From 2d51b64758db57e7d549639d12af35d03a0bd3d8 Mon Sep 17 00:00:00 2001 From: Cliff Click Date: Tue, 29 Sep 2026 17:10:17 -0700 Subject: Forward port lazy mem eval to ch18 --- .../com/seaofnodes/simple/GlobalCodeMotion.java | 29 +-- .../main/java/com/seaofnodes/simple/Parser.java | 87 ++++---- .../com/seaofnodes/simple/node/BulkMemPhiNode.java | 237 +++++++++++++++++++++ .../java/com/seaofnodes/simple/node/LoadNode.java | 96 +++++++-- .../com/seaofnodes/simple/node/MemMergeNode.java | 89 ++++++++ .../java/com/seaofnodes/simple/node/MemOpNode.java | 7 + .../com/seaofnodes/simple/node/MemPhiNode.java | 42 ++++ .../java/com/seaofnodes/simple/node/NewNode.java | 48 ++--- .../main/java/com/seaofnodes/simple/node/Node.java | 36 +++- .../java/com/seaofnodes/simple/node/PhiNode.java | 91 ++++++-- .../com/seaofnodes/simple/node/RegionNode.java | 2 +- .../com/seaofnodes/simple/node/ReturnNode.java | 23 +- .../com/seaofnodes/simple/node/ScopeMinNode.java | 150 ------------- .../java/com/seaofnodes/simple/node/ScopeNode.java | 106 +++++---- .../java/com/seaofnodes/simple/node/StartNode.java | 12 +- .../java/com/seaofnodes/simple/node/StoreNode.java | 27 +-- .../simple/print/SimpleGraphAdapter.java | 2 +- .../java/com/seaofnodes/simple/Chapter04Test.java | 5 +- .../java/com/seaofnodes/simple/Chapter10Test.java | 119 ++++++++++- .../java/com/seaofnodes/simple/Chapter10bTest.java | 64 ++++++ .../java/com/seaofnodes/simple/Chapter13Test.java | 2 +- .../java/com/seaofnodes/simple/Chapter14Test.java | 2 +- .../java/com/seaofnodes/simple/Chapter15Test.java | 4 +- .../java/com/seaofnodes/simple/Chapter16Test.java | 23 ++ .../com/seaofnodes/simple/evaluator/Evaluator.java | 16 +- .../com/seaofnodes/simple/evaluator/Scheduler.java | 33 ++- 26 files changed, 950 insertions(+), 402 deletions(-) create mode 100644 chapter17/src/main/java/com/seaofnodes/simple/node/BulkMemPhiNode.java create mode 100644 chapter17/src/main/java/com/seaofnodes/simple/node/MemMergeNode.java create mode 100644 chapter17/src/main/java/com/seaofnodes/simple/node/MemPhiNode.java delete mode 100644 chapter17/src/main/java/com/seaofnodes/simple/node/ScopeMinNode.java create mode 100644 chapter17/src/test/java/com/seaofnodes/simple/Chapter10bTest.java (limited to 'chapter17/src') diff --git a/chapter17/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java b/chapter17/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java index ee6e1f1a..eba99128 100644 --- a/chapter17/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java +++ b/chapter17/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java @@ -109,12 +109,8 @@ public abstract class GlobalCodeMotion { // Loads need their memory inputs' uses also done if( n instanceof LoadNode ld ) - for( Node memuse : ld.mem()._outputs ) - if( late[memuse._nid]==null && - // 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)) ) + for( Node memuse : ld.antiDeps() ) + if( late[memuse._nid]==null ) continue outer; // All uses done, schedule @@ -126,9 +122,7 @@ 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 ) @@ -185,13 +179,24 @@ 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 work, BitSet visit) { + if( visit.get(def._nid) ) return; + visit.set(def._nid); + for( Node out : def._outputs ) + if( out instanceof LoadNode ld && late[ld._nid]==null ) work.push(ld); + if( def instanceof MemMergeNode ) + for( int i=1; iearly, 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: assert late[st._nid]!