From 3507afb1602bcf159dede6bde2c70d620d6563da Mon Sep 17 00:00:00 2001 From: Cliff Click Date: Tue, 26 Nov 2024 11:01:47 -0800 Subject: Initial push --- .../src/main/java/com/seaofnodes/simple/Ary.java | 217 +++ .../com/seaofnodes/simple/GlobalCodeMotion.java | 228 ++++ .../com/seaofnodes/simple/GraphVisualizer.java | 288 ++++ .../main/java/com/seaofnodes/simple/IRPrinter.java | 286 ++++ .../main/java/com/seaofnodes/simple/IterPeeps.java | 205 +++ .../main/java/com/seaofnodes/simple/Parser.java | 1427 ++++++++++++++++++++ .../src/main/java/com/seaofnodes/simple/Utils.java | 53 + .../java/com/seaofnodes/simple/node/AddFNode.java | 44 + .../java/com/seaofnodes/simple/node/AddNode.java | 187 +++ .../java/com/seaofnodes/simple/node/AndNode.java | 68 + .../java/com/seaofnodes/simple/node/BoolNode.java | 120 ++ .../java/com/seaofnodes/simple/node/CFGNode.java | 132 ++ .../java/com/seaofnodes/simple/node/CProjNode.java | 60 + .../java/com/seaofnodes/simple/node/CastNode.java | 40 + .../com/seaofnodes/simple/node/ConstantNode.java | 54 + .../java/com/seaofnodes/simple/node/DivFNode.java | 46 + .../java/com/seaofnodes/simple/node/DivNode.java | 47 + .../java/com/seaofnodes/simple/node/IfNode.java | 66 + .../java/com/seaofnodes/simple/node/LoadNode.java | 172 +++ .../java/com/seaofnodes/simple/node/LoopNode.java | 61 + .../java/com/seaofnodes/simple/node/MemOpNode.java | 55 + .../com/seaofnodes/simple/node/MinusFNode.java | 36 + .../java/com/seaofnodes/simple/node/MinusNode.java | 42 + .../java/com/seaofnodes/simple/node/MulFNode.java | 51 + .../java/com/seaofnodes/simple/node/MulNode.java | 63 + .../java/com/seaofnodes/simple/node/MultiNode.java | 20 + .../java/com/seaofnodes/simple/node/NeverNode.java | 21 + .../java/com/seaofnodes/simple/node/NewNode.java | 75 + .../main/java/com/seaofnodes/simple/node/Node.java | 649 +++++++++ .../java/com/seaofnodes/simple/node/NotNode.java | 44 + .../java/com/seaofnodes/simple/node/OrNode.java | 65 + .../java/com/seaofnodes/simple/node/OutNode.java | 7 + .../java/com/seaofnodes/simple/node/PhiNode.java | 160 +++ .../java/com/seaofnodes/simple/node/ProjNode.java | 47 + .../com/seaofnodes/simple/node/ReadOnlyNode.java | 29 + .../com/seaofnodes/simple/node/RegionNode.java | 120 ++ .../com/seaofnodes/simple/node/ReturnNode.java | 50 + .../com/seaofnodes/simple/node/RoundF32Node.java | 38 + .../java/com/seaofnodes/simple/node/SarNode.java | 61 + .../com/seaofnodes/simple/node/ScopeMinNode.java | 150 ++ .../java/com/seaofnodes/simple/node/ScopeNode.java | 368 +++++ .../java/com/seaofnodes/simple/node/ShlNode.java | 64 + .../java/com/seaofnodes/simple/node/ShrNode.java | 52 + .../java/com/seaofnodes/simple/node/StartNode.java | 44 + .../java/com/seaofnodes/simple/node/StopNode.java | 85 ++ .../java/com/seaofnodes/simple/node/StoreNode.java | 123 ++ .../com/seaofnodes/simple/node/StructNode.java | 45 + .../java/com/seaofnodes/simple/node/SubFNode.java | 44 + .../java/com/seaofnodes/simple/node/SubNode.java | 60 + .../com/seaofnodes/simple/node/ToFloatNode.java | 30 + .../java/com/seaofnodes/simple/node/XCtrlNode.java | 13 + .../java/com/seaofnodes/simple/node/XorNode.java | 65 + .../java/com/seaofnodes/simple/type/Field.java | 77 ++ .../main/java/com/seaofnodes/simple/type/Type.java | 218 +++ .../java/com/seaofnodes/simple/type/TypeFloat.java | 120 ++ .../com/seaofnodes/simple/type/TypeInteger.java | 133 ++ .../java/com/seaofnodes/simple/type/TypeMem.java | 67 + .../com/seaofnodes/simple/type/TypeMemPtr.java | 98 ++ .../com/seaofnodes/simple/type/TypeStruct.java | 256 ++++ .../java/com/seaofnodes/simple/type/TypeTuple.java | 95 ++ 60 files changed, 7641 insertions(+) create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/Ary.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/GraphVisualizer.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/IRPrinter.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/IterPeeps.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/Parser.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/Utils.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/AddFNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/AddNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/AndNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/BoolNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/CFGNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/CProjNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/CastNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/ConstantNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/DivFNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/DivNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/IfNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/LoadNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/LoopNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/MemOpNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/MinusFNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/MinusNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/MulFNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/MulNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/MultiNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/NeverNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/NewNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/Node.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/NotNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/OrNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/OutNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/PhiNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/ProjNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/ReadOnlyNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/RegionNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/ReturnNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/RoundF32Node.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/SarNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/ScopeMinNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/ScopeNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/ShlNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/ShrNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/StartNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/StopNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/StoreNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/StructNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/SubFNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/SubNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/ToFloatNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/XCtrlNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/node/XorNode.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/type/Field.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/type/Type.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/type/TypeFloat.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/type/TypeInteger.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/type/TypeMem.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/type/TypeMemPtr.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/type/TypeStruct.java create mode 100644 chapter18/src/main/java/com/seaofnodes/simple/type/TypeTuple.java (limited to 'chapter18/src/main/java') diff --git a/chapter18/src/main/java/com/seaofnodes/simple/Ary.java b/chapter18/src/main/java/com/seaofnodes/simple/Ary.java new file mode 100644 index 00000000..8654d41e --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/Ary.java @@ -0,0 +1,217 @@ +package com.seaofnodes.simple; + +import java.lang.StringBuilder; +import java.lang.reflect.Array; +import java.util.*; + +// ArrayList with saner syntax +public class Ary extends AbstractList implements List { + public E[] _es; + public int _len; + + public Ary(E[] es) { this(es,es.length); } + public Ary(E[] es, int len) { if( es.length==0 ) es=Arrays.copyOf(es,1); _es=es; _len=len; } + @SuppressWarnings("unchecked") + public Ary(Class clazz) { this((E[]) Array.newInstance(clazz, 1),0); } + + /** @return list is empty */ + public boolean isEmpty() { return _len==0; } + /** @return active list length */ + public int len() { return _len; } + public int size() { return len(); } + + /** @param i element index + * @return element being returned; throws if OOB + * @exception AIOOBE if !(0 <= i < _len) + */ + public E get( int i ) { return at(i); } + public E at( int i ) { + range_check(i); + return _es[i]; + } + + /** @return last element */ + public E last( ) { return at(_len-1); } + + /** Add element in amortized constant time + * @param e Element to add at end of list + **/ + public E push( E e ) { + if( _len >= _es.length ) _es = Arrays.copyOf(_es,Math.max(1,_es.length<<1)); + return (_es[_len++] = e); + } + /** @return remove and return last element */ + public E pop( ) { + range_check(0); + return _es[--_len]; + } + @Override public E removeLast() { return pop(); } + + + /** Add element in amortized constant time + * @param e Element to add at end of list + * @return true */ + public boolean add( E e ) { + if( _len >= _es.length ) _es = Arrays.copyOf(_es,Math.max(1,_es.length<<1)); + _es[_len++] = e; + return true; + } + + ///** Slow, linear-time, element insert. Preserves order. + // * @param i index to insert at, between 0 and _len inclusive. + // * @param e Element to insert + // */ + //public void insert( int i, E e ) { + // if( i < 0 || i>_len ) + // throw new ArrayIndexOutOfBoundsException(""+i+" >= "+_len); + // if( _len >= _es.length ) _es = Arrays.copyOf(_es,Math.max(1,_es.length<<1)); + // System.arraycopy(_es,i,_es,i+1,(_len++)-i); + // _es[i] = e; + //} + + /** Fast, constant-time, element removal. Does not preserve order + * @param i element to be removed + * @return element removed */ + public E del( int i ) { + range_check(i); + E tmp = _es[i]; + _es[i]=_es[--_len]; + return tmp; + } + + ///** Element removal, using '=='. Does not preserve order. + // * @param e element to be removed + // * @return element removed */ + //public E del( E e ) { + // for( int i=0; i<_len; i++ ) { + // E tmp = _es[i]; + // if( tmp==e ) { + // _es[i]=_es[--_len]; + // return tmp; + // } + // } + // return null; + //} + + /** Slow, linear-time, element removal. Preserves order. + * @param i element to be removed + * @return element removed */ + @Override public E remove( int i ) { + range_check(i); + E e = _es[i]; + System.arraycopy(_es,i+1,_es,i,(--_len)-i); + return e; + } + + /** Remove all elements */ + public void clear( ) { _len=0; } + + /** Set existing element + * @param i element to set + * @param e value to set + * @return old value + * @exception AIOOBE if !(0 <= i < _len) + */ + public E set( int i, E e ) { + range_check(i); + E old = _es[i]; + _es[i] = e; + return old; + } + + public Ary setLen( int len ) { + if( len > _len ) + while( len>= _es.length ) _es = Arrays.copyOf(_es,_es.length<<1); + _len = len; + while( _es.length > (len<<1) ) // Shrink if hugely too large + _es = Arrays.copyOf(_es,_es.length>>1); + Arrays.fill(_es,len,_es.length,null); + return this; + } + + /** @param c Collection to be added */ + public boolean addAll( Collection c ) { + if( c==null ) return false; + for( E e : c ) add(e); + return true; + } + public Ary addAll( Iterable it ) { for( E e : it ) add(e); return this; } + + /** @param es Array to be added */ + public Ary addAll( F[] es ) { + if( es==null || es.length==0 ) return this; + while( _len+es.length > _es.length ) _es = Arrays.copyOf(_es,_es.length<<1); + System.arraycopy(es,0,_es,_len,es.length); + _len += es.length; + return this; + } + + /** @param c Collection to be added */ + public Ary addAll( Ary c ) { + if( c==null || c._len==0 ) return this; + while( _len+c._len > _es.length ) _es = Arrays.copyOf(_es,_es.length<<1); + System.arraycopy(c._es,0,_es,_len,c._len); + _len += c._len; + return this; + } + + /** @return compact array version */ + public E[] asAry() { return Arrays.copyOf(_es,_len); } + + /** Sorts in-place + * @param c Comparator to sort by */ + public void sort_update(Comparator c ) { Arrays.sort(_es, 0, _len, c); } + + /** Find the first matching element using ==, or -1 if none. Note that + * most del calls shuffle the list, so the first element might be random. + * @param e Element to find + * @return index of first matching element, or -1 if none */ + public int find( E e ) { + for( int i=0; i<_len; i++ ) if( _es[i]==e ) return i; + return -1; + } + + + /** @return an iterator */ + @Override public Iterator iterator() { return new Iter(); } + private class Iter implements Iterator { + int _i=0; + @Override public boolean hasNext() { return _i<_len; } + @Override public E next() { return _es[_i++]; } + } + + @Override public String toString() { + StringBuilder sb = new StringBuilder().append('{'); + for( int i=0; i<_len; i++ ) { + if( i>0 ) sb.append(','); + if( _es[i] != null ) sb.append(_es[i].toString()); + } + return sb.append('}').toString(); + } + + private void range_check( int i ) { + if( i < 0 || i>=_len ) + throw new ArrayIndexOutOfBoundsException(""+i+" >= "+_len); + } + + // Note that the hashCode() and equals() are not invariant to changes in the + // underlying array. If the hashCode() is used (e.g., inserting into a + // HashMap) and the then the array changes, the hashCode() will change also. + @Override public boolean equals( Object o ) { + if( this==o ) return true; + if( !(o instanceof Ary ary) ) return false; + if( _len != ary._len ) return false; + if( _es == ary._es ) return true; + for( int i=0; i<_len; i++ ) + if( !(_es[i]==null ? (ary._es[i] == null) : _es[i].equals(ary._es[i])) ) + return false; + return true; + } + @Override public int hashCode( ) { + int sum=_len; + for( int i=0; i<_len; i++ ) + sum += _es[i]==null ? 0 : _es[i].hashCode(); + return sum; + } + +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java b/chapter18/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java new file mode 100644 index 00000000..9698abd9 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java @@ -0,0 +1,228 @@ + package com.seaofnodes.simple; + +import com.seaofnodes.simple.IterPeeps.WorkList; +import com.seaofnodes.simple.node.*; +import com.seaofnodes.simple.type.*; +import java.util.*; + +public abstract class GlobalCodeMotion { + + // Arrange that the existing isCFG() Nodes form a valid CFG. The + // Node.use(0) is always a block tail (either IfNode or head of the + // following block). There are no unreachable infinite loops. + public static void buildCFG( StopNode stop ) { + schedEarly(); + Parser.SCHEDULED = true; + schedLate( stop); + } + + + // ------------------------------------------------------------------------ + // Visit all nodes in CFG Reverse Post-Order, essentially defs before uses + // (except at loops). Since defs are visited first - and hoisted as early + // as possible, when we come to a use we place it just after its deepest + // input. + private static void schedEarly() { + ArrayList rpo = new ArrayList<>(); + BitSet visit = new BitSet(); + _rpo_cfg(Parser.START, visit, rpo); + // Reverse Post-Order on CFG + for( int j=rpo.size()-1; j>=0; j-- ) { + CFGNode cfg = rpo.get(j); + cfg.loopDepth(); + for( Node n : cfg._inputs ) + _schedEarly(n,visit); + if( cfg instanceof RegionNode ) + for( Node phi : cfg._outputs ) + if( phi instanceof PhiNode ) + _schedEarly(phi,visit); + } + } + + // Post-Order of CFG + private static void _rpo_cfg(Node n, BitSet visit, ArrayList rpo) { + if( !(n instanceof CFGNode cfg) || visit.get(cfg._nid) ) + return; // Been there, done that + visit.set(cfg._nid); + for( Node use : cfg._outputs ) + _rpo_cfg(use,visit,rpo); + rpo.add(cfg); + } + + private static void _schedEarly(Node n, BitSet visit) { + if( n==null || visit.get(n._nid) ) return; // Been there, done that + assert !(n instanceof CFGNode); + visit.set(n._nid); + // Schedule not-pinned not-CFG inputs before self. Since skipping + // Pinned, this never walks the backedge of Phis (and thus spins around + // a data-only loop, eventually attempting relying on some pre-visited- + // not-post-visited data op with no scheduled control. + for( Node def : n._inputs ) + if( def!=null && !(def instanceof PhiNode) ) + _schedEarly(def,visit); + // If not-pinned (e.g. constants, projections, phi) and not-CFG + if( !n.isPinned() ) { + // Schedule at deepest input + CFGNode early = Parser.START; // Maximally early, lowest idepth + if( n.in(0) instanceof CFGNode cfg ) early = cfg; + for( int i=1; i early.idepth() ) + early = n.in(i).cfg0(); // Latest/deepest input + n.setDef(0,early); // First place this can go + } + } + + // ------------------------------------------------------------------------ + private static void schedLate( StopNode stop) { + CFGNode[] late = new CFGNode[Node.UID()]; + Node[] ns = new Node[Node.UID()]; + // Breadth-first scheduling + breadth(stop,ns,late); + + // Copy the best placement choice into the control slot + for( int i=0; i work = new WorkList<>(); + work.push(stop); + Node n; + outer: + while( (n = work.pop()) != null ) { + assert late[n._nid]==null; // No double visit + // These I know the late schedule of, and need to set early for loops + if( n instanceof CFGNode cfg ) late[n._nid] = cfg.blockHead() ? cfg : cfg.cfg(0); + else if( n instanceof PhiNode phi ) late[n._nid] = phi.region(); + else if( n instanceof ProjNode && n.in(0) instanceof CFGNode cfg ) late[n._nid] = cfg; + else { + + // All uses done? + for( Node use : n._outputs ) + if( use!=null && late[use._nid]==null ) + 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 && + // 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)) ) + continue outer; + + // All uses done, schedule + _doSchedLate(n,ns,late); + } + + // Walk all inputs and put on worklist, as their last-use might now be done + for( Node def : n._inputs ) + if( def!=null && late[def._nid]==null ) { + work.push(def); + // if the def has a load use, maybe the load can fire + for( Node ld : def._outputs ) + if( ld instanceof LoadNode && late[ld._nid]==null ) + work.push(ld); + } + } + } + + private static void _doSchedLate(Node n, Node[] ns, CFGNode[] late) { + // Walk uses, gathering the LCA (Least Common Ancestor) of uses + CFGNode early = n.in(0) instanceof CFGNode cfg ? cfg : n.in(0).cfg0(); + assert early != null; + CFGNode lca = null; + for( Node use : n._outputs ) + if( use != null ) + lca = use_block(n,use, late)._idom(lca,null); + + // Loads may need anti-dependencies, raising their LCA + if( n instanceof LoadNode load ) + lca = find_anti_dep(lca,load,early,late); + + // Walk up from the LCA to the early, looking for best place. This is + // the lowest execution frequency, approximated by least loop depth and + // deepest control flow. + CFGNode best = lca; + lca = lca.idom(); // Already found best for starting LCA + for( ; lca != early.idom(); lca = lca.idom() ) + if( better(lca,best) ) + best = lca; + assert !(best instanceof IfNode); + ns [n._nid] = n; + late[n._nid] = best; + } + + // Block of use. Normally from late[] schedule, except for Phis, which go + // to the matching Region input. + private static CFGNode use_block(Node n, Node use, CFGNode[] late) { + if( !(use instanceof PhiNode phi) ) + return late[use._nid]; + CFGNode found=null; + for( int i=1; i best.idepth() || best instanceof IfNode); + } + + private static CFGNode find_anti_dep(CFGNode lca, LoadNode load, CFGNode early, CFGNode[] late) { + // We could skip final-field loads here. + // Walk LCA->early, flagging Load's block location choices + for( CFGNode cfg=lca; early!=null && cfg!=early.idom(); cfg = cfg.idom() ) + cfg._anti = load._nid; + // Walk load->mem uses, looking for Stores causing an anti-dep + for( Node mem : load.mem()._outputs ) { + switch( mem ) { + case StoreNode st: + assert late[st._nid]!=null; + lca = anti_dep(load,late[st._nid],st.cfg0(),lca,st); + break; + case NewNode st: + assert late[st._nid]!=null; + lca = anti_dep(load,late[st._nid],st.cfg0(),lca,st); + break; + case PhiNode phi: + // Repeat anti-dep for matching Phi inputs. + // No anti-dep edges but may raise the LCA. + for( int i=1; i xScopes) { + + // Since the graph has cycles, we need to create a flat list of all the + // nodes in the graph. + Collection all = findAll(stop, scope); + StringBuilder sb = new StringBuilder(); + sb.append("digraph chapter11 {\n"); + sb.append("/*\n"); + sb.append(stop._src); + sb.append("\n*/\n"); + + // To keep the Scopes below the graph and pointing up into the graph we + // need to group the Nodes in a subgraph cluster, and the scopes into a + // different subgraph cluster. THEN we can draw edges between the + // scopes and nodes. If we try to cross subgraph cluster borders while + // still making the subgraphs DOT gets confused. + sb.append("\trankdir=BT;\n"); // Force Nodes before Scopes + + // CNC - turned off Apr/8/2024, gives more flex in the layout and + // removes some of the more ludicrous layout choices. + // Preserve node input order + //sb.append("\tordering=\"in\";\n"); + + // Merge multiple edges hitting the same node. Makes common shared + // nodes much prettier to look at. + sb.append("\tconcentrate=\"true\";\n"); + + // Force nested scopes to order + sb.append("\tcompound=\"true\";\n"); + + // Just the Nodes first, in a cluster no edges + nodes(sb, all); + + // Now the scopes, in a cluster no edges + if( xScopes != null ) + for( ScopeNode xscope : xScopes ) + if( !xscope.isDead() ) + scope( sb, xscope ); + + // Walk the Node edges + nodeEdges(sb, all); + + // Walk the active Scope edges + if( xScopes != null ) + for( ScopeNode xscope : xScopes ) + if( !xscope.isDead() ) + scopeEdges( sb, xscope ); + + sb.append("}\n"); + return sb.toString(); + } + + private void nodesByCluster(StringBuilder sb, boolean doCtrl, Collection all) { + if (!_separateControlCluster && doCtrl) // all nodes in 1 cluster + return; + // Just the Nodes first, in a cluster no edges + sb.append(doCtrl ? "\tsubgraph cluster_Controls {\n" : "\tsubgraph cluster_Nodes {\n"); // Magic "cluster_" in the subgraph name + for (Node n : all) { + if (n instanceof ProjNode || n instanceof CProjNode || n instanceof ScopeMinNode || n==Parser.XCTRL) + continue; // Do not emit, rolled into MultiNode or Scope cluster already + if( _separateControlCluster && doCtrl && !(n instanceof CFGNode) ) continue; + if( _separateControlCluster && !doCtrl && (n instanceof CFGNode) ) continue; + sb.append("\t\t").append(n.uniqueName()).append(" [ "); + String lab = n.glabel(); + if (n instanceof MultiNode) { + // Make a box with the MultiNode on top, and all the projections on the bottom + sb.append("shape=plaintext label=<\n"); + sb.append("\t\t\t\n"); + sb.append("\t\t\t\n"); + sb.append("\t\t\t"); + boolean doProjTable = false; + n._outputs.sort((x,y) -> x instanceof ProjNode xp && y instanceof ProjNode yp ? (xp._idx - yp._idx) : (x._nid - y._nid)); + for (Node use : n._outputs) { + if (use instanceof ProjNode proj) { + if (!doProjTable) { + doProjTable = true; + sb.append("
").append(lab).append("
").append("\n"); + sb.append("\t\t\t\t").append("\n"); + sb.append("\t\t\t\t"); + } + sb.append(""); + } + if (use instanceof CProjNode proj) { + if (!doProjTable) { + doProjTable = true; + sb.append(""); + } + sb.append("\n"); + sb.append("\t\t\t
").append(proj.glabel()).append("").append("\n"); + sb.append("\t\t\t\t").append("\n"); + sb.append("\t\t\t\t"); + } + sb.append(""); + } + } + if (doProjTable) { + sb.append("").append("\n"); + sb.append("\t\t\t\t
").append(proj.glabel()).append("
").append("\n"); + sb.append("\t\t\t
>\n\t\t"); + + } else { + // control nodes have box shape + // other nodes are ellipses, i.e. default shape + if( n instanceof CFGNode ) sb.append("shape=box style=filled fillcolor=yellow "); + else if (n instanceof PhiNode) sb.append("style=filled fillcolor=lightyellow "); + sb.append("label=\"").append(lab).append("\" "); + } + sb.append("];\n"); + } + if (!_separateControlCluster) { + // Force Region & Phis to line up + for (Node n : all) { + if (n instanceof RegionNode region) { + sb.append("\t\t{ rank=same; "); + sb.append(region).append(";"); + for (Node phi : region._outputs) + if (phi instanceof PhiNode) sb.append(phi.uniqueName()).append(";"); + sb.append("}\n"); + } + } + } + sb.append("\t}\n"); // End Node cluster + } + + private void nodes(StringBuilder sb, Collection all) { + nodesByCluster(sb, true, all); + nodesByCluster(sb, false, all); + } + + // Build a nested scope display, walking the _prev edge + private void scope( StringBuilder sb, ScopeNode scope ) { + sb.append("\tnode [shape=plaintext];\n"); + int last = scope.nIns(); + int max = scope._lexSize.size(); + for( int i = 0; i < max; i++ ) { + int level = max-i-1; + String scopeName = makeScopeName(scope, level); + sb.append("\tsubgraph cluster_").append(scopeName).append(" {\n"); // Magic "cluster_" in the subgraph name + // Special for memory ScopeMinNode + if( level==0 ) { + ScopeMinNode n = scope.mem(); + sb.append("\t\t").append(n.uniqueName()).append(" [label=\"").append(n.glabel()).append("\"];\n"); + } + sb.append("\t\t").append(scopeName).append(" [label=<\n"); + sb.append("\t\t\t\n"); + // Add the scope level + sb.append("\t\t\t"); + int lexStart=scope._lexSize.at(level); + for( int j=lexStart; j").append(v._name).append(""); + } + last = lexStart; + sb.append("\n"); + sb.append("\t\t\t
").append(level).append("
>];\n"); + } + // Scope clusters nest, so the graphics shows the nested scopes, so + // they are not closed as they are printed; so they just keep nesting. + // We close them all at once here. + sb.append( "\t}\n".repeat( max ) ); // End all Scope clusters + } + + private String makeScopeName(ScopeNode sn, int level) { return sn.uniqueName() + "_" + level; } + private String makePortName(String scopeName, String varName) { return scopeName + "_" + varName; } + + // Walk the node edges + private void nodeEdges(StringBuilder sb, Collection all) { + // All them edge labels + sb.append("\tedge [ fontname=Helvetica, fontsize=8 ];\n"); + for( Node n : all ) { + // Do not display the Constant->Start edge; + // ProjNodes handled by Multi; + // ScopeNodes are done separately + if( n instanceof ConstantNode || n instanceof ProjNode || n instanceof CProjNode || n instanceof ScopeNode || n==Parser.XCTRL ) + continue; + for( int i=0; idef edge from Phi to Region. + sb.append('\t').append(n.uniqueName()); + sb.append(" -> "); + sb.append(def.uniqueName()); + sb.append(" [style=dotted taillabel=").append(i).append("];\n"); + } else if( def != null ) { + // Most edges land here use->def + sb.append('\t').append(n.uniqueName()).append(" -> "); + if( def instanceof CProjNode proj ) { + String mname = proj.ctrl().uniqueName(); + sb.append(mname).append(":p").append(proj._idx); + } else if( def instanceof ProjNode proj ) { + String mname = proj.in(0).uniqueName(); + sb.append(mname).append(":p").append(proj._idx); + } else sb.append(def.uniqueName()); + // Number edges, so we can see how they track + sb.append("[taillabel=").append(i); + // The edge from New to ctrl is just for anchoring the New + if ( n instanceof NewNode ) + sb.append(" color=green"); + // control edges are colored red + else if( def instanceof CFGNode ) + sb.append(" color=red"); + else if( def.isMem() ) + sb.append(" color=blue"); + // Backedges do not add a ranking constraint + if( i==2 && (n instanceof PhiNode || n instanceof LoopNode) ) + sb.append(" constraint=false"); + sb.append("];\n"); + } + } + } + } + + // Walk the scope edges + private void scopeEdges( StringBuilder sb, ScopeNode scope ) { + sb.append("\tedge [style=dashed color=cornflowerblue];\n"); + int level=0; + for( var v : scope._vars ) { + Node def = scope.in(v._idx); + while( def instanceof ScopeNode lazy ) + def = lazy.in(v._idx); + if( def==null ) continue; + while( level < scope._lexSize.size() && v._idx >= scope._lexSize.at(level) ) + level++; + String scopeName = makeScopeName(scope, level-1); + sb.append("\t") + .append(scopeName).append(":") + .append('"').append(makePortName(scopeName, v._name)).append('"') // wrap port name with quotes because $ctrl is not valid unquoted + .append(" -> "); + if( def instanceof CProjNode proj ) { + sb.append(def.in(0).uniqueName()).append(":p").append(proj._idx); + } else if( def instanceof ProjNode proj ) { + sb.append(def.in(0).uniqueName()).append(":p").append(proj._idx); + } else sb.append(def.uniqueName()); + sb.append(";\n"); + } + } + + /** + * Finds all nodes in the graph. + */ + private Collection findAll(Node stop, Node scope) { + final HashMap all = new HashMap<>(); + for( Node n : stop._inputs ) + walk(all, n); + if( scope != null ) + for( Node n : scope._inputs ) + walk(all, n); + return all.values(); + } + + /** + * Walk a subgraph and populate distinct nodes in the all list. + */ + private void walk(HashMap all, Node n) { + if(n == null ) return; + if (all.get(n._nid) != null) return; // Been there, done that + all.put(n._nid, n); + for (Node c : n._inputs) + walk(all, c); + for( Node c : n._outputs ) + walk(all, c); + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/IRPrinter.java b/chapter18/src/main/java/com/seaofnodes/simple/IRPrinter.java new file mode 100644 index 00000000..7ef55a3f --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/IRPrinter.java @@ -0,0 +1,286 @@ +package com.seaofnodes.simple; + +import com.seaofnodes.simple.node.*; +import java.lang.StringBuilder; +import java.util.ArrayList; +import java.util.BitSet; +import java.util.HashMap; + +public class IRPrinter { + + // Print a node on 1 line, columnar aligned, as: + // NNID NNAME DDEF DDEF [[ UUSE UUSE ]] TYPE + // 1234 sssss 1234 1234 1234 1234 1234 1234 tttttt + public static void _printLine( Node n, StringBuilder sb ) { + sb.append("%4d %-7.7s ".formatted(n._nid,n.label())); + if( n._inputs==null ) { + sb.append("DEAD\n"); + return; + } + for( Node def : n._inputs ) + sb.append(def==null ? "____ " : "%4d ".formatted(def._nid)); + for( int i = n._inputs.size(); i<4; i++ ) + sb.append(" "); + sb.append(" [[ "); + for( Node use : n._outputs ) + sb.append(use==null ? "____ " : "%4d ".formatted(use._nid)); + int lim = 6 - Math.max(n._inputs.size(),4); + for( int i = n._outputs.size(); i 0) + sb.append(", "); + if (def == null) sb.append("_"); + else nodeId(sb, def); + } + sb.append(")").append("\n"); + } + + public static void printLine( Node n, StringBuilder sb, boolean llvmFormat ) { + if (llvmFormat) _printLineLlvmFormat( n, sb ); + else _printLine ( n, sb ); + } + + public static String prettyPrint(Node node, int depth) { + return Parser.SCHEDULED + ? prettyPrintScheduled( node, depth, false ) + : prettyPrint( node, depth, false ); + } + + // Another bulk pretty-printer. Makes more effort at basic-block grouping. + public static String prettyPrint( Node node, int depth, boolean llvmFormat ) { + // First, a Breadth First Search at a fixed depth. + BFS bfs = new BFS(node,depth); + // Convert just that set to a post-order + ArrayList rpos = new ArrayList<>(); + BitSet visit = new BitSet(); + for( int i=bfs._lim; i< bfs._bfs.size(); i++ ) + postOrd( bfs._bfs.get(i), rpos, visit, bfs._bs); + // Reverse the post-order walk + StringBuilder sb = new StringBuilder(); + boolean gap=false; + for( int i=rpos.size()-1; i>=0; i-- ) { + Node n = rpos.get(i); + if( n instanceof CFGNode || n.isMultiHead() ) { + if( !gap ) sb.append("\n"); // Blank before multihead + printLine( n, sb, llvmFormat ); // Print head + while( --i >= 0 ) { + Node t = rpos.get(i); + if( !t.isMultiTail() ) { i++; break; } + printLine( t, sb, llvmFormat ); + } + sb.append("\n"); // Blank after multitail + gap = true; + } else { + printLine( n, sb, llvmFormat ); + gap = false; + } + } + return sb.toString(); + } + + private static void postOrd(Node n, ArrayList rpos, BitSet visit, BitSet bfs) { + if( !bfs.get(n._nid) ) + return; // Not in the BFS visit + if( visit.get(n._nid) ) return; // Already post-order walked + visit.set(n._nid); + // First walk the CFG, then everything + if( n instanceof CFGNode ) { + for( Node use : n._outputs ) + if( use instanceof CFGNode && use.nOuts() >= 1 && !(use._outputs.get( 0 ) instanceof LoopNode) ) + postOrd(use, rpos,visit,bfs); + for( Node use : n._outputs ) + if( use instanceof CFGNode ) + postOrd(use,rpos,visit,bfs); + } + for( Node use : n._outputs ) + if( use != null ) + postOrd(use, rpos,visit,bfs); + // Post-order + rpos.add(n); + } + + // Breadth-first search, broken out in a class to keep in more independent. + // Maintains a root-set of Nodes at the limit (or past by 1 if MultiHead). + public static class BFS { + // A breadth first search, plus MultiHeads for any MultiTails + public final ArrayList _bfs; + public final BitSet _bs; // Visited members by node id + public final int _depth; // Depth limit + public final int _lim; // From here to _bfs._len can be roots for a reverse search + public BFS( Node base, int d ) { + _depth = d; + _bfs = new ArrayList<>(); + _bs = new BitSet(); + + add(base); // Prime the pump + int idx=0, lim=1; // Limit is where depth counter changes + while( idx < _bfs.size() ) { // Ran out of nodes below depth + Node n = _bfs.get(idx++); + for( Node def : n._inputs ) + if( def!=null && !_bs.get(def._nid) ) + add(def); + if( idx==lim ) { // Depth counter changes at limit + if( --d < 0 ) + break; // Ran out of depth + lim = _bfs.size(); // New depth limit + } + } + // Toss things past the limit except multi-heads + while( idx < _bfs.size() ) { + Node n = _bfs.get(idx); + if( n.isMultiHead() ) idx++; + else del(idx); + } + // Root set is any node with no inputs in the visited set + lim = _bfs.size(); + for( int i=_bfs.size()-1; i>=0; i-- ) + if( !any_visited(_bfs.get(i)) ) + swap( i,--lim); + _lim = lim; + } + void swap( int x, int y ) { + if( x==y ) return; + Node tx = _bfs.get(x); + Node ty = _bfs.get(y); + _bfs.set(x,ty); + _bfs.set(y,tx); + } + void add(Node n) { + _bfs.add(n); + _bs.set(n._nid); + } + void del(int idx) { + _bs.clear(_bfs.get(idx)._nid); + Utils.del(_bfs, idx); + } + boolean any_visited( Node n ) { + for( Node def : n._inputs ) + if( def!=null && _bs.get(def._nid) ) + return true; + return false; + } + } + + // Bulk pretty printer, knowing scheduling information is available + public static String prettyPrintScheduled( Node node, int depth, boolean llvmFormat ) { + // Backwards DFS walk to depth. + HashMap ds = new HashMap<>(); + ArrayList ns = new ArrayList<>(); + _walk(ds,ns,node,depth); + // Remove data projections, these are force-printed behind their multinode head + for( int i=0; i bns = new ArrayList<>(); + while( !ds.isEmpty() ) { + CFGNode blk = null; + for( Node n : ns ) { + CFGNode cfg = n instanceof CFGNode cfg0 && cfg0.blockHead() ? cfg0 : n.cfg0(); + if( blk==null || cfg.idepth() < blk.idepth() ) + blk = cfg; + } + ds.remove(blk._nid); + ns.remove(blk); + + // Print block header + sb.append("%-13.13s".formatted(label(blk)+":")); + sb.append( " ".repeat(4) ).append(" [[ "); + if( blk instanceof StartNode ) ; + else if( blk instanceof RegionNode || blk instanceof StopNode ) + for( int i=(blk instanceof StopNode ? 0 : 1); i ds, ArrayList ns, Node node, int d ) { + Integer nd = ds.get(node._nid); + if( nd!=null && d <= nd ) return; // Been there, done that + if( ds.put(node._nid,d) == null ) + ns.add(node); + if( d == 0 ) return; // Depth cutoff + for( Node def : node._inputs ) + if( def != null ) + _walk(ds,ns,def,d-1); + } + + static String label( CFGNode blk ) { + if( blk instanceof StartNode ) return "START"; + return (blk instanceof LoopNode ? "LOOP" : "L")+blk._nid; + } + static void label( StringBuilder sb, CFGNode blk ) { + if( !blk.blockHead() ) blk = blk.cfg(0); + sb.append( "%-9.9s ".formatted( label( blk ) ) ); + } + static void printLine( Node n, StringBuilder sb, boolean llvmFormat, ArrayList bns, int i, HashMap ds, ArrayList ns ) { + printLine( n, sb, llvmFormat ); + if( i != -1 ) Utils.del(bns,i); + ds.remove(n._nid); + ns.remove(n); + } + +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/IterPeeps.java b/chapter18/src/main/java/com/seaofnodes/simple/IterPeeps.java new file mode 100644 index 00000000..36f053f9 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/IterPeeps.java @@ -0,0 +1,205 @@ +package com.seaofnodes.simple; + +import com.seaofnodes.simple.node.*; + +import java.util.Arrays; +import java.util.BitSet; +import java.util.Random; + +/** + * The IterPeeps runs after parsing. It iterates the peepholes to a fixed point + * so that no more peepholes apply. This should be linear because peepholes rarely + * (never?) increase code size. The graph should monotonically reduce in some + * dimension, which is usually size. It might also reduce in e.g. number of + * MulNodes or Load/Store nodes, swapping out more "expensive" Nodes for cheaper + * ones. + *
+ * The theoretical overall worklist is mindless just grabbing the next thing and + * doing it. If the graph changes, put the neighbors on the worklist. + *
+ * Lather, Rinse, Repeat until the worklist runs dry. + *

