ch 20 forward port memory

27 files changed, 688 insertions, 391 deletions

diff --git a/chapter20/src/main/java/com/seaofnodes/simple/Parser.java b/chapter20/src/main/java/com/seaofnodes/simple/Parser.java
index af047850..16018fba 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/Parser.java
+++ b/chapter20/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 whole program, as-if function header "{ int arg -> body }"
         parseFunctionBody(_code._main,loc(),"arg");
@@ -178,7 +178,7 @@ public class Parser {
         // Build a multi-exit return point for all function returns
         RegionNode r = new RegionNode((Lexer)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);
@@ -190,11 +190,8 @@ public class Parser {
         // external calls.
         _scope.push(ScopeNode.Kind.Function);
         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);
@@ -208,7 +205,7 @@ public class Parser {
 
         // Last expression is the return
         if( ctrl()._type==Type.CONTROL )
-            fun.addReturn(ctrl(), _scope.mem().merge(), last);
+            fun.addReturn(ctrl(), _scope.mem(), last);
 
         // Pop off the inProgress node on the multi-exit Region merge
         assert r.inProgress();
@@ -568,9 +565,8 @@ 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;
     }
@@ -1183,29 +1179,31 @@ 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++ )
-            ns[2+i] = memAlias(fs[i]._alias);
-        Node nnn = new NewNode(TypeMemPtr.make(obj), ns).peephole().keep();
-        for( int i = 0; i < len; i++ )
-            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 ) 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 ) 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( val._type != val._type.makeZero() ) {
-                Node mem = memAlias(fs[i]._alias);
-                Node st = new StoreNode(loc(),fs[i]._fname,fs[i]._alias,fs[i]._type,mem,ptr,con(obj.offset(i)),val,true).peephole();
-                memAlias(fs[i]._alias,st);
+                Field f = fs[i];
+                Node st = new StoreNode(loc(),f._fname,f._alias,f._type,mem(),ptr,con(obj.offset(i)),val,true).peephole();
+                mem(new MemMergeNode(mem(),f._alias,st).peephole());
             }
         }
-
         return ptr.unkeep();
     }
 
@@ -1218,13 +1216,18 @@ public class Parser {
         return newStruct(ary,size,0,ALTMP);
     }
 
-    // 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
@@ -1267,6 +1270,7 @@ 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
@@ -1279,16 +1283,11 @@ public class Parser {
             Node val = parseAsgn().keep();
             Node lift = liftExpr( val, tf, f._final );
 
-            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
         }
 
-        Node load = new LoadNode(loc(),name, f._alias, tf, memAlias(f._alias), expr.keep(), off);
+        Node load = 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.
@@ -1299,12 +1298,7 @@ public class Parser {
         char ch = _lexer.matchOperAssign();
         if( ch!=0 ) {
             Node op = opAssign(ch,load,f._final,tf,name);
-            Node st = new StoreNode(loc(), name, f._alias, tf, memAlias(f._alias), expr.unkeep(), off, 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));
+            store(name,f._alias,tf,expr.unkeep(),off,op,false,base.isAry() ? ctrl() : null);
             load = postfix(ch) ? load.unkeep() : op;
             // And use the original loaded value as the result
         } else expr.unkill();
@@ -1385,7 +1379,7 @@ 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(expr);        // Function pointer
         // Unkeep them all
         for( Node arg : args )
@@ -1405,11 +1399,8 @@ public class Parser {
         CallEndNode cend = (CallEndNode)new CallEndNode(call).peephole();
         // 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/chapter20/src/main/java/com/seaofnodes/simple/codegen/GlobalCodeMotion.java b/chapter20/src/main/java/com/seaofnodes/simple/codegen/GlobalCodeMotion.java
index 4ebd86c0..52906afb 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/codegen/GlobalCodeMotion.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/codegen/GlobalCodeMotion.java
@@ -51,7 +51,7 @@ 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.
             }
@@ -172,15 +172,9 @@ public abstract class GlobalCodeMotion {
                         continue outer; // Nope, await all uses done
 
                 // Loads need their memory inputs' uses also done
-                if( n instanceof LoadNode ld )
-                    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 indirectly defines memory
-                             (memuse._type instanceof TypeTuple tt && tt._types[ld._alias] instanceof TypeMem)) )
+                if( n instanceof MemOpNode ld && !ld.isMem() )
+                    for( Node memuse : ld.antiDeps() )
+                        if( late[memuse._nid]==null )
                             continue outer;
 