=null; @@ -208,10 +213,6 @@ 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 LoadNode ld: break; // Loads do not cause anti-deps on other loads - case ReturnNode ret: break; // Load must already be ahead of Return - case ScopeMinNode ret: break; // Mem uses now on ScopeMin - case NeverNode never: break; default: throw Utils.TODO(); } } diff --git a/chapter17/src/main/java/com/seaofnodes/simple/Parser.java b/chapter17/src/main/java/com/seaofnodes/simple/Parser.java index 86b59dc3..9f81d556 100644 --- a/chapter17/src/main/java/com/seaofnodes/simple/Parser.java +++ b/chapter17/src/main/java/com/seaofnodes/simple/Parser.java @@ -164,11 +164,8 @@ public class Parser { _xScopes.push(_scope); // Enter a new scope for the initial control and arguments _scope.push(); - ScopeMinNode mem = new ScopeMinNode(); - mem.addDef(null); _scope.define(ScopeNode.CTRL, Type.CONTROL , false, new CProjNode(START, 0, ScopeNode.CTRL).peephole()); - mem.addDef( new ProjNode(START, 1, ScopeNode.MEM0).peephole()); - _scope.define(ScopeNode.MEM0, TypeMem.TOP , false, mem .peephole()); + _scope.define(ScopeNode.MEM0, TypeMem.BOT , false, new ProjNode(START, 1, ScopeNode.MEM0).peephole()); _scope.define(ScopeNode.ARG0, TypeInteger.BOT, false, new ProjNode(START, 2, ScopeNode.ARG0).peephole()); // Parse whole program @@ -525,7 +522,7 @@ public class Parser { Node expr = parseAsgn(); // Final variable to update - ScopeMinNode.Var def = _scope.lookup(name); + ScopeNode.Var def = _scope.lookup(name); if( def==null ) throw error("Undefined name '" + name + "'"); @@ -654,7 +651,7 @@ public class Parser { Field[] fs = new Field[varlen-lexlen]; for( int i=lexlen; i ALTMP = new Ary<>(Node.class); @@ -1036,13 +1037,18 @@ 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(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. For now this is just a field @@ -1085,6 +1091,7 @@ 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 @@ -1097,16 +1104,11 @@ public class Parser { Node val = parseAsgn().keep(); Node lift = liftExpr( val, tf, f._final ); - Node st = new StoreNode(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(name, f._alias, tf.glb(), memAlias(f._alias), expr.keep(), off); + Node load = new LoadNode(name, f._alias, tf.glb(), 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. @@ -1119,12 +1121,7 @@ public class Parser { throw error("Cannot reassign final '"+f._fname+"'"); Node inc = peep(new AddNode(load,con( _lexer.peek(-1)=='+' ? 1 : -1))); Node val = zsMask(inc,tf); - Node st = new StoreNode(name, f._alias, tf, memAlias(f._alias), expr.unkeep(), off, val, 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,val,false,base.isAry() ? ctrl() : null); // And use the original loaded value as the result } else expr.unkill(); off.unkill(); diff --git a/chapter17/src/main/java/com/seaofnodes/simple/node/BulkMemPhiNode.java b/chapter17/src/main/java/com/seaofnodes/simple/node/BulkMemPhiNode.java new file mode 100644 index 00000000..84f97189 --- /dev/null +++ b/chapter17/src/main/java/com/seaofnodes/simple/node/BulkMemPhiNode.java @@ -0,0 +1,237 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.IterPeeps; +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); } + + @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= 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 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 + : "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= 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= 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/chapter17/src/main/java/com/seaofnodes/simple/node/LoadNode.java b/chapter17/src/main/java/com/seaofnodes/simple/node/LoadNode.java index 17ed127c..2b430df4 100644 --- a/chapter17/src/main/java/com/seaofnodes/simple/node/LoadNode.java +++ b/chapter17/src/main/java/com/seaofnodes/simple/node/LoadNode.java @@ -1,9 +1,10 @@ package com.seaofnodes.simple.node; import com.seaofnodes.simple.Utils; -import com.seaofnodes.simple.Parser; +import com.seaofnodes.simple.IterPeeps; import com.seaofnodes.simple.type.