+ * The main issues we have to deal with: + * + *

    + *
  • Nodes have uses; replacing some set of Nodes with another requires more graph + * reworking. Not rocket science, but it can be fiddly. Its helpful to have a + * small set of graph munging utilities, and the strong invariant that the graph + * is stable and correct between peepholes. In our case `Node.subsume` does + * most of the munging, building on our prior stable Node utilities.
    • + * + *
  • Changing a Node also changes the graph "neighborhood". The neighbors need to + * be checked to see if THEY can also peephole, and so on. After any peephole + * or graph update we put a Nodes uses and defs on the worklist.
    • + * + *
  • Our strong invariant is that for all Nodes, either they are on the worklist + * OR no peephole applies. This invariant is easy to check, although expensive. + * Basically the normal "iterate peepholes to a fixed point" is linear, and this + * check is linear at each peephole step... so quadratic overall. Its a useful + * assert, but one we can disable once the overall algorithm is stable - and + * then turn it back on again when some new set of peepholes is misbehaving. + * The code for this is turned on in `IterPeeps.iterate` as `assert + * progressOnList(stop);`
    • + *
+ */ +public abstract class IterPeeps { + + private static final WorkList WORK = new WorkList<>(); + + public static N add( N n ) { + return (N)WORK.push(n); + } + + public static void addAll( Ary ary ) { + WORK.addAll(ary); + } + + /** + * Iterate peepholes to a fixed point + */ + public static StopNode iterate(StopNode stop) { + assert progressOnList(stop); + int cnt=0; + + Node n; + while( (n=WORK.pop()) != null ) { + if( n.isDead() ) continue; + cnt++; // Useful for debugging, searching which peephole broke things + Node x = n.peepholeOpt(); + if( x != null ) { + if( x.isDead() ) continue; + // peepholeOpt can return brand-new nodes, needing an initial type set + if( x._type==null ) x.setType(x.compute()); + // Changes require neighbors onto the worklist + if( x != n || !(x instanceof ConstantNode) ) { + // All outputs of n (changing node) not x (prior existing node). + for( Node z : n._outputs ) WORK.push(z); + // Everybody gets a free "go again" in case they didn't get + // made in their final form. + WORK.push(x); + // If the result is not self, revisit all inputs (because + // there's a new user), and replace in the graph. + if( x != n ) { + for( Node z : n. _inputs ) WORK.push(z); + n.subsume(x); + } + } + // If there are distant neighbors, move to worklist + n.moveDepsToWorklist(); + assert progressOnList(stop); // Very expensive assert + } + if( n.isUnused() && !(n instanceof StopNode) ) + n.kill(); // Just plain dead + } + + return stop; + } + + // Visit ALL nodes and confirm the invariant: + // Either you are on the WORK worklist OR running `iter()` makes no progress. + + // This invariant ensures that no progress is missed, i.e., when the + // worklist is empty we have indeed done all that can be done. To help + // with debugging, the {@code assert} is broken out in a place where it is easy to + // stop if a change is found. + + // Also, the normal usage of `iter()` may attempt peepholes with distance + // neighbors and these should fail, but will then try to add dependencies + // {@link #Node.addDep} which is a side effect in an assert. The {@link + // #midAssert} is used to stop this side effect. + private static boolean MID_ASSERT; + public static boolean midAssert() { return MID_ASSERT; } + private static boolean progressOnList(Node stop) { + MID_ASSERT = true; + int old_cnt = Node.ITER_CNT, old_nop = Node.ITER_NOP_CNT; + Node changed = stop.walk( n -> { + Node m = n; + if( n.compute().isa(n._type) && (!n.iskeep() || n._nid<=5) ) { // Types must be forwards, even if on worklist + if( WORK.on(n) ) return null; + m = n.peepholeOpt(); + if( m==null ) return null; + } + System.err.println("BREAK HERE FOR BUG"); + return m; + }); + Node.ITER_CNT = old_cnt; Node.ITER_NOP_CNT = old_nop; + MID_ASSERT = false; + return changed==null; + } + + public static void reset() { + WORK.clear(); + } + + /** + * Classic WorkList, with a fast add/remove, dup removal, random pull. + * The Node's nid is used to check membership in the worklist. + */ + public static class WorkList { + + private Node[] _es; + private int _len; + private final BitSet _on; // Bit set if Node._nid is on WorkList + private final Random _R; // For randomizing pull from the WorkList + private final long _seed; + + /* Useful stat - how many nodes are processed in the post parse iterative opt */ + private long _totalWork = 0; + + WorkList() { this(123); } + WorkList(long seed) { + _es = new Node[1]; + _len=0; + _on = new BitSet(); + _seed = seed; + _R = new Random(); + _R.setSeed(_seed); + } + + /** + * Pushes a Node on the WorkList, ensuring no duplicates + * If Node is null it will not be added. + */ + public E push( E x ) { + if( x==null ) return null; + int idx = x._nid; + if( !_on.get(idx) ) { + _on.set(idx); + if( _len==_es.length ) + _es = Arrays.copyOf(_es,_len<<1); + _es[_len++] = x; + _totalWork++; + } + return x; + } + + public void addAll( Ary ary ) { + for( E n : ary ) + push(n); + } + + /** + * True if Node is on the WorkList + */ + boolean on( E x ) { return _on.get(x._nid); } + + /** + * Removes a random Node from the WorkList; null if WorkList is empty + */ + E pop() { + if( _len == 0 ) return null; + int idx = _R.nextInt(_len); + E x = (E)_es[idx]; + _es[idx] = _es[--_len]; // Compress array + _on.clear(x._nid); + return x; + } + + public void clear() { + _len = 0; + _on.clear(); + _R.setSeed(_seed); + _totalWork = 0; + } + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/Parser.java b/chapter18/src/main/java/com/seaofnodes/simple/Parser.java new file mode 100644 index 00000000..1cdf27fb --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/Parser.java @@ -0,0 +1,1427 @@ +package com.seaofnodes.simple; + +import com.seaofnodes.simple.node.*; +import com.seaofnodes.simple.type.*; + +import java.util.*; + +/** + * The Parser converts a Simple source program to the Sea of Nodes intermediate + * representation directly in one pass. There is no intermediate Abstract + * Syntax Tree structure. + *

+ * This is a simple recursive descent parser. All lexical analysis is done here as well. + */ +public class Parser { + + /** + * A Global Static, unique to each compilation. This is a public, so we + * can make constants everywhere without having to thread the StartNode + * through the entire parser and optimizer. + *

+ * To make the compiler multithreaded, this field will have to move into a TLS. + */ + public static StartNode START; + + public static ConstantNode ZERO; // Very common node, cached here + public static XCtrlNode XCTRL; // Very common node, cached here + + // Next available memory alias number + static int ALIAS; + + public StopNode STOP; + + // Debugger Printing. + public static boolean SCHEDULED; // True if debug printer can use schedule info + + // The Lexer. Thin wrapper over a byte[] buffer with a cursor. + private final Lexer _lexer; + + /** + * Current ScopeNode - ScopeNodes change as we parse code, but at any point of time + * there is one current ScopeNode. The reason the current ScopeNode can change is to do with how + * we handle branching. See {@link #parseIf()}. + *

+ * Each ScopeNode contains a stack of lexical scopes, each scope is a symbol table that binds + * variable names to Nodes. The top of this stack represents current scope. + *

+ * We keep a list of all ScopeNodes so that we can show them in graphs. + * @see #parseIf() + * @see #_xScopes + */ + public ScopeNode _scope; + + /** + * List of keywords disallowed as identifiers + */ + private static final HashSet KEYWORDS = new HashSet<>(){{ + add("bool"); + add("break"); + add("byte"); + add("continue"); + add("else"); + add("f32"); + add("f64"); + add("false"); + add("flt"); + add("i16"); + add("i32"); + add("i64"); + add("i8"); + add("if"); + add("int"); + add("new"); + add("null"); + add("return"); + add("struct"); + add("true"); + add("u1"); + add("u16"); + add("u32"); + add("u8"); + add("while"); + }}; + + + /** + * We clone ScopeNodes when control flows branch; it is useful to have + * a list of all active ScopeNodes for purposes of visualization of the SoN graph + */ + public final Stack _xScopes = new Stack<>(); + + ScopeNode _continueScope; + ScopeNode _breakScope; // Merge all the while-breaks here + + // Mapping from a type name to a Type. The string name matches + // `type.str()` call. No TypeMemPtrs are in here, because Simple does not + // have C-style '*ptr' references. + public static HashMap TYPES; + + // Mapping from a type name to the constructor for a Type. + public final HashMap INITS; + + + public Parser(String source, TypeInteger arg) { + Node.reset(); + IterPeeps.reset(); + SCHEDULED = false; + _lexer = new Lexer(source); + _scope = new ScopeNode(); + _continueScope = _breakScope = null; + START = new StartNode(arg); + STOP = new StopNode(source); + ZERO = con(0).keep(); + XCTRL= new XCtrlNode().peephole().keep(); + ALIAS = 2; // alias 0 for the control, 1 for memory + TYPES = defaultTypes(); + INITS = new HashMap<>(); + } + + public Parser(String source) { + this(source, TypeInteger.BOT); + } + + @Override + public String toString() { return _lexer.toString(); } + + public static HashMap defaultTypes() { + return new HashMap<>() {{ + put("bool",TypeInteger.U1 ); + put("byte",TypeInteger.U8 ); + put("f32" ,TypeFloat .B32); + put("f64" ,TypeFloat .BOT); + put("flt" ,TypeFloat .BOT); + put("i16" ,TypeInteger.I16); + put("i32" ,TypeInteger.I32); + put("i64" ,TypeInteger.BOT); + put("i8" ,TypeInteger.I8 ); + put("int" ,TypeInteger.BOT); + put("u1" ,TypeInteger.U1 ); + put("u16" ,TypeInteger.U16); + put("u32" ,TypeInteger.U32); + put("u8" ,TypeInteger.U8 ); + put("val" ,Type.TOP); // Marker type, indicates type inference + put("var" ,Type.BOTTOM); // Marker type, indicates type inference + }}; + } + + // Debugging utility to find a Node by index + public static Node find(int nid) { return START.find(nid); } + + private Node ctrl() { return _scope.ctrl(); } + private N ctrl(N n) { return _scope.ctrl(n); } + + public StopNode parse() { return parse(false); } + public StopNode parse(boolean show) { + _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.ARG0, TypeInteger.BOT, false, new ProjNode(START, 2, ScopeNode.ARG0).peephole()); + + // Parse whole program + parseBlock(false); + + if( ctrl()._type==Type.CONTROL ) + STOP.addReturn(new ReturnNode(ctrl(), ZERO, _scope).peephole()); + _scope.kill(); + _xScopes.pop(); + for( StructNode init : INITS.values() ) + init.unkeep().kill(); + INITS.clear(); + if (!_lexer.isEOF()) throw error("Syntax error, unexpected " + _lexer.getAnyNextToken()); + STOP.peephole(); + if( show ) showGraph(); + return STOP; + } + + /** + * Parses a block + * + * 