                 // All uses done, schedule
@@ -192,9 +186,7 @@ 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 LoadNode ld && late[ld._nid]==null )
-                        work.push(ld);
+                wakeLoads(def,late,work,new BitSet());
             }
             if( n instanceof LoopNode loop )
                 for( Node phi : loop._outputs )
@@ -213,7 +205,7 @@ public abstract class GlobalCodeMotion {
               lca = use_block(n,use, late).domLCA(lca,null);
 
         // Loads may need anti-dependencies, raising their LCA
-        if( n instanceof LoadNode load )
+        if( n instanceof MemOpNode load && !load.isMem() )
             lca = find_anti_dep(lca,load,early,late,anti);
 
         // Walk up from the LCA to the early, looking for best place.  This is
@@ -253,15 +245,26 @@ public abstract class GlobalCodeMotion {
             best instanceof IfNode;
     }
 
-    private static CFGNode find_anti_dep(CFGNode lca, LoadNode load, CFGNode early, CFGNode[] late, int[] anti) {
+    // 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.isMem() && 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 StoreNode st:
+            case MemOpNode st:
                 assert late[st._nid]!=null;
                 lca = anti_dep(load,late[st._nid],st.cfg0(),lca,st,anti);
                 break;
@@ -276,11 +279,10 @@ public abstract class GlobalCodeMotion {
                     if( phi.in(i)==load.mem() )
                         lca = anti_dep(load,phi.region().cfg(i),load.mem().cfg0(),lca,null,anti);
                 break;
-            case NewNode st: 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
-            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],st.cfg0(),lca,st,anti);
+                break;
             default: throw Utils.TODO();
             }
         }
@@ -288,15 +290,13 @@ public abstract class GlobalCodeMotion {
     }
 
     //
-    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.
+    private static CFGNode anti_dep( MemOpNode load, CFGNode stblk, CFGNode defblk, CFGNode lca, Node st, int[] anti ) {
         for( ; stblk != defblk.idom(); stblk = stblk.idom() ) {
             // Store and Load overlap, need anti-dependence
             if( 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
+                if( lca == stblk && st != null && !(st instanceof CallNode) && st._inputs.find(load) == -1 ) // And if something moved,
+                    st.addDef(load);   // Add anti-dep as well; Calls already end their block.
                 return lca;            // Cap this stores' anti-dep to here
             }
         }
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/BulkMemPhiNode.java b/chapter20/src/main/java/com/seaofnodes/simple/node/BulkMemPhiNode.java
new file mode 100644
index 00000000..e9d4b897
--- /dev/null
+++ b/chapter20/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.
+        use.addDepForwards(this);
+        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/chapter20/src/main/java/com/seaofnodes/simple/node/LoadNode.java b/chapter20/src/main/java/com/seaofnodes/simple/node/LoadNode.java
index 94c3e337..76125e39 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/LoadNode.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/LoadNode.java
@@ -1,5 +1,6 @@
 package com.seaofnodes.simple.node;
 
+import com.seaofnodes.simple.codegen.CodeGen;
 import com.seaofnodes.simple.Parser;
 import com.seaofnodes.simple.Utils;
 import com.seaofnodes.simple.type.*;
@@ -23,6 +24,25 @@ public class LoadNode extends MemOpNode {
         super(loc, name, alias, 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();
+        mem.addDepForwards(dep);
+        for( Node use : mem._outputs ) {
+            if( use==null ) continue;
+            use.addDepForwards(dep);
+            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
@@ -30,32 +50,32 @@ public class LoadNode extends MemOpNode {
 
     @Override
     public Type compute() {
-        if( mem()._type instanceof TypeMem mem ) {
-            // Update declared forward ref to the actual
-            if( _declaredType.isFRef() && mem._t instanceof TypeMemPtr tmp && !tmp.isFRef() )
-                _declaredType = tmp;
-            // No lifting if ptr might null-check
-            if( err()==null )
-                return _declaredType.join(mem._t);
-        }
-        return _declaredType;
+        Type t = MemMergeNode.contents(mem(),_alias,this);
+        // Update declared forward ref to the actual.
+        if( _declaredType.isFRef() && t instanceof TypeMemPtr tmp && !tmp.isFRef() )
+            _declaredType = tmp;
+        return err()==null ? _declaredType.join(t) : _declaredType;
     }
 