*; import java.util.BitSet; +import java.util.ArrayList; /** * Load represents extracting a value from inside a memory object, @@ -23,6 +24,25 @@ public class LoadNode extends MemOpNode { super(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. Stores 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 ) return true; + } + return false; + } + // Debugger label @Override public String label() { return "ld_"+mlabel(); } @Override @@ -30,31 +50,31 @@ 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,IterPeeps.add(merge.alias(_alias))); + return this; + } Node ptr = ptr(); // Simple Load-after-Store on same address. if( mem() instanceof StoreNode st && - ptr == st.ptr() && off() == st.off() ) { // Must check same object + _alias==st._alias && ptr == st.ptr() && off() == st.off() ) { // Must check same object assert _name.equals(st._name); // Equiv class aliasing is perfect return 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 nnn.in(nnn.findAlias(_alias)); // Load from New init // Load-after-Store on same address, but bypassing provably unrelated @@ -65,12 +85,16 @@ public class LoadNode extends MemOpNode { Node mem = mem(); outer: while( true ) { + mem.addDep(this); switch( mem ) { + case MemMergeNode merge: + mem = merge.alias(_alias); + break; case StoreNode st: - if( ptr == st.ptr() && off() == st.off() ) + if( _alias==st._alias && ptr == st.ptr() && off() == st.off() ) return 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 @@ -84,13 +108,11 @@ public class LoadNode extends MemOpNode { return castRO(nnn1.in(nnn1.findAlias(_alias))); // Load from New init if( !(ptr instanceof ProjNode pproj && pproj.in(0) instanceof NewNode nnn2) ) break outer; // Cannot tell, ptr not related to New - mem = nnn1.in(_alias);// Bypass unrelated New + mem = nnn1.mem();// Bypass unrelated New break; - } else if( mproj.in(0) instanceof StartNode ) { - break outer; - } else throw Utils.TODO(); + } else break outer; default: - throw Utils.TODO(); + break outer; } } @@ -101,7 +123,7 @@ 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 @@ -122,7 +144,9 @@ public class LoadNode extends MemOpNode { private Node ld( int idx ) { Node mem = mem(), ptr = ptr(); - return new LoadNode(_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(_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 ) { @@ -155,8 +179,9 @@ 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() ) { px.addDep(this); return true; } - if( px instanceof StoreNode st1 && ptr()==st1.ptr() && off()==st1.off() ) { px.addDep(this); 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; } @@ -166,4 +191,29 @@ public class LoadNode extends MemOpNode { return new ReadOnlyNode(rez).peephole(); return rez; } + + // 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 antiDeps() { + ArrayList deps = new ArrayList<>(); + antiDeps(mem(),deps,new BitSet()); + return deps; + } + + private void antiDeps(Node mem, ArrayList 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 1; + while( nIns()<=alias ) addDef(null); + setDef(alias,mem); + } + + // Scalar contents of one alias. New's field inputs supply the initialized type; + // 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); + Node init = nnn.in(nnn.findAlias(alias)).addDep(dep); + return contents(nnn.mem(),alias,dep).meet(init._type); + } + // There are no heap arguments in this chapter. Before the first New, + // the set of allocated objects (and hence stored values) is empty. + if( mem instanceof ProjNode proj && proj.in(0) instanceof StartNode ) return Type.TOP; + 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; + } + + @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() || in(i)==null) ) + alias(i,mem.in(i)); + IterPeeps.add(mem.in(1)); + setDef(1,mem.in(1)); + return this; + } + assert BulkMemPhiNode.checkMerge(this); + return progress ? this : null; + } +} diff --git a/chapter17/src/main/java/com/seaofnodes/simple/node/MemOpNode.java b/chapter17/src/main/java/com/seaofnodes/simple/node/MemOpNode.java index 82aab8ac..0bf9b6ac 100644 --- a/chapter17/src/main/java/com/seaofnodes/simple/node/MemOpNode.java +++ b/chapter17/src/main/java/com/seaofnodes/simple/node/MemOpNode.java @@ -33,6 +33,13 @@ public abstract class MemOpNode extends Node { public Node ptr() { return in(2); } public Node off() { return in(3); } + // 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 diff --git a/chapter17/src/main/java/com/seaofnodes/simple/node/MemPhiNode.java b/chapter17/src/main/java/com/seaofnodes/simple/node/MemPhiNode.java new file mode 100644 index 00000000..6aae00f9 --- /dev/null +++ b/chapter17/src/main/java/com/seaofnodes/simple/node/MemPhiNode.java @@ -0,0 +1,42 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.IterPeeps; +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; + } + + @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 _inputs; + public Ary _inputs; /** * Outputs reference Nodes that are not null and have this Node as an @@ -43,7 +43,7 @@ public abstract class Node implements OutNode { * walked in either direction. These outputs are typically used for * efficient optimizations but otherwise have no semantics meaning. */ - public final Ary _outputs; + public Ary _outputs; /** @@ -214,6 +214,7 @@ public abstract class Node implements OutNode { // 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(); } @@ -274,6 +275,7 @@ public abstract class Node implements OutNode { 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); IterPeeps.addAll(n._outputs); } @@ -460,6 +462,14 @@ public abstract class Node implements OutNode { * 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( IterPeeps.midAssert() ) return this; @@ -467,8 +477,8 @@ public abstract class Node implements OutNode { if( obs != null ) obs.dep(this, dep); 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; } @@ -594,6 +604,20 @@ public abstract class Node implements OutNode { // Does not need to be implemented in isCFG() nodes. Node copy(Node lhs, Node rhs) { throw Utils.TODO("Binary ops need to implement copy"); } + // Exact-class copy preserving operation attributes, with fresh identity + // and no edges, dependencies, or GVN membership. + public final Node copyEmpty() { + Node n; + try { n = (Node)clone(); } + catch( CloneNotSupportedException e ) { throw new AssertionError(e); } + n._nid = UNIQUE_ID++; + n._inputs = new Ary<>(Node.class); + n._outputs = new Ary<>(Node.class); + n._deps = null; + n._hash = 0; + return n; + } + // Report any post-optimize errors String err() { return null; } diff --git a/chapter17/src/main/java/com/seaofnodes/simple/node/PhiNode.java b/chapter17/src/main/java/com/seaofnodes/simple/node/PhiNode.java index 0532d2b7..c6cbd3dd 100644 --- a/chapter17/src/main/java/com/seaofnodes/simple/node/PhiNode.java +++ b/chapter17/src/main/java/com/seaofnodes/simple/node/PhiNode.java @@ -16,6 +16,10 @@ public class PhiNode extends Node { public PhiNode(String label, Type declaredType, Node... inputs) { super(inputs); _label = label; assert declaredType!=null; _declaredType = declaredType; } + 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 @@ -65,24 +69,9 @@ public class PhiNode extends Node { if (live != null) return live; - // 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( 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 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 ) { @@ -112,12 +101,72 @@ public class PhiNode extends Node { } private boolean same_op() { - for( int i=2; i1 ) { + 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 - * In this presentation, Return functions as a Stop node, since multiple return statements are not possible. - * The Stop node will be introduced in Chapter 6 when we implement if statements. - *