+     *     '{' statements '}'
+     *
+ * Does not parse the opening or closing '{}' + * @return a {@link Node} or {@code null} + */ + private Node parseBlock(boolean inCon) { + // Enter a new scope + _scope.push(inCon); + while (!peek('}') && !_lexer.isEOF()) + parseStatement(); + // Exit scope + _scope.pop(); + return null; + } + + /** + * Parses a statement + * + * 
+     *     returnStatement | declStatement | blockStatement | ifStatement | expressionStatement
+     *
+ * @return a {@link Node} or {@code null} + */ + private Node parseStatement() { + if( false ) return null; + else if (matchx("return") ) return parseReturn(); + else if (match ("{") ) return require(parseBlock(false),"}"); + else if (matchx("if") ) return parseIf(); + else if (matchx("while") ) return parseWhile(); + else if (matchx("for") ) return parseFor(); + else if (matchx("break") ) return parseBreak(); + else if (matchx("continue")) return parseContinue(); + else if (matchx("struct") ) return parseStruct(); + else if (matchx("#showGraph")) return require(showGraph(),";"); + else if (matchx(";") ) return null; // Empty statement + // Declaration or normal assignment/expression + else return parseDeclarationStatement(); + } + + /** + * Parses a while statement + * + * 
+     *     while ( expression ) statement
+     *
+ * @return a {@link Node}, never {@code null} + */ + private Node parseWhile() { + require("("); + return parseLooping(false); + } + + + /** + * Parses a for statement + * + * 
+     *     for( var x=init; test; incr ) body
+     *
+ * @return a {@link Node}, never {@code null} + */ + private Node parseFor() { + // { var x=init,y=init,...; + // while( pred ) { + // body; + // next; + // } + // } + require("("); + _scope.push(); // Scope for the index variables + if( !match(";") ) // Can be empty init "for(;test;next) body" + parseDeclarationStatement(); // Non-empty init + Node rez = parseLooping(true); + _scope.pop(); // Exit index variable scope + return rez; + } + + // Shared by `for` and `while` + private Node parseLooping( boolean doFor ) { + + var savedContinueScope = _continueScope; + var savedBreakScope = _breakScope; + + // Loop region has two control inputs, the first is the entry + // point, and second is back edge that is set after loop is parsed + // (see end_loop() call below). Note that the absence of back edge is + // used as an indicator to switch off peepholes of the region and + // associated phis; see {@code inProgress()}. + + ctrl(new LoopNode(ctrl()).peephole()); // Note we set back edge to null here + + // At loop head, we clone the current Scope (this includes all + // names in every nesting level within the Scope). + // We create phis eagerly for all the names we find, see dup(). + + // Save the current scope as the loop head + ScopeNode head = _scope.keep(); + // Clone the head Scope to create a new Scope for the body. + // Create phis eagerly as part of cloning + _xScopes.push(_scope = _scope.dup(true)); // The true argument triggers creating phis + + // Parse predicate + var pred = peek(';') ? con(1) : parseAsgn(); + require( doFor ? ";" : ")" ); + + // IfNode takes current control and predicate + Node ifNode = new IfNode(ctrl(), pred.keep()).peephole(); + // Setup projection nodes + Node ifT = new CProjNode(ifNode. keep(), 0, "True" ).peephole().keep(); + Node ifF = new CProjNode(ifNode.unkeep(), 1, "False").peephole(); + + // for( ;;next ) body + int nextPos = -1, nextEnd = -1; + if( doFor ) { + // Skip the next expression and parse it later + nextPos = pos(); + skipAsgn(); + nextEnd = pos(); + require(")"); + } + + // Clone the body Scope to create the break/exit Scope which accounts for any + // side effects in the predicate. The break/exit Scope will be the final + // scope after the loop, and its control input is the False branch of + // the loop predicate. Note that body Scope is still our current scope. + ctrl(ifF); + _xScopes.push(_breakScope = _scope.dup()); + _breakScope.addGuards(ifF,pred,true); // Up-cast predicate + + // No continues yet + _continueScope = null; + + // Parse the true side, which corresponds to loop body + // Our current scope is the body Scope + ctrl(ifT.unkeep()); // set ctrl token to ifTrue projection + _scope.addGuards(ifT,pred.unkeep(),false); // Up-cast predicate + parseStatement(); // Parse loop body + _scope.removeGuards(ifT); + + // Merge the loop bottom into other continue statements + if (_continueScope != null) { + _continueScope = jumpTo(_continueScope); + _scope.kill(); + _scope = _continueScope; + } + + // Now append the next code onto the body code + if( doFor ) { + int old = pos(nextPos); + if( !peek(')') ) + parseAsgn(); + if( pos() != nextEnd ) + throw errorSyntax( "Unexpected code after expression" ); + pos(old); + } + + // The true branch loops back, so whatever is current _scope.ctrl gets + // added to head loop as input. endLoop() updates the head scope, and + // goes through all the phis that were created earlier. For each phi, + // it sets the second input to the corresponding input from the back + // edge. If the phi is redundant, it is replaced by its sole input. + var exit = _breakScope; + head.endLoop(_scope, exit); + head.unkeep().kill(); + + _xScopes.pop(); // Cleanup + _xScopes.pop(); // Cleanup + + _continueScope = savedContinueScope; + _breakScope = savedBreakScope; + + // At exit the false control is the current control, and + // the scope is the exit scope after the exit test. + _xScopes.pop(); + _xScopes.push(exit); + _scope = exit; + return _scope; + } + + private ScopeNode jumpTo(ScopeNode toScope) { + ScopeNode cur = _scope.dup(); + ctrl(XCTRL); // Kill current scope + // Prune nested lexical scopes that have depth > than the loop head + // We use _breakScope as a proxy for the loop head scope to obtain the depth + while( cur._lexSize.size() > _breakScope._lexSize.size() ) + cur.pop(); + // If this is a continue then first time the target is null + // So we just use the pruned current scope as the base for the + // "continue" + if( toScope == null ) + return cur; + // toScope is either the break scope, or a scope that was created here + assert toScope._lexSize.size() <= _breakScope._lexSize.size(); + toScope.ctrl(toScope.mergeScopes(cur).peephole()); + return toScope; + } + + private void checkLoopActive() { if (_breakScope == null) throw Parser.error("No active loop for a break or continue"); } + + private Node parseContinue() { checkLoopActive(); return (_continueScope = require(jumpTo( _continueScope ),";")); } + private Node parseBreak () { + checkLoopActive(); + // At the time of the break, and loop-exit conditions are only valid if + // they are ALSO valid at the break. It is the intersection of + // conditions here, not the union. + _breakScope.removeGuards(_breakScope.ctrl()); + _breakScope = require(jumpTo(_breakScope ),";"); + _breakScope.addGuards(_breakScope.ctrl(), null, false); + return _breakScope; + } + + // Look for an unbalanced `)`, skipping balanced + private Type skipAsgn() { + int paren=0; + while( true ) + // Next X char handles skipping complex comments + switch( _lexer.nextXChar() ) { + case Character.MAX_VALUE: + throw Utils.TODO(); + case ')': + if( --paren<0 ) + return posT(pos()-1); // Leave the `)` behind + break; + case '(': paren ++; break; + default: break; + } + } + + + /** + * Parses a statement + * + * 
+     *     if ( expression ) statement [else statement]
+     *
+ * @return a {@link Node}, never {@code null} + */ + private Node parseIf() { + // Parse predicate + require("("); + var pred = require(parseAsgn(), ")"); + return parseTrinary(pred,true,"else"); + } + + // Parse a conditional expression, merging results. + private Node parseTrinary( Node pred, boolean stmt, String fside ) { + pred.keep(); + + // IfNode takes current control and predicate + Node ifNode = new IfNode(ctrl(), pred).peephole(); + // Setup projection nodes + Node ifT = new CProjNode(ifNode. keep(), 0, "True" ).peephole().keep(); + Node ifF = new CProjNode(ifNode.unkeep(), 1, "False").peephole().keep(); + // In if true branch, the ifT proj node becomes the ctrl + // But first clone the scope and set it as current + int ndefs = _scope.nIns(); + ScopeNode fScope = _scope.dup(); // Duplicate current scope + _xScopes.push(fScope); // For graph visualization we need all scopes + + // Parse the true side + ctrl(ifT.unkeep()); // set ctrl token to ifTrue projection + _scope.addGuards(ifT,pred,false); // Up-cast predicate + Node lhs = stmt ? parseStatement() : parseAsgn().keep(); // Parse true-side + _scope.removeGuards(ifT); + + ScopeNode tScope = _scope; + + // Parse the false side + _scope = fScope; // Restore scope, then parse else block if any + ctrl(ifF.unkeep()); // Ctrl token is now set to ifFalse projection + // Up-cast predicate, even if not else clause, because predicate can + // remain true if the true clause exits: `if( !ptr ) return 0; return ptr.fld;` + _scope.addGuards(ifF,pred,true); + boolean doRHS = match(fside); + Node rhs = doRHS + ? (stmt ? parseStatement() : parseAsgn()) + : (stmt ? null : con(lhs._type.makeZero())); + _scope.removeGuards(ifF); + if( doRHS ) + fScope = _scope; + pred.unkeep(); + + // Check for `if(pred) int x=17;` + if( tScope.nIns() != ndefs || fScope.nIns() != ndefs ) + throw error("Cannot define a new name on one arm of an if"); + + // Check the trinary widening int/flt + if( !stmt ) { + rhs = widenInt( rhs, lhs._type ).keep(); + lhs = widenInt( lhs.unkeep(), rhs._type ).keep(); + rhs.unkeep(); + } + + // Merge results + _scope = tScope; + _xScopes.pop(); // Discard pushed from graph display + + RegionNode r = ctrl(tScope.mergeScopes(fScope)); + Node ret = stmt ? r : peep(new PhiNode("",lhs._type.meet(rhs._type),r,lhs.unkeep(),rhs)); + r.peephole(); + return ret; + } + + /** + * Parses a return statement; "return" already parsed. + * The $ctrl edge is killed. + * + * 
+     *     'return' expr ;
+     *
+ * @return an expression {@link Node}, never {@code null} + */ + private Node parseReturn() { + var expr = require(parseAsgn(), ";"); + Node ret = STOP.addReturn(new ReturnNode(ctrl(), expr, _scope).peephole()); + ctrl(XCTRL); // Kill control + return ret; + } + + /** + * Dumps out the node graph + * @return {@code null} + */ + Node showGraph() { + System.out.println(new GraphVisualizer().generateDotOutput(STOP,_scope,_xScopes)); + return null; + } + + /** Parse: [name '='] expr + */ + private Node parseAsgn() { + int old = pos(); + String name = _lexer.matchId(); + // Just a plain expression, no assignment. + // Distinguish `var==expr` from `var=expr` + if( name==null || KEYWORDS.contains(name) || !matchOpx('=','=') ) + { pos(old); return parseExpression(); } + + // Parse assignment expression + Node expr = parseAsgn(); + + // Final variable to update + ScopeMinNode.Var def = _scope.lookup(name); + if( def==null ) + throw error("Undefined name '" + name + "'"); + + // TOP fields are for late-initialized fields; these have never + // been written to, and this must be the final write. Other writes + // outside the constructor need to check the final bit. + if( _scope.in(def._idx)._type!=Type.TOP && def._final && !_scope.inCon() ) + throw error("Cannot reassign final '"+name+"'"); + + // Lift expression, based on type + Node lift = liftExpr(expr, def.type(), def._final); + // Update + _scope.update(name,lift); + // Return un-lifted expr + return expr; + } + + // Make finals deep; widen ints to floats; narrow wide int types. + // Early error if types do not match variable. + private Node liftExpr( Node expr, Type t, boolean xfinal ) { + assert !(expr._type instanceof TypeMemPtr tmp) || !tmp.isFRef(); + // Final is deep on ptrs + if( xfinal && t instanceof TypeMemPtr tmp ) { + t = tmp.makeRO(); + expr = peep(new ReadOnlyNode(expr)); + } + // Auto-widen int to float + expr = widenInt( expr, t ); + // Auto-narrow wide ints to narrow ints + expr = zsMask(expr,t); + // Type is sane + if( !expr._type.isa(t) ) + throw error("Type " + expr._type.str() + " is not of declared type " + t.str()); + return expr; + } + + private Node widenInt( Node expr, Type t ) { + return expr._type instanceof TypeInteger && t instanceof TypeFloat + ? peep(new ToFloatNode(expr)) : expr; + } + + /** + * Parse declaration or expression statement + * declStmt = type var['=' exprAsgn][, var['=' exprAsgn]]* ';' | exprAsgn ';' + *

+ * exprAsgn = var '=' exprAsgn | expr + */ + private Node parseDeclarationStatement() { + Type t = type(); + if( t == null ) + return require(parseAsgn(),";"); + + // now parse var['=' asgnexpr] in a loop + Node n = parseDeclaration(t); + while( match(",") ) + n = parseDeclaration(t); + return require(n,";"); + } + + /** Parse final: [!]var['=' asgn] + */ + private Node parseDeclaration(Type t) { + assert t!=null; + // Has var/val instead of a user-declared type + boolean inferType = t==Type.TOP || t==Type.BOTTOM; + boolean hasBang = match("!"); + String name = requireId(); + + // Optional initializing expression follows + boolean xfinal = false; + Node expr; + if( match("=") ) { + expr = parseAsgn(); + // `val` is always final + xfinal = (t==Type.TOP) || + // var is always not-final, final if no Bang AND TMP since primitives are not-final by default + (t!=Type.BOTTOM && !hasBang && (t instanceof TypeMemPtr)); + // var/val, then type comes from expression + if( inferType ) + t = expr._type.glb(); + + } else { + if( inferType && !_scope.inCon() ) + throw errorSyntax("=expression"); + expr = con(t.makeInit()); + } + + // Lift expression, based on type + Node lift = liftExpr(expr, t, xfinal); + if( xfinal && t instanceof TypeMemPtr tmp ) + t = tmp.makeRO(); + + // Define a new name, + if( !_scope.define(name,t,xfinal,lift) ) + throw error("Redefining name '" + name + "'"); + return null; + } + + + + /** + * Parse a struct declaration, and return the following statement. + * Only allowed in top level scope. + * Structs cannot be redefined. + * + * @return The statement following the struct + */ + private Node parseStruct() { + if (_xScopes.size() > 1) throw errorSyntax("struct declarations can only appear in top level scope"); + String typeName = requireId(); + Type t = TYPES.get(typeName); + if( t!=null && !(t instanceof TypeMemPtr tmp && tmp.isFRef() ) ) + throw errorSyntax("struct '" + typeName + "' cannot be redefined"); + + // A Block scope parse, and inspect the scope afterward for fields. + _scope.push(true); + _xScopes.push(_scope); + require("{"); + while (!peek('}') && !_lexer.isEOF()) + parseStatement(); + + // Grab the declarations and build fields and a Struct + int lexlen = _scope._lexSize.last(); + int varlen = _scope._vars._len; + StructNode s = new StructNode(); + Field[] fs = new Field[varlen-lexlen]; + for( int i=lexlen; i + * expr : bitwise [? expr [: expr]] + * + * @return an expression {@link Node}, never {@code null} + */ + private Node parseExpression() { + Node expr = parseBitwise(); + return match("?") ? parseTrinary(expr,false,":") : expr; + } + + /** + * Parse an bitwise expression + * + * 

