfg0().domLCA(ld.cfg0(),this) == ld.cfg0() ) // Higher control means load is legal earlier
                 return ld;
 
-        // Load-after-Store on same address, but bypassing provably unrelated
-        // stores.  This is a more complex superset of the above two peeps.
-        // "Provably unrelated" is really weak.
-        if( ptr instanceof ReadOnlyNode ro )
-            ptr = ro.in(1);
-        outer:
-        while( true ) {
-            switch( mem ) {
-            case StoreNode st:
-                if( ptr == addDep(st.ptr()) && off() == st.off() )
-                    return extend(castRO(st.val())); // Proved equal
-                // Can we prove unequal?  Offsets do not overlap?
-                if( !off()._type.join(st.off()._type).isHigh() && // Offsets overlap
-                    !neverAlias(ptr,st.ptr()) )                   // And might alias
-                    break outer; // Cannot tell, stop trying
-                // Pointers cannot overlap
-                mem = st.mem(); // Proved never equal
-                break;
-            case PhiNode phi:
-                // Assume related
-                addDep(phi);
-                break outer;
-            case ConstantNode top: break outer;  // Assume shortly dead
-            case ProjNode mproj: // Memory projection
-                switch( mproj.in(0) ) {
-                case NewNode nnn1:
-                    if( ptr instanceof ProjNode pproj && pproj.in(0) == mproj.in(0) )
-                        return zero(nnn1);
-                    if( !(ptr instanceof ProjNode pproj && pproj.in(0) instanceof NewNode) )
-                        break outer; // Cannot tell, ptr not related to New
-                    mem = nnn1.in(nnn1.findAlias(_alias));// Bypass unrelated New
-                    break;
-                case StartNode  start: break outer;
-                case CallEndNode cend: break outer; // TODO: Bypass no-alias call
-                default: throw Utils.TODO();
-                }
-                break;
-            case CastNode cast: mem = cast.in(1); break;
-            case MemMergeNode merge:  mem = merge.alias(_alias);  break;
-
-            default:
-                throw Utils.TODO();
-            }
-        }
+        if( ptr instanceof ReadOnlyNode ro ) ptr=ro.in(1);
 
         // Push a Load up through a Phi, as long as it collapses on at least
         // one arm.  If at a Loop, the backedge MUST collapse - else we risk
@@ -139,19 +99,27 @@ public class LoadNode extends MemOpNode {
         //   if( pred ) ptr.x = e0;         val = pred ? e0
         //   else       ptr.x = e1;                    : e1;
         //   val = ptr.x;                   ptr.x = val;
-        if( mem() instanceof PhiNode memphi && memphi.region()._type == Type.CONTROL && memphi.nIns()== 3 &&
+        if( mem() instanceof MemPhiNode memphi && memphi.region()._type == Type.CONTROL && memphi.nIns()== 3 &&
+            // Array access control must not be moved across the merge.
+            in(0)==null &&
             // Offset can be hoisted
             off() instanceof ConstantNode &&
             // Pointer can be hoisted
             hoistPtr(ptr,memphi)  ) {
 
-            // Profit on RHS/Loop backedge
-            if( profit(memphi,2) ||
-                // Else must not be a loop to count profit on LHS.
-                (!(memphi.region() instanceof LoopNode) && profit(memphi,1)) ) {
-                Node ld1 = ld(1);
-                Node ld2 = ld(2);
-                return new PhiNode(_name,_declaredType,memphi.region(),ld1,ld2);
+            // Returning to this memory means no change only if the pointer
+            // is also loop invariant. A pointer Phi may select another object.
+            Node stop = memphi.region() instanceof LoopNode &&
+                !(ptr instanceof PhiNode p && p.region()==memphi.region()) ? memphi : null;
+            if( find(memphi.in(2),ptr,stop,new BitSet())!=null ||
+                (!(memphi.region() instanceof LoopNode) && find(memphi.in(1),ptr,null,new BitSet())!=null) ) {
+                PhiNode phi = new PhiNode(_name,_declaredType,memphi.region()).keep();
+                phi.setType(_type); // Backedge searches may return this value Phi.
+                for( int i=1; i<memphi.nIns(); i++ ) {
+                    Node p = ptr instanceof PhiNode pp && pp.region()==memphi.region() ? pp.in(i) : ptr;
+                    phi.addDef(load(memphi.in(i),p,stop,phi));
+                }
+                return CodeGen.CODE.add(phi.unkeep());
             }
         }
 
@@ -166,9 +134,67 @@ public class LoadNode extends MemOpNode {
         return castRO(new ConstantNode(zero).peephole());
     }
 
-    private Node ld( int idx ) {
-        Node mem = mem(), ptr = ptr();
-        return new LoadNode(_loc,_name,_alias,_declaredType,mem.in(idx),ptr instanceof PhiNode && ptr.in(0)==mem.in(0) ? ptr.in(idx) : ptr, off()).peephole();
+    // Search only; no new nodes or rewiring. Return a Store/New that folds,
+    // the unchanged loop memory, or a merge whose every arm folds. Null fails.
+    // Register dependencies so a later pointer/offset rewrite retries the query.
+    private Node find(Node mem, Node ptr, Node stop, BitSet visit) {
+        if( ptr instanceof ReadOnlyNode ro ) ptr=ro.in(1);
+        while( mem!=null ) {
+            addDep(mem);
+            if( mem==stop ) return mem;
+            switch( mem ) {
+            case MemMergeNode merge: mem=merge.alias(_alias); break;
+            case CastNode cast: mem=cast.in(1); break;
+            case StoreNode st:
+                addDep(st.ptr());
+                if( _alias==st._alias && ptr==st.ptr() && off()==st.off() ) return st;
+                if( _alias==st._alias && !neverAlias(ptr,st.ptr()) &&
+                    !addDep(off())._type.join(addDep(st.off())._type).isHigh() ) return null;
+                mem=st.mem();
+                break;
+            case ProjNode proj:
+                if( !(proj.in(0) instanceof NewNode nnn) ) return null;
+                if( ptr instanceof ProjNode p && p.in(0)==nnn ) return nnn;
+                if( !(ptr instanceof ProjNode p && p.in(0) instanceof NewNode) ) return null;
+                mem=nnn.mem(); // Distinct allocations cannot overlap.
+                break;
+            case MemPhiNode phi:
+                addDep(phi.region());
+                // Only the original loop closes a successful cycle. Another
+                // loop or an unfinished merge needs a separate proof.
+                if( visit==null || !(phi.region() instanceof RegionNode r) || r instanceof LoopNode ||
+                    r.inProgress() || phi.inProgress() || r._type!=Type.CONTROL ) return null;
+                if( visit.get(phi._nid) ) return phi; // Already proved every arm.
+                for( int i=1; i<phi.nIns(); i++ )
+                    if( find(phi.in(i),ptr,stop,visit)==null ) return null;
+                visit.set(phi._nid);
+                return phi;
+            default: return null;
+            }
+        }
+        return null;
+    }
+
+    private Node folded(Node win) {
+        return win instanceof StoreNode st ? extend(castRO(st.val())) : zero((NewNode)win);
+    }
+
+    // Materialize only after the profitability search succeeds. The value Phi
+    // stands for a load which comes back around the loop with unchanged memory.
+    private Node load(Node mem, Node ptr, Node stop, PhiNode value) {
+        Node win=find(mem,ptr,stop,new BitSet());
+        if( win!=null ) {
+            if( win==stop ) return value;
+            if( win instanceof MemPhiNode mp ) {
+                PhiNode phi=new PhiNode(_name,_declaredType,mp.region()).keep();
+                for( int i=1; i<mp.nIns(); i++ ) phi.addDef(load(mp.in(i),ptr,stop,value));
+                return phi.unkeep().peephole();
+            }
+            return folded(win).peephole();
+        }
+        LoadNode ld=new LoadNode(_loc,_name,_alias,_declaredType,mem,ptr,off());
+        ld.setDef(0,in(0));
+        return ld.peephole();
     }
 
     private static boolean neverAlias( Node ptr1, Node ptr2 ) {
@@ -197,16 +223,6 @@ public class LoadNode extends MemOpNode {
         return false;
     }
 