- * The Return's output is the value from the data node. - */ +/** Keeps all preceding memory effects and the return value alive: {ctrl, $mem, value}. */ public class ReturnNode extends CFGNode { public ReturnNode(Node ctrl, Node data, ScopeNode scope) { - // Add memory slices to Return, so all memory updates are live-on-exit. - super(ctrl, data); - if( scope!=null ) - for( int i=2; i= 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 ) { - ScopeMinNode 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); - } - // Memory projections are made lazily; expand as needed - return st==null ? old : alias(alias,st); // Not lazy, so this is the answer - } - - - void _merge(ScopeMinNode 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 name as it will trigger a phi creation - //Var v = _vars.at(i); - Node lhs = this._mem(i,null); - Node rhs = that._mem(i,null); - alias(i, new PhiNode(Parser.memName(i), TypeMem.BOT, r, lhs, rhs).peephole()); - } - } - - // Fill in the backedge of any inserted Phis - void _endLoopMem( ScopeNode scope, ScopeMinNode back, ScopeMinNode exit ) { - for( int i=2; i _vars; - // Since of each nested lexical scope + // Size of each nested lexical scope public final Ary _lexSize; // True if parsing inside of a constructor @@ -35,6 +67,7 @@ public class ScopeNode extends ScopeMinNode { // A new ScopeNode public ScopeNode() { + _type = Type.BOTTOM; _vars = new Ary<>(Var .class); _lexSize= new Ary<>(Integer.class); _inCons = new Ary<>(Boolean.class); @@ -70,7 +103,8 @@ public class ScopeNode extends ScopeMinNode { public Node ctrl() { return in(0); } - public ScopeMinNode mem() { return (ScopeMinNode)in(1); } + public Node mem() { return in(lookup(MEM0)); } + public void mem(Node n) { update(MEM0,n); } /** * The ctrl of a ScopeNode is always bound to the currently active @@ -124,12 +158,6 @@ public class ScopeNode extends ScopeMinNode { 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. * @@ -165,7 +193,7 @@ public class ScopeNode extends ScopeMinNode { // 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); } if( st!=null ) setDef(v._idx,st); // Set new value @@ -203,20 +231,8 @@ public class ScopeNode extends ScopeMinNode { n.keep(); dup.addDef(ctrl()); // Control input is just copied - // Memory input is a shallow copy - ScopeMinNode memdup = new ScopeMinNode(), mem = mem(); - memdup.addDef(null); - memdup.addDef(loop ? this : mem.in(1)); - for( int i=2; i { 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( i == 1 && (n instanceof MemOpNode || n instanceof ReturnNode) ) return Role.MEM; + if( i == 1 && (n instanceof MemOpNode || n instanceof ReturnNode || n instanceof NewNode) ) return Role.MEM; Node def = n.in(i); if( def != null && isMem(def) ) return Role.MEM; return Role.DATA; diff --git a/chapter17/src/test/java/com/seaofnodes/simple/Chapter04Test.java b/chapter17/src/test/java/com/seaofnodes/simple/Chapter04Test.java index dfd3569f..36727a8f 100644 --- a/chapter17/src/test/java/com/seaofnodes/simple/Chapter04Test.java +++ b/chapter17/src/test/java/com/seaofnodes/simple/Chapter04Test.java @@ -84,7 +84,10 @@ public class Chapter04Test { StopNode stop = parser.parse(); ReturnNode ret = stop.ret(); assertTrue(ret.in(0) instanceof CProjNode); - assertTrue(ret.in(1) instanceof ProjNode); + assertTrue(ret.in(1) instanceof ProjNode); + assertEquals(1, ((ProjNode)ret.in(1))._idx); + assertTrue(ret.in(2) instanceof ProjNode); + assertEquals(2, ((ProjNode)ret.in(2))._idx); } @Test diff --git a/chapter17/src/test/java/com/seaofnodes/simple/Chapter10Test.java b/chapter17/src/test/java/com/seaofnodes/simple/Chapter10Test.java index e3d8b177..86a9937f 100644 --- a/chapter17/src/test/java/com/seaofnodes/simple/Chapter10Test.java +++ b/chapter17/src/test/java/com/seaofnodes/simple/Chapter10Test.java @@ -1,7 +1,8 @@ package com.seaofnodes.simple; import com.seaofnodes.simple.evaluator.Evaluator; -import com.seaofnodes.simple.node.StopNode; +import com.seaofnodes.simple.node.