+     *     bitwise : compareExpr (('&' | '|' | '^') compareExpr)*
+     *
+ * @return a bitwise expression {@link Node}, never {@code null} + */ + private Node parseBitwise() { + Node lhs = parseComparison(); + while( true ) { + if( false ) ; + else if( match("&") ) lhs = new AndNode(lhs,null); + else if( match("|") ) lhs = new OrNode(lhs,null); + else if( match("^") ) lhs = new XorNode(lhs,null); + else break; + lhs.setDef(2,parseComparison()); + lhs = peep(lhs); + } + return lhs; + } + + /** + * Parse an expression of the form: + * + * 
+     *     expr : additiveExpr op additiveExpr
+     *
+ * @return an comparator expression {@link Node}, never {@code null} + */ + private Node parseComparison() { + var lhs = parseShift(); + while( true ) { + int idx=0; boolean negate=false; + // Test for any local nodes made, and "keep" lhs during peepholes + if( false ) ; + else if( match("==") ) { idx=2; lhs = new BoolNode.EQ(lhs, null); } + else if( match("!=") ) { idx=2; lhs = new BoolNode.EQ(lhs, null); negate=true; } + else if( match("<=") ) { idx=2; lhs = new BoolNode.LE(lhs, null); } + else if( match("<" ) ) { idx=2; lhs = new BoolNode.LT(lhs, null); } + else if( match(">=") ) { idx=1; lhs = new BoolNode.LE(null, lhs); } + else if( match(">" ) ) { idx=1; lhs = new BoolNode.LT(null, lhs); } + else break; + // Peepholes can fire, but lhs is already "hooked", kept alive + lhs.setDef(idx,parseShift()); + lhs = peep(lhs.widen()); + if( negate ) // Extra negate for != + lhs = peep(new NotNode(lhs)); + } + return lhs; + } + + /** + * Parse an additive expression + * + * 
+     *     shiftExpr : additiveExpr (('<<' | '>>' | '>>>') additiveExpr)*
+     *
+ * @return a shift expression {@link Node}, never {@code null} + */ + private Node parseShift() { + Node lhs = parseAddition(); + while( true ) { + if( false ) ; + else if( match("<<") ) lhs = new ShlNode(lhs,null); + else if( match(">>>")) lhs = new ShrNode(lhs,null); + else if( match(">>") ) lhs = new SarNode(lhs,null); + else break; + lhs.setDef(2,parseAddition()); + lhs = peep(lhs.widen()); + } + return lhs; + } + + /** + * Parse an additive expression + * + * 
+     *     additiveExpr : multiplicativeExpr (('+' | '-') multiplicativeExpr)*
+     *
+ * @return an add expression {@link Node}, never {@code null} + */ + private Node parseAddition() { + Node lhs = parseMultiplication(); + while( true ) { + if( false ) ; + else if( match("+") ) lhs = new AddNode(lhs,null); + else if( match("-") ) lhs = new SubNode(lhs,null); + else break; + lhs.setDef(2,parseMultiplication()); + lhs = peep(lhs.widen()); + } + return lhs; + } + + /** + * Parse an multiplicativeExpr expression + * + * 
+     *     multiplicativeExpr : unaryExpr (('*' | '/') unaryExpr)*
+     *
+ * @return a multiply expression {@link Node}, never {@code null} + */ + private Node parseMultiplication() { + var lhs = parseUnary(); boolean mul; + while( true ) { + if( false ) ; + else if( match("*") ) lhs = new MulNode(lhs,null); + else if( match("/") ) lhs = new DivNode(lhs,null); + else break; + lhs.setDef(2,parseUnary()); + lhs = peep(lhs.widen()); + } + return lhs; + } + + /** + * Parse a unary minus expression. + * + * 
+     *     unaryExpr : ('-') unaryExpr | '!') unaryExpr | postfixExpr | primaryExpr | '--' Id | '++' Id
+     *
+ * @return a unary expression {@link Node}, never {@code null} + */ + private Node parseUnary() { + // Pre-dec/pre-inc + int old = pos(); + if( match("--") || match("++") ) { + int delta = _lexer.peek(-1)=='+' ? 1 : -1; // Pre vs post + String name = _lexer.matchId(); + if( name!=null ) { + ScopeMinNode.Var n = _scope.lookup(name); + if( n != null ) { + if( n._final ) + throw error("Cannot reassign final '"+n._name+"'"); + Node expr = zsMask(peep(new AddNode(_scope.in(n),con(delta))),n.type()); + _scope.update(n,expr); + return expr; + } + } + // Reset, try again + pos(old); + } + if (match("-")) return peep(new MinusNode(parseUnary()).widen()); + if (match("!")) return peep(new NotNode(parseUnary())); + return parsePostfix(parsePrimary()); + } + + /** + * Parse a primary expression: + * + * 
+     *     primaryExpr : integerLiteral | true | false | null | new Type | '(' expression ')' | Id['++','--']
+     *
+ * @return a primary {@link Node}, never {@code null} + */ + private Node parsePrimary() { + if( _lexer.isNumber(_lexer.peek()) ) return parseLiteral(); + if( matchx("true" ) ) return con(1); + if( matchx("false") ) return ZERO; + if( matchx("null" ) ) return con(TypeMemPtr.NULLPTR); + if( match("(") ) return require(parseAsgn(), ")"); + if( matchx("new" ) ) return alloc(); + // Expect an identifier now + ScopeMinNode.Var n = requireLookupId("an identifier or expression"); + Node rvalue = _scope.in(n); + if( rvalue._type == Type.BOTTOM ) + throw error("Cannot read uninitialized field '"+n._name+"'"); + // Check for assign-update, x += e0; + char ch = _lexer.matchOperAssign(); + if( ch==0 ) return rvalue; + if( n._final ) + throw error("Cannot reassign final '"+n._name+"'"); + Node op = switch(ch) { + case '+' -> new AddNode(rvalue,null); + case '-' -> new SubNode(rvalue,null); + case '*' -> new MulNode(rvalue,null); + case '/' -> new DivNode(rvalue,null); + case 1 -> new AddNode(rvalue,con( 1)); + case (char)-1 -> new AddNode(rvalue,con(-1)); + default -> throw Utils.TODO(); + }; + // Return pre-value (x+=1) or post-value (x++) + boolean pre = op.in(2)==null; + // Parse RHS argument as needed + if( pre ) + { op.keep().setDef(2,parseAsgn()); op.unkeep(); } + else rvalue.keep(); // Keep post-value across peeps + op = zsMask(peep(op),n.type()); + _scope.update(n,op); + return pre ? op : rvalue.unkeep(); + } + + ScopeMinNode.Var requireLookupId(String msg) { + String id = _lexer.matchId(); + if( id == null || KEYWORDS.contains(id) ) + throw errorSyntax(msg); + ScopeMinNode.Var n = _scope.lookup(id); + if( n==null ) + throw error("Undefined name '" + id + "'"); + return n; + } + + /** + Parse an allocation + */ + private Node alloc() { + Type t = type(); + if( t==null ) throw error("Expected a type"); + // Parse ary[ length_expr ] + if( match("[") ) { + Node len = parseAsgn(); + if( !(len._type instanceof TypeInteger) ) + throw error("Cannot allocate an array with length "+len._type); + require("]"); + TypeMemPtr tmp = typeAry(t); + return newArray(tmp._obj,len); + } + + if( !(t instanceof TypeMemPtr tmp) ) + throw error("Cannot allocate a "+t.str()); + + // Parse new struct { default_initialization } + StructNode s = INITS.get(tmp._obj._name); + if( s==null ) throw error("Unknown struct type '" + tmp._obj._name + "'"); + + Field[] fs = s._ts._fields; + // if the object is fully initialized, we can skip a block here. + // Check for constructor block: + boolean hasConstructor = match("{"); + Ary init=s._inputs; int idx=0; + if( hasConstructor ) { + idx = _scope.nIns(); + // Push a scope, and pre-assign all struct fields. + _scope.push(); + for( int i=0; i init ) { + Field[] fs = obj._fields; + if( fs==null ) + throw error("Unknown struct type '" + obj._name + "'"); + int len = fs.length; + Node[] ns = new Node[2+len+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); + // Initial values for every field + for( int i = 0; i < len; i++ ) + ns[2+len+i] = init.get(i+idx); + Node nnn = new NewNode(TypeMemPtr.make(obj), ns).peephole(); + for( int i = 0; i < len; i++ ) + memAlias(fs[i]._alias, new ProjNode(nnn,i+2,memName(fs[i]._alias)).peephole()); + return new ProjNode(nnn,1,obj._name).peephole(); + } + + private static final Ary ALTMP = new Ary<>(Node.class); + private Node newArray(TypeStruct ary, Node len) { + int base = ary.aryBase (); + int scale= ary.aryScale(); + Node size = peep(new AddNode(con(base),peep(new ShlNode(len.keep(),con(scale))))); + ALTMP.clear(); ALTMP.add(len.unkeep()); ALTMP.add(con(ary._fields[1]._type.makeInit())); + 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(); } + + /** + * Parse postfix expression. For now this is just a field + * expression, but in future could be array index too. + * + * 
+     *     expr ('.' IDENTIFIER)* [ = expr ]
+     *     expr #
+     *     expr ('[' expr ']')* = [ = expr ]
+     *
+ */ + private Node parsePostfix(Node expr) { + String name = null; + if( match(".") ) name = requireId(); + else if( match("#") ) name = "#"; + else if( match("[") ) name = "[]"; + else return expr; // No postfix + + // Sanity check expr for being a reference + if( !(expr._type instanceof TypeMemPtr ptr) ) + throw error("Expected reference but got " + expr._type.str()); + + // Happens when parsing known dead code, which often has other typing + // issues. Since the code is dead, possibly due to inlining, lets not + // spoil the user experience with error messages. + if( ctrl()._type==Type.XCONTROL ) + // Exit out via parsing the trailing expression + return matchOpx('=','=') ? parseAsgn() : parsePostfix(con(Type.TOP)); + + // Sanity check field name for existing + TypeMemPtr tmp = (TypeMemPtr)TYPES.get(ptr._obj._name); + if( tmp == null ) throw error("Accessing unknown field '" + name + "' from '" + ptr + "'"); + TypeStruct base = tmp._obj; + int fidx = base.find(name); + if( fidx == -1 ) throw error("Accessing unknown field '" + name + "' from '" + ptr.str() + "'"); + + // Get field type and layout offset from base type and field index fidx + Field f = base._fields[fidx]; // Field from field index + Type tf = f._type; + if( tf instanceof TypeMemPtr ftmp && ftmp.isFRef() ) + tf = ftmp.makeFrom(((TypeMemPtr)(TYPES.get(ftmp._obj._name)))._obj); + + // Field offset; fixed for structs, computed for arrays + Node off = (name.equals("[]") // If field is an array body + // Array index math + ? peep(new AddNode(con(base.aryBase()),peep(new ShlNode(require(parseAsgn(),"]"),con(base.aryScale()))))) + // Struct field offsets are hardwired + : con(base.offset(fidx))).keep(); + + // Disambiguate "obj.fld==x" boolean test from "obj.fld=x" field assignment + if( matchOpx('=','=') ) { + 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()); + 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); + // 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() ) load.setDef(0,ctrl()); + load = peep(load); + + // ary[idx]++ or ptr.fld++ + if( matchx("++") || matchx("--") ) { + if( f._final && !f._fname.equals("[]") ) + 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)); + // And use the original loaded value as the result + } else expr.unkill(); + off.unkill(); + + return parsePostfix(load); + } + + + // zero/sign extend. "i" is limited to either classic unsigned (min==0) or + // classic signed (min=minus-power-of-2); max=power-of-2-minus-1. + private Node zsMask(Node val, Type t ) { + if( !(val._type instanceof TypeInteger tval && t instanceof TypeInteger t0 && !tval.isa(t0)) ) { + if( !(val._type instanceof TypeFloat tval && t instanceof TypeFloat t0 && !tval.isa(t0)) ) + return val; + // Float rounding + return peep(new RoundF32Node(val)); + } + if( t0._min==0 ) // Unsigned + return peep(new AndNode(val,con(t0._max))); + // Signed extension + int shift = Long.numberOfLeadingZeros(t0._max)-1; + Node shf = con(shift); + if( shf._type==TypeInteger.ZERO ) + return val; + return peep(new SarNode(peep(new ShlNode(val,shf.keep())),shf.unkeep())); + } + + + /** + * Parse integer literal + * + * 
+     *     integerLiteral: [1-9][0-9]* | [0]
+     *     floatLiteral: [digits].[digits]?[e [digits]]?
+     *
+ */ + private ConstantNode parseLiteral() { return con(_lexer.parseNumber()); } + public static Node con( long con ) { return con(TypeInteger.constant(con)); } + public static ConstantNode con( Type t ) { return (ConstantNode)new ConstantNode(t).peephole(); } + public Node peep( Node n ) { + // Peephole, then improve with lexically scoped guards + return _scope.upcastGuard(n.peephole()); + } + + ////////////////////////////////// + // Utilities for lexical analysis + + // Return true and skip if "syntax" is next in the stream. + private boolean match (String syntax) { return _lexer.match (syntax); } + // Match must be "exact", not be followed by more id letters + private boolean matchx(String syntax) { return _lexer.matchx(syntax); } + private boolean matchOpx(char c0, char c1) { return _lexer.matchOpx(c0,c1); } + // Return true and do NOT skip if 'ch' is next + private boolean peek(char ch) { return _lexer.peek(ch); } + private boolean peekIsId() { return _lexer.peekIsId(); } + + private int pos() { return _lexer._position; } + private int pos(int pos) { + int old = _lexer._position; + _lexer._position = pos; + return old; + } + private Type posT(int pos) { _lexer._position = pos; return null; } + + // Require and return an identifier + private String requireId() { + String id = _lexer.matchId(); + if (id != null && !KEYWORDS.contains(id) ) return id.intern(); + throw error("Expected an identifier, found '"+id+"'"); + } + + private String matchId() { + int old = pos(); + String id = _lexer.matchId(); + if( id==null ) return null; + if( !KEYWORDS.contains(id) ) return id; + pos(old); + return null; + } + + // Require an exact match + private Parser require(String syntax) { require(null, syntax); return this; } + private N require(N n, String syntax) { + if (match(syntax)) return n; + throw errorSyntax(syntax); + } + + RuntimeException errorSyntax(String syntax) { + return error("Syntax error, expected " + syntax + ": " + _lexer.getAnyNextToken()); + } + + public static RuntimeException error( String errorMessage ) { + return new RuntimeException(errorMessage); + } + + //////////////////////////////////// + // Lexer components + + private static class Lexer { + + // Input buffer; an array of text bytes read from a file or a string + private final byte[] _input; + // Tracks current position in input buffer + private int _position = 0; + + /** + * Record the source text for lexing + */ + public Lexer(String source) { + this(source.getBytes()); + } + + /** + * Direct from disk file source + */ + public Lexer(byte[] buf) { + _input = buf; + } + + // Very handy in the debugger, shows the unparsed program + @Override + public String toString() { + return new String(_input, _position, _input.length - _position); + } + + // True if at EOF + private boolean isEOF() { + return _position >= _input.length; + } + + // Peek next character, or report EOF + private char peek() { + return isEOF() ? Character.MAX_VALUE // Special value that causes parsing to terminate + : (char) _input[_position]; + } + // Just crash if misused + public byte peek(int off) { return _input[_position+off]; } + + private char nextChar() { + char ch = peek(); + _position++; + return ch; + } + + // True if a white space + private boolean isWhiteSpace() { + return peek() <= ' '; // Includes all the use space, tab, newline, CR + } + + /** + * Return the next non-white-space character + */ + private void skipWhiteSpace() { + while( true ) { + if( isWhiteSpace() ) _position++; + // Skip // to end of line + else if( _position+2 < _input.length && + _input[_position ] == '/' && + _input[_position+1] == '/') { + _position += 2; + while( !isEOF() && _input[_position] != '\n' ) _position++; + } else break; + } + } + + // Next non-white-space character, or EOF + public char nextXChar() { skipWhiteSpace(); return nextChar(); } + + // Return true, if we find "syntax" after skipping white space; also + // then advance the cursor past syntax. + // Return false otherwise, and do not advance the cursor. + boolean match(String syntax) { + skipWhiteSpace(); + int len = syntax.length(); + if (_position + len > _input.length) return false; + for (int i = 0; i < len; i++) + if ((char) _input[_position + i] != syntax.charAt(i)) + return false; + _position += len; + return true; + } + + // Match must be exact and not followed by more ID characters. + // Prevents identifier "ifxy" from matching an "if" statement. + boolean matchx(String syntax) { + if( !match(syntax) ) return false; + if( !isIdLetter(peek()) ) return true; + _position -= syntax.length(); + return false; + } + boolean matchOpx(char c0, char c1) { + skipWhiteSpace(); + if( _position+1 >= _input.length || _input[_position]!=c0 || _input[_position+1]==c1 ) + return false; + _position++; + return true; + } + + private boolean peek(char ch) { + skipWhiteSpace(); + return peek()==ch; + } + + boolean peekIsId() { + skipWhiteSpace(); + return isIdStart(peek()); + } + + // Return an identifier or null + String matchId() { + return peekIsId() ? parseId() : null; + } + + // Used for errors + String getAnyNextToken() { + if (isEOF()) return ""; + if (isIdStart(peek())) return parseId(); + if (isNumber(peek())) return parseNumberString(); + if (isPunctuation(peek())) return parsePunctuation(); + return String.valueOf(peek()); + } + + + boolean isNumber(char ch) {return Character.isDigit(ch);} + + // Return a constant Type, either TypeInteger or TypeFloat + private Type parseNumber() { + int old = _position; + int len = isLongOrDouble(); + if( len > 0 ) { + if( len > 1 && _input[old]=='0' ) + throw error("Syntax error: integer values cannot start with '0'"); + return TypeInteger.constant(Long.parseLong(new String(_input,old,len))); + } + return TypeFloat.constant(Double.parseDouble(new String(_input,old,-len))); + } + private String parseNumberString() { + int old = _position; + int len = Math.abs(isLongOrDouble()); + _position += len; + return new String(_input,old,len); + } + + // Return +len that ends a long + // Return -len that ends a double + private int isLongOrDouble() { + int old = _position; + char c; + while( Character.isDigit(c=nextChar()) ) ; + if( !(c=='e' || c=='.') ) + return --_position - old; + while( Character.isDigit(c=nextChar()) || c=='e' || c=='.' ) ; + return -(--_position - old); + } + + // First letter of an identifier + private boolean isIdStart(char ch) { + return Character.isAlphabetic(ch) || ch == '_'; + } + + // All characters of an identifier, e.g. "_x123" + private boolean isIdLetter(char ch) { + return Character.isLetterOrDigit(ch) || ch == '_'; + } + + private String parseId() { + int start = _position; + while (isIdLetter(nextChar())) ; + return new String(_input, start, --_position - start); + } + + // + private boolean isPunctuation(char ch) { + return "=;[]<>()+-/*&|^".indexOf(ch) != -1; + } + + private String parsePunctuation() { + int start = _position; + return new String(_input, start, 1); + } + + // Next oper= character, or 0. + // As a convenience, mark "++" as a char 1 and "--" as char -1 (65535) + public char matchOperAssign() { + skipWhiteSpace(); + if( _position+2 >= _input.length ) return 0; + char ch0 = (char)_input[_position]; + if( "+-/*&|^".indexOf(ch0) == -1 ) return 0; + char ch1 = (char)_input[_position+1]; + if( ch1 == '=' ) { _position += 2; return ch0; } + if( ch0 == '+' && ch1 == '+' ) { _position += 2; return (char) 1; } + if( ch0 == '-' && ch1 == '-' ) { _position += 2; return (char)-1; } + return 0; + } + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/Utils.java b/chapter18/src/main/java/com/seaofnodes/simple/Utils.java new file mode 100644 index 00000000..1346c3ea --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/Utils.java @@ -0,0 +1,53 @@ +package com.seaofnodes.simple; + +import java.util.ArrayList; + +public class Utils { + public static RuntimeException TODO() { return TODO("Not yet implemented"); } + public static RuntimeException TODO(String msg) { return new RuntimeException(msg); } + + /** + * Fast, constant-time, element removal. Does not preserve order + * + * @param array ArrayList to modify + * @param i element to be removed + * @return element removed + */ + public static void del(ArrayList array, int i) { + E last = array.removeLast(); + if (i < array.size()) array.set(i, last); + } + + /** + * Search a list for an element by reference + * + * @param ary List to search in + * @param x Object to be searched + * @return >= 0 on success, -1 on failure + */ + public static int find( ArrayList ary, E x ) { + for( int i=0; i= 0 on success, -1 on failure + */ + public static int find( E[] ary, E x ) { + for( int i=0; i>>n); } + + // Fold a 64bit hash into 32 bits + public static int fold( long x ) { return (int)((x>>32) ^ x); } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/AddFNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/AddFNode.java new file mode 100644 index 00000000..e5a2440c --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/AddFNode.java @@ -0,0 +1,44 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeFloat; + +import java.util.BitSet; + +public class AddFNode extends Node { + public AddFNode(Node lhs, Node rhs) { super(null, lhs, rhs); } + + @Override public String label() { return "AddF"; } + + @Override public String glabel() { return "+"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + in(1)._print0(sb.append("("), visited); + in(2)._print0(sb.append("+"), visited); + return sb.append(")"); + } + + @Override + public Type compute() { + if (in(1)._type instanceof TypeFloat i0 && + in(2)._type instanceof TypeFloat i1) { + if (i0.isConstant() && i1.isConstant()) + return TypeFloat.constant(i0.value()+i1.value()); + } + return in(1)._type.meet(in(2)._type); + } + + @Override + public Node idealize() { + Node lhs = in(1); + Type t2 = in(2)._type; + + // Add of 1. + if ( t2.isConstant() && t2 instanceof TypeFloat i && i.value()==0 ) + return lhs; + + return null; + } + @Override Node copy(Node lhs, Node rhs) { return new AddFNode(lhs,rhs); } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/AddNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/AddNode.java new file mode 100644 index 00000000..0834a381 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/AddNode.java @@ -0,0 +1,187 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.*; + +import java.util.BitSet; + +public class AddNode extends Node { + public AddNode(Node lhs, Node rhs) { super(null, lhs, rhs); } + + @Override public String label() { return "Add"; } + + @Override public String glabel() { return "+"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + in(1)._print0(sb.append("("), visited); + in(2)._print0(sb.append("+"), visited); + return sb.append(")"); + } + + + @Override + public Type compute() { + Type t1 = in(1)._type, t2 = in(2)._type; + if( t1.isHigh() || t2.isHigh() ) + return TypeInteger.TOP; + if( t1 instanceof TypeInteger i1 && + t2 instanceof TypeInteger i2 ) { + if (i1.isConstant() && i2.isConstant()) + return TypeInteger.constant(i1.value()+i2.value()); + // Fold ranges like {0-1} + {2-3} into {2-4}. + if( !overflow(i1._min,i2._min) && + !overflow(i1._max,i2._max) ) + return TypeInteger.make(i1._min+i2._min,i1._max+i2._max); + } + return TypeInteger.BOT; + } + + static boolean overflow( long x, long y ) { + if( (x ^ y ) < 0 ) return false; // unequal signs, never overflow + return (x ^ (x + y)) < 0; // sum has unequal signs, so overflow + } + @Override + public Node idealize () { + Node lhs = in(1); + Node rhs = in(2); + Type t2 = rhs._type; + + // Add of 0. We do not check for (0+x) because this will already + // canonicalize to (x+0) + if( t2 instanceof TypeInteger i && i.isConstant() && i.value()==0 ) + return lhs; + + // Add of same to a multiply by 2 + if( lhs==rhs ) + return new MulNode(lhs,new ConstantNode(TypeInteger.constant(2)).peephole()); + + // Goal: a left-spine set of adds, with constants on the rhs (which then fold). + + // Move non-adds to RHS + if( !(lhs instanceof AddNode) && rhs instanceof AddNode ) + return swap12(); + + // x+(-y) becomes x-y + if( rhs instanceof MinusNode minus ) + return new SubNode(lhs,minus.in(1)); + + // Now we might see (add add non) or (add non non) or (add add add) but never (add non add) + + // Do we have x + (y + z) ? + // Swap to (x + y) + z + // Rotate (add add add) to remove the add on RHS + if( rhs instanceof AddNode add ) + return new AddNode(new AddNode(lhs,add.in(1)).peephole(), add.in(2)); + + // Now we might see (add add non) or (add non non) but never (add non add) nor (add add add) + if( !(lhs instanceof AddNode) ) + // Rotate; look for (add (phi cons) con/(phi cons)) + return spine_cmp(lhs,rhs,this) ? swap12() : phiCon(this,true); + + // Now we only see (add add non) + + // Dead data cycle; comes about from dead infinite loops. Do nothing, + // the loop will peep as dead after a bit. + if( lhs.in(1) == lhs ) + return null; + + // Do we have (x + con1) + con2? + // Replace with (x + (con1+con2) which then fold the constants + // lhs.in(2) is con1 here + // If lhs.in(2) is not a constant, we add ourselves as a dependency + // because if it later became a constant then we could make this + // transformation. + if( lhs.in(2).addDep(this)._type.isConstant() && rhs._type.isConstant() ) + return new AddNode(lhs.in(1),new AddNode(lhs.in(2),rhs).peephole()); + + + // Do we have ((x + (phi cons)) + con) ? + // Do we have ((x + (phi cons)) + (phi cons)) ? + // Push constant up through the phi: x + (phi con0+con0 con1+con1...) + Node phicon = phiCon(this,true); + if( phicon!=null ) return phicon; + + // Now we sort along the spine via rotates, to gather similar things together. + + // Do we rotate (x + y) + z + // into (x + z) + y ? + if( spine_cmp(lhs.in(2),rhs,this) ) + return new AddNode(new AddNode(lhs.in(1),rhs).peephole(),lhs.in(2)); + + return null; + } + + // Rotation is only valid for associative ops, e.g. Add, Mul, And, Or, Xor. + // Do we have ((phi cons)|(x + (phi cons)) + con|(phi cons)) ? + // Push constant up through the phi: x + (phi con0+con0 con1+con1...) + static Node phiCon(Node op, boolean rotate) { + Node lhs = op.in(1); + Node rhs = op.in(2); + if( rhs._type==TypeInteger.TOP ) return null; + // LHS is either a Phi of constants, or another op with Phi of constants + PhiNode lphi = pcon(lhs,op); + if( rotate && lphi==null && lhs.nIns() > 2 ) { + // Only valid to rotate constants if both are same associative ops + if( lhs.getClass() != op.getClass() ) return null; + lphi = pcon(lhs.in(2),op); // Will rotate with the Phi push + } + if( lphi==null ) return null; + + // RHS is a constant or a Phi of constants + if( !(rhs instanceof ConstantNode) && pcon(rhs,op)==null ) + return null; + + // If both are Phis, must be same Region + if( rhs instanceof PhiNode && lphi.in(0) != rhs.in(0) ) + return null; + + // Note that this is the exact reverse of Phi pulling a common op down + // to reduce total op-count. We don't get in an endless push-up + // push-down peephole cycle because the constants all fold first. + Node[] ns = new Node[lphi.nIns()]; + ns[0] = lphi.in(0); + // Push constant up through the phi: x + (phi con0+con0 con1+con1...) + for( int i=1; i hi._nid; + } + + @Override Node copy(Node lhs, Node rhs) { return new AddNode(lhs,rhs); } + @Override Node copyF() { return new AddFNode(null,null); } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/AndNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/AndNode.java new file mode 100644 index 00000000..17a4e0ca --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/AndNode.java @@ -0,0 +1,68 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeInteger; + +import java.util.BitSet; + +public class AndNode extends Node { + public AndNode(Node lhs, Node rhs) { super(null, lhs, rhs); } + + @Override public String label() { return "And"; } + + @Override public String glabel() { return "&"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + in(1)._print0(sb.append("("), visited); + in(2)._print0(sb.append("&"), visited); + return sb.append(")"); + } + + @Override + public Type compute() { + Type t1 = in(1)._type, t2 = in(2)._type; + if( t1.isHigh() || t2.isHigh() ) + return TypeInteger.TOP; + if( t1 instanceof TypeInteger i0 && + t2 instanceof TypeInteger i1 ) { + if( i0.isConstant() && i1.isConstant() ) + return TypeInteger.constant(i0.value()&i1.value()); + // Sharpen allowed bits if either value is narrowed + long mask = i0.mask() & i1.mask(); + return mask < 0 ? TypeInteger.BOT : TypeInteger.make(0,mask); + } + return TypeInteger.BOT; + } + + @Override + public Node idealize() { + Node lhs = in(1); + Node rhs = in(2); + Type t1 = lhs._type; + Type t2 = rhs._type; + + // And of -1. We do not check for (-1&x) because this will already + // canonicalize to (x&-1) + if( t2.isConstant() && t2 instanceof TypeInteger i && i.value()==-1 ) + return lhs; + + // Move constants to RHS: con*arg becomes arg*con + if ( t1.isConstant() && !t2.isConstant() ) + return swap12(); + + // Do we have ((x & (phi cons)) & con) ? + // Do we have ((x & (phi cons)) & (phi cons)) ? + // Push constant up through the phi: x & (phi con0&con0 con1&con1...) + Node phicon = AddNode.phiCon(this,true); + if( phicon!=null ) return phicon; + + return null; + } + @Override Node copy(Node lhs, Node rhs) { return new AndNode(lhs,rhs); } + @Override String err() { + if( !(in(1)._type instanceof TypeInteger) ) return "Cannot '&' " + in(1)._type; + if( !(in(2)._type instanceof TypeInteger) ) return "Cannot '&' " + in(2)._type; + return null; + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/BoolNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/BoolNode.java new file mode 100644 index 00000000..238be00e --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/BoolNode.java @@ -0,0 +1,120 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.Parser; +import com.seaofnodes.simple.Utils; +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeInteger; +import com.seaofnodes.simple.type.TypeFloat; +import com.seaofnodes.simple.type.TypeMemPtr; + +import java.util.BitSet; +import static com.seaofnodes.simple.type.TypeInteger.*; + +abstract public class BoolNode extends Node { + + public BoolNode(Node lhs, Node rhs) { + super(null, lhs, rhs); + } + + abstract String op(); // String opcode name + + @Override + public String label() { return getClass().getSimpleName(); } + + @Override + public String glabel() { return op(); } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + in(1)._print0(sb.append("("), visited); + in(2)._print0(sb.append(op()), visited); + return sb.append(")"); + } + + @Override + public TypeInteger compute() { + Type t1 = in(1)._type; + Type t2 = in(2)._type; + if( t1.isHigh() || t2.isHigh() ) + return BOOL.dual(); + if( t1 instanceof TypeInteger i1 && + t2 instanceof TypeInteger i2 ) + return doOp(i1,i2); + if( t1 instanceof TypeFloat f1 && + t2 instanceof TypeFloat f2 && + f1.isConstant() && f2.isConstant() ) + return TypeInteger.constant(doOp(f1.value(), f2.value()) ? 1 : 0); + return BOOL; + } + + TypeInteger doOp(TypeInteger t1, TypeInteger t2) { throw Utils.TODO(); } + boolean doOp(double lhs, double rhs) { throw Utils.TODO(); } + Node copyF(Node lhs, Node rhs) { return null; } + + @Override + public Node idealize() { + // Compare of same + if( in(1)==in(2) ) + return new ConstantNode(doOp(ZERO,ZERO)); + + // Equals pushes constant to the right; 5==X becomes X==5. + if( this instanceof EQ ) { + if( !(in(2) instanceof ConstantNode) ) { + // con==noncon becomes noncon==con + if( in(1) instanceof ConstantNode || in(1)._nid > in(2)._nid ) + // Equals sorts by NID otherwise: non.high == non.low becomes non.low == non.high + return in(1)._type instanceof TypeFloat ? new EQF(in(2),in(1)) : new EQ(in(2),in(1)); + } + // Equals X==0 becomes a !X + if( (in(2)._type == ZERO || in(2)._type == TypeMemPtr.NULLPTR) ) + return new NotNode(in(1)); + } + + // Do we have ((x * (phi cons)) * con) ? + // Do we have ((x * (phi cons)) * (phi cons)) ? + // Push constant up through the phi: x * (phi con0*con0 con1*con1...) + Node phicon = AddNode.phiCon(this,this instanceof EQ); + if( phicon!=null ) return phicon; + + return null; + } + + public static class EQ extends BoolNode { + public EQ(Node lhs, Node rhs) { super(lhs,rhs); } + String op() { return "=="; } + TypeInteger doOp(TypeInteger i1, TypeInteger i2) { + if( i1==i2 && i1.isConstant() ) return TRUE; + if( i1._max < i2._min || i1._min > i2._max ) return FALSE; + return BOOL; + } + Node copy(Node lhs, Node rhs) { return new EQ(lhs,rhs); } + Node copyF() { return new EQF(null,null); } + } + public static class LT extends BoolNode { + public LT(Node lhs, Node rhs) { super(lhs,rhs); } + String op() { return "<" ; } + TypeInteger doOp(TypeInteger i1, TypeInteger i2) { + if( i1._max < i2._min ) return TRUE; + if( i1._min >= i2._max ) return FALSE; + return BOOL; + } + Node copy(Node lhs, Node rhs) { return new LT(lhs,rhs); } + Node copyF() { return new LTF(null,null); } + } + public static class LE extends BoolNode { + public LE(Node lhs, Node rhs) { super(lhs,rhs); } + String op() { return "<="; } + TypeInteger doOp(TypeInteger i1, TypeInteger i2) { + if( i1._max <= i2._min ) return TRUE; + if( i1._min > i2._max ) return FALSE; + return BOOL; + } + Node copy(Node lhs, Node rhs) { return new LE(lhs,rhs); } + Node copyF() { return new LEF(null,null); } + } + + public static class EQF extends EQ { public EQF(Node lhs, Node rhs) { super(lhs,rhs); } boolean doOp(double lhs, double rhs) { return lhs == rhs; } } + public static class LTF extends LT { public LTF(Node lhs, Node rhs) { super(lhs,rhs); } boolean doOp(double lhs, double rhs) { return lhs < rhs; } } + public static class LEF extends LE { public LEF(Node lhs, Node rhs) { super(lhs,rhs); } boolean doOp(double lhs, double rhs) { return lhs <= rhs; } } + +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/CFGNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/CFGNode.java new file mode 100644 index 00000000..052ad8fc --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/CFGNode.java @@ -0,0 +1,132 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.Parser; +import com.seaofnodes.simple.type.*; +import com.seaofnodes.simple.Utils; + +import java.util.BitSet; +import java.util.HashSet; +import java.util.HashMap; + +/** Control Flow Graph Nodes + *