-    // Profitable if we find a matching Store on this Phi arm.
-    private boolean profit(PhiNode phi, int idx) {
-        Node px = phi.in(idx);
-        if( px==null ) return false;
-        if( px._type instanceof TypeMem mem && mem._t.isHighOrConst() ) return true;
-        if( px instanceof StoreNode st1 && ptr()==addDep(st1.ptr() )&& off()==st1.off() ) return true;
-        addDep(px);
-        return false;
-    }
-
     // Read-Only is a deep property, and cannot be cast-away
     private Node castRO(Node rez) {
         if( ptr()._type.isFinal() && !rez._type.isFinal() )
@@ -228,4 +244,5 @@ public class LoadNode extends MemOpNode {
         Node shl = new ShlNode(null,val,shf.keep()).peephole();
         return new SarNode(null,shl,shf.unkeep());
     }
+
 }
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/LoopNode.java b/chapter23/src/main/java/com/seaofnodes/simple/node/LoopNode.java
index 6c69a56e..db52c156 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/node/LoopNode.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/LoopNode.java
@@ -57,11 +57,12 @@ public class LoopNode extends RegionNode {
         // Now fold control into the exit.  Might have 1 valid exit, or an
         // XCtrl or a bunch of prior NeverNode exits.
         Node top = new ConstantNode(Type.TOP).peephole();
-        Node memout = new MemMergeNode(false);
-        memout.addDef(f); // placeholder for control
+        MemMergeNode memout = new MemMergeNode(new ConstantNode(TypeMem.BOT).peephole());
+        memout.setDef(0,f);
         for( Node u : _outputs )
-            if( u instanceof PhiNode phi && phi._type.isa(TypeMem.BOT) )
-                memout.addDef(phi);
+            if( u instanceof BulkMemPhiNode ) memout.setDef(1,u);
+            else if( u instanceof MemPhiNode phi ) memout.alias(phi._alias,phi);
+        memout.init();
 
         Node ctrl = ret.ctrl(), mem = ret.mem(), expr = ret.expr();
         if( ctrl!=null && ctrl._type != Type.XCONTROL ) {
@@ -72,7 +73,7 @@ public class LoopNode extends RegionNode {
                 // Nope, insert an aligned exit layer
                 RegionNode r = new RegionNode(_loc,null,ctrl).init();
                 ctrl = r;  r._ltree = stop._ltree;
-                mem  = new PhiNode(r,mem ).init();
+                mem  = new BulkMemPhiNode(ScopeNode.MEM0,r,mem).init();
                 expr = new PhiNode(r,expr).init();
             }
             // Append new Never exit
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/MemMergeNode.java b/chapter23/src/main/java/com/seaofnodes/simple/node/MemMergeNode.java
index 20aeb454..e56eeb06 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/node/MemMergeNode.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/MemMergeNode.java
@@ -5,176 +5,86 @@ import com.seaofnodes.simple.codegen.CodeGen;
 import com.seaofnodes.simple.type.*;
 import java.util.*;
 
-/**
- *  Memory Merge - a merge of many aliases into a "fat memory".  All aliases
- *  are here, but most will be lazy - take the default fat memory.
+/** A default slice and explicit alias overrides. A null default means partial
+ *  memory: absent aliases are not covered, rather than having unknown contents.
  */
 public class MemMergeNode extends Node {
+    public MemMergeNode(MemMergeNode mem) { super(mem); }
 
-    /*
-     *  In-Progress means this is being used by the Parser to track memory
-     *  aliases.  No optimizations are allowed.  When no longer "in progress"
-     *  normal peeps work.
-     */
-    public final boolean _inProgress;
-
-    public MemMergeNode( boolean inProgress) { _type = TypeMem.BOT; _inProgress = inProgress; }
-    public MemMergeNode(MemMergeNode mem) { super(mem); _inProgress = false; }
-
-
-    // If being used by a Scope, this is "in progress" from the Parser.
-    // Otherwise, it's a memory merge
-    boolean inProgress() { return _inProgress; }
+    public MemMergeNode(Node bulk) { super(null,bulk); _type = TypeMem.BOT; }
+    public MemMergeNode(Node bulk, int alias, Node precise) {
+        this(bulk);
+        alias(alias,precise);
+    }
+    public Node alias(int alias) {
+        Node n = alias < nIns() ? in(alias) : null;
+        assert n!=null || !(in(1) instanceof BulkMemPhiNode bulk) || !bulk.isSplit(alias);
+        n = n==null ? in(1) : n;
+        assert n!=null : "Alias not covered: "+alias;
+        return n;
+    }
 
-    @Override public String label() { return "ALLMEM"; }
-    @Override public boolean isMem() { return true; }
+    public void alias(int alias, Node mem) {
+        assert alias > 1;
+        while( nIns()<=alias ) addDef(null);
+        setDef(alias,mem);
+    }
 
-    @Override
-    public StringBuilder _print1(StringBuilder sb, BitSet visited) {
-        sb.append(_inProgress ? "Merge[" : "MEM[ ");
-        for( int j=2; j<nIns(); j++ ) {
-            sb.append(j);
-            sb.append(":");
-            Node n = in(j);
-            while( n instanceof ScopeNode loop ) {
-                sb.append("Lazy_");
-                n = loop.mem().alias(j);
-            }
-            if( n==null ) sb.append("___ ");
-            else n._print0(sb, visited).append(" ");
+    // Scalar contents of one alias. New zeroes the allocated fields;
+    // join it with the incoming contents for all previously allocated objects.
+    // Phis use their cached types, so this query does not recurse around loops.
+    static Type contents(Node mem, int alias, Node dep) {
+        dep.addDep(mem);
+        if( mem instanceof CastNode cast ) return contents(cast.in(1),alias,dep);
+        if( mem instanceof MemMergeNode merge )
+            return contents(merge.alias(alias),alias,dep);
+        if( mem instanceof ProjNode proj && proj.in(0) instanceof NewNode nnn ) {
+            assert proj._idx==1 && nnn.field(alias)!=null;
+            dep.addDep(nnn);
+            return contents(nnn.mem(),alias,dep).meet(nnn.field(alias)._t.makeZero());
         }
-        sb.setLength(sb.length()-1);
-        return sb.append("]");
+        // Function parameters and call results may contain arbitrary heap values.
+        // Their bulk memory types do not imply empty or zero-filled storage.
+        if( mem._type==Type.TOP || mem._type==TypeMem.TOP ) return Type.TOP;
+        if( mem._type instanceof TypeMem mt && mt._alias==alias ) return mt._t;
+        return Type.BOTTOM;
     }
 