*; +import java.util.BitSet; import org.junit.Test; import static org.junit.Assert.assertEquals; import static org.junit.Assert.fail; @@ -452,4 +453,120 @@ return 0; assertEquals("return 0;", stop.toString()); } + @Test public void testReadBeforeStores() { + var stop = new Parser(""" + struct S { int x; int y; }; + S !a = new S; S !b = new S; + a.x = 11; b.x = 22; + S !p = a; if (arg) p = b; + int before = p.x; + a.y = 55; b.y = 66; + a.x = 33; b.x = 44; + return before*100 + p.x; + """).parse().iterate(); + assertEquals(1133L, com.seaofnodes.simple.evaluator.Evaluator.evaluate(stop,0)); + assertEquals(2244L, com.seaofnodes.simple.evaluator.Evaluator.evaluate(stop,1)); + } + + static final String NESTED_MEMORY = """ + struct S { int x; int y; int z; }; + S !s = new S; + s.x = 5; s.y = 7; s.z = 11; + while (arg > 0) { + s.x = s.x + 1; + int j = 3; + while (j > 0) { + j = j - 1; + if (j == 1) continue; + s.y = s.y + arg; + if (arg == 2) break; + s.z = s.z + 1; + } + arg = arg - 1; + } + return s; + """; + + @Test public void testMemoryAcrossNestedLoops() { + var stop = new Parser(NESTED_MEMORY).parse().iterate(); + for (int arg=0; arg<7; arg++) { + var obj = (com.seaofnodes.simple.evaluator.Evaluator.Obj) + com.seaofnodes.simple.evaluator.Evaluator.evaluate(stop,arg); + org.junit.Assert.assertArrayEquals(new Object[] { + 5L+arg, 7L+arg*(arg+1)-(arg>=2 ? 2 : 0), 11L+2*arg-(arg>=2 ? 2 : 0) + }, obj.fields()); + } + } + + @Test public void testMemoryAtEarlyReturns() { + var stop = new Parser(""" + struct S { int x; int y; int z; }; + S !s = new S; + s.x = 3; s.y = 5; s.z = 7; + if (arg) { s.x = 11; s.z = 13; return s; } + s.y = 17; + return s; + """).parse().iterate(); + for (int arg=0; arg<2; arg++) { + var obj = (com.seaofnodes.simple.evaluator.Evaluator.Obj) + com.seaofnodes.simple.evaluator.Evaluator.evaluate(stop,arg); + org.junit.Assert.assertArrayEquals(arg==0 ? new Object[]{3L,17L,7L} : new Object[]{11L,5L,13L}, obj.fields()); + } + } + + @Test public void testKeepLoadsAtMerge() { + StopNode stop = new Parser(""" + struct S { int x; }; + S !a = new S; S !b = new S; + a.x = arg; b.x = arg+1; + S !p = a; S !q = b; + if (arg<0) { p=b; q=a; } + int v; + if (arg>1) v=p.x; else v=q.x; + return v; + """).parse().iterate(); + // The one-step safety check stops at the memory aggregate. + assertEquals(2,countMemoryNodes(stop,LoadNode.class,new BitSet())); + assertEquals(-1L,Evaluator.evaluate(stop,-1)); + assertEquals(1L,Evaluator.evaluate(stop,0)); + assertEquals(3L,Evaluator.evaluate(stop,3)); + } + + @Test public void testDropStores() { + StopNode stop = new Parser(""" + struct S { int x; }; + S !s = new S; + if (arg) s.x=arg+1; else s.x=arg+2; + return s; + """).parse().iterate(); + assertEquals(1,countMemoryNodes(stop,StoreNode.class,new BitSet())); + assertEquals(2L,((Evaluator.Obj)Evaluator.evaluate(stop,0)).fields()[0]); + assertEquals(4L,((Evaluator.Obj)Evaluator.evaluate(stop,3)).fields()[0]); + } + + @Test public void testKeepReadsBeforeWrites() { + StopNode stop = new Parser(""" + struct S { int x; int y; }; + S !a = new S; S !b = new S; + a.x=arg; b.x=arg+1; + S !p=a; S !q=b; + if (arg<0) { p=b; q=a; } + int v; + if (arg>1) { v=p.x; p.x=41; p.y=5; } + else { v=q.x; q.x=42; } + return v; + """).parse().iterate(); + assertEquals(2,countMemoryNodes(stop,LoadNode.class,new BitSet())); + assertEquals(-1L,Evaluator.evaluate(stop,-1)); + assertEquals(1L,Evaluator.evaluate(stop,0)); + assertEquals(3L,Evaluator.evaluate(stop,3)); + } + + private static int countMemoryNodes(Node n, Class kind, BitSet seen) { + if (n==null || seen.get(n._nid)) return 0; + seen.set(n._nid); + int cnt=kind.isInstance(n) ? 1 : 0; + for (Node def : n._inputs) cnt+=countMemoryNodes(def,kind,seen); + return cnt; + } } diff --git a/chapter17/src/test/java/com/seaofnodes/simple/Chapter10bTest.java b/chapter17/src/test/java/com/seaofnodes/simple/Chapter10bTest.java new file mode 100644 index 00000000..fd6fdf36 --- /dev/null +++ b/chapter17/src/test/java/com/seaofnodes/simple/Chapter10bTest.java @@ -0,0 +1,64 @@ +package com.seaofnodes.simple; + +import com.seaofnodes.simple.node.