+ * CFG nodes have a immediate dominator depth (idepth) and a loop nesting + * depth(loop_depth). + *

+ * idepth is computed lazily upon first request, and is valid even in the + * Parser, and is used by peepholes during parsing and afterward. + *

+ * loop_depth is computed after optimization as part of scheduling. + * + */ +public abstract class CFGNode extends Node { + + public CFGNode(Node... nodes) { super(nodes); } + + public CFGNode cfg(int idx) { return (CFGNode)in(idx); } + + // Block head is Start, Region, CProj, but not e.g. If, Return, Stop + public boolean blockHead() { return false; } + + // ------------------------------------------------------------------------ + /** + * Immediate dominator tree depth, used to approximate a real IDOM during + * parsing where we do not have the whole program, and also peepholes + * change the CFG incrementally. + *

+ * See {@link ...} + */ + public int _idepth; + public int idepth() { return _idepth==0 ? (_idepth=idom().idepth()+1) : _idepth; } + + // Return the immediate dominator of this Node and compute dom tree depth. + public CFGNode idom(Node dep) { return cfg(0); } + public final CFGNode idom() { return idom(null); } + + // Return the LCA of two idoms + public CFGNode _idom(CFGNode rhs, Node dep) { + if( rhs==null ) return this; + CFGNode lhs = this; + while( lhs != rhs ) { + var comp = lhs.idepth() - rhs.idepth(); + if( comp >= 0 ) lhs = ((CFGNode)lhs.addDep(dep)).idom(); + if( comp <= 0 ) rhs = ((CFGNode)rhs.addDep(dep)).idom(); + } + return lhs; + } + + // Anti-dependence field support + public int _anti; // Per-CFG field to help find anti-deps + + // ------------------------------------------------------------------------ + // Loop nesting + public LoopNode loop() { return _ltree._head; } + public int loopDepth() { return _ltree==null ? 0 : _ltree.depth(); } + + LoopTree _ltree; + int _pre; // Pre-order numbers for loop tree finding + private static class LoopTree { + LoopTree _par; + final LoopNode _head; + int _depth; + LoopTree(LoopNode head) { _head = head; } + @Override public String toString() { return "LOOP"+_head._nid; } + int depth() { + return _depth==0 ? (_par==null ? 0 : (_depth = _par.depth()+1)) : _depth; + } + } + + // ------------------------------------------------------------------------ + // Tag all CFG Nodes with their containing LoopNode; LoopNodes themselves + // also refer to *their* containing LoopNode, as well as have their depth. + // Start is a LoopNode which contains all at depth 1. + public void buildLoopTree(StopNode stop) { + _ltree = stop._ltree = Parser.XCTRL._ltree = new LoopTree((StartNode)this); + _bltWalk(2,stop, new BitSet()); + } + private int _bltWalk(int pre, StopNode stop, BitSet post) { + // Pre-walked? + if( _pre!=0 ) return pre; + _pre = pre++; + // Pre-walk + for( Node use : _outputs ) + if( use instanceof CFGNode usecfg ) + pre = usecfg._bltWalk(pre,stop,post); + + // Post-order work: find innermost loop + LoopTree inner = null, ltree; + for( Node use : _outputs ) { + if( !(use instanceof CFGNode usecfg) ) continue; + // Child visited but not post-visited? + if( !post.get(usecfg._nid) ) { + // Must be a backedge to a LoopNode then + ltree = usecfg._ltree = new LoopTree((LoopNode)usecfg); + } else { + // Take child's loop choice, which must exist + ltree = usecfg._ltree; + // If falling into a loop, use the target loop's parent instead + if( ltree._head == usecfg ) { + if( ltree._par == null ) + // This loop never had an If test choose to take its + // exit, i.e. it is a no-exit infinite loop. + ltree._par = ltree._head.forceExit(stop)._ltree; + ltree = ltree._par; + } + } + // Sort inner loops. The decision point is some branch far removed + // from either loop head OR either backedge so requires pre-order + // numbers to figure out innermost. + if( inner == null ) { inner = ltree; continue; } + if( inner == ltree ) continue; // No change + LoopTree outer = ltree._head._pre > inner._head._pre ? inner : ltree; + inner = ltree._head._pre > inner._head._pre ? ltree : inner; + inner._par = outer; + } + // Set selected loop + if( inner!=null ) + _ltree = inner; + // Tag as post-walked + post.set(_nid); + return pre; + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/CProjNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/CProjNode.java new file mode 100644 index 00000000..d2f1bbe1 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/CProjNode.java @@ -0,0 +1,60 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.Parser; +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeMem; +import com.seaofnodes.simple.type.TypeTuple; +import java.util.BitSet; + +public class CProjNode extends CFGNode { + + // Which slice of the incoming multipart value + public final int _idx; + + // Debugging label + public final String _label; + + public CProjNode(Node ctrl, int idx, String label) { + super(ctrl); + _idx = idx; + _label = label; + } + + @Override public String label() { return _label; } + + @Override StringBuilder _print1(StringBuilder sb, BitSet visited) { return sb.append(_label); } + + @Override public boolean isMultiTail() { return in(0).isMultiHead(); } + @Override public boolean blockHead() { return true; } + + public CFGNode ctrl() { return cfg(0); } + + @Override + public Type compute() { + Type t = ctrl()._type; + return t instanceof TypeTuple tt ? tt._types[_idx] : Type.BOTTOM; + } + + @Override + public Node idealize() { + if( ctrl()._type instanceof TypeTuple tt ) { + if( tt._types[_idx]==Type.XCONTROL ) + return Parser.XCTRL; // We are dead + if( ctrl() instanceof IfNode && tt._types[1-_idx]==Type.XCONTROL ) // Only true for IfNodes + return ctrl().in(0); // We become our input control + } + + // Flip a negating if-test, to remove the not + if( ctrl() instanceof IfNode iff && iff.pred().addDep(this) instanceof NotNode not ) + return new CProjNode(new IfNode(iff.ctrl(),not.in(1)).peephole(),1-_idx,_idx==0 ? "False" : "True"); + + return null; + } + + @Override + boolean eq( Node n ) { return _idx == ((CProjNode)n)._idx; } + + @Override + int hash() { return _idx; } + +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/CastNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/CastNode.java new file mode 100644 index 00000000..462d6cba --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/CastNode.java @@ -0,0 +1,40 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeInteger; +import com.seaofnodes.simple.IterPeeps; + +import java.util.BitSet; + +// Upcast (join) the input to a t. Used after guard test to lift an input. +// Can also be used to make a type-assertion if ctrl is null. +public class CastNode extends Node { + private final Type _t; + public CastNode(Type t, Node ctrl, Node in) { + super(ctrl, in); + _t = t; + setType(compute()); + } + + @Override public String label() { return "("+_t.str()+")"; } + + @Override + public String uniqueName() { return "Cast_" + _nid; } + + @Override public boolean isPinned() { return true; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + return in(1)._print0(sb.append(label()), visited); + } + + @Override + public Type compute() { + return in(1)._type.join(_t); + } + + @Override + public Node idealize() { + return in(1)._type.isa(_t) ? in(1) : null; + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/ConstantNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/ConstantNode.java new file mode 100644 index 00000000..845b9d08 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/ConstantNode.java @@ -0,0 +1,54 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.Parser; +import com.seaofnodes.simple.type.Type; + +import java.util.BitSet; + +/** + * A Constant node represents a constant value. At present, the only constants + * that we allow are integer literals; therefore Constants contain an integer + * value. As we add other types of constants, we will refactor how we represent + * Constants. + *

+ * Constants have no semantic inputs. However, we set Start as an input to + * Constants to enable a forward graph walk. This edge carries no semantic + * meaning, and it is present solely to allow visitation. + *

+ * The Constant's value is the value stored in it. + */ +public class ConstantNode extends Node { + Type _con; + public ConstantNode( Type type ) { + super(Parser.START); + _con = type; + } + + @Override + public String label() { return "#"+_con; } + + @Override + public String uniqueName() { return "Con_" + _nid; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + return _con.print(sb); + } + + @Override public boolean isMultiTail() { return true; } + + @Override + public Type compute() { return _con; } + + @Override + public Node idealize() { return null; } + + @Override + boolean eq(Node n) { + ConstantNode con = (ConstantNode)n; // Contract + return _con==con._con; + } + + @Override + int hash() { return _con.hashCode(); } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/DivFNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/DivFNode.java new file mode 100644 index 00000000..36fd4524 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/DivFNode.java @@ -0,0 +1,46 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeFloat; + +import java.util.BitSet; + +public class DivFNode extends Node { + public DivFNode(Node lhs, Node rhs) { super(null, lhs, rhs); } + + @Override public String label() { return "DivF"; } + + @Override public String glabel() { return "/"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + in(1)._print0(sb.append("("), visited); + in(2)._print0(sb.append("/"), visited); + return sb.append(")"); + } + + @Override + public Type compute() { + if (in(1)._type instanceof TypeFloat i0 && + in(2)._type instanceof TypeFloat i1) { + if (i0.isConstant() && i1.isConstant()) + return TypeFloat.constant(i0.value()/i1.value()); + } + return in(1)._type.meet(in(2)._type); + } + + @Override + public Node idealize() { + Node lhs = in(1); + Node rhs = in(2); + Type t1 = lhs._type; + Type t2 = rhs._type; + + // Div of 1. + if ( t2.isConstant() && t2 instanceof TypeFloat i && i.value()==1 ) + return lhs; + + return null; + } + @Override Node copy(Node lhs, Node rhs) { return new DivFNode(lhs,rhs); } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/DivNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/DivNode.java new file mode 100644 index 00000000..76e98678 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/DivNode.java @@ -0,0 +1,47 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeInteger; + +import java.util.BitSet; + +public class DivNode extends Node { + public DivNode(Node lhs, Node rhs) { super(null, lhs, rhs); } + + @Override public String label() { return "Div"; } + + @Override public String glabel() { return "//"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + in(1)._print0(sb.append("("), visited); + in(2)._print0(sb.append("/"), visited); + return sb.append(")"); + } + + @Override + public Type compute() { + Type t1 = in(1)._type, t2 = in(2)._type; + if( t1.isHigh() || t2.isHigh() ) + return TypeInteger.TOP; + if( t1 instanceof TypeInteger i1 && + t2 instanceof TypeInteger i2 ) { + if (i1.isConstant() && i2.isConstant()) + return i2.value() == 0 + ? TypeInteger.ZERO + : TypeInteger.constant(i1.value()/i2.value()); + } + return TypeInteger.BOT; + } + + @Override + public Node idealize() { + // Div of 1. + if( in(2)._type == TypeInteger.TRUE ) + return in(1); + return null; + } + + @Override Node copy(Node lhs, Node rhs) { return new DivNode(lhs,rhs); } + @Override Node copyF() { return new DivFNode(null,null); } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/IfNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/IfNode.java new file mode 100644 index 00000000..513b4721 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/IfNode.java @@ -0,0 +1,66 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.IterPeeps; +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeInteger; +import com.seaofnodes.simple.type.TypeTuple; +import java.util.BitSet; +import java.util.HashSet; + +public class IfNode extends CFGNode implements MultiNode { + + public IfNode(Node ctrl, Node pred) { + super(ctrl, pred); + IterPeeps.add(this); // Because idoms are complex, just add it + } + + @Override + public String label() { return "If"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + sb.append("if( "); + return in(1)._print0(sb, visited).append(" )"); + } + + @Override public boolean isMultiHead() { return true; } + + public Node ctrl() { return in(0); } + public Node pred() { return in(1); } + + @Override + public Type compute() { + // If the If node is not reachable then neither is any following Proj + if (ctrl()._type != Type.CONTROL && ctrl()._type != Type.BOTTOM ) + return TypeTuple.IF_NEITHER; + Node pred = pred(); + Type t = pred._type; + // High types mean NEITHER side is reachable. + // Wait until the type falls to decide which way to go. + if( t == Type.TOP || t == TypeInteger.TOP ) + return TypeTuple.IF_NEITHER; + // If constant is 0 then false branch is reachable + // Else true branch is reachable + if( t.isConstant() ) + return t.makeInit()==t ? TypeTuple.IF_FALSE : TypeTuple.IF_TRUE; + // Integers allow non-zero ranges: "1-65535" + if( t.makeZero().meet(t) != t ) + return TypeTuple.IF_TRUE; + + return TypeTuple.IF_BOTH; + } + + @Override + public Node idealize() { + // Hunt up the immediate dominator tree. If we find an identical if + // test on either the true or false branch, that side wins. + if( !pred()._type.isHighOrConst() ) + for( CFGNode dom = idom(), prior=this; dom!=null; prior = dom, dom = dom.idom() ) + if( dom.addDep(this) instanceof IfNode iff && iff.pred().addDep(this)==pred() && prior instanceof CProjNode prj ) { + setDef(1,new ConstantNode(TypeInteger.make(true,prj._idx==0?1:0)).peephole()); + return this; + } + return null; + } + +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/LoadNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/LoadNode.java new file mode 100644 index 00000000..1528f2e9 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/LoadNode.java @@ -0,0 +1,172 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.Utils; +import com.seaofnodes.simple.Parser; +import com.seaofnodes.simple.type.*; +import java.util.BitSet; + +/** + * Load represents extracting a value from inside a memory object, + * in chapter 10 this means Struct fields. + */ +public class LoadNode extends MemOpNode { + + /** + * Load a value from a ptr.field. + * + * @param name The field we are loading + * @param mem The memory alias node - this is updated after a Store + * @param ptr The ptr to the struct base from where we load a field + * @param off The offset inside the struct base + */ + public LoadNode(String name, int alias, Type glb, Node mem, Node ptr, Node off) { + super(name, alias, glb, mem, ptr, off); + } + + // GraphVis DOT code (must be valid Java identifiers) and debugger labels + @Override public String label() { return "ld_"+mlabel(); } + // GraphVis node-internal labels + @Override public String glabel() { return "." +_name; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { return sb.append(".").append(_name); } + + @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; + } + + @Override + public Node idealize() { + 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 + 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 + return nnn.in(nnn.findAlias(_alias)); // Load from New init + + // Load-after-Store on same address, but bypassing provably unrelated + // stores. This is a more complex superset of the above two peeps. + // "Provably unrelated" is really weak. + if( ptr instanceof ReadOnlyNode ro ) + ptr = ro.in(1); + Node mem = mem(); + outer: + while( true ) { + switch( mem ) { + case StoreNode st: + if( 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 + !neverAlias(ptr,st.ptr()) ) // And might alias + break outer; // Cannot tell, stop trying + // Pointers cannot overlap + mem = st.mem(); // Proved never equal + break; + case PhiNode phi: break outer; // Assume related + case ConstantNode top: break outer; // Assume shortly dead + case ProjNode mproj: + if( mproj.in(0) instanceof NewNode nnn1 ) { + if( ptr instanceof ProjNode pproj && pproj.in(0) == mproj.in(0) ) + 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 + break; + } else if( mproj.in(0) instanceof StartNode ) { + break outer; + } else throw Utils.TODO(); + default: + throw Utils.TODO(); + } + } + + // Push a Load up through a Phi, as long as it collapses on at least + // one arm. If at a Loop, the backedge MUST collapse - else we risk + // spinning the same transform around the loop indefinitely. + // BEFORE (2 Sts, 1 Ld): AFTER (1 St, 0 Ld): + // 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 && + // Offset can be hoisted + off() instanceof ConstantNode && + // Pointer can be hoisted + hoistPtr(ptr,memphi) ) { + + // Profit on RHS/Loop backedge + if( profit(memphi,2) || + // Else must not be a loop to count profit on LHS. + (!(memphi.region() instanceof LoopNode) && profit(memphi,1)) ) { + Node ld1 = ld(1); + Node ld2 = ld(2); + return new PhiNode(_name,_declaredType,memphi.region(),ld1,ld2); + } + } + + return null; + } + + 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(); + } + + private static boolean neverAlias( Node ptr1, Node ptr2 ) { + return ptr1.in(0) != ptr2.in(0) && + // Unrelated allocations + ptr1 instanceof ProjNode && ptr1.in(0) instanceof NewNode && + ptr2 instanceof ProjNode && ptr2.in(0) instanceof NewNode; + } + + private static boolean hoistPtr(Node ptr, PhiNode memphi ) { + // Can I hoist ptr above the Region? + if( !(memphi.region() instanceof RegionNode r) ) + return false; // Dead or dying Region/Phi + // If ptr from same Region, then yes, just use hoisted split pointers + if( ptr instanceof PhiNode pphi && pphi.region() == r ) + return true; + + // No, so can we lift this ptr? + CFGNode cptr = ptr.cfg0(); + if( cptr != null ) + // Pointer is controlled high + // TODO: Really needs to be the LCA of all inputs is high + return cptr.idepth() <= r.idepth(); + + // Dunno without a longer walk + return false; + } + + // Profitable if we find a matching Store on this Phi arm. + private boolean profit(PhiNode phi, int idx) { + Node px = phi.in(idx); + if( px==null ) return false; + if( px._type instanceof TypeMem mem && mem._t.isHighOrConst() ) { px.addDep(this); return true; } + if( px instanceof StoreNode st1 && ptr()==st1.ptr() && off()==st1.off() ) { px.addDep(this); return true; } + return false; + } + + // Read-Only is a deep property, and cannot be cast-away + private Node castRO(Node rez) { + if( ptr()._type.isFinal() && !rez._type.isFinal() ) + return new ReadOnlyNode(rez).peephole(); + return rez; + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/LoopNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/LoopNode.java new file mode 100644 index 00000000..ecda9c63 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/LoopNode.java @@ -0,0 +1,61 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.Parser; +import com.seaofnodes.simple.Utils; +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeInteger; +import java.util.BitSet; +import java.util.HashSet; + +public class LoopNode extends RegionNode { + public LoopNode( Node entry ) { super(null,entry,null); } + + public CFGNode entry() { return cfg(1); } + public CFGNode back () { return cfg(2); } + + @Override + public String label() { return "Loop"; } + + @Override + public Type compute() { + if( inProgress() ) return Type.CONTROL; + return entry()._type; + } + + @Override + public Node idealize() { + return inProgress() ? null : super.idealize(); + } + + // Bypass Region idom, same as the default idom() using use in(1) instead of in(0) + @Override public int idepth() { return _idepth==0 ? (_idepth=idom().idepth()+1) : _idepth; } + // Bypass Region idom, same as the default idom() using use in(1) instead of in(0) + @Override public CFGNode idom(Node dep) { return entry(); } + + // If this is an unreachable loop, it may not have an exit. If it does not + // (i.e., infinite loop), force an exit to make it reachable. + public StopNode forceExit( StopNode stop ) { + // Walk the backedge, then immediate dominator tree util we hit this + // Loop again. If we ever hit a CProj from an If (as opposed to + // directly on the If) we found our exit. + CFGNode x = back(); + while( x != this ) { + if( x instanceof CProjNode exit && exit.in(0) instanceof IfNode iff ) { + CFGNode other = iff.cproj(1-exit._idx); + if( other.loopDepth() < loopDepth() ) + return stop; // Found an exit, not an infinite loop + } + x = x.idom(); + } + // Found a no-exit loop. Insert an exit + NeverNode iff = new NeverNode(back()); + for( Node use : _outputs ) + if( use instanceof PhiNode phi ) + iff.addDef(use); + CProjNode t = new CProjNode(iff,0,"True" ); + CProjNode f = new CProjNode(iff,1,"False"); + setDef(2,f); + stop.addDef(new ReturnNode(t,Parser.ZERO,null)); + return stop; + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/MemOpNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/MemOpNode.java new file mode 100644 index 00000000..82aab8ac --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/MemOpNode.java @@ -0,0 +1,55 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeMemPtr; + +/** + * Convenience common base for Load and Store. + */ +public abstract class MemOpNode extends Node { + + public final String _name; + public final int _alias; + // Declared type; not final because it might be a forward-reference + // which will be lazily improved when the reference is declared. + public Type _declaredType; + + public MemOpNode(String name, int alias, Type glb, Node mem, Node ptr, Node off) { + super(null, mem, ptr, off); + _name = name; + _alias = alias; + _declaredType = glb; + } + public MemOpNode(String name, int alias, Type glb, Node mem, Node ptr, Node off, Node value) { + this(name, alias, glb, mem, ptr, off); + addDef(value); + } + + // + static String mlabel(String name) { return name.equals("[]") ? "ary" : (name.equals("#") ? "len" : name); } + String mlabel() { return mlabel(_name); } + + public Node mem() { return in(1); } + public Node ptr() { return in(2); } + public Node off() { return in(3); } + + @Override + boolean eq(Node n) { + MemOpNode mem = (MemOpNode)n; // Invariant + return _alias==mem._alias; // When comparing types error to use "equals"; always use "==" + } + + @Override + int hash() { return _alias; } + + @Override + String err() { + Type ptr = ptr()._type; + // Already an error, but better error messages come from elsewhere + if( ptr == Type.BOTTOM ) return null; + // Better be a not-nil TMP + if( ptr instanceof TypeMemPtr tmp && !tmp._nil ) + return null; + return "Might be null accessing '" + _name + "'"; + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/MinusFNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/MinusFNode.java new file mode 100644 index 00000000..9727f58b --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/MinusFNode.java @@ -0,0 +1,36 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeFloat; + +import java.util.BitSet; + +public class MinusFNode extends Node { + public MinusFNode(Node in) { super(null, in); } + + @Override public String label() { return "MinusF"; } + + @Override public String glabel() { return "-"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + in(1)._print0(sb.append("(-"), visited); + return sb.append(")"); + } + + @Override + public Type compute() { + if (in(1)._type instanceof TypeFloat i0) + return i0.isConstant() ? TypeFloat.constant(-i0.value()) : i0; + return TypeFloat.BOT; + } + + @Override + public Node idealize() { + // -(-x) is x + if( in(1) instanceof MinusFNode minus ) + return minus.in(1); + + return null; + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/MinusNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/MinusNode.java new file mode 100644 index 00000000..760a4274 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/MinusNode.java @@ -0,0 +1,42 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeInteger; +import com.seaofnodes.simple.Utils; + +import java.util.BitSet; + +public class MinusNode extends Node { + public MinusNode(Node in) { super(null, in); } + + @Override public String label() { return "Minus"; } + + @Override public String glabel() { return "-"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + in(1)._print0(sb.append("(-"), visited); + return sb.append(")"); + } + + @Override + public Type compute() { + if( in(1)._type instanceof TypeInteger i0 ) { + if( i0.isHigh() ) return TypeInteger.TOP; + if( i0==TypeInteger.BOT || i0._min == Long.MIN_VALUE || i0._max == Long.MIN_VALUE ) + return TypeInteger.BOT; + return TypeInteger.make(-i0._max,-i0._min); + } + return TypeInteger.TOP.meet(in(1)._type); + } + + @Override + public Node idealize() { + // -(-x) is x + if( in(1) instanceof MinusNode minus ) + return minus.in(1); + + return null; + } + @Override Node copyF() { return new MinusFNode(null); } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/MulFNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/MulFNode.java new file mode 100644 index 00000000..2d44a9fb --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/MulFNode.java @@ -0,0 +1,51 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeFloat; + +import java.util.BitSet; + +public class MulFNode extends Node { + public MulFNode(Node lhs, Node rhs) { super(null, lhs, rhs); } + + @Override public String label() { return "MulF"; } + + @Override public String glabel() { return "*"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + in(1)._print0(sb.append("("), visited); + in(2)._print0(sb.append("*"), visited); + return sb.append(")"); + } + + @Override + public Type compute() { + if (in(1)._type instanceof TypeFloat i0 && + in(2)._type instanceof TypeFloat i1) { + if (i0.isConstant() && i1.isConstant()) + return TypeFloat.constant(i0.value()*i1.value()); + } + return in(1)._type.meet(in(2)._type); + } + + @Override + public Node idealize() { + Node lhs = in(1); + Node rhs = in(2); + Type t1 = lhs._type; + Type t2 = rhs._type; + + // Mul of 1. We do not check for (1*x) because this will already + // canonicalize to (x*1) + if ( t2.isConstant() && t2 instanceof TypeFloat i && i.value()==1 ) + return lhs; + + // Move constants to RHS: con*arg becomes arg*con + if ( t1.isConstant() && !t2.isConstant() ) + return swap12(); + + return null; + } + @Override Node copy(Node lhs, Node rhs) { return new MulFNode(lhs,rhs); } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/MulNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/MulNode.java new file mode 100644 index 00000000..b5ef91e9 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/MulNode.java @@ -0,0 +1,63 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeInteger; + +import java.util.BitSet; + +public class MulNode extends Node { + public MulNode(Node lhs, Node rhs) { super(null, lhs, rhs); } + + @Override public String label() { return "Mul"; } + + @Override public String glabel() { return "*"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + in(1)._print0(sb.append("("), visited); + in(2)._print0(sb.append("*"), visited); + return sb.append(")"); + } + + @Override + public Type compute() { + Type t1 = in(1)._type, t2 = in(2)._type; + if( t1.isHigh() || t2.isHigh() ) + return TypeInteger.TOP; + if( t1 instanceof TypeInteger i1 && + t2 instanceof TypeInteger i2 ) { + if( i1==TypeInteger.ZERO || i2==TypeInteger.ZERO) + return TypeInteger.ZERO; + if (i1.isConstant() && i2.isConstant()) + return TypeInteger.constant(i1.value()*i2.value()); + } + return TypeInteger.BOT; + } + + @Override + public Node idealize() { + Node lhs = in(1); + Node rhs = in(2); + Type t1 = lhs._type; + Type t2 = rhs._type; + + // Mul of 1. We do not check for (1*x) because this will already + // canonicalize to (x*1) + if ( t2.isConstant() && t2 instanceof TypeInteger i && i.value()==1 ) + return lhs; + + // Move constants to RHS: con*arg becomes arg*con + if ( t1.isConstant() && !t2.isConstant() ) + return swap12(); + + // Do we have ((x * (phi cons)) * con) ? + // Do we have ((x * (phi cons)) * (phi cons)) ? + // Push constant up through the phi: x * (phi con0*con0 con1*con1...) + Node phicon = AddNode.phiCon(this,true); + if( phicon!=null ) return phicon; + + return null; + } + @Override Node copy(Node lhs, Node rhs) { return new MulNode(lhs,rhs); } + @Override Node copyF() { return new MulFNode(null,null); } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/MultiNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/MultiNode.java new file mode 100644 index 00000000..ac9c0045 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/MultiNode.java @@ -0,0 +1,20 @@ +package com.seaofnodes.simple.node; + +public interface MultiNode extends OutNode { + + // Find a projection by index + default ProjNode proj( int idx ) { + for( Node out : outs() ) + if( out instanceof ProjNode prj && prj._idx==idx ) + return prj; + return null; + } + + // Find a projection by index + default CProjNode cproj( int idx ) { + for( Node out : outs() ) + if( out instanceof CProjNode prj && prj._idx==idx ) + return prj; + return null; + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/NeverNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/NeverNode.java new file mode 100644 index 00000000..e77ed653 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/NeverNode.java @@ -0,0 +1,21 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.Parser; +import com.seaofnodes.simple.type.TypeInteger; +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeTuple; + +import java.util.BitSet; + +// "Never true" for infinite loop exits +public class NeverNode extends IfNode { + public NeverNode(Node ctrl) { super(ctrl,Parser.ZERO); } + + @Override public String label() { return "Never"; } + + @Override StringBuilder _print1(StringBuilder sb, BitSet visited) { return sb.append("Never"); } + + @Override public Type compute() { return TypeTuple.IF_BOTH; } + + @Override public Node idealize() { return null; } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/NewNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/NewNode.java new file mode 100644 index 00000000..ddb1e09a --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/NewNode.java @@ -0,0 +1,75 @@ +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. + */ +public class NewNode extends Node implements MultiNode { + + public final TypeMemPtr _ptr; + public final int _len; + + public NewNode(TypeMemPtr ptr, Node... nodes) { + super(nodes); + _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; + for( int i=0; i<_len; i++ ) { + // Memory slices for all fields. + assert nodes[2+ i]._type instanceof TypeMem; + // Value slices for all fields. + assert nodes[2 + _len + i]._type != null; + } + } + + @Override public String label() { + return "new_"+glabel(); + } + @Override public String glabel() { + return _ptr._obj.isAry() ? "ary_"+_ptr._obj._fields[1]._type.str() : _ptr._obj.str(); + } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + sb.append("new "); + return sb.append(_ptr._obj.str()); + } + + // Find matching alias input + int findAlias(int alias) { + return 2+_ptr._obj.findAlias(alias)+_len; + } + + + @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 + * Generally fixed length, ordered, nulls allowed, no unused trailing space. + * Ordering is required because e.g. "a/b" is different from "b/a". + * The first input (offset 0) is often a {@link CFGNode} node. + */ + public final Ary _inputs; + + /** + * Outputs reference Nodes that are not null and have this Node as an + * input. These nodes are users of this node, thus these are def-use + * references to Nodes. + *