-    // Make a memory merge: no longer a Scope really, tracking memory state but
-    // not related to the parser in any way.
-    public Node merge() {
-        // Force default memory to not be lazy
-        MemMergeNode merge = new MemMergeNode(false);
-        for( Node n : _inputs )
-            merge.addDef(n);
+    @Override public String label() { return "MemMerge"; }
+    @Override public boolean isMem() { return true; }
+    @Override public Type compute() { return TypeMem.BOT; }
+    @Override public StringBuilder _print1(StringBuilder sb, BitSet visited) {
+        sb.append("MEM[");
         for( int i=1; i<nIns(); i++ )
-            merge._mem(i,null);
-        return merge.peephole();
+            if( in(i)!=null ) {
+                sb.append(i==1 ? "default:" : i+":");
+                in(i)._print0(sb,visited).append(" ");
+            }
+        return sb.append("]");
     }
 
-
-    @Override public Type compute() { return TypeMem.BOT; }
-
     @Override public Node idealize() {
-        if( inProgress() ) return null;
-
-        // Fold defaults into the default
         boolean progress=false, allDefault=true;
         for( int i=2; i<nIns(); i++ ) {
-            if( in(i) instanceof CastNode cast )
-                { setDef(i,cast.in(1)); progress=true; }
-            if( in(1) == in(i) ) { setDef(i,null); progress=true; }
-            else                 { allDefault=false; }
+            if( in(i) instanceof CastNode cast ) { setDef(i,cast.in(1)); progress=true; }
+            if( in(i)!=null && in(i)==in(1) ) { setDef(i,null); progress=true; }
+            if( in(i) instanceof MemMergeNode mem ) {
+                setDef(i,CodeGen.CODE.add(mem.alias(i)));
+                progress=true;
+            }
+            if( in(i)!=null ) allDefault=false;
         }
+        if( allDefault && in(1)!=null ) return in(1);
 
-        // If not merging any memory (all memory is just the default)
-        if( allDefault )
-            return in(1);       // Become default memory
-
-        // Collapse stacked all-mem
         if( in(1) instanceof MemMergeNode mem ) {
-            // Goal is to swap my default mem with mem's default mem
-            for( int i=2; i<mem.nIns(); i++ ) {
-                if( mem.in(i) != null ) {
-                    // deeper default mem has a non-default
-                    if( i>=nIns() || in(i)==null )
-                        setDefX(i,mem.in(i));
-                }
-            }
+            for( int i=2; i<mem.nIns(); i++ )
+                if( mem.in(i)!=null && (i>=nIns() || in(i)==null) )
+                    alias(i,mem.in(i));
+            CodeGen.CODE.add(mem.in(1));
             setDef(1,mem.in(1));
             return this;
         }
-
+        assert BulkMemPhiNode.checkMerge(this);
         return progress ? this : null;
     }
-
-
-    public Node in( Var v ) { return in(v._idx); }
-
-    public Node alias( int alias ) {
-        return alias < nIns() && in(alias)!=null ? in(alias) : in(1);
-    }
-
-    Node alias( int alias, Node st ) {
-        while( alias >= nIns() ) addDef(null);
-        return setDef(alias,st);
-    }
-
-    // Read or update from memory.
-    // A shared implementation allows us to create lazy phis both during
-    // lookups and updates; the lazy phi creation is part of chapter 8.
-    Node _mem( int alias, Node st ) {
-        // Memory projections are made lazily; if one does not exist
-        // then it must be START.proj(1)
-        Node old = alias(alias);
-        if( old instanceof ScopeNode loop ) {
-            MemMergeNode loopmem = loop.mem();
-            Node memdef = loopmem.alias(alias);
-            // Lazy phi!
-            old = memdef instanceof PhiNode phi && loop.ctrl()==phi.region()
-                // Loop already has a real Phi, use it
-                ? memdef
-                // Set real Phi in the loop head
-                // The phi takes its one input (no backedge yet) from a recursive
-                // lookup, which might have insert a Phi in every loop nest.
-                : loopmem.alias(alias, new PhiNode(Parser.memName(alias), TypeMem.make(alias,Type.BOTTOM),loop.ctrl(),loopmem._mem(alias,null),null).peephole() );
-            alias(alias,old);
-        }
-        // Memory projections are made lazily; expand as needed
-        return st==null ? old : alias(alias,st); // Not lazy, so this is the answer
-    }
-
-
-    void _merge( MemMergeNode that, RegionNode r) {
-        int len = Math.max(nIns(),that.nIns());
-        for( int i = 1; i < len; i++)
-            if( alias(i) != that.alias(i) ) { // No need for redundant Phis
-                // If we are in lazy phi mode we need to a lookup
-                // by alias as it will trigger a phi creation
-                Node lhs = this._mem(i,null);
-                Node rhs = that._mem(i,null);
-                alias(i, new PhiNode(Parser.memName(i), lhs._type.meet(rhs._type).glb(true), r, lhs, rhs).peephole());
-            }
-    }
-
-    // Fill in the backedge of any inserted Phis
-    void _endLoopMem( ScopeNode scope, MemMergeNode back, MemMergeNode exit ) {
-        Node exit_def = exit.alias(1);
-        for( int i=1; i<nIns(); i++ ) {
-            if( in(i) instanceof PhiNode phi && phi.region()==scope.ctrl() ) {
-                assert phi.in(2)==null;
-                phi.setDef(2,back.alias(i)==scope ? phi : back.alias(i)); // Fill backedge
-            }
-            if( exit_def == scope ) // Replace a lazy-phi on the exit path also
-                exit.alias(i,in(i));
-        }
-    }
-
-    // Now one-time do a useless-phi removal
-    void _useless( ) {
-        for( int i=2; i<nIns(); i++ ) {
-            if( in(i) instanceof PhiNode phi ) {
-                // Do an eager useless-phi removal
-                Node in = phi.peephole();
-                CodeGen.CODE.addAll(phi._outputs);
-                phi.moveDepsToWorklist();
-                if( in != phi ) {
-                    if( !phi.iskeep() ) // Keeping phi around for parser elsewhere
-                        phi.subsume(in);
-                    setDef(i,in); // Set the update back into Scope
-                }
-            }
-        }
-    }
-
-    @Override public boolean eq( Node n ) {
-        return this==n || !_inProgress;
-    }
 }
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/MemOpNode.java b/chapter23/src/main/java/com/seaofnodes/simple/node/MemOpNode.java
index ead82322..65021110 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/node/MemOpNode.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/MemOpNode.java
@@ -6,6 +6,7 @@ import com.seaofnodes.simple.type.TypeMemPtr;
 import com.seaofnodes.simple.util.Utils;
 import java.lang.StringBuilder;
 import java.util.BitSet;
+import java.util.ArrayList;
 
 /**
  * Convenience common base for Load and Store.
@@ -78,6 +79,8 @@ public abstract class MemOpNode extends Node {
     static String mlabel(String name) { return "[]".equals(name) ? "ary" : ("#".equals(name) ? "len" : name); }
     String mlabel() { return mlabel(_name); }
 
+    @Override public boolean isMem() { return !_isLoad; }
+
     public Node mem() { return in(1); }
     public Node ptr() { return in(2); }
     public Node off() { return in(3); }
@@ -86,6 +89,13 @@ public abstract class MemOpNode extends Node {
     public StringBuilder _printMach( StringBuilder sb, BitSet visited ) { throw Utils.TODO(); }
     public int log_size() { return _declaredType.log_size();  }
 