*; +import com.seaofnodes.simple.evaluator.Evaluator; +import org.junit.Test; +import static org.junit.Assert.*; + +public class Chapter10bTest { + @Test public void testIndependentField() { + var stop = new Parser(""" + struct S { int x; int y; }; + S !s = new S; + s.x = 42; + while (arg > 0) { s.y = s.y + arg; arg = arg - 1; } + return s.x; + """).parse().iterate(); + assertEquals("return 42;",stop.toString()); + assertEquals(42L,Evaluator.evaluate(stop,5)); + assertNull(stop.walk(n -> n instanceof LoadNode ? n : null)); + } + + @Test public void testParallelMemoryPhis() { + var stop = new Parser(""" + struct S { int x; int y; int z; }; + S !a = new S; S !b = new S; + S !p = a; if (arg) p = b; + while (arg > 0) { + a.x = p.x + arg; + a.y = p.y + 2; + a.z = p.z + 3; + arg = arg - 1; + } + return a; + """).parse().iterate(); + int[] phis = {0}; + stop.walk(n -> { + if (n instanceof MemPhiNode) phis[0]++; + if (n instanceof MemMergeNode m && m.in(1) instanceof BulkMemPhiNode b) + for(int a=b._aliases.nextSetBit(0); a>=0; a=b._aliases.nextSetBit(a+1)) + assertTrue(a=3); + assertArrayEquals(new Object[]{1L,2L,3L},((Evaluator.Obj)Evaluator.evaluate(stop,4)).fields()); + } + + @Test public void testLateSliceWorklist() throws ReflectiveOperationException { + // Seed 97 exposed a store selecting a bulk predecessor without queuing + // it. Keep both the worklist assertion and the final heap check. + for (int seed : new int[]{0,97,123,456}) { + var parser = new Parser(Chapter10Test.NESTED_MEMORY); + var field = IterPeeps.class.getDeclaredField("WORK"); + field.setAccessible(true); + Object work = field.get(null); + field = work.getClass().getDeclaredField("_R"); + field.setAccessible(true); + ((java.util.Random)field.get(work)).setSeed(seed); + var stop = parser.parse().iterate(); + assertArrayEquals(new Object[]{9L,25L,17L},((Evaluator.Obj)Evaluator.evaluate(stop,4)).fields()); + } + } +} diff --git a/chapter17/src/test/java/com/seaofnodes/simple/Chapter13Test.java b/chapter17/src/test/java/com/seaofnodes/simple/Chapter13Test.java index 862b75e3..a1fdeb3a 100644 --- a/chapter17/src/test/java/com/seaofnodes/simple/Chapter13Test.java +++ b/chapter17/src/test/java/com/seaofnodes/simple/Chapter13Test.java @@ -183,7 +183,7 @@ return new S1.s=new S2; struct I {int i;}; struct P { I? pi; }; P !p1 = new P; -P p2 = new P; +P !p2 = new P; p2.pi = new I; p2.pi.i = 2; if (arg) p1 = new P; diff --git a/chapter17/src/test/java/com/seaofnodes/simple/Chapter14Test.java b/chapter17/src/test/java/com/seaofnodes/simple/Chapter14Test.java index 23c82358..6846e7af 100644 --- a/chapter17/src/test/java/com/seaofnodes/simple/Chapter14Test.java +++ b/chapter17/src/test/java/com/seaofnodes/simple/Chapter14Test.java @@ -175,7 +175,7 @@ while(v1) break; return v1; """); StopNode stop = parser.parse().iterate(); - assertEquals("return (Phi(Loop,((1<{side=3.0,diag=1.7320508075688772}", Evaluator.evaluate(stop, 3).toString()); assertEquals("Obj{side=4.0,diag=2.0}", Evaluator.evaluate(stop, 4).toString()); } + static final String CONSTRUCTOR_MEMORY = """ + struct S { int x; int y; }; + struct T { int z=arg+40; }; + T !t = new T; + S !a = new S { x=11; y=7; }; + S !b = new S { x=22; y=9; }; + S !p=a; + if (arg) p=b; + int before=p.x; + S !c = new S { + x=p.x+1; + { int i=0; while (i<2) { p.y=p.y+1; i=i+1; } } + y=p.y; + }; + p.x=33; + return before*10000+c.x*100+c.y+t.z; + """; + + @Test public void testConstructorMemory() { + StopNode stop = new Parser(CONSTRUCTOR_MEMORY).parse().iterate(); + assertEquals(111249L,Evaluator.evaluate(stop,0)); + assertEquals(222352L,Evaluator.evaluate(stop,1)); + } } diff --git a/chapter17/src/test/java/com/seaofnodes/simple/evaluator/Evaluator.java b/chapter17/src/test/java/com/seaofnodes/simple/evaluator/Evaluator.java