+ * Outputs directly match inputs, making a directed graph that can be + * walked in either direction. These outputs are typically used for + * efficient optimizations but otherwise have no semantics meaning. + */ + public final Ary _outputs; + + + /** + * Current computed type for this Node. This value changes as the graph + * changes and more knowledge is gained about the program. + */ + public Type _type; + + + /** + * A private Global Static mutable counter, for unique node id generation. + * To make the compiler multithreaded, this field will have to move into a TLS. + * Starting with value 1, to avoid bugs confusing node ID 0 with uninitialized values. + * */ + private static int UNIQUE_ID = 1; + public static int UID() { return UNIQUE_ID; } + + protected Node(Node... inputs) { + _nid = UNIQUE_ID++; // allocate unique dense ID + _inputs = new Ary<>(Node.class); + Collections.addAll(_inputs,inputs); + _outputs = new Ary<>(Node.class); + for( Node n : _inputs ) + if( n != null ) + n.addUse( this ); + } + + // Easy reading label for debugger, e.g. "Add" or "Region" or "EQ" + public abstract String label(); + + // Unique label for graph visualization, e.g. "Add12" or "Region30" or "EQ99" + public String uniqueName() { + // Get rid of $ as graphviz doesn't like it + String label = label().replaceAll("\\$", ""); + return label + _nid; + } + + // Graphical label, e.g. "+" or "Region" or "==" + public String glabel() { return label(); } + + + // ------------------------------------------------------------------------ + + // Debugger Printing. + + // {@code toString} is what you get in the debugger. It has to print 1 + // line (because this is what a debugger typically displays by default) and + // has to be robust with broken graph/nodes. + @Override + public final String toString() { return print(); } + + // This is a *deep* print. We print with a tik-tok style; the common + // _print0 calls the per-Node _print1, which calls back to _print0; + public final String print() { + return _print0(new StringBuilder(), new BitSet()).toString(); + } + + // This is the common print: check for repeats, check for DEAD and print + // "DEAD" else call the per-Node print1. + final StringBuilder _print0(StringBuilder sb, BitSet visited) { + if (visited.get(_nid) && !(this instanceof ConstantNode) ) + return sb.append(label()); + visited.set(_nid); + return isDead() + ? sb.append(uniqueName()).append(":DEAD") + : _print1(sb, visited); + } + // Every Node implements this; a partial-line recursive print + abstract StringBuilder _print1(StringBuilder sb, BitSet visited); + + public String p(int depth) { return IRPrinter.prettyPrint(this,depth); } + + public boolean isMultiHead() { return false; } + public boolean isMultiTail() { return false; } + + // ------------------------------------------------------------------------ + // Graph Node & Edge manipulation + + /** + * Gets the ith input node + * @param i Offset of the input node + * @return Input node or null + */ + public Node in(int i) { return _inputs.get(i); } + public Node out(int i) { return _outputs.get(i); } + @Override public Ary outs() { return _outputs; } + + public int nIns() { return _inputs.size(); } + + public int nOuts() { return _outputs.size(); } + + public boolean isUnused() { return nOuts() == 0; } + + public CFGNode cfg0() { return (CFGNode)in(0); } + + /** + * Change a def into a Node. Keeps the edges correct, by removing + * the corresponding use->def edge. This may make the original + * def go dead. This function is co-recursive with {@link #kill}. + *

+ + * This method is the normal path for altering a Node, because it does the + * proper default edge maintenance. It also immediately kills + * Nodes that lose their last use; at times care must be taken to avoid + * killing Nodes that are being used without having an output Node. This + * definitely happens in the middle of recursive {@link #peephole} calls. + * + * @param idx which def to set + * @param new_def the new definition + * @return new_def for flow coding + */ + public N setDef(int idx, N new_def ) { + unlock(); + Node old_def = in(idx); + if( old_def == new_def ) return new_def; // No change + // If new def is not null, add the corresponding def->use edge + // This needs to happen before removing the old node's def->use edge as + // the new_def might get killed if the old node kills it recursively. + if( new_def != null ) + new_def.addUse(this); + // Set the new_def over the old (killed) edge + _inputs.set(idx,new_def); + if( old_def != null && // If the old def exists, remove a def->use edge + old_def.delUse(this) ) // If we removed the last use, the old def is now dead + old_def.kill(); // Kill old def + moveDepsToWorklist(); + // Return new_def for easy flow-coding + return new_def; + } + + // Remove the numbered input, compressing the inputs in-place. This + // shuffles the order deterministically - which is suitable for Region and + // Phi, but not for every Node. If the def goes dead, it is recursively + // killed, which may include 'this' Node. + Node delDef(int idx) { + unlock(); + Node old_def = in(idx); + _inputs.del(idx); + if( old_def.delUse(this) ) // If we removed the last use, the old def is now dead + old_def.kill(); // Kill old def + old_def.moveDepsToWorklist(); + return this; + } + + // Insert the numbered input, sliding other inputs to the right + Node insertDef(int idx, Node new_def) { + _inputs.add(idx,null); + return setDef(idx,new_def); + } + + + /** + * Add a new def to an existing Node. Keep the edges correct by + * adding the corresponding def->use edge. + * + * @param new_def the new definition, appended to the end of existing definitions + * @return new_def for flow coding + */ + public Node addDef(Node new_def) { + unlock(); + // Add use->def edge + _inputs.add(new_def); + // If new def is not null, add the corresponding def->use edge + if( new_def != null ) + new_def.addUse(this); + return new_def; + } + + // Breaks the edge invariants, used temporarily + protected N addUse(Node n) { _outputs.add(n); return (N)this; } + + // Remove node 'use' from 'def's (i.e. our) output list, by compressing the list in-place. + // Return true if the output list is empty afterward. + // Error is 'use' does not exist; ok for 'use' to be null. + protected boolean delUse( Node use ) { + _outputs.del(_outputs.find(use)); + return _outputs.isEmpty(); + } + + // Shortcut for "popping" until n nodes. A "pop" is basically a + // setDef(last,null) followed by lowering the nIns() count. + void popUntil(int n) { + unlock(); + while( nIns() > n ) { + Node old_def = _inputs.pop(); + if( old_def != null && // If it exists and + old_def.delUse(this) ) // If we removed the last use, the old def is now dead + old_def.kill(); // Kill old def + } + } + + /** + * Kill a Node with no uses, by setting all of its defs + * to null. This may recursively kill more Nodes, and is basically dead + * code elimination. + */ + public void kill( ) { + unlock(); + moveDepsToWorklist(); + assert isUnused(); // Has no uses, so it is dead + _type=null; // Flag as dead + while( nIns()>0 ) { // Set all inputs to null, recursively killing unused Nodes + Node old_def = _inputs.removeLast(); + if( old_def != null ) { + IterPeeps.add(old_def);// Revisit neighbor because removed use + if( old_def.delUse(this) ) // If we removed the last use, the old def is now dead + old_def.kill(); // Kill old def + } + } + assert isDead(); // Really dead now + } + + // Mostly used for asserts and printing. + public boolean isDead() { return isUnused() && nIns()==0 && _type==null; } + + // Shortcuts to stop DCE mid-parse + // Add bogus null use to keep node alive + public N keep() { return addUse(null); } + // Remove bogus null. + public N unkeep() { + delUse(null); + return (N)this; + } + // Test "keep" status + public boolean iskeep() { return _outputs.find(null) != -1; } + public void unkill() { + if( unkeep().isUnused() ) + kill(); + } + + + // Replace self with nnn in the graph, making 'this' go dead + public void subsume( Node nnn ) { + assert nnn!=this; + while( nOuts() > 0 ) { + Node n = _outputs.removeLast(); + n.unlock(); + int idx = n._inputs.find(this); + n._inputs.set(idx,nnn); + nnn.addUse(n); + IterPeeps.addAll(n._outputs); + } + kill(); + } + + // ------------------------------------------------------------------------ + // Graph-based optimizations + + /** + * We allow disabling peephole opt so that we can observe the + * full graph, vs the optimized graph. + */ + public static boolean _disablePeephole = false; + + /** + * Try to peephole at this node and return a better replacement Node. + * Always returns some not-null Node (often this). + */ + public final Node peephole( ) { + if (_disablePeephole) { + _type = compute(); + return this; // Peephole optimizations turned off + } + Node n = peepholeOpt(); + return n==null ? this : deadCodeElim(n.peephole()); // Cannot return null for no-progress + } + + /** + * Try to peephole at this node and return a better replacement Node if + * possible. We compute a {@link Type} and then check and replace: + *

    + *
  • if the Type {@link Type#isConstant}, we replace with a {@link ConstantNode}
    • + *
  • in a future chapter we will look for a + * Common Subexpression + * to eliminate.
    • + *
  • we ask the Node for a better replacement. The "better replacement" + * is things like {@code (1+2)} becomes {@code 3} and {@code (1+(x+2))} becomes + * {@code (x+(1+2))}. By canonicalizing expressions we fold common addressing + * math constants, remove algebraic identities and generally simplify the + * code.
    • + * + * Unlike peephole above, this explicitly returns null for no-change, or not-null + * for a better replacement (which can be this). + *
+ */ + public final Node peepholeOpt( ) { + ITER_CNT++; + // Compute initial or improved Type + Type old = setType(compute()); + + // Replace constant computations from non-constants with a constant node + if( !(this instanceof ConstantNode) && !(this instanceof XCtrlNode) && _type.isHighOrConst() ) + return (_type==Type.XCONTROL ? new XCtrlNode() : new ConstantNode(_type)).peepholeOpt(); + + // Global Value Numbering + if( _hash==0 ) { + Node n = GVN.get(this); // Will set _hash as a side effect + if( n==null ) + GVN.put(this,this); // Put in table now + else { + // Because of random worklist ordering, the two equal nodes + // might have different types. Because of monotonicity, both + // types are valid. To preserve monotonicity, the resulting + // shared Node has to have the best of both types. + n.setType(n._type.join(_type)); + _hash = 0; // Clear, since it never went in the table + return deadCodeElim(n);// Return previous; does Common Subexpression Elimination + } + } + + // Ask each node for a better replacement + Node n = idealize(); + if( n != null ) // Something changed + return n; // Report progress + + if( old==_type ) ITER_NOP_CNT++; + return old==_type ? null : this; // Report progress + } + public static int ITER_CNT, ITER_NOP_CNT; + + // m is the new Node, self is the old. + // Return 'm', which may have zero uses but is alive nonetheless. + // If self has zero uses (and is not 'm'), {@link #kill} self. + private Node deadCodeElim(Node m) { + // If self is going dead and not being returned here (Nodes returned + // from peephole commonly have no uses (yet)), then kill self. + if( m != this && isUnused() && !isDead() ) { + // Killing self - and since self recursively kills self's inputs we + // might end up killing 'm', which we are returning as a live Node. + // So we add a bogus extra null output edge to stop kill(). + m.keep(); // Keep m alive + kill(); // Kill self because replacing with 'm' + m.unkeep(); // Okay to peephole m + } + return m; + } + + /** + * This function needs to be + * Monotonic + * as it is part of a Monotone Analysis Framework. + * See for example this set of slides. + *

+ * For Chapter 2, all our Types are really integer constants, and so all + * the needed properties are trivially true, and we can ignore the high + * theory. Much later on, this will become important and allow us to do + * many fancy complex optimizations trivially... because theory. + *

+ * {@link #compute()} needs to be stand-alone, and cannot recursively call compute + * on its inputs, because programs are cyclic (have loops!) and this will just + * infinitely recurse until stack overflow. Instead, compute typically + * computes a new type from the {@link #_type} field of its inputs. + */ + public abstract Type compute(); + + // Set the type. Assert monotonic progress. + // If changing, add users to worklist. + public Type setType(Type type) { + Type old = _type; + assert old==null || type.isa(old); // Since _type not set, can just re-run this in assert in the debugger + if( old == type ) return old; + _type = type; // Set _type late for easier assert debugging + IterPeeps.addAll(_outputs); + moveDepsToWorklist(); + return old; + } + + + /** + * This function rewrites the current Node into a more "idealized" form. + * This is the bulk of our peephole rewrite rules, and we use this to + * e.g. turn arbitrary collections of adds and multiplies with mixed + * constants into a normal form that's easy for hardware to implement. + * Example: An array addressing expression: + * 

   ary[idx+1]
+ * might turn into Sea-of-Nodes IR: + * 
   (ary+12)+((idx+1) * 4)
+ * This expression can then be idealized into: + * 
   ary + ((idx*4) + (12 + (1*4)))
+ * And more folding: + * 
   ary + ((idx<<2) + 16)
+ * And during code-gen: + * 
   MOV4 Rary,[Ridx<<2 +16] // or some such hardware-specific notation
+ *

+ * {@link #idealize} has a very specific calling convention: + *

    + *
  • If NO change is made, return {@code null} + *
  • If ANY change is made, return not-null; this can be {@code this} + *
  • The returned Node does NOT call {@link #peephole} on itself; the {@link #peephole} call will recursively peephole it. + *
  • Any NEW nodes that are not directly returned DO call {@link #peephole}. + *
+ *

+ * Examples: + * + * + * + * + * + *
before after return comment
{@code (x+5) }{@code (x+5) }{@code null }No change
{@code (5+x) }{@code (x+5) }{@code this }Swapped arguments
{@code ((x+1)+2)}{@code (x+(1+2))}{@code (x+_) }Returns 2 new Nodes
+ * + * The last entry deserves more discussion. The new Node {@code (1+2)} + * created in {@link #idealize} calls {@link #peephole} (which then folds + * into a constant). The other new Node {@code (x+3)} does not call + * peephole, because it is returned and peephole itself will recursively + * call peephole. + *