     @Override
     public Node idealize() {
+        if( mem() instanceof MemMergeNode merge ) {
+            setDef(1,CodeGen.CODE.add(merge.alias(_alias)));
+            return this;
+        }
         Node ptr = ptr();
         Node mem = mem();
 
         // Simple Load-after-Store on same address.
-        if( mem instanceof StoreNode st &&
-            ptr == st.ptr() && off() == st.off() ) { // Must check same object
+        if( mem() instanceof StoreNode st &&
+            _alias==st._alias && 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-New on same address.
         if( mem instanceof ProjNode p && p.in(0) instanceof NewNode nnn &&
-            ptr == nnn.proj(1) ) // Must check same object
+            ptr == nnn.proj(0) ) // Must check same object
             return zero(nnn);   // Load zero from new
 
         // Uplift control to a prior dominating load.
@@ -72,12 +92,16 @@ public class LoadNode extends MemOpNode {
             ptr = ro.in(1);
         outer:
         while( true ) {
+            mem.addDep(this);
             switch( mem ) {
+            case MemMergeNode merge:
+                mem = merge.alias(_alias);
+                break;
             case StoreNode st:
-                if( ptr == st.ptr().addDep(this) && off() == st.off() )
+                if( _alias==st._alias && ptr == st.ptr().addDep(this) && 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
+                if( _alias==st._alias && !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
@@ -95,17 +119,16 @@ public class LoadNode extends MemOpNode {
                         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
+                    mem = nnn1.mem();// 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 MemMergeNode merge:  mem = merge.alias(_alias);  break;
 
             default:
-                throw Utils.TODO();
+                break outer;
             }
         }
 
@@ -116,7 +139,7 @@ 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 &&
             // Offset can be hoisted
             off() instanceof ConstantNode &&
             // Pointer can be hoisted
@@ -145,7 +168,9 @@ public class LoadNode extends MemOpNode {
 
     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();
+        LoadNode ld = new LoadNode(_loc,_name,_alias,_declaredType,mem.in(idx),ptr instanceof PhiNode && ptr.in(0)==mem.in(0) ? ptr.in(idx) : ptr,off());
+        ld.setDef(0,in(0));
+        return ld.peephole();
     }
 
     private static boolean neverAlias( Node ptr1, Node ptr2 ) {
@@ -178,9 +203,9 @@ public class LoadNode extends MemOpNode {
     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()==st1.ptr() && off()==st1.off() ) return true;
         px.addDep(this);
+        if( px._type instanceof TypeMem mem && mem._t.isHighOrConst() ) return true;
+        if( px instanceof StoreNode st1 && _alias==st1._alias && ptr()==st1.ptr() && off()==st1.off() ) return true;
         return false;
     }
 
@@ -205,4 +230,5 @@ public class LoadNode extends MemOpNode {
         Node shl = new ShlNode(null,val,shf.keep()).peephole();
         return new SarNode(null,shl,shf.unkeep());
     }
+
 }
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/MemMergeNode.java b/chapter20/src/main/java/com/seaofnodes/simple/node/MemMergeNode.java
index adf83aab..31c76a93 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/MemMergeNode.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/MemMergeNode.java
@@ -5,151 +5,85 @@ 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; }
-
-    @Override public String label() { return "ALLMEM"; }
-    @Override public boolean isMem() { return true; }
-
-    @Override
-    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(" ");
-        }
-        sb.setLength(sb.length()-1);
-        return sb.append("]");
-    }
-
-    // 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);
-        merge._mem(1,null);
-        return merge.peephole();
-    }
-
-
-    @Override public Type compute() { return TypeMem.BOT; }
-
-    @Override public Node idealize() {
-        if( inProgress() ) return null;
-
-        // If not merging any memory (all memory is just the default)
-        if( nIns()==2 )
-            return in(1);       // Become default memory
-
-        return null;
+    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 in( Var v ) { return in(v._idx); }
-
-    public Node alias( int alias ) {
-        return in(alias<nIns() && in(alias)!=null ? alias : 1);
+    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;
     }
 
-    Node alias( int alias, Node st ) {
-        while( alias >= nIns() ) addDef(null);
-        return setDef(alias,st);
+    public void alias(int alias, Node mem) {
+        assert alias > 1;
+        while( nIns()<=alias ) addDef(null);
+        setDef(alias,mem);
     }
 