index 72f84fa2..30af9895 100644 --- a/chapter17/src/test/java/com/seaofnodes/simple/evaluator/Evaluator.java +++ b/chapter17/src/test/java/com/seaofnodes/simple/evaluator/Evaluator.java @@ -187,7 +187,7 @@ public class Evaluator { Object[] body=null; int num; if( type.isAry() ) { - long sz = (Long)val(alloc.in(1)); + long sz = (Long)val(alloc.size()); long n = offToIdx(sz, type); if( n < 0 ) throw new NegativeArraySizeException(""+n); @@ -201,17 +201,13 @@ public class Evaluator { assert elem instanceof TypeMemPtr; } // Length value - body[0] = vall(alloc.in(2+2)); + body[0] = vall(alloc.in(3)); } else { body = new Object[num = type._fields.length]; for (int i=0; i (double)vall(cast.in(1)); case LoadNode load -> load(load); case StoreNode store -> store(store); + case MemMergeNode merge -> MEMORY; case NewNode alloc -> alloc(alloc); case CProjNode cproj -> ((Object[])val(cproj.ctrl()))[cproj._idx]; case ProjNode proj -> ((Object[])val( proj.in(0) ))[ proj._idx]; - case ScopeMinNode mem -> null; case ReadOnlyNode ro -> val(ro.in(1)); default -> throw new AssertionError("Unexpected node " + node); }; @@ -331,7 +327,7 @@ public class Evaluator { case null: return Status.FALLTHROUGH; case ReturnNode ret: - return val(ret.in(1)); + return val(ret.expr()); case IfNode ifn: block = block.next()[isTrue(val(ifn.in(1))) ? 0 : 1]; if (block == null) return Status.FALLTHROUGH; diff --git a/chapter17/src/test/java/com/seaofnodes/simple/evaluator/Scheduler.java b/chapter17/src/test/java/com/seaofnodes/simple/evaluator/Scheduler.java index 06169e2c..afcc700f 100644 --- a/chapter17/src/test/java/com/seaofnodes/simple/evaluator/Scheduler.java +++ b/chapter17/src/test/java/com/seaofnodes/simple/evaluator/Scheduler.java @@ -31,6 +31,8 @@ import java.util.*; */ public class Scheduler { + private final HashMap> readers = new HashMap<>(); + /** * A basic block with a schedule of the containing nodes. * @@ -252,7 +254,7 @@ public class Scheduler { // but they did not refine the placement // so do that now. var mem = l.in(1); - for(var out : mem._outputs) { + for(var out : l.antiDeps()) { if (out instanceof PhiNode p) { var r = p.in(0); for (int i = 1; i < p.nIns(); i++) { @@ -270,13 +272,9 @@ public class Scheduler { for(var in:data.node._inputs) { if (in!=null) update(d(in), data.block); } - if (data.node instanceof StoreNode s) { - // Store nodes have anti-deps to load nodes. - // So decrease the uses of these loads when the store is placed. - for (var out: s.in(1)._outputs) { - if (out instanceof LoadNode) od(out).ifPresent(this::decUsers); - } - } + var loads = readers.get(data.node); + if (loads != null) + for (var load : loads) decUsers(load); } // Now all nodes should be placed and have a block assigned @@ -442,7 +440,6 @@ public class Scheduler { private void doMarkAlive(Node node) { var cfgQueue = new Stack(); var dataQueue = new Stack(); - var mem = new Stack(); markAlive(cfgQueue, node, true); // Mark all CFG nodes. while (!cfgQueue.isEmpty()) { @@ -467,16 +464,16 @@ public class Scheduler { } else { for (var in : node._inputs) if (in != null && !(in instanceof CFGNode)) markAlive(dataQueue, in, false); } - if (node instanceof StoreNode) mem.push(data); - } - // Handle store nodes and increase load with an anti-dep to the store. - while (!mem.isEmpty()) { - var data = mem.pop(); - node = data.node; - for(var out:node.in(1)._outputs) { - if (out instanceof LoadNode) od(out).ifPresent(d->d.users++); - } } + // Writes, including New behind an aggregate, schedule before their + // anti-dependent reads in this reverse scheduling pass. + for (var load : data.values()) + if (load.node instanceof LoadNode l) + for (var use : l.antiDeps()) + if ((use instanceof StoreNode || use instanceof NewNode) && data.containsKey(use)) { + load.users++; + readers.computeIfAbsent(use,k->new ArrayList<>()).add(load); + } } /** -- cgit v1.2.3