+ * Since idealize calls peephole and peephole calls idealize, you must be + * careful that all idealizations are monotonic: all transforms + * remove some feature, so that the set of available transforms always + * shrinks. If you don't, you risk an infinite peephole loop! + * + * @return Either a new or changed node, or null for no changes. + */ + public abstract Node idealize(); + + + // Some of the peephole rules get complex, and search further afield than + // just the nearest neighbor. These peepholes can fail the pattern match + // on a node some distance away, and if that node ever changes we should + // retry the peephole. Track a set of Nodes dependent on `this`, and + // revisit them if `this` changes. + Ary _deps; + + /** + * Add a node to the list of dependencies. Only add it if its not an input + * or output of this node, that is, it is at least one step away. The node + * being added must benefit from this node being peepholed. + */ + Node addDep( Node dep ) { + // Running peepholes during the big assert cannot have side effects + // like adding dependencies. + if( IterPeeps.midAssert() ) 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; + _deps.add(dep); + return this; + } + + // Move the dependents onto a worklist, and clear for future dependents. + public void moveDepsToWorklist( ) { + if( _deps==null ) return; + IterPeeps.addAll(_deps); + _deps.clear(); + } + + + // Global Value Numbering. Hash over opcode and inputs; hits in this table + // are structurally equal. + public static final HashMap GVN = new HashMap<>(); + + // Two nodes are equal if they have the same inputs and the same "opcode" + // which means the same Java class, plus same internal parts. + @Override public final boolean equals( Object o ) { + if( o==this ) return true; + if( o.getClass() != getClass() ) return false; + Node n = (Node)o; + int len = _inputs.size(); + if( len != n._inputs.size() ) return false; + for( int i=0; i>13) ^ n._nid; + if( hash==0 ) hash = 0xDEADBEEF; // Bad hash, so use some junky thing + return (_hash=hash); + } + // Subclasses add extra hash info (such as ConstantNodes constant) + int hash() { return 0; } + + // ------------------------------------------------------------------------ + // + + /** Is this Node Memory related */ + public boolean isMem() { return false; } + + /** Pinned in the schedule; these are data nodes whose input#0 is not allowed to change */ + public boolean isPinned() { return false; } + + // Semantic change to the graph (so NOT a peephole), used by the Parser. + // If any input is a float, flip to a float-flavored opcode and widen any + // non-float input. + public final Node widen() { + if( !hasFloatInput() ) return this; + Node flt = copyF(); + if( flt==null ) return this; + for( int i=1; i E walk( Function pred ) { + assert WVISIT.isEmpty(); + E rez = _walk(pred); + WVISIT.clear(); + return rez; + } + + private E _walk( Function pred ) { + if( WVISIT.get(_nid) ) return null; // Been there, done that + WVISIT.set(_nid); + E x = pred.apply(this); + if( x != null ) return x; + for( Node def : _inputs ) if( def != null && (x = def._walk(pred)) != null ) return x; + for( Node use : _outputs ) if( use != null && (x = use._walk(pred)) != null ) return x; + return null; + } + + /** + * Debugging utility to find a Node by index + */ + public Node find(int nid) { + return walk( n -> n._nid==nid ? n : null ); + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/NotNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/NotNode.java new file mode 100644 index 00000000..664d7c7b --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/NotNode.java @@ -0,0 +1,44 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.Utils; +import com.seaofnodes.simple.type.*; +import java.util.BitSet; + +public class NotNode extends Node { + public NotNode(Node in) { super(null, in); } + + @Override public String label() { return "Not"; } + + @Override public String glabel() { return "!"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + in(1)._print0(sb.append("(!"), visited); + return sb.append(")"); + } + + @Override + public Type compute() { + Type t0 = in(1)._type; + if( t0.isHigh() ) return TypeInteger.BOOL.dual(); + switch( t0 ) { + case TypeInteger i0: return i0._max < 0 || i0._min > 0 ? TypeInteger.FALSE : (i0==TypeInteger.ZERO ? TypeInteger.TRUE : TypeInteger.BOOL); + case TypeFloat i0: return i0.isConstant() ? TypeInteger.constant(i0.value()==0 ? 1 : 0) : i0; + case TypeMemPtr p0: + // top->top, bot->bot, null->1, *void->0, not-null ptr->0, ptr/nil->bot + // If input in null then true + // If input is not null ptr then false + if( p0 == TypeMemPtr.NULLPTR ) return TypeInteger.TRUE; + if( !p0._nil ) return TypeInteger.FALSE; + return TypeInteger.BOOL; + case Type t: + if( t0.getClass() != Type.class ) + // Only doing NOT on ints and ptrs + throw Utils.TODO(); + return TypeInteger.BOOL; + } + } + + @Override + public Node idealize() { return null; } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/OrNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/OrNode.java new file mode 100644 index 00000000..b572346f --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/OrNode.java @@ -0,0 +1,65 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeInteger; + +import java.util.BitSet; + +public class OrNode extends Node { + public OrNode(Node lhs, Node rhs) { super(null, lhs, rhs); } + + @Override public String label() { return "Or"; } + + @Override public String glabel() { return "|"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + in(1)._print0(sb.append("("), visited); + in(2)._print0(sb.append("|"), visited); + return sb.append(")"); + } + + @Override + public Type compute() { + Type t1 = in(1)._type, t2 = in(2)._type; + if( t1.isHigh() || t2.isHigh() ) + return TypeInteger.TOP; + if( t1 instanceof TypeInteger i0 && + t2 instanceof TypeInteger i1 ) { + if( i0.isConstant() && i1.isConstant() ) + return TypeInteger.constant(i0.value()|i1.value()); + } + return TypeInteger.BOT; + } + + @Override + public Node idealize() { + Node lhs = in(1); + Node rhs = in(2); + Type t1 = lhs._type; + Type t2 = rhs._type; + + // Or of 0. We do not check for (0|x) because this will already + // canonicalize to (x|0) + if( t2.isConstant() && t2 instanceof TypeInteger i && i.value()==0 ) + return lhs; + + // Move constants to RHS: con*arg becomes arg*con + if ( t1.isConstant() && !t2.isConstant() ) + return swap12(); + + // Do we have ((x | (phi cons)) | con) ? + // Do we have ((x | (phi cons)) | (phi cons)) ? + // Push constant up through the phi: x | (phi con0|con0 con1|con1...) + Node phicon = AddNode.phiCon(this,true); + if( phicon!=null ) return phicon; + + return null; + } + @Override Node copy(Node lhs, Node rhs) { return new OrNode(lhs,rhs); } + @Override String err() { + if( !(in(1)._type instanceof TypeInteger) ) return "Cannot '|' " + in(1)._type; + if( !(in(2)._type instanceof TypeInteger) ) return "Cannot '|' " + in(2)._type; + return null; + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/OutNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/OutNode.java new file mode 100644 index 00000000..dbac136b --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/OutNode.java @@ -0,0 +1,7 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.Ary; + +public interface OutNode { + Ary outs(); +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/PhiNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/PhiNode.java new file mode 100644 index 00000000..cb266ca3 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/PhiNode.java @@ -0,0 +1,160 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeMem; +import com.seaofnodes.simple.Utils; + +import java.util.BitSet; + +public class PhiNode extends Node { + + final String _label; + + // The Phi type we compute must stay within the domain of the Phi. + // Example Int stays Int, Ptr stays Ptr, Control stays Control, Mem stays Mem. + final Type _declaredType; + + public PhiNode(String label, Type declaredType, Node... inputs) { super(inputs); _label = label; assert declaredType!=null; _declaredType = declaredType; } + + @Override public String label() { return "Phi_"+MemOpNode.mlabel(_label); } + + @Override public String glabel() { return "φ_"+_label; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + if( !(region() instanceof RegionNode r) || r.inProgress() ) + sb.append("Z"); + sb.append("Phi("); + for( Node in : _inputs ) { + if (in == null) sb.append("____"); + else in._print0(sb, visited); + sb.append(","); + } + sb.setLength(sb.length()-1); + sb.append(")"); + return sb; + } + + public CFGNode region() { return (CFGNode)in(0); } + @Override public boolean isMultiTail() { return true; } + @Override public boolean isMem() { return _declaredType instanceof TypeMem; } + @Override public boolean isPinned() { return true; } + + @Override + public Type compute() { + if( !(region() instanceof RegionNode r) ) + return region()._type==Type.XCONTROL ? (_type instanceof TypeMem ? TypeMem.TOP : Type.TOP) : _type; + // During parsing Phis have to be computed type pessimistically. + if( r.inProgress() ) return _declaredType; + // Set type to local top of the starting type + Type t = _declaredType.lub(); + for (int i = 1; i < nIns(); i++) + // If the region's control input is live, add this as a dependency + // to the control because we can be peeped should it become dead. + if( r.in(i).addDep(this)._type != Type.XCONTROL ) + t = t.meet(in(i)._type); + return t; + } + + @Override + public Node idealize() { + if( !(region() instanceof RegionNode r ) ) + return in(1); // Input has collapse to e.g. starting control. + if( r.inProgress() || r.nIns()<=1 ) + return null; // Input is in-progress + + // If we have only a single unique input, become it. + Node live = singleUniqueInput(); + 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 0 && idom.in(0) != iff ) idom = idom.idom(); + if( idom instanceof CProjNode proj && proj._idx==1 ) + return val; + } + } + } + + return null; + } + + private boolean same_op() { + for( int i=2; i1 && in(nIns()-1) == null; + } + + // Never equal if inProgress + @Override boolean eq( Node n ) { + return !inProgress(); + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/ReturnNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/ReturnNode.java new file mode 100644 index 00000000..b586a44f --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/ReturnNode.java @@ -0,0 +1,50 @@ +package com.seaofnodes.simple.node; + +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. + *

+ * 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. + */ +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>"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + in(1)._print0(sb.append("("), visited); + in(2)._print0(sb.append(">>"), visited); + return sb.append(")"); + } + + @Override + public Type compute() { + Type t1 = in(1)._type, t2 = in(2)._type; + if( t1.isHigh() || t2.isHigh() ) + return TypeInteger.TOP; + if (t1 instanceof TypeInteger i0 && + t2 instanceof TypeInteger i1) { + if( i0 == TypeInteger.ZERO ) + return TypeInteger.ZERO; + if( i0.isConstant() && i1.isConstant() ) + return TypeInteger.constant(i0.value()>>i1.value()); + if( i1.isConstant() ) { + int log = (int)i1.value(); + return TypeInteger.make(-1L<<(63-log),(1L<<(63-log))-1); + } + } + return TypeInteger.BOT; + } + + @Override + public Node idealize() { + Node lhs = in(1); + Node rhs = in(2); + Type t2 = rhs._type; + + // Sar of 0. + if( t2.isConstant() && t2 instanceof TypeInteger i && (i.value()&63)==0 ) + return lhs; + + // TODO: x >> 3 >> (y ? 1 : 2) ==> x >> (y ? 4 : 5) + + return null; + } + @Override Node copy(Node lhs, Node rhs) { return new SarNode(lhs,rhs); } + @Override String err() { + if( !(in(1)._type instanceof TypeInteger) ) return "Cannot '>>' " + in(1)._type; + if( !(in(2)._type instanceof TypeInteger) ) return "Cannot '>>' " + in(2)._type; + return null; + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/ScopeMinNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/ScopeMinNode.java new file mode 100644 index 00000000..9134f9c3 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/ScopeMinNode.java @@ -0,0 +1,150 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.IterPeeps; +import com.seaofnodes.simple.Parser; +import com.seaofnodes.simple.Utils; +import com.seaofnodes.simple.type.*; +import java.util.*; + +import static com.seaofnodes.simple.Utils.TODO; + +/* */ +public class ScopeMinNode extends Node { + + /** The tracked fields are now complex enough to deserve a array-of-structs layout + */ + public static class Var { + public final int _idx; // index in containing scope + public final String _name; // Declared name + private Type _type; // Declared type + public final boolean _final; // Final field + public Var(int idx, String name, Type type, boolean xfinal) { + _idx = idx; + _name = name; + _type = type; + _final = xfinal; + } + public Type type() { + if( !_type.isFRef() ) return _type; + // Update self to no longer use the forward ref type + Type def = Parser.TYPES.get(((TypeMemPtr)_type)._obj._name); + return (_type=_type.meet(def)); + } + public Type lazyGLB() { + Type t = type(); + return t instanceof TypeMemPtr ? t : t.glb(); + } + @Override public String toString() { + return _type.toString()+(_final ? " ": " !")+_name; + } + } + + public ScopeMinNode() { _type = TypeMem.BOT; } + + + @Override public String label() { return "MEM"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + sb.append("MEM[ "); + for( int j=2; j= 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 + public final Ary _lexSize; + + // True if parsing inside of a constructor + public final Ary _inCons; + + // Extra guards; tested predicates and casted results + private final Ary _guards; + + // A new ScopeNode + public ScopeNode() { + _vars = new Ary<>(Var .class); + _lexSize= new Ary<>(Integer.class); + _inCons = new Ary<>(Boolean.class); + _guards = new Ary<>(Node .class); + } + + @Override public String label() { return "Scope"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + sb.append("Scope[ "); + int j=1; + for( int i=0; i N ctrl(N n) { return setDef(0,n); } + + public void push() { push(false); } + public void push(boolean inCon) { + assert _lexSize._len==_inCons._len; + _lexSize.push(_vars.size()); + _inCons.push(inCon); + } + public void pop() { + assert _lexSize._len==_inCons._len; + int n = _lexSize.pop(); + _inCons.pop(); + popUntil(n); + _vars.setLen(n); + } + + public boolean inCon() { return _inCons.last(); } + + // Find name in reverse, return an index into _vars or -1. Linear scan + // instead of hashtable, but probably doesn't matter until the scan + // typically hits many dozens of variables. + int find( String name ) { + for( int i=_vars.size()-1; i>=0; i-- ) + if( _vars.get(i)._name.equals(name) ) + return i; + return -1; + } + + /** + * Create a new variable name in the current scope + */ + public boolean define( String name, Type declaredType, boolean xfinal, Node init ) { + assert name.charAt(0)!='$' || _lexSize.size()==1; // Later scopes do not define memory + if( _lexSize._len > 1 ) + for( int i=_vars.size()-1; i>=_lexSize.last(); i-- ) + if( _vars.get(i)._name.equals(name) ) + return false; // Double define + _vars.add(new Var(nIns(),name,declaredType,xfinal)); + addDef(init); + 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. + * + * @param name Name to be looked up + * @return null if not found, or the implementing Node + */ + public Var lookup( String name ) { + int idx = find(name); + // -1 is missed in all scopes, not found + return idx == -1 ? null : update(_vars.at(idx),null); + } + + /** + * If the name is present in any scope, then redefine else null + * + * @param name Name being redefined + * @param n The node to bind to the name + */ + public void update( String name, Node n ) { + int idx = find(name); + assert idx>=0; + update(_vars.at(idx),n); + } + + public Var update( ScopeNode.Var v, Node st ) { + Node old = in(v._idx); + if( old instanceof ScopeNode loop ) { + // Lazy Phi! + Node def = loop.in(v._idx); + old = def instanceof PhiNode phi && loop.ctrl()==phi.region() + // Loop already has a real Phi, use it + ? def + // 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()); + setDef(v._idx,old); + } + if( st!=null ) setDef(v._idx,st); // Set new value + return v; + } + + /** + * Duplicate a ScopeNode; including all levels, up to Nodes. So this is + * neither shallow (would dup the Scope but not the internal HashMap + * tables), nor deep (would dup the Scope, the HashMap tables, but then + * also the program Nodes). + *

+ * If the {@code loop} flag is set, the edges are filled in as the original + * Scope, as an indication of Lazy Phis at loop heads. The goal here is to + * not make Phis at loop heads for variables which are never touched in the + * loop body. + *

+ * The new Scope is a full-fledged Node with proper use<->def edges. + */ + public ScopeNode dup() { return dup(false); } + public ScopeNode dup(boolean loop) { + ScopeNode dup = new ScopeNode(); + // Our goals are: + // 1) duplicate the name bindings of the ScopeNode across all stack levels + // 2) Make the new ScopeNode a user of all the nodes bound + // 3) Ensure that the order of defs is the same to allow easy merging + dup._vars .addAll(_vars ); + dup._lexSize.addAll(_lexSize); + dup._inCons .addAll(_inCons ); + dup._guards .addAll(_guards ); + // The dup'd guards all need dup'd keepers, to keep proper accounting + // when later removing all guards + for( Node n : _guards ) + if( !(n instanceof CFGNode) ) + 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 0 ) { + Node cast = _guards.at(--i); + if( cast instanceof CFGNode ) continue; // Marker between guard sets + Node xpred = _guards.at(--i); + if( xpred == pred ) + return cast; + } + return pred; + } + + // Kill guards also + @Override public void kill() { + for( Node n : _guards ) + if( !(n instanceof CFGNode) ) + n.unkill(); + _guards.clear(); + super.kill(); + } + + + private Node replace( Node old, Node cast ) { + assert old!=null && old!=cast; + for( int i=0; i (x << i) + (c << i) + if( lhs instanceof AddNode add && add.addDep(this).in(2)._type instanceof TypeInteger c && c.isConstant() && c!=TypeInteger.ZERO ) { + long sum = c.value() << shl.value(); + if( Integer.MIN_VALUE <= sum && sum <= Integer.MAX_VALUE ) + return new AddNode( new ShlNode(add.in(1),rhs).peephole(), Parser.con(sum) ); + } + } + + // TODO: x << 3 << (y ? 1 : 2) ==> x << (y ? 4 : 5) + + return null; + } + @Override Node copy(Node lhs, Node rhs) { return new ShlNode(lhs,rhs); } + @Override String err() { + if( !(in(1)._type instanceof TypeInteger) ) return "Cannot '<<' " + in(1)._type; + if( !(in(2)._type instanceof TypeInteger) ) return "Cannot '<<' " + in(2)._type; + return null; + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/ShrNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/ShrNode.java new file mode 100644 index 00000000..22eb45a0 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/ShrNode.java @@ -0,0 +1,52 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeInteger; + +import java.util.BitSet; + +public class ShrNode extends Node { + public ShrNode(Node lhs, Node rhs) { super(null, lhs, rhs); } + + @Override public String label() { return "Shr"; } + + @Override public String glabel() { return ">>>"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + in(1)._print0(sb.append("("), visited); + in(2)._print0(sb.append(">>>"), visited); + return sb.append(")"); + } + + @Override + public Type compute() { + Type t1 = in(1)._type, t2 = in(2)._type; + if( t1.isHigh() || t2.isHigh() ) + return TypeInteger.TOP; + if (t1 instanceof TypeInteger i0 && + t2 instanceof TypeInteger i1 ) { + if( i0 == TypeInteger.ZERO ) + return TypeInteger.ZERO; + if( i0.isConstant() && i1.isConstant() ) + return TypeInteger.constant(i0.value()>>>i1.value()); + } + return TypeInteger.BOT; + } + + @Override + public Node idealize() { + Node lhs = in(1); + Node rhs = in(2); + Type t2 = rhs._type; + + // Shr of 0. + if( t2.isConstant() && t2 instanceof TypeInteger i && (i.value()&63)==0 ) + return lhs; + + // TODO: x >>> 3 >>> (y ? 1 : 2) ==> x >>> (y ? 4 : 5) + + return null; + } + @Override Node copy(Node lhs, Node rhs) { return new ShrNode(lhs,rhs); } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/StartNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/StartNode.java new file mode 100644 index 00000000..ab6e4931 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/StartNode.java @@ -0,0 +1,44 @@ +package com.seaofnodes.simple.node; + +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. + */ +public class StartNode extends LoopNode implements MultiNode { + + final Type _arg; + + public StartNode(Type arg) { super(null); _arg = arg; _type = compute(); } + + @Override public String label() { return "Start"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + return sb.append(label()); + } + + @Override public boolean isMultiHead() { return true; } + @Override public boolean blockHead() { return true; } + @Override public CFGNode cfg0() { return this; } + + @Override public TypeTuple compute() { + return TypeTuple.make(Type.CONTROL,TypeMem.TOP,_arg); + } + + @Override public Node idealize() { return null; } + + // No immediate dominator, and idepth==0 + @Override public int idepth() { return 0; } + @Override public CFGNode idom(Node dep) { return null; } + +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/node/StopNode.java b/chapter18/src/main/java/com/seaofnodes/simple/node/StopNode.java new file mode 100644 index 00000000..867225cb --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/node/StopNode.java @@ -0,0 +1,85 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.Parser; +import com.seaofnodes.simple.IterPeeps; +import com.seaofnodes.simple.GlobalCodeMotion; +import com.seaofnodes.simple.GraphVisualizer; + +import java.util.BitSet; + +public class StopNode extends CFGNode { + + public final String _src; + + public StopNode(String src) { + super(); + _src = src; + } + + @Override + public String label() { + return "Stop"; + } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + if( ret()!=null ) return ret()._print0(sb, visited); + sb.append("Stop[ "); + for( Node ret : _inputs ) + ret._print0(sb, visited).append(" "); + return sb.append("]"); + } + + @Override public boolean blockHead() { return true; } + + // If a single Return, return it. + // Otherwise, null because ambiguous. + public ReturnNode ret() { + return nIns()==1 ? (ReturnNode)in(0) : null; + } + + @Override + public Type compute() { + return Type.BOTTOM; + } + + @Override + public Node idealize() { + int len = nIns(); + for( int i=0; i ts) { ts.add(TEST); } + + @Override Field xmeet( Type that ) { + Field fld = (Field)that; // Invariant + assert _fname.equals(fld._fname); + assert _alias==fld._alias; + return make(_fname,_type.meet(fld._type),_alias,_final | fld._final); + } + + @Override + public Field dual() { return make(_fname,_type.dual(),_alias,!_final); } + + @Override public Field glb() { + Type glb = _type.glb(); + return glb==_type ? this : make(_fname,glb,_alias,_final); + } + + @Override public Field lub() { + Type lub = _type.lub(); + return lub==_type ? this : make(_fname,lub,_alias,_final); + } + + // Override in subclasses + int hash() { return _fname.hashCode() ^ _type.hashCode() ^ _alias ^ (_final ? 1024 : 0); } + + boolean eq(Type t) { + Field f = (Field)t; + return _fname.equals(f._fname) && _type==f._type && _alias==f._alias && _final==f._final; + } + + + @Override + public StringBuilder print( StringBuilder sb ) { + return _type.print(sb.append(_final?"":"!").append(_fname).append(":").append(_alias).append(" : ")); + } + + @Override public String str() { return _fname; } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/type/Type.java b/chapter18/src/main/java/com/seaofnodes/simple/type/Type.java new file mode 100644 index 00000000..4ebb2b6d --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/type/Type.java @@ -0,0 +1,218 @@ +package com.seaofnodes.simple.type; + +import com.seaofnodes.simple.Utils; + +import java.util.ArrayList; +import java.util.HashMap; + +/** + * These types are part of a Monotone Analysis Framework, + * @see see for example this set of slides. + *

+ * The types form a lattice; @see a symmetric complete bounded (ranked) lattice. + *

+ * This wild lattice theory will be needed later to allow us to easily beef up + * the analysis and optimization of the Simple compiler... but we don't need it + * now, just know that it is coming along in a later Chapter. + *