-    // 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.BOT,loop.ctrl(),loopmem._mem(alias,null),null).peephole() );
-            alias(alias,old);
+    // 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) {
+        mem.addDep(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;
+            nnn.addDep(dep);
+            return contents(nnn.mem(),alias,dep).meet(nnn.field(alias)._type.makeZero());
         }
-        // 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 = 2; 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.glb().meet(rhs._type.glb()), r, lhs, rhs).peephole());
-            }
+        // 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;
     }
 
-    // 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.in(i)==scope ? phi : back.in(i)); // Fill backedge
+    @Override public String label() { return "MemMerge"; }
+    @Override public boolean isMem() { return true; }
+    @Override public Type compute() { return TypeMem.BOT; }
+    @Override StringBuilder _print1(StringBuilder sb, BitSet visited) {
+        sb.append("MEM[");
+        for( int i=1; i<nIns(); i++ )
+            if( in(i)!=null ) {
+                sb.append(i==1 ? "default:" : i+":");
+                in(i)._print0(sb,visited).append(" ");
             }
-            if( exit_def == scope ) // Replace a lazy-phi on the exit path also
-                exit.alias(i,in(i));
-        }
+        return sb.append("]");
     }
 
-    // Now one-time do a useless-phi removal
-    void _useless( ) {
+    @Override public Node idealize() {
+        boolean progress=false, allDefault=true;
         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
-                }
+            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;
         }
-    }
-
-    @Override boolean eq( Node n ) {
-        return this==n || !_inProgress;
+        if( allDefault && in(1)!=null ) return in(1);
+
+        if( in(1) instanceof MemMergeNode mem ) {
+            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;
     }
 }
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/MemOpNode.java b/chapter20/src/main/java/com/seaofnodes/simple/node/MemOpNode.java
index cba4dc1b..41760124 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/MemOpNode.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/MemOpNode.java
@@ -6,6 +6,7 @@ import com.seaofnodes.simple.type.Type;
 import com.seaofnodes.simple.type.TypeMemPtr;
 import java.lang.StringBuilder;
 import java.util.BitSet;
+import java.util.ArrayList;
 
 /**
  * Convenience common base for Load and Store.
@@ -51,6 +52,13 @@ public abstract class MemOpNode extends Node {
     public StringBuilder _printMach( StringBuilder sb, BitSet visited ) { throw Utils.TODO(); }
 
 
+    // 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) {
+        ptr().addDep(dep);
+        return _alias==other._alias && _declaredType==other._declaredType && err()==null;
+    }
+
     @Override
     boolean eq(Node n) {
         MemOpNode mem = (MemOpNode)n; // Invariant
@@ -71,4 +79,29 @@ 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.isMem() && 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/chapter20/src/main/java/com/seaofnodes/simple/node/MemPhiNode.java b/chapter20/src/main/java/com/seaofnodes/simple/node/MemPhiNode.java
new file mode 100644
index 00000000..26572ce1
--- /dev/null
+++ b/chapter20/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 _declaredType;
+        Type t = Type.TOP;
+        for( int i=1; i<nIns(); i++ )
+            if( r.in(i).addDep(this)._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 boolean eq(Node n) { return _alias==((MemPhiNode)n)._alias && super.eq(n); }
+    @Override int hash() { return _alias; }
+}
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/NewNode.java b/chapter20/src/main/java/com/seaofnodes/simple/node/NewNode.java
index abce1601..b9c9de71 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/NewNode.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/NewNode.java
@@ -1,37 +1,37 @@
 package com.seaofnodes.simple.node;
 
 import com.seaofnodes.simple.type.*;
-import com.seaofnodes.simple.Utils;
 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++ )
-            // Memory slices for all fields.
-            assert 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._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]._type.str() : _ptr._obj.str()); }
     @Override
@@ -40,32 +40,10 @@ 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++ ) {
-            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]._type);
-            ts[i+2] = TypeMem.make(fs[i]._alias,tfld2);
-        }
-        return TypeTuple.make(ts);
+        return TypeTuple.make(_ptr,TypeMem.BOT);
     }
 
     @Override
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/Node.java b/chapter20/src/main/java/com/seaofnodes/simple/node/Node.java
index 65dbbd76..aecac75c 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/Node.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/Node.java
@@ -230,6 +230,7 @@ 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();
     }
 