+    // Extra conditions for factoring matching memory operations through a Phi.
+    // The caller has already checked the opcode, input shape and control.
+    boolean canDrop(MemOpNode other, Node dep) {
+        dep.addDep(ptr());
+        return _alias==other._alias && _declaredType==other._declaredType && err()==null;
+    }
+
     @Override
     public boolean eq(Node n) {
         MemOpNode mem = (MemOpNode)n; // Invariant
@@ -106,4 +116,29 @@ public abstract class MemOpNode extends Node {
             return null;
         return Parser.error( "Might be null accessing '" + _name + "'",_loc);
     }
+    // Memory writers which must follow this read. A packaging node is not a
+    // write, but can hide New's input, so follow this alias through aggregates.
+    public ArrayList<Node> antiDeps() {
+        ArrayList<Node> deps = new ArrayList<>();
+        antiDeps(mem(),deps,new BitSet());
+        return deps;
+    }
+
+    private void antiDeps(Node mem, ArrayList<Node> deps, BitSet visit) {
+        if( visit.get(mem._nid) ) return;
+        visit.set(mem._nid);
+        for( Node use : mem._outputs ) {
+            if( use instanceof MemMergeNode merge ) {
+                Node slice = _alias<merge.nIns() ? merge.in(_alias) : null;
+                if( (slice==mem) || (slice==null && merge.in(1)==mem) )
+                    antiDeps(merge,deps,visit);
+            } else if( use instanceof CallNode ||
+                       (use instanceof MemOpNode st && !st._isLoad && st._alias==_alias) ||
+                       (use instanceof NewNode nnn && nnn.field(_alias)!=null) ||
+                       (use instanceof MemPhiNode phi && phi._alias==_alias) ||
+                       (use instanceof BulkMemPhiNode phi && !phi.isSplit(_alias)) ) {
+                if( !deps.contains(use) ) deps.add(use);
+            }
+        }
+    }
 }
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/MemPhiNode.java b/chapter23/src/main/java/com/seaofnodes/simple/node/MemPhiNode.java
new file mode 100644
index 00000000..7e54f50f
--- /dev/null
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/MemPhiNode.java
@@ -0,0 +1,44 @@
+package com.seaofnodes.simple.node;
+
+import com.seaofnodes.simple.codegen.CodeGen;
+import com.seaofnodes.simple.type.Type;
+import com.seaofnodes.simple.type.TypeMem;
+
+/** A Phi for one field alias, parallel to a bulk Phi at the same Region. */
+public class MemPhiNode extends PhiNode {
+    public final int _alias;
+
+    public MemPhiNode(String label, int alias, Node... inputs) {
+        super(label,TypeMem.make(alias,Type.BOTTOM),inputs);
+        assert alias > 1;
+        _alias = alias;
+    }
+
+    public MemPhiNode(MemPhiNode phi) { super(phi); _alias = phi._alias; }
+
+    @Override public String label() { return "MemPhi_"+_alias; }
+
+    @Override public Type compute() {
+        assert BulkMemPhiNode.checkMem(this,_alias);
+        if( !(region() instanceof RegionNode r) )
+            return region()._type==Type.XCONTROL ? TypeMem.TOP : _type;
+        if( r.inProgress() ) return _minType;
+        Type t = Type.TOP;
+        for( int i=1; i<nIns(); i++ )
+            if( addDep(r.in(i))._type!=Type.XCONTROL )
+                t = t.meet(MemMergeNode.contents(in(i),_alias,this));
+        return TypeMem.make(_alias,t);
+    }
+
+    @Override public Node idealize() {
+        for( int i=1; i<nIns(); i++ )
+            if( in(i) instanceof MemMergeNode mem ) {
+                setDef(i,CodeGen.CODE.add(mem.alias(_alias)));
+                return this;
+            }
+        return super.idealize();
+    }
+
+    @Override public boolean eq(Node n) { return _alias==((MemPhiNode)n)._alias && super.eq(n); }
+    @Override int hash() { return _alias; }
+}
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/NewNode.java b/chapter23/src/main/java/com/seaofnodes/simple/node/NewNode.java
index b0b4103b..49a82a87 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/node/NewNode.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/NewNode.java
@@ -6,33 +6,35 @@ import com.seaofnodes.simple.util.SB;
 
 import java.util.BitSet;
 
-/**
- *  Allocation!  Allocate a chunk of memory, and pre-zero it.
- *  The inputs include control and size, and ALL aliases being set.
- *  The output is large tuple, one for every alias plus the created pointer.
- *  New is expected to be followed by projections for every alias.
+/** Allocate a zeroed object. Inputs {ctrl, $mem, size};
+ *  results {ptr, $mem}.
+ *  The memory input and result cover only the aliases in the allocated struct.
  */
 public class NewNode extends Node implements MultiNode {
 
     public final TypeMemPtr _ptr;
-    public final int _len;
 
     public NewNode(TypeMemPtr ptr, Node... nodes) {
         super(nodes);
         assert !ptr.nullable();
         _ptr = ptr;
-        _len = ptr._obj._fields.length;
-        // Control in slot 0
-        assert nodes[0]._type==Type.CONTROL || nodes[0]._type == Type.XCONTROL;
-        // Malloc-length in slot 1
-        assert nodes[1]._type instanceof TypeInteger || nodes[1]._type==Type.NIL;
-        for( int i=0; i<_len; i++ )
-          assert ptr._obj._fields[i]._one || nodes[2 + i]._type.isa( TypeMem.BOT );
+        assert nodes.length==3;
+        assert nodes[0]._type==Type.CONTROL || nodes[0]._type==Type.XCONTROL;
+        assert nodes[1]._type instanceof TypeMem;
+        assert nodes[2]._type instanceof TypeInteger || nodes[2]._type==Type.NIL;
     }
 
-    public NewNode(NewNode nnn) { super(nnn); _ptr = nnn._ptr; _len = nnn._len; }
+    public Node mem() { return in(1); }
+    public Node size() { return in(2); }
+
+    public Field field(int alias) {
+        for( Field f : _ptr._obj._fields )
+            if( !f._one && f._alias==alias ) return f;
+        return null;
+    }
+
+    public NewNode(NewNode nnn) { super(nnn); _ptr = nnn._ptr; }
 
-    public Node mem (int idx) { return in(idx+2); }
 
     @Override public String label() { return "new_"+(_ptr._obj.isAry() ? "ary_"+_ptr._obj._fields[1]._t.str() : _ptr._obj.str()); }
     @Override
@@ -41,38 +43,10 @@ public class NewNode extends Node implements MultiNode {
         return sb.append(_ptr._obj.str());
     }
 
-    int findAlias(int alias) {
-        int idx = _ptr._obj.findAlias(alias);
-        assert idx!= -1;        // Error, caller should be calling
-        return idx+2;           // Skip ctrl, size
-    }
-
-    // 0 - ctrl
-    // 1 - byte size
-    // 2-len+2 - aliases, one per field
-    // len+2 - 2*len+2 - initial values, one per field
-    public Node size() { return in(1); }
 
     @Override
     public TypeTuple compute() {
-        Field[] fs = _ptr._obj._fields;
-        Type[] ts = new Type[fs.length+2];
-        ts[0] = Type.CONTROL;
-        ts[1] = _ptr;
-        for( int i=0; i<fs.length; i++ ) {
-            if( _ptr._obj._fields[i]._one ) {
-                // Once-only fields use the declared type
-                ts[i+2] = _ptr._obj._fields[i]._t;
-            } else {
-                // Others take from the inputs
-                Type mt = in(i+2)._type;
-                TypeMem mem = mt==Type.TOP ? TypeMem.TOP : (TypeMem)mt;
-                Type tfld = mem._t.meet(mem._t.makeZero());
-                Type tfld2 = tfld.join(fs[i]._t );
-                ts[i+2] = TypeMem.make(fs[i]._alias,tfld2);
-            }
-        }
-        return TypeTuple.make(ts);
+        return TypeTuple.make(_ptr,TypeMem.BOT);
     }
 