@@ -288,6 +289,7 @@ 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.addAll(n._outputs);
         }
@@ -539,6 +541,14 @@ public abstract class Node implements Cloneable {
      * being added must benefit from this node being peepholed.
      */
     Node addDep( Node dep ) {
+        return addDep(dep,false);
+    }
+
+    // A bulk Phi inspects its users' partitions. Def-to-use propagation does
+    // not revisit that producer when an immediate user's partition changes.
+    Node addDepForwards(Node dep) { return addDep(dep,true); }
+
+    private Node addDep(Node dep, boolean forwards) {
         // Running peepholes during the big assert cannot have side effects
         // like adding dependencies.
         if( CODE._midAssert ) return this;
@@ -547,8 +557,8 @@ public abstract class Node implements Cloneable {
         if( dep == null ) return this;
         if( _deps==null ) _deps = new Ary<>(Node.class);
         if( _deps   .find(dep) != -1 ) return this; // Already on list
-        if( _inputs .find(dep) != -1 ) return this; // No need for deps on immediate neighbors
-        if( _outputs.find(dep) != -1 ) return this;
+        if( !forwards && _inputs .find(dep) != -1 ) return this; // No need for deps on immediate neighbors
+        if( !forwards && _outputs.find(dep) != -1 ) return this;
         _deps.add(dep);
         return this;
     }
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/PhiNode.java b/chapter20/src/main/java/com/seaofnodes/simple/node/PhiNode.java
index 944e7ed6..52298345 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/PhiNode.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/PhiNode.java
@@ -24,6 +24,10 @@ 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
@@ -77,32 +81,9 @@ public class PhiNode extends Node {
             if( r.in(i)._type == Type.XCONTROL )
                 return null;
 
-        // Pull "down" a common data op.  One less op in the world.  One more
-        // Phi, but Phis do not make code.
-        //   Phi(op(A,B),op(Q,R),op(X,Y)) becomes
-        //     op(Phi(A,Q,X), Phi(B,R,Y)).
-        Node op = in(1);
-        if( !isMem() && op.nIns()==3 && op.in(0)==null && same_op() ) {
-            assert !(op instanceof CFGNode);
-            Node[] lhss = new Node[nIns()];
-            Node[] rhss = new Node[nIns()];
-            lhss[0] = rhss[0] = in(0); // Set Region
-            for( int i=1; i<nIns(); i++ ) {
-                lhss[i] = in(i).in(1);
-                rhss[i] = in(i).in(2);
-            }
-            Node phi_lhs = new PhiNode(_label, _declaredType,lhss).peephole();
-            Node phi_rhs = new PhiNode(_label, _declaredType,rhss).peephole();
-            Node down = op.copy(phi_lhs,phi_rhs);
-            // Test not running backwards, which can happen for e.g. And's
-            if( down.compute().isa(compute()) )
-                return down;
-            in(1).in(1).addDep(this);
-            in(1).in(2).addDep(this);
-            in(2).in(1).addDep(this);
-            in(2).in(2).addDep(this);
-            down.kill();
-        }
+        // 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 ) {
@@ -134,12 +115,75 @@ public class PhiNode extends Node {
     }
 
     private boolean same_op() {
-        for( int i=2; i<nIns(); i++ )
-            if( in(1).getClass() != in(i).getClass() )
-                return false;
+        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 ) {
+                    use.addDepForwards(this);
+                    return false;
+                }
+        Node busy = null;
+        for( int i=1; i<nIns(); i++ ) {
+            Node n = in(i);
+            if( region().in(i).addDep(this)._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.
+            n.addDepForwards(this);
+            if( n.nOuts()>1 ) {
+                for( Node use : n._outputs )
+                    if( use!=null && use!=this ) use.addDepForwards(this);
+                if( n instanceof StoreNode || busy!=null ) return false;
+                busy=n;
+            }
+            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() {
+        Node op = in(1), cp = op.copyEmpty();
+        cp._type = null;
+        cp.addDef(null);
+        for( int j=1; j<op.nIns(); j++ ) {
+            Node x = op.in(j);
+            boolean different = false;
+            for( int i=2; i<nIns(); i++ )
+                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);
+                }
+                PhiNode phi = j==1 && op instanceof MemOpNode mem
+                    ? new MemPhiNode(_label,mem._alias,ins)
+                    : PhiNode.make(_label,t.glb(),ins);
+                x = phi.peephole();
+            }
+            cp.addDef(x);
+        }
+        // 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 ) in(i).in(j).addDep(this);
+        cp.kill();
+        return null;
+    }
+
     /**
      * If only single unique input, return it
      */
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/RegionNode.java b/chapter20/src/main/java/com/seaofnodes/simple/node/RegionNode.java
index 166ef4bf..f28952e0 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/RegionNode.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/RegionNode.java
@@ -47,7 +47,7 @@ public class RegionNode extends CFGNode {
         if( !hasPhi() &&         // No Phi users, just a control user
             in(1) instanceof CProjNode p1 &&
             in(2) instanceof CProjNode p2 &&
-            p1.in(0).addDep(this)==p2.in(0).addDep(this) &&
+            p1.addDepForwards(this).in(0)==p2.addDepForwards(this).in(0) &&
             p1.in(0) instanceof IfNode iff ) {
             // Replace with the iff.ctrl directly
             if( nIns()==3 ) return iff.ctrl();
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/ReturnNode.java b/chapter20/src/main/java/com/seaofnodes/simple/node/ReturnNode.java
index b767b69b..7858671c 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/ReturnNode.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/ReturnNode.java
@@ -6,15 +6,7 @@ import com.seaofnodes.simple.Utils;
 import com.seaofnodes.simple.type.*;
 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;
@@ -49,7 +41,7 @@ public class ReturnNode extends CFGNode {
     public Type compute() {
         if( inProgress () ) return TypeTuple.RET; // In progress
         // Inlining can delete the entry while callers still use this return.
-        return TypeTuple.make(ctrl()._type,mem()._type,expr()._type);
+        return TypeTuple.make(ctrl()._type,TypeMem.BOT,expr()._type);
     }
 
     @Override public Node idealize() {
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/ScopeNode.java b/chapter20/src/main/java/com/seaofnodes/simple/node/ScopeNode.java
index 13ddc53a..406be11f 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/ScopeNode.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/ScopeNode.java
@@ -7,9 +7,9 @@ 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
@@ -19,6 +19,11 @@ public class ScopeNode extends MemMergeNode {
     public static final String ARG0 = "arg";
     public static final String MEM0 = "$mem";
 
+    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;
 
@@ -34,7 +39,7 @@ public class ScopeNode extends MemMergeNode {
 
     // A new ScopeNode
     public ScopeNode() {
-        super(true);
+        _type = Type.BOTTOM;
         _vars   = new Ary<>(Var    .class);
         _lexSize= new Ary<>(Integer.class);
         _kinds  = new Ary<>(Kind   .class);
@@ -70,7 +75,7 @@ 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)); }
 
     /**
      * The ctrl of a ScopeNode is always bound to the currently active
@@ -83,7 +88,7 @@ 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) {
         assert _lexSize._len==_kinds._len;
@@ -166,12 +171,6 @@ 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.
      *
@@ -207,7 +206,7 @@ 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.lazyGLB(), loop.ctrl(), loop.in(loop.update(v,null)._idx),null).peephole());
+                : loop.setDef(v._idx,PhiNode.make(v._name, v.lazyGLB(), loop.ctrl(), loop.in(loop.update(v,null)._idx),null).peephole());
             setDef(v._idx,old);
         }
         assert !v._final || st==null;
@@ -246,20 +245,8 @@ 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
+        // 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;
     }
@@ -274,22 +261,21 @@ 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 = _vars.at(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());
             }
     }
 
@@ -314,21 +300,15 @@ 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++ ) {
+        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 ) {
@@ -342,6 +322,23 @@ 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/chapter20/src/main/java/com/seaofnodes/simple/node/StartNode.java b/chapter20/src/main/java/com/seaofnodes/simple/node/StartNode.java
index 22a1750d..2783ab93 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/StartNode.java
+++ b/chapter20/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.Utils.TODO;
 
-/**
- * The Start node represents the start of the function.
- *
- * 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;
@@ -45,7 +38,7 @@ 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/chapter20/src/main/java/com/seaofnodes/simple/node/StoreNode.java b/chapter20/src/main/java/com/seaofnodes/simple/node/StoreNode.java
index cb374c58..4833df39 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/StoreNode.java
+++ b/chapter20/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; }
 
@@ -38,28 +43,25 @@ public class StoreNode extends MemOpNode {
     @Override
     public Type compute() {
         Type val = val()._type;
-        TypeMem mem = (TypeMem)mem()._type; // 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 ) {
-            // Update declared forward ref to the actual
-            if( _declaredType.isFRef() && val instanceof TypeMemPtr tmp && !tmp.isFRef() )
-                _declaredType = tmp;
-            val = val.join(_declaredType);
-            t = val.meet(mem._t);
-        }
+        if( _declaredType.isFRef() && val instanceof TypeMemPtr tmp && !tmp.isFRef() )
+            _declaredType = tmp;
+        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;
+        }
+
 