     @Override
@@ -117,9 +91,9 @@ public class NewNode extends Node implements MultiNode {
         _kills = kills;
     }
     public String op() { return "alloc"; }
-    public RegMask regmap(int i) { return i==1 ? _arg3Mask : null; }
+    public RegMask regmap(int i) { return i==2 ? _arg3Mask : null; }
     public RegMask outregmap() { return null; }
-    public RegMask outregmap(int idx) { return idx==1  ? _retMask : null; }
+    public RegMask outregmap(int idx) { return idx==0  ? _retMask : null; }
     public RegMask killmap() { return _kills; }
     public void asm(CodeGen code, SB sb) {
         sb.p("#calloc, ").p(code.reg(size()));
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/Node.java b/chapter23/src/main/java/com/seaofnodes/simple/node/Node.java
index ab221789..c27270b6 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/node/Node.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/Node.java
@@ -232,6 +232,7 @@ public abstract class Node implements Cloneable {
     // Error is 'use' does not exist; ok for 'use' to be null.
     protected boolean delUse( Node use ) {
         _outputs.del(_outputs.find(use));
+        moveDepsToWorklist(); // User-count and anti-dependence queries can now change.
         return _outputs.isEmpty();
     }
 
@@ -305,6 +306,7 @@ public abstract class Node implements Cloneable {
             n.unlock();
             int idx = n._inputs.find(this);
             n._inputs.set(idx,nnn);
+            n.moveDepsToWorklist(); // Rewiring can change a dependent query without changing type.
             nnn.addUse(n);
             CODE.add(n);
             CODE.addAll(n._outputs);
@@ -552,7 +554,12 @@ public abstract class Node implements Cloneable {
      * or output of this node, that is, it is at least one step away.  The node
      * being added must benefit from this node being peepholed.
      */
-    <N extends Node> N addDep( N dep ) {
+    <N extends Node> N addDep( N dep ) { return addDep(dep,false); }
+
+    // Keep an explicit dependency when inspecting an immediate user's shape.
+    <N extends Node> N addDepForwards(N dep) { return addDep(dep,true); }
+
+    private <N extends Node> N addDep(N dep, boolean forwards) {
         // Running peepholes during the big assert cannot have side effects
         // like adding dependencies.
         if( CODE._midAssert ) return dep;
@@ -560,8 +567,8 @@ public abstract class Node implements Cloneable {
         if( obs != null ) obs.dep(this, dep);
         if( dep._deps==null ) dep._deps = new Ary<>(Node.class);
         if( dep._deps   .find(this) != -1 ) return dep; // Already on list
-        if( dep._inputs .find(this) != -1 ) return dep; // No need for deps on immediate neighbors
-        if( dep._outputs.find(this) != -1 ) return dep;
+        if( !forwards && dep._inputs .find(this) != -1 ) return dep; // No need for deps on immediate neighbors
+        if( !forwards && dep._outputs.find(this) != -1 ) return dep;
         dep._deps.add(this);
         return dep;
     }
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/PhiNode.java b/chapter23/src/main/java/com/seaofnodes/simple/node/PhiNode.java
index a711fe81..8ae94260 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/node/PhiNode.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/PhiNode.java
@@ -33,6 +33,10 @@ public class PhiNode extends Node {
             addDef(sample);
     }
 
+    public static PhiNode make(String label, Type type, Node... inputs) {
+        return type instanceof TypeMem ? new BulkMemPhiNode(label,inputs) : new PhiNode(label,type,inputs);
+    }
+
     @Override public String label() { return "Phi_"+MemOpNode.mlabel(_label); }
 
     @Override
@@ -94,6 +98,8 @@ public class PhiNode extends Node {
         if( live != null ) {
             if( live._type.isa(_type) )
                 return live;
+            // Let the bulk input split this alias before collapsing a precise Phi.
+            if( live instanceof BulkMemPhiNode ) return null;
             // Keep the Phi upcast
             return new CastNode(_type,null,live);
         }
@@ -103,19 +109,9 @@ public class PhiNode extends Node {
             if( r.in(i)._type == Type.XCONTROL )
                 return null;
 
-        // Simple Phi-after-MemMerge to a known alias can bypass.  Happens when inlining.
-        if( _type instanceof TypeMem tmem && tmem._alias!=1 ) {
-            for( int i=1; i<nIns(); i++ )
-                if( in(i) instanceof MemMergeNode mem ) {
-                    setDef(i,mem.alias(tmem._alias));
-                    return this;
-                }
-        }
-
-        // Generic "pull down op"
-        Node progress;
-        if( same_op() && (progress = drop_same_op()) != null )
-            return progress;
+        // Phi(op(a,b,...),op(c,d,...)) -> op(Phi(a,c),Phi(b,d),...).
+        Node down;
+        if( same_op() && (down=drop_same_op())!=null ) return down;
 
         // If merging Phi(N, cast(N)) - we are losing the cast JOIN effects, so just remove.
         if( nIns()==3 ) {
@@ -149,50 +145,77 @@ public class PhiNode extends Node {
     // Same op on all Phi paths; all ops have only the Phi as a use.
     // None have a control input.
     private boolean same_op() {
+        Node op = in(1);
+        if( op instanceof CFGNode || op instanceof ConstantNode || op instanceof PhiNode ||
+            op instanceof ProjNode || op instanceof NewNode || op instanceof ScopeNode || op instanceof MemMergeNode ) return false;
+        // A bulk Phi must split its aliases before factoring precise Stores.
+        if( this instanceof BulkMemPhiNode ) return false;
+        // Bulk splitting currently identifies parallel slices by Region/alias.
+        // Factoring can introduce another slice at a different memory point;
+        // wait until that Region's bulk partitioning has finished.
+        if( op instanceof MemOpNode )
+            for( Node use : region()._outputs )
+                if( use instanceof BulkMemPhiNode ) {
+                    addDepForwards(use);
+                    return false;
+                }
+        Node busy = null;
         for( int i=1; i<nIns(); i++ ) {
-            Node op = in(i);
-            if( in(1).getClass() != op.getClass() || op.in(0)!=null || in(1).nIns() != op.nIns() )
-                return false;      // Wrong class or CFG bound or mismatched inputs
-            if( in(1) instanceof MemOpNode mem && mem._alias != ((MemOpNode)op)._alias )
-                return false;
-            if( op.nOuts() > 1 ) { // Too many users, but addDep in case lose users
-                for( Node out : op._outputs )
-                    if( out!=null && out!=this )
-                        addDep(out);
-                return false;
+            Node n = in(i);
+            if( addDep(region().in(i))._type==Type.XCONTROL ||
+                op.getClass()!=n.getClass() || n.nIns()!=op.nIns() || n.in(0)!=null || !op.eq(n) ) return false;
+            if( n instanceof MemOpNode mem && !mem.canDrop((MemOpNode)op,this) ) return false;
+            // Moving a Store must remove the old effect, not duplicate it.
+            // Allow one shared arm: the other arms disappear into the single
+            // factored operation, usually reducing the total operation count.
+            addDepForwards(n);
+            if( n.nOuts()>1 ) {
+                for( Node use : n._outputs )
+                    if( use!=null && use!=this ) addDepForwards(use);
+                if( n instanceof StoreNode || busy!=null ) return false;
+                busy=n;
             }
-            for( int j=1; j<in(1).nIns(); j++ )
-                if( op.in(j) instanceof ScopeNode || (op.in(j)==null ^ in(1).in(j)==null) )
-                    return false; // Lazy Phi input
+            for( int j=1; j<n.nIns(); j++ )
+                if( n.in(j) instanceof ScopeNode || ((n.in(j)==null) != (op.in(j)==null)) ) return false;
         }
         return true;
     }
 