         // 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.
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/arm/MemOpARM.java b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/arm/MemOpARM.java
index adffa501..e874ce15 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/arm/MemOpARM.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/arm/MemOpARM.java
@@ -34,6 +34,8 @@ public class MemOpARM extends MemOpNode implements MachNode {
 
     @Override public  StringBuilder _printMach(StringBuilder sb, BitSet visited) { return sb.append(".").append(_name); }
 
+    @Override public boolean isMem() { return _type instanceof TypeMem; }
+
     @Override public String label() { return op(); }
     @Override public Type compute() { throw Utils.TODO(); }
     @Override public Node idealize() { throw Utils.TODO(); }
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/arm/NewARM.java b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/arm/NewARM.java
index bd88fa83..c145f348 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/arm/NewARM.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/arm/NewARM.java
@@ -12,9 +12,9 @@ public class NewARM extends NewNode implements MachNode {
     // A pre-zeroed chunk of memory.
     NewARM(NewNode nnn) { super(nnn); }
     // Size and pointer result in standard calling convention; null for all the
-    // memory aliases edges
-    @Override public RegMask    regmap(int i) { return i == 1 ? arm. X0_MASK : null; }
-    @Override public RegMask outregmap(int i) { return i == 1 ? arm. X0_MASK : null; }
+    // memory and ordering edges
+    @Override public RegMask    regmap(int i) { return i == 2 ? arm. X0_MASK : null; }
+    @Override public RegMask outregmap(int i) { return i == 0 ? arm. X0_MASK : null; }
     @Override public RegMask outregmap() { return null; }
 
     // Encoding is appended into the byte array; size is returned
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/arm/StoreARM.java b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/arm/StoreARM.java
index 49045b28..d78fdcad 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/arm/StoreARM.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/arm/StoreARM.java
@@ -34,7 +34,7 @@ public class StoreARM extends MemOpARM {
         if( i==2 ) return arm.RMASK;
         if( i==3 ) return arm.RMASK;
         if( i==4 ) return arm.RMASK;
-        throw Utils.TODO();
+        return null; // Trailing anti-dependence edges
     }
     // Register mask allowed as a result.  0 for no register.
     @Override public RegMask outregmap() { return null; }
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/arm/arm.java b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/arm/arm.java
index fd207157..009af342 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/arm/arm.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/arm/arm.java
@@ -204,6 +204,8 @@ public class arm extends Machine {
         case NewNode      nnn   -> new NewARM(nnn);
         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 ReadOnlyNode(read);
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/riscv/MemOpRISC.java b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/riscv/MemOpRISC.java
index bcd7924e..01f51b57 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/riscv/MemOpRISC.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/riscv/MemOpRISC.java
@@ -33,6 +33,8 @@ public class MemOpRISC extends MemOpNode implements MachNode {
 
     @Override public  StringBuilder _printMach(StringBuilder sb, BitSet visited) { return sb.append(".").append(_name); }
 
+    @Override public boolean isMem() { return _type instanceof TypeMem; }
+
     @Override public String label() { return op(); }
     @Override public Type compute() { throw Utils.TODO(); }
     @Override public Node idealize() { throw Utils.TODO(); }
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/riscv/NewRISC.java b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/riscv/NewRISC.java
index 53ad0af2..e8d32570 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/riscv/NewRISC.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/riscv/NewRISC.java
@@ -10,9 +10,9 @@ public class NewRISC extends NewNode implements MachNode {
     // A pre-zeroed chunk of memory.
     NewRISC( NewNode nnn ) { super(nnn); }
     // Size and pointer result in standard calling convention; null for all the
-    // memory aliases edges
-    @Override public RegMask    regmap(int i) { return i == 1 ? riscv.A0_MASK : null; }
-    @Override public RegMask outregmap(int i) { return i == 1 ? riscv.A0_MASK : null; }
+    // memory and ordering edges
+    @Override public RegMask    regmap(int i) { return i == 2 ? riscv.A0_MASK : null; }
+    @Override public RegMask outregmap(int i) { return i == 0 ? riscv.A0_MASK : null; }
     @Override public RegMask outregmap() { return null; }
 
     // Encoding is appended into the byte array; size is returned
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/riscv/StoreRISC.java b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/riscv/StoreRISC.java
index c57cf11d..22e443fd 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/riscv/StoreRISC.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/riscv/StoreRISC.java
@@ -34,7 +34,7 @@ public class StoreRISC extends MemOpRISC {
         if( i==2 ) return riscv.RMASK;
         if( i==3 ) return riscv.RMASK;
         if( i==4 ) return riscv.RMASK;
-        throw Utils.TODO();
+        return null; // Trailing anti-dependence edges
     }
 
 
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/riscv/riscv.java b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/riscv/riscv.java
index 83706cbf..daf33f29 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/riscv/riscv.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/riscv/riscv.java
@@ -206,6 +206,8 @@ public class riscv extends Machine {
         case NewNode nnn -> new NewRISC(nnn);
         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 ReadOnlyNode(read);
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/MemOpX86.java b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/MemOpX86.java
index f65fe054..1a771081 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/MemOpX86.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/MemOpX86.java
@@ -57,6 +57,8 @@ public abstract class MemOpX86 extends MemOpNode implements MachNode {
         return sb.append(".").append(_name);
     }
 
+    @Override public boolean isMem() { return _type instanceof TypeMem; }
+
     @Override public String label() { return op(); }
     @Override public Type compute() { throw Utils.TODO(); }
     @Override public Node idealize() { throw Utils.TODO(); }
@@ -67,7 +69,7 @@ public abstract class MemOpX86 extends MemOpNode implements MachNode {
         if( i==2 ) return x86_64_v2.RMASK;    // base  in GPR
         if( i==3 ) return x86_64_v2.RMASK;    // index in GPR
         if( i==4 ) return _sz >= '4' ? x86_64_v2.MEM_MASK : x86_64_v2.RMASK; // Narrow stores need GPRs
-        throw Utils.TODO();
+        return null; // Trailing anti-dependence edges
     }
 
     @Override public int encoding(ByteArrayOutputStream bytes) { throw Utils.TODO(); }
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/NewX86.java b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/NewX86.java
index cd47c64b..f4803b2e 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/NewX86.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/NewX86.java
@@ -10,9 +10,9 @@ public class NewX86 extends NewNode implements MachNode {
     // A pre-zeroed chunk of memory.
     NewX86( NewNode nnn ) { super(nnn); }
     // Size and pointer result in standard calling convention; null for all the
-    // memory aliases edges
-    @Override public RegMask    regmap(int i) { return i == 1 ? x86_64_v2.RDI_MASK : null; }
-    @Override public RegMask outregmap(int i) { return i == 1 ? x86_64_v2.RAX_MASK : null; }
+    // memory and ordering edges
+    @Override public RegMask    regmap(int i) { return i == 2 ? x86_64_v2.RDI_MASK : null; }
+    @Override public RegMask outregmap(int i) { return i == 0 ? x86_64_v2.RAX_MASK : null; }
     @Override public RegMask outregmap() { return null; }
 
     // Encoding is appended into the byte array; size is returned
diff --git a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java
index aacd2352..57771d6a 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java
+++ b/chapter20/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java
@@ -235,6 +235,8 @@ public class x86_64_v2 extends Machine {
         case NewNode      nnn   -> new NewX86(nnn);
         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 ReadOnlyNode(read);
@@ -261,11 +263,12 @@ public class x86_64_v2 extends Machine {
     private Node add( AddNode add ) {
         Node lhs = add.in(1);
         Node rhs = add.in(2);
+        // The other addend must occupy the result register, even if constant.
         if( lhs instanceof LoadNode ld && ld.nOuts()==1 )
-            return new AddMemX86(add,address(ld),ld.ptr(),idx,off,scale, imm(rhs),val);
+            return new AddMemX86(add,address(ld),ld.ptr(),idx,off,scale, 0,rhs);
 
         if( rhs instanceof LoadNode ld && ld.nOuts()==1 )
-            throw Utils.TODO(); // Swap load sides
+            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/chapter20/src/main/java/com/seaofnodes/simple/print/SimpleGraphAdapter.java b/chapter20/src/main/java/com/seaofnodes/simple/print/SimpleGraphAdapter.java
index d0df57fa..7a2025da 100644
--- a/chapter20/src/main/java/com/seaofnodes/simple/print/SimpleGraphAdapter.java
+++ b/chapter20/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;