     private Node drop_same_op() {
-        assert !(in(1) instanceof CFGNode);
-        Node op = in(1);
-        Node cp = op.copy();
-        cp._type = null;    // Fresh type
-        cp.addDef(null);    // No control
-
+        Node op = in(1), cp = op.copyEmpty();
+        cp._type = null;
+        cp.addDef(null);
         for( int j=1; j<op.nIns(); j++ ) {
-            boolean needsPhi = false;
-            Node x = op.in(j); // Jth input from sample #1
+            Node x = op.in(j);
+            boolean different = false;
             for( int i=2; i<nIns(); i++ )
-                if( in(i).in(j) != x )
-                    { needsPhi=true; break; }
-            if( needsPhi ) {
-                x = new PhiNode(_label,op.in(j)._type.glb(false));
-                x.addDef(region());
-                for( int i=1; i<nIns(); i++ )
-                    x.addDef(in(i).in(j));
-                x = x.peephole();
+                if( in(i).in(j)!=x ) different=true;
+            if( different ) {
+                Node[] ins = new Node[nIns()];
+                ins[0] = region();
+                Type t = Type.TOP;
+                for( int i=1; i<nIns(); i++ ) {
+                    ins[i] = in(i).in(j);
+                    t = t.meet(ins[i]._type);
+                }
+                // Not accepts both pointers and integers, but a Phi cannot mix them.
+                if( t==Type.BOTTOM ) {
+                    for( int i=1; i<ins.length; i++ ) addDep(ins[i]);
+                    cp.kill();
+                    return null;
+                }
+                PhiNode phi = j==1 && op instanceof MemOpNode mem
+                    ? new MemPhiNode(_label,mem._alias,ins)
+                    : PhiNode.make(_label,t.glb(false),ins);
+                x = phi.peephole();
             }
             cp.addDef(x);
         }
-        // Test not running backwards, which can happen for e.g. And's
-        if( cp.compute().isa(compute()) )
-            return cp;
+        // Factoring must not widen the result (e.g. correlated And operands).
+        if( cp.compute().isa(compute()) ) return cp;
+        for( int i=1; i<nIns(); i++ )
+            for( int j=1; j<in(i).nIns(); j++ )
+                if( in(i).in(j)!=null ) addDep(in(i).in(j));
         cp.kill();
         return null;
     }
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/RegionNode.java b/chapter23/src/main/java/com/seaofnodes/simple/node/RegionNode.java
index 42b23633..14ba493e 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/node/RegionNode.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/RegionNode.java
@@ -48,7 +48,7 @@ public class RegionNode extends CFGNode {
         if( !hasPhi() && nIns()>3 &&  // No Phi users, just a control user
             in(1) instanceof CProjNode p1 &&
             in(2) instanceof CProjNode p2 &&
-            addDep(p1.in(0))==addDep(p2.in(0)) &&
+            addDepForwards(p1).in(0)==addDepForwards(p2).in(0) &&
             p1.in(0) instanceof IfNode iff ) {
             // Replace with the iff.ctrl directly
             if( nIns()==3 ) return iff.ctrl();
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/ReturnNode.java b/chapter23/src/main/java/com/seaofnodes/simple/node/ReturnNode.java
index 95303994..212bb875 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/node/ReturnNode.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/ReturnNode.java
@@ -9,15 +9,7 @@ import com.seaofnodes.simple.util.SB;
 import com.seaofnodes.simple.util.Utils;
 import java.util.BitSet;
 
-/**
- * The Return node has two inputs.  The first input is a control node and the
- * second is the data node that supplies the return value.
- * <p>
- * In this presentation, Return functions as a Stop node, since multiple <code>return</code> statements are not possible.
- * The Stop node will be introduced in Chapter 6 when we implement <code>if</code> statements.
- * <p>
- * The Return's output is the value from the data node.
- */
+/** Keeps all preceding memory effects and the return value alive: {ctrl, $mem, value, rpc}. */
 public class ReturnNode extends CFGNode {
 
     public FunNode _fun;
@@ -51,7 +43,8 @@ public class ReturnNode extends CFGNode {
     @Override
     public Type compute() {
         if( inProgress () ) return TypeTuple.RET; // In progress
-        return TypeTuple.make(ctrl()._type,mem()._type,expr()._type);
+        // Inlining can delete the entry while callers still use this return.
+        return TypeTuple.make(ctrl()._type,TypeMem.BOT,expr()._type);
     }
 
     @Override public Node idealize() {
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/ScopeNode.java b/chapter23/src/main/java/com/seaofnodes/simple/node/ScopeNode.java
index 04666956..cddb9f24 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/node/ScopeNode.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/ScopeNode.java
@@ -12,9 +12,9 @@ import java.util.*;
 
 /**
  * The Scope node is purely a parser helper - it tracks names to nodes with a
- * stack of variables.
+ * stack of variable records.
  */
-public class ScopeNode extends MemMergeNode {
+public class ScopeNode extends Node {
 
     /**
      * The control is a name that binds to the currently active control
@@ -24,8 +24,12 @@ public class ScopeNode extends MemMergeNode {
     public static final String ARG0 = "arg";
     public static final String MEM0 = "$mem";
 
-    // All active/live variables in all nested scopes, all run together.
-    // Maps 1-to-1 to the _inputs array.
+    public Node in(Var v) { return in(v._idx); }
+
+    @Override public Type compute() { return Type.BOTTOM; }
+    @Override public Node idealize() { return null; }
+
+    // All active/live variables in all nested scopes, all run together
     public final Ary<Var> _vars;
 
     // Lexical scope is typed:
@@ -50,10 +54,10 @@ public class ScopeNode extends MemMergeNode {
 
     // A new ScopeNode
     public ScopeNode() {
-        super(true);
-        _vars   = new Ary<>(Var .class);
-        _kinds  = new Ary<>(Kind.class);
-        _guards = new Ary<>(Node.class);
+        _type = Type.BOTTOM;
+        _vars   = new Ary<>(Var    .class);
+        _kinds  = new Ary<>(Kind   .class);
+        _guards = new Ary<>(Node   .class);
     }
 
     @Override public String label() { return "Scope"; }
@@ -85,7 +89,7 @@ public class ScopeNode extends MemMergeNode {
 
 
     public Node ctrl() { return in(0); }
-    public MemMergeNode mem() { return (MemMergeNode)in(1); }
+    public Node mem() { return in(lookup(MEM0)); }
     public Var var(int i) { return _vars.at(i); }
 
     /**
@@ -99,7 +103,7 @@ public class ScopeNode extends MemMergeNode {
      * @return Node that was bound
      */
     public <N extends Node> N ctrl(N n) { return setDef(0,n); }
-    public Node mem(Node n) { return setDef(1,n); }
+    public Node mem(Node n) { update(MEM0,n); return n; }
 
     public void push( Kind kind ) {
         kind._lexSize = _vars.size();
@@ -206,12 +210,6 @@ public class ScopeNode extends MemMergeNode {
         return true;
     }
 
-    // Read from memory
-    public Node mem( int alias ) { return mem()._mem(alias,null); }
-    // Write to memory
-    public void mem( int alias, Node st ) { mem()._mem(alias,st); }
-
-
     /**
      * Lookup a name in all scopes starting from most deeply nested.
      *
@@ -247,7 +245,7 @@ public class ScopeNode extends MemMergeNode {
                 // Set real Phi in the loop head
                 // The phi takes its one input (no backedge yet) from a recursive
                 // lookup, which might have insert a Phi in every loop nest.
-                : loop.setDef(v._idx,new PhiNode(v._name, v.type(), loop.ctrl(), loop.in(loop.update(v,null)._idx),null).peephole());
+                : loop.setDef(v._idx,PhiNode.make(v._name, v.type(), loop.ctrl(), loop.in(loop.update(v,null)._idx),null).peephole());
             setDef(v._idx,old);
         }
         //assert !v._final || st==null;
@@ -285,21 +283,9 @@ public class ScopeNode extends MemMergeNode {
                 n.keep();
         dup.addDef(ctrl());     // Control input is just copied
 
-        // Memory input is a shallow copy
-        MemMergeNode memdup = new MemMergeNode(true), mem = mem();
-        memdup.addDef(null);
-        memdup.addDef(loop ? this : mem.in(1));
-        for( int i=2; i<mem.nIns(); i++ )
-            // For lazy phis on loops we use a sentinel
-            // that will trigger phi creation on update
-            memdup.addDef(loop ? this : mem.in(i));
-        dup.addDef(memdup);
-
-        // Copy of other inputs
-        for( int i=2; i<nIns(); i++ )
-            // For lazy phis on loops we use a sentinel
-            // that will trigger phi creation on update
-            dup.addDef(loop && !var(i)._final ? this : in(i));
+        // Memory is one binding, with the same lazy loop Phi as scalar names.
+        for( int i=1; i<nIns(); i++ )
+            dup.addDef(loop && !_vars.at(i)._final ? this : in(i));
         return dup;
     }
 
@@ -313,22 +299,21 @@ public class ScopeNode extends MemMergeNode {
      */
     public RegionNode mergeScopes(ScopeNode that, Parser.Lexer loc) {
         RegionNode r = ctrl(new RegionNode(loc,null,ctrl(), that.ctrl()).keep());
-        mem()._merge(that.mem(),r);
-        this ._merge(that      ,r);
+        _merge(that,r);
         that.kill();            // Kill merged scope
         CodeGen.CODE.add(r);
         return r.unkeep();
     }
 
     private void _merge(ScopeNode that, RegionNode r) {
-        for( int i = 2; i < nIns(); i++)
+        for( int i = 1; i < nIns(); i++)
             if( in(i) != that.in(i) ) { // No need for redundant Phis
                 // If we are in lazy phi mode we need to a lookup
                 // by name as it will trigger a phi creation
                 Var v = var(i);
                 Node lhs = this.in(this.update(v,null));
                 Node rhs = that.in(that.update(v,null));
-                setDef(i, new PhiNode(v._name, v.type(), r, lhs, rhs).peephole());
+                setDef(i, PhiNode.make(v._name, v.type(), r, lhs, rhs).peephole());
             }
     }
 
@@ -353,23 +338,17 @@ public class ScopeNode extends MemMergeNode {
         assert ctrl instanceof LoopNode loop && loop.inProgress();
         ctrl.setDef(2,back.ctrl());
 
-        mem()._endLoopMem( this, back.mem(), exit.mem() );
-        this ._endLoop   ( this, back      , exit       );
+        _endLoop(this,back,exit);
         back.kill();            // Loop backedge is dead
         // Now one-time do a useless-phi removal
-        mem()._useless();
-        this ._useless();
-
-        // The exit mem's lazy default value had been the loop top,
-        // now it goes back to predating the loop.
-        exit.mem().setDef(1,mem().in(1));
+        _useless();
     }
 
     // Fill in the backedge of any inserted Phis
     void _endLoop( ScopeNode scope, Node back, Node exit ) {
-        for( int i=2; i<nIns(); i++ ) {
-            if( var(i)._final ) continue; // Final vars did not get modified in the loop
-            if( var(i).type().isHighOrConst() ) continue; // Cannot lift higher than a constant, so no Phi
+        for( int i=1; i<nIns(); i++ ) {
+            if( _vars.at(i)._final ) continue; // Final vars did not get modified in the loop
+            if( _vars.at(i).type().isHighOrConst() ) continue; // Cannot lift higher than a constant, so no Phi
             if( back.in(i) != scope ) {
                 PhiNode phi = (PhiNode)in(i);
                 assert phi.region()==scope.ctrl() && phi.in(2)==null;
@@ -381,6 +360,23 @@ public class ScopeNode extends MemMergeNode {
     }
 
 
+    // Now one-time do a useless-phi removal
+    void _useless( ) {
+        for( int i=1; i<nIns(); i++ ) {
+            if( in(i) instanceof PhiNode phi ) {
+                // Do an eager useless-phi removal
+                Node in = phi.peephole();
+                CodeGen.CODE.addAll(phi._outputs);
+                phi.moveDepsToWorklist();
+                if( in != phi ) {
+                    if( !phi.iskeep() ) // Keeping phi around for parser elsewhere
+                        phi.subsume(in);
+                    setDef(i,in); // Set the update back into Scope
+                }
+            }
+        }
+    }
+
     // Up-casting: using the results of an If to improve a value.
     // E.g. "if( ptr ) ptr.field;" is legal because ptr is known not-null.
     public void addGuards( Node ctrl, Node pred, boolean invert ) {
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/StartNode.java b/chapter23/src/main/java/com/seaofnodes/simple/node/StartNode.java
index 6ed2acd3..186adfb7 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/node/StartNode.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/StartNode.java
@@ -3,18 +3,11 @@ package com.seaofnodes.simple.node;
 import com.seaofnodes.simple.codegen.CodeGen;
 import com.seaofnodes.simple.Parser;
 import com.seaofnodes.simple.type.*;
-import java.util.Arrays;
 import java.util.BitSet;
 import java.util.HashSet;
 import static com.seaofnodes.simple.util.Utils.TODO;
 
-/**
- * The Start node represents the start of the function.
- * <p>
- * Start initially has 1 input (arg) from outside and the initial control.
- * In ch10 we also add mem aliases as structs get defined; each field in struct
- * adds a distinct alias to Start's tuple.
- */
+/** Start supplies control, whole memory, and the argument, in that order. */
 public class StartNode extends LoopNode implements MultiNode {
 
     final Type _arg;
@@ -44,7 +37,7 @@ public class StartNode extends LoopNode implements MultiNode {
 
 
     @Override public TypeTuple compute() {
-        return TypeTuple.make(Type.CONTROL,TypeMem.TOP,_arg);
+        return TypeTuple.make(Type.CONTROL,TypeMem.BOT,_arg);
     }
 
     @Override public Node idealize() { return null; }
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/StoreNode.java b/chapter23/src/main/java/com/seaofnodes/simple/node/StoreNode.java
index 74c27dc8..ac437788 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/node/StoreNode.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/StoreNode.java
@@ -1,5 +1,6 @@
 package com.seaofnodes.simple.node;
 
+import com.seaofnodes.simple.codegen.*;
 import com.seaofnodes.simple.*;
 import com.seaofnodes.simple.type.*;
 import java.util.BitSet;
@@ -25,6 +26,10 @@ public class StoreNode extends MemOpNode {
     }
 
     // Debugger label
+    @Override boolean canDrop(MemOpNode other, Node dep) {
+        return super.canDrop(other,dep) && _init==((StoreNode)other)._init;
+    }
+
     @Override public String  label() { return "st_"+mlabel(); }
     @Override public boolean isMem() { return true; }
 
@@ -38,24 +43,18 @@ public class StoreNode extends MemOpNode {
     @Override
     public Type compute() {
         Type val = val()._type;
-        Type mem0 = mem()._type;
-        if( mem0 == Type.TOP ) return TypeMem.TOP;
-        TypeMem mem = (TypeMem)mem0; // Invariant
-        if( mem == TypeMem.TOP ) return TypeMem.TOP;
-        Type t = Type.BOTTOM;               // No idea on field contents
-        // Same alias, lift val to the declared type and then meet into other fields
-        if( mem._alias == _alias ) {
-            assert !_declaredType.isFRef();
-            // Sharpen memory value; required for narrowing stores where the parser inserts
-            // zero/sign masking and somebody reads the TypeMem type.
-            val = val.join(_declaredType);
-            t = val.meet(mem._t);
-        }
+        assert !_declaredType.isFRef();
+        Type t = val.join(_declaredType).meet(MemMergeNode.contents(mem(),_alias,this));
         return TypeMem.make(_alias,t);
     }
 
     @Override
     public Node idealize() {
+        if( mem() instanceof MemMergeNode merge ) {
+            setDef(1,CodeGen.CODE.add(merge.alias(_alias)));
+            return this;
+        }
+
 
         if( mem() instanceof CastNode cast ) {
             setDef(1,cast.in(1));
@@ -65,8 +64,8 @@ public class StoreNode extends MemOpNode {
         // Simple store-after-store on same address.  Should pick up the
         // required init-store being stomped by a first user store.
         if( mem() instanceof StoreNode st &&
-            ptr()==st.ptr() &&  // Must check same object
-            off()==st.off() &&  // And same offset (could be "same alias" but this handles arrays to same index)
+            ptr()==st.ptr() && _alias==st._alias &&  // Must check same object
+            off()==st.off() &&  // And same offset
             ptr()._type instanceof TypeMemPtr && // No bother if weird dead pointers
             // Must have exactly one use of "this" or you get weird
             // non-serializable memory effects in the worse case.
@@ -76,12 +75,6 @@ public class StoreNode extends MemOpNode {
             return this;
         }
 
-        // Simple store-after-MemMerge to a known alias can bypass.  Happens when inlining.
-        if( mem() instanceof MemMergeNode mem ) {
-            setDef(1,mem.alias(_alias));
-            return this;
-        }
-
         // Value is automatically truncated by narrow store
         if( val() instanceof AndNode and && and.in(2)._type.isConstant()  ) {
             int log = _declaredType.log_size();
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/cpus/arm/arm.java b/chapter23/src/main/java/com/seaofnodes/simple/node/cpus/arm/arm.java
index bca3dfa4..c39c5f74 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/node/cpus/arm/arm.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/cpus/arm/arm.java
@@ -673,6 +673,8 @@ public class arm extends Machine {
         case NotNode not     -> new NotARM(not);
         case OrNode or       -> or(or);
         case ParmNode parm   -> new ParmARM(parm);
+        case BulkMemPhiNode phi -> new BulkMemPhiNode(phi);
+        case MemPhiNode phi -> new MemPhiNode(phi);
         case PhiNode phi     -> new PhiNode(phi);
         case ProjNode prj    -> prj(prj);
         case ReadOnlyNode read -> new ReadOnlyMach(read);
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/cpus/riscv/riscv.java b/chapter23/src/main/java/com/seaofnodes/simple/node/cpus/riscv/riscv.java
index 48cb5d23..88ec9ebd 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/node/cpus/riscv/riscv.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/cpus/riscv/riscv.java
@@ -345,6 +345,8 @@ public class riscv extends Machine {
         case NotNode      not -> new NotRISC(not);
         case OrNode        or -> or(or);
         case ParmNode    parm -> new ParmRISC(parm);
+        case BulkMemPhiNode phi -> new BulkMemPhiNode(phi);
+        case MemPhiNode phi -> new MemPhiNode(phi);
         case PhiNode      phi -> new PhiNode(phi);
         case ProjNode     prj -> prj(prj);
         case ReadOnlyNode read-> new ReadOnlyMach(read);
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/LoadX86.java b/chapter23/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/LoadX86.java
index 441cf693..d0fb0081 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/LoadX86.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/LoadX86.java
@@ -24,6 +24,11 @@ public class LoadX86 extends MemOpX86 {
     }
 
     static void enc( Encoding enc, Type decl, short dst, short ptr, short idx, int off, int scale ) {
+        // Allocation may put a full-width integer in an XMM register (or a
+        // floating value in a GPR). Move its bits using the selected register bank.
+        if( dst>=x86_64_v2.XMM_OFFSET && decl.log_size()==3 ) decl=TypeFloat.F64;
+        else if( dst<x86_64_v2.XMM_OFFSET && decl instanceof TypeFloat )
+            decl = decl==TypeFloat.F32 ? TypeInteger.U32 : TypeInteger.BOT;
         if( decl == TypeFloat.F32) enc.add1(0xF3);
         if( decl == TypeFloat.F64) enc.add1(0xF2);
 
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java b/chapter23/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java
index 7ac7fc19..cf9288b8 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java
@@ -363,6 +363,8 @@ public class x86_64_v2 extends Machine {
         case NotNode      not -> new NotX86(not);
         case OrNode        or -> or(or);
         case ParmNode    parm -> new ParmX86(parm);
+        case BulkMemPhiNode phi -> new BulkMemPhiNode(phi);
+        case MemPhiNode phi -> new MemPhiNode(phi);
         case PhiNode      phi -> new PhiNode(phi);
         case ProjNode     prj -> prj(prj);
         case ReadOnlyNode read-> new ReadOnlyMach(read);
@@ -396,9 +398,8 @@ public class x86_64_v2 extends Machine {
         if( lhs instanceof LoadNode ld && ld.nOuts() == 1 && ld._declaredType.log_size() >= 3)
             return new AddMemX86(add, address(ld), ld.ptr(), idx, off, scale, 0, rhs);
 
-//        if(rhs instanceof LoadNode ld && ld.nOuts() == 1 && ld._declaredType.log_size() >= 3) {
-//            throw Utils.TODO(); // Swap load sides
-//        }
+        if( rhs instanceof LoadNode ld && ld.nOuts()==1 && ld._declaredType.log_size()>=3 )
+            return new AddMemX86(add,address(ld),ld.ptr(),idx,off,scale, 0,lhs);
 
         // Attempt a full LEA-style break down.
         // Returns one of AddX86, AddIX86, LeaX86, or LHS
diff --git a/chapter23/src/main/java/com/seaofnodes/simple/print/SimpleGraphAdapter.java b/chapter23/src/main/java/com/seaofnodes/simple/print/SimpleGraphAdapter.java
index 7f7107ec..6f19fa9d 100644
--- a/chapter23/src/main/java/com/seaofnodes/simple/print/SimpleGraphAdapter.java
+++ b/chapter23/src/main/java/com/seaofnodes/simple/print/SimpleGraphAdapter.java
@@ -67,7 +67,7 @@ public class SimpleGraphAdapter extends GraphAdapter<Node> {
             return n instanceof CFGNode ? Role.CTRL : isMem(n) ? Role.MEM : Role.DATA;
         if( n instanceof RegionNode && i == 0 ) return Role.ASSOC;
         if( i == 0 || n instanceof RegionNode || n instanceof StopNode ) return Role.CTRL;
-        if( n instanceof MemMergeNode || i == 1 && (n instanceof MemOpNode || n instanceof ReturnNode || n instanceof CallNode) ) return Role.MEM;
+        if( n instanceof MemMergeNode || (i == 1 && (n instanceof MemOpNode || n instanceof ReturnNode || n instanceof CallNode || n instanceof NewNode)) ) return Role.MEM;
         Node def = n.in(i);
         if( def != null && isMem(def) ) return Role.MEM;
         return Role.DATA;