g + * One of the fun things here is that while the theory is deep and subtle, the + * actual implementation is darn near trivial and is generally really obvious + * what we're doing with it. Right now, it's just simple integer math to do + * simple constant folding e.g. 1+2 == 3 stuff. + */ + +public class Type { + static final HashMap INTERN = new HashMap<>(); + + // ---------------------------------------------------------- + // Simple types are implemented fully here. "Simple" means: the code and + // type hierarchy are simple, not that the Type is conceptually simple. + static final byte TBOT = 0; // Bottom (ALL) + static final byte TTOP = 1; // Top (ANY) + static final byte TCTRL = 2; // Ctrl flow bottom + static final byte TXCTRL = 3; // Ctrl flow top (mini-lattice: any-xctrl-ctrl-all) + static final byte TSIMPLE = 4; // End of the Simple Types + static final byte TINT = 5; // All Integers; see TypeInteger + static final byte TFLT = 6; // All Integers; see TypeInteger + static final byte TTUPLE = 7; // Tuples; finite collections of unrelated Types, kept in parallel + static final byte TMEM = 8; // All memory (alias 0) or A slice of memory - with specific alias + static final byte TMEMPTR = 9; // Memory pointer type + static final byte TSTRUCT =10; // Structs; tuples with named fields + static final byte TFLD =11; // Fields into struct + static final byte TARRAY =12; // Array + + public final byte _type; + + public boolean is_simple() { return _type < TSIMPLE; } + private static final String[] STRS = new String[]{"Bot","Top","Ctrl","~Ctrl"}; + protected Type(byte type) { _type = type; } + + public static final Type BOTTOM = new Type( TBOT ).intern(); // ALL + public static final Type TOP = new Type( TTOP ).intern(); // ANY + public static final Type CONTROL = new Type( TCTRL ).intern(); // Ctrl + public static final Type XCONTROL = new Type( TXCTRL ).intern(); // ~Ctrl + public static Type[] gather() { + ArrayList ts = new ArrayList<>(); + ts.add(BOTTOM); + ts.add(CONTROL); + Field.gather(ts); + TypeInteger.gather(ts); + TypeFloat.gather(ts); + TypeMem.gather(ts); + TypeMemPtr.gather(ts); + TypeStruct.gather(ts); + TypeTuple.gather(ts); + int sz = ts.size(); + for( int i = 0; i < sz; i++ ) + ts.add(ts.get(i).dual()); + return ts.toArray(new Type[ts.size()]); + } + + // Is high or on the lattice centerline. + public boolean isHigh () { return _type==TTOP || _type==TXCTRL; } + public boolean isHighOrConst() { return _type==TTOP || _type==TXCTRL; } + + // Strict constant values, things on the lattice centerline. + // Excludes both high and low values + public boolean isConstant() { return false; } + + /** + * Display Type name in a format that's good for IR printer + */ + public StringBuilder typeName( StringBuilder sb) { return print(sb); } + + + // ---------------------------------------------------------- + + // Notes on Type interning: At the moment it is not easy to reset the + // interned types because we hold static references to several types and + // these are scattered around. This means the INTERN cache will retain all + // types from every run of the Parser. For this to work correctly types + // must be rigorous about defining when they are the same. Also types need + // to be immutable once defined. The rationale for interning is + // *correctness* with cyclic type definitions. Simple structural recursive + // checks go exponential with merely sharing, but with cycles they will + // stack overflow and crash. Intering means we do not need to have sharing + // checks with every type compare, only during interning. + + // Factory method which interns "this" + public T intern() { + T nnn = (T)INTERN.get(this); + if( nnn==null ) + INTERN.put(nnn=(T)this,this); + return nnn; + } + + private int _hash; // Hash cache; not-zero when set. + @Override + public final int hashCode() { + if( _hash!=0 ) return _hash; + _hash = hash(); + if( _hash==0 ) _hash = 0xDEADBEEF; // Bad hash from subclass; use some junk thing + return _hash; + } + // Override in subclasses + int hash() { return _type; } + + @Override + public final boolean equals( Object o ) { + if( o==this ) return true; + if( !(o instanceof Type t)) return false; + if( _type != t._type ) return false; + return eq(t); + } + // Overridden in subclasses; subclass can assume "this!=t" and java classes are same + boolean eq(Type t) { return true; } + + + // ---------------------------------------------------------- + public final Type meet(Type t) { + // Shortcut for the self case + if( t == this ) return this; + // Same-type is always safe in the subclasses + if( _type==t._type ) return xmeet(t); + // Reverse; xmeet 2nd arg is never "is_simple" and never equal to "this". + if( is_simple() ) return this.xmeet(t ); + if( t.is_simple() ) return t .xmeet(this); + return Type.BOTTOM; // Mixing 2 unrelated types + } + + // Compute meet right now. Overridden in subclasses. + // Handle cases where 'this.is_simple()' and unequal to 't'. + // Subclassed xmeet calls can assert that '!t.is_simple()'. + Type xmeet(Type t) { + assert is_simple(); // Should be overridden in subclass + // ANY meet anything is thing; thing meet ALL is ALL + if( _type==TBOT || t._type==TTOP ) return this; + if( _type==TTOP || t._type==TBOT ) return t; + // 'this' is {TCTRL,TXCTRL} + if( !t.is_simple() ) return BOTTOM; + // 't' is {TCTRL,TXCTRL} + return _type==TCTRL || t._type==TCTRL ? CONTROL : XCONTROL; + } + + public Type dual() { + return switch( _type ) { + case TBOT -> TOP; + case TTOP -> BOTTOM; + case TCTRL -> XCONTROL; + case TXCTRL -> CONTROL; + default -> throw Utils.TODO(); // Should not reach here + }; + } + + // ---------------------------------------------------------- + // Our lattice is defined with a MEET and a DUAL. + // JOIN is dual of meet of both duals. + public final Type join(Type t) { + if( this==t ) return this; + return dual().meet(t.dual()).dual(); + } + + // True if this "isa" t; e.g. 17 isa TypeInteger.BOT + public boolean isa( Type t ) { return meet(t)==t; } + + /** Compute greatest lower bound in the lattice */ + public Type glb() { return _type==TCTRL ? CONTROL : BOTTOM; } + /** Compute least upper bound in the lattice */ + public Type lub() { return _type==TCTRL ? XCONTROL : TOP; } + + // Make an initial/default version of this type. Typically, 0 for integers + // and null for nullable pointers. + public Type makeInit() { return this; } + // Make a zero version of this type, 0 for integers and null for pointers. + public Type makeZero() { return null; } + // Make a non-zero version of this type, if possible. Integers attempt to + // exclude zero from their range and pointers become not-null. + public Type nonZero() { return glb(); } + + // Is forward-reference + public boolean isFRef() { return false; } + + // All reachable struct Fields are final + public boolean isFinal() { return true; } + + // Make all reachable struct Fields final + public Type makeRO() { return this; } + + // ---------------------------------------------------------- + + // Size in bits to hold an instance of this type. + // Sizes are expected to be between 1 and 64 bits. + // Size 0 means this either takes no space (such as a known-zero field) + // or isn't a scalar to be stored in memory. + public int log_size() { throw Utils.TODO(); } + + // ---------------------------------------------------------- + // Useful in the debugger, which calls toString everywhere. + // This is a more verbose dev-friendly print. + @Override + public final String toString() { + return print(new StringBuilder()).toString(); + } + + public StringBuilder print(StringBuilder sb) { return is_simple() ? sb.append(STRS[_type]) : sb;} + + // This is used by error messages, and is a shorted print. + public String str() { return STRS[_type]; } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/type/TypeFloat.java b/chapter18/src/main/java/com/seaofnodes/simple/type/TypeFloat.java new file mode 100644 index 00000000..f85abbed --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/type/TypeFloat.java @@ -0,0 +1,120 @@ +package com.seaofnodes.simple.type; + +import java.util.ArrayList; + +/** + * Float Type + */ +public class TypeFloat extends Type { + + public final static TypeFloat TOP = make((byte)-64, 0); + public final static TypeFloat T32 = make((byte)-32, 0); + public final static TypeFloat ZERO= make((byte) 0, 0); + public final static TypeFloat B32 = make((byte) 32, 0); + public final static TypeFloat BOT = make((byte) 64, 0); + + /** +/-64 for double; +/-32 for float, or 0 for constants */ + public final byte _sz; + + /** + * The constant value or 0 + */ + public final double _con; + + private TypeFloat(byte sz, double con) { + super(TFLT); + _sz = sz ; + _con = con; + } + private static TypeFloat make(byte sz, double con) { + return new TypeFloat(sz,con).intern(); + } + + public static TypeFloat constant(double con) { return make((byte)0, con); } + + public static void gather(ArrayList ts) { ts.add(ZERO); ts.add(BOT); ts.add(B32); ts.add(constant(3.141592653589793)); } + + // FIXME this display format is problematic + // In visualizer '#' gets prepended if its a constant + @Override + public StringBuilder print(StringBuilder sb) { + if( this==TOP ) return sb.append("FltTop"); + if( this==T32 ) return sb.append("F32Top"); + if( this==B32 ) return sb.append("F32Bot"); + if( this==BOT ) return sb.append("FltBot"); + return sb.append(_con); + } + + @Override public String str() { + if( this==TOP ) return "~flt"; + if( this==T32 ) return "~f32"; + if( this==B32 ) return "f32"; + if( this==BOT ) return "flt"; + return ""+_con+(isF32() ? "f" : ""); + } + + /** + * Display Type name in a format that's good for IR printer + */ + @Override + public StringBuilder typeName(StringBuilder sb) { + if( this==TOP ) return sb.append("FltTop"); + if( this==T32 ) return sb.append("F32Top"); + if( this==B32 ) return sb.append("F32Bot"); + if( this==BOT ) return sb.append("FltBot"); + return sb.append(isF32() ? "F32" : "Flt"); + } + + boolean isF32() { return ((float)_con)==_con; } + @Override public boolean isHigh() { return _sz< 0; } + @Override public boolean isHighOrConst() { return _sz<=0; } + @Override public boolean isConstant() { return _sz==0; } + + @Override public int log_size() { + int sz = _sz==0 ? (isF32() ? 32 : 64) : Math.abs(_sz); + return _sz==32 ? 2 : 3; + } + + public double value() { return _con; } + + @Override + public Type xmeet(Type other) { + // Invariant from caller: 'this' != 'other' and same class (TypeFloat) + TypeFloat i = (TypeFloat)other; // Contract + // Larger size in i1, smaller in i0 + TypeFloat i0 = _sz < i._sz ? this : i; + TypeFloat i1 = _sz < i._sz ? i : this; + + if( i1._sz== 64 ) return BOT; + if( i0._sz==-64 ) return i1; + if( i1._sz== 32 ) + return i0._sz==0 && !i0.isF32() ? BOT : B32; + if( i1._sz!= 0 ) return i1; + // i1 is a constant + if( i0._sz==-32 ) + return i1.isF32() ? i1 : BOT; + // Since both are constants, and are never equals (contract) unequals + // constants fall to bottom + return i0.isF32() && i1.isF32() ? B32 : BOT; + } + + @Override + public Type dual() { + if( _sz==0 ) return this; // Constants are a self-dual + return make((byte)-_sz,0); + } + + @Override public Type glb() { return BOT; } + @Override public Type lub() { return TOP; } + @Override public TypeFloat makeInit() { return ZERO; } + @Override public TypeFloat makeZero() { return ZERO; } + + @Override + int hash() { return (int)(Double.hashCode(_con) ^ _sz); } + @Override + public boolean eq( Type t ) { + TypeFloat i = (TypeFloat)t; // Contract + return _con==i._con && _sz==i._sz; + } + +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/type/TypeInteger.java b/chapter18/src/main/java/com/seaofnodes/simple/type/TypeInteger.java new file mode 100644 index 00000000..b2979d7b --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/type/TypeInteger.java @@ -0,0 +1,133 @@ +package com.seaofnodes.simple.type; + +import com.seaofnodes.simple.Utils; +import java.util.ArrayList; + +/** + * Integer Type + */ +public class TypeInteger extends Type { + + public final static TypeInteger TOP = make(false, 0); + public final static TypeInteger BOT = make(false, 1); + public final static TypeInteger ZERO= make(0,0); + public final static TypeInteger U1 = make(0,1); + public final static TypeInteger BOOL= U1; + public final static TypeInteger FALSE=ZERO; + public final static TypeInteger TRUE= make(1,1); + public final static TypeInteger I8 = make(-128,127); + public final static TypeInteger I16 = make(-32768,32767); + public final static TypeInteger I32 = make(-1L<<31,(1L<<31)-1); + + public final static TypeInteger U8 = make(0,255); + public final static TypeInteger U16 = make(0,65535); + public final static TypeInteger U32 = make(0,(1L<<32)-1); + + /** + * Describes an integer *range* - everything from min to max; both min and + * max are inclusive. If min==max, this is a constant. + * + * If min <= max, this is a below center (towards bottom). + * If min > max, this is an above center (towards top). + */ + public final long _min, _max; + + private TypeInteger(long min, long max) { super(TINT); _min = min; _max = max; } + public static TypeInteger make(long lo, long hi) { return new TypeInteger(lo,hi).intern(); } + public static TypeInteger make(boolean is_con, long con) { + return make(is_con ? con : (con==0 ? Long.MAX_VALUE : Long.MIN_VALUE), + is_con ? con : (con==0 ? Long.MIN_VALUE : Long.MAX_VALUE)); + } + + public static TypeInteger constant(long con) { return make(true, con); } + + public static void gather(ArrayList ts) { ts.add(ZERO); ts.add(BOT); ts.add(U1); ts.add(U8); } + + // FIXME this display format is problematic + // In visualizer '#' gets prepended if its a constant + @Override + public StringBuilder print(StringBuilder sb) { return sb.append(str()); } + + @Override public String str() { + if( this==TOP ) return "~int"; + if( this==BOT ) return "int"; + if( this==BOOL) return ("bool"); + if( this==I8 ) return ("i8"); + if( this==I16 ) return ("i16"); + if( this==I32 ) return ("i32"); + if( this==U8 ) return ("u8"); + if( this==U16 ) return ("u16"); + if( this==U32 ) return ("u32"); + if( isConstant() ) return ""+_min; + return "["+_min+"-"+_max+"]"; + } + + /** + * Display Type name in a format that's good for IR printer + */ + @Override + public StringBuilder typeName(StringBuilder sb) { + if( this==TOP ) return sb.append("IntTop"); + if( this==BOT ) return sb.append("IntBot"); + return sb.append("Int"); + } + + @Override public boolean isHigh () { return _min > _max; } + @Override public boolean isHighOrConst() { return _min >= _max; } + @Override public boolean isConstant () { return _min == _max; } + + @Override public int log_size() { + if( this==I8 || this==U8 || this==BOOL ) return 0; // 1<<0 == 1 bytes + if( this==I16 || this==U16 ) return 1; // 1<<1 == 2 bytes + if( this==I32 || this==U32 ) return 2; // 1<<2 == 4 bytes + if( this==BOT ) return 3; // 1<<3 == 8 bytes + if( isHighOrConst() ) return 0; + throw Utils.TODO(); + } + + public long value() { assert isConstant(); return _min; } + + // AND-mask of forced zeros. e.g. unsigned types will return their mask; + // u8 will return 0xFF. But also a range of 16-18 (0x10-0x12) will return + // 0x13 - no value in the range {16,17,18} will allow bit 0x04 to be set. + public long mask() { + if( isHigh() ) return 0; + if( isConstant() ) return _min; + //if( _min<0 ) return -1L; + //if( _max==Long.MAX_VALUE ) return -1L; + // Those bit positions which differ min to max + long x = _min ^ _max; + // Highest '1' bit in the differ set. Since the range is from min to + // max, all values below here are possible. + long ff1 = Long.highestOneBit(x); + // Make a all-1's mask from ff1, and set over the same bits (either min + // or max is ok). + long mask = _min | (ff1-1) | ff1; + return mask; + } + + @Override + public Type xmeet(Type other) { + // Invariant from caller: 'this' != 'other' and same class (TypeInteger) + TypeInteger i = (TypeInteger)other; // Contract + return make(Math.min(_min,i._min), Math.max(_max,i._max)); + } + + @Override public TypeInteger dual() { return make(_max,_min); } + @Override public TypeInteger glb() { return BOT; } + @Override public TypeInteger lub() { return TOP; } + @Override public TypeInteger makeInit() { return ZERO; } + @Override public TypeInteger makeZero() { return ZERO; } + @Override public TypeInteger nonZero() { + if( isHigh() ) return this; + if( _min==0 ) return make(1,Math.max(_max,1)); // specifically good on BOOL + if( _max==0 ) return make(_min,-1); + return this; + } + @Override int hash() { return (int)((_min ^ (_min>>32)) *(_max ^ (_max>>32))); } + @Override + public boolean eq( Type t ) { + TypeInteger i = (TypeInteger)t; // Contract + return _min==i._min && _max==i._max; + } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/type/TypeMem.java b/chapter18/src/main/java/com/seaofnodes/simple/type/TypeMem.java new file mode 100644 index 00000000..33c6a8ce --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/type/TypeMem.java @@ -0,0 +1,67 @@ +package com.seaofnodes.simple.type; + +import com.seaofnodes.simple.Utils; +import java.lang.Long; +import java.util.BitSet; +import java.util.ArrayList; + + +/** + * Represents a slice of memory corresponding to a set of aliases + */ +public class TypeMem extends Type { + + // Which slice of memory? + // 0 means TOP, no slice. + // -1 means BOT, all memory. + // N means slice#N. + public final int _alias; + public final Type _t; // Memory contents, some scalar type + + private TypeMem(int alias, Type t) { super(TMEM); _alias = alias; _t = t; } + + public static TypeMem make(int alias, Type t) { return new TypeMem(alias,t).intern(); } + public static final TypeMem TOP = make(-1, Type.TOP ); + public static final TypeMem BOT = make(-1, Type.BOTTOM); + + public static void gather(ArrayList ts) { ts.add(make(1,TypeInteger.ZERO)); ts.add(BOT); } + + @Override + TypeMem xmeet(Type t) { + TypeMem that = (TypeMem) t; // Invariant: TypeMem and unequal + if( this==TOP ) return that; + if( that==TOP ) return this; + if( this==BOT ) return BOT; + if( that==BOT ) return BOT; + int alias = _alias==that._alias ? _alias : -1; + Type mt = _t.meet(that._t); + return make(alias,mt); + } + + @Override + public Type dual() { + return make(_alias,_t.dual()); + } + + @Override public Type glb() { return make(_alias,_t.glb()); } + @Override public Type lub() { return make(_alias,_t.lub()); } + + @Override + int hash() { return 9876543 + _alias + _t.hashCode(); } + + @Override + boolean eq(Type t) { + TypeMem that = (TypeMem) t; // Invariant + return _alias == that._alias && _t == that._t; + } + + @Override + public StringBuilder print(StringBuilder sb) { + sb.append("MEM#"); + if( _alias== 0 ) return sb.append("TOP"); + if( _alias==-1 ) return sb.append("BOT"); + return _t.print(sb.append(_alias).append(":")); + } + + @Override public String str() { return toString(); } +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/type/TypeMemPtr.java b/chapter18/src/main/java/com/seaofnodes/simple/type/TypeMemPtr.java new file mode 100644 index 00000000..786a7be5 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/type/TypeMemPtr.java @@ -0,0 +1,98 @@ +package com.seaofnodes.simple.type; + +import java.util.ArrayList; + +/** + * Represents a Pointer to memory. + * + * Null is generic pointer to non-existent memory. + * *void is a non-Null pointer to all possible refs, both structs and arrays. + * Pointers can be to specific struct and array types, or a union with Null. + * The distinguished *$BOT ptr represents union of *void and Null. + * The distinguished *$TOP ptr represents the dual of *$BOT. + */ +public class TypeMemPtr extends Type { + // A TypeMemPtr is pair (obj,nil) + // where obj is a TypeStruct, possibly TypeStruct.BOT + // where nil is an explicit null is allowed or not + + // Examples: + // (Person,false) - a not-nil Person + // (Person,true ) - a Person or a nil + // (BOT ,false) - a not-nil void* (unspecified struct) + // (TOP ,true ) - a nil + + public final TypeStruct _obj; + public final boolean _nil; + + private TypeMemPtr(TypeStruct obj, boolean nil) { + super(TMEMPTR); + assert obj!=null; + _obj = obj; + _nil = nil; + } + public static TypeMemPtr make(TypeStruct obj, boolean nil) { return new TypeMemPtr(obj, nil).intern(); } + public static TypeMemPtr make(TypeStruct obj) { return make(obj, false); } + public TypeMemPtr makeFrom(TypeStruct obj) { return obj==_obj ? this : make(obj, _nil); } + public TypeMemPtr makeFrom(boolean nil) { return nil==_nil ? this : make(_obj, nil); } + @Override public TypeMemPtr makeRO() { return makeFrom(_obj.makeRO()); } + @Override public boolean isFinal() { return _obj.isFinal(); } + + // An abstract pointer, pointing to either a Struct or an Array. + // Can also be null or not. + public static TypeMemPtr BOT = make(TypeStruct.BOT,true); + public static TypeMemPtr TOP = BOT.dual(); + // An abstract null (can be Struct or Array) or not-null C void* + public static TypeMemPtr NULLPTR = make(TypeStruct.TOP,true); + public static TypeMemPtr VOIDPTR = NULLPTR.dual(); // A bottom mix of not-null ptrs, like C's void* but not null + + public static TypeMemPtr TEST= make(TypeStruct.TEST,false); + public static void gather(ArrayList ts) { ts.add(NULLPTR); ts.add(BOT); ts.add(TEST); } + + @Override + Type xmeet(Type t) { + TypeMemPtr that = (TypeMemPtr) t; + return TypeMemPtr.make((TypeStruct)_obj.meet(that._obj), _nil | that._nil); + } + + @Override + public TypeMemPtr dual() { return TypeMemPtr.make(_obj.dual(), !_nil); } + + @Override public TypeMemPtr glb() { return make(_obj.glb(),true ); } + @Override public TypeMemPtr lub() { return make(_obj.lub(),false); } + // Is forward-reference + @Override public boolean isFRef() { return _obj.isFRef(); } + @Override public Type makeInit() { return _nil ? NULLPTR : Type.TOP; } + @Override public Type makeZero() { return NULLPTR; } + @Override public Type nonZero() { return VOIDPTR; } + + @Override public boolean isHigh() { return this==TOP; } + @Override public boolean isConstant() { return this==NULLPTR; } + @Override public boolean isHighOrConst() { return this==TOP || this==NULLPTR; } + + @Override public int log_size() { return 2; } // (1<<2)==4-byte pointers + + @Override + int hash() { return _obj.hashCode() ^ (_nil ? 1024 : 0); } + + @Override + boolean eq(Type t) { + TypeMemPtr ptr = (TypeMemPtr)t; // Invariant + return _obj == ptr._obj && _nil == ptr._nil; + } + + // [void,name,MANY]*[,?] + @Override + public StringBuilder print(StringBuilder sb) { + if( this== NULLPTR) return sb.append("null"); + if( this== VOIDPTR) return sb.append("*void"); + return _obj.print(sb.append("*")).append(_nil ? "?" : ""); + } + + @Override public String str() { + if( this== NULLPTR) return "null"; + if( this== VOIDPTR) return "*void"; + return "*"+_obj.str()+(_nil ? "?" : ""); + } + +} diff --git a/chapter18/src/main/java/com/seaofnodes/simple/type/TypeStruct.java b/chapter18/src/main/java/com/seaofnodes/simple/type/TypeStruct.java new file mode 100644 index 00000000..b3dbcad0 --- /dev/null +++ b/chapter18/src/main/java/com/seaofnodes/simple/type/TypeStruct.java @@ -0,0 +1,256 @@ +package com.seaofnodes.simple.type; + +import com.seaofnodes.simple.Utils; +import java.util.ArrayList; + +/** + * Represents a struct type. + */ +public class TypeStruct extends Type { + + // A Struct has a name and a set of fields; the fields themselves have + // names, types and aliases. The name has no semantic meaning, but is + // useful for debugging. Briefly during parsing a Struct can be a + // forward-ref; in this case the _fields array is null. + + // Its illegal to attempt to load a field from a forward-ref struct. + + // Example: "int rez = new S.x; struct S { int x; } return rez;" // Error, S not defined + // Rewrite: "struct S { int x; } int rez = new S.x; return rez;" // Ok, no forward ref + // + // During the normal optimization run, struct types "bottom out" at further + // struct references, so we don't have to handle e.g. cyclic types. The + // "bottom out" is again the forward-ref struct. + public final String _name; + public final Field[] _fields; + + TypeStruct(String name, Field[] fields) { + super(TSTRUCT); + _name = name; + _fields = fields; + } + + // All fields directly listed + public static TypeStruct make(String name, Field... fields) { return new TypeStruct(name, fields).intern(); } + public static final TypeStruct TOP = make("$TOP",new Field[0]); + public static final TypeStruct BOT = make("$BOT",new Field[0]); + public static final TypeStruct TEST = make("test",new Field[]{Field.TEST}); + // Forward-ref version + public static TypeStruct makeFRef(String name) { return make(name, (Field[])null); } + // Make a read-only version + @Override public TypeStruct makeRO() { + if( isFinal() ) return this; + Field[] flds = new Field[_fields.length]; + for( int i=0; i ts) { ts.add(TEST); ts.add(BOT); ts.add(S1); ts.add(S2); ts.add(ARY); } + + // Find field index by name + public int find(String fname) { + for( int i=0; i<_fields.length; i++ ) + if( _fields[i]._fname.equals(fname) ) + return i; + return -1; + } + public Field field(String fname) { + int idx = find(fname); + return idx== -1 ? null : _fields[idx]; + } + + // Find field index by alias + public int findAlias( int alias ) { + for( int i=0; i<_fields.length; i++ ) + if( _fields[i]._alias==alias ) + return i; + return -1; + } + + + @Override + Type xmeet(Type t) { + TypeStruct that = (TypeStruct) t; + if( this==TOP ) return that; + if( that==TOP ) return this; + if( this==BOT ) return BOT; + if( that==BOT ) return BOT; + + // Within the same compilation unit, struct names are unique. If the + // names differ, its different structs. Across many compilation units, + // structs with the same name but different field layouts can be + // interned... which begs the question: + // "What is the meet of structs from two different compilation units?" + // And the answer is: "don't ask". + if( !_name.equals(that._name) ) + return BOT; // It's a struct; that's about all we know + if( this._fields==null ) return that; + if( that._fields==null ) return this; + if( _fields.length != that._fields.length ) return BOT; + // Just do field meets + Field[] flds = new Field[_fields.length]; + for( int i=0; i<_fields.length; i++ ) { + Field f0 = _fields[i], f1 = that._fields[i]; + if( !f0._fname.equals(f1._fname) || f0._alias != f1._alias ) + return BOT; + flds[i] = (Field)f0.meet(f1); + } + return make(_name,flds); + } + + @Override + public TypeStruct dual() { + if( this==TOP ) return BOT; + if( this==BOT ) return TOP; + if( _fields == null ) return this; + Field[] flds = new Field[_fields.length]; + for( int i=0; i<_fields.length; i++ ) + flds[i] = _fields[i].dual(); + return make(_name,flds); + } + + // Keeps the same struct, but lower-bounds all fields. + @Override public TypeStruct glb() { + if( _glb() ) return this; + // Need to glb each field + Field[] flds = new Field[_fields.length]; + for( int i=0; i<_fields.length; i++ ) + flds[i] = _fields[i].glb(); + return make(_name,flds); + } + private boolean _glb() { + if( _fields!=null ) + for( Field f : _fields ) + if( f.glb() != f ) + return false; + return true; + } + @Override public TypeStruct lub() { + if( _lub() ) return this; + // Need to glb each field + Field[] flds = new Field[_fields.length]; + for( int i=0; i<_fields.length; i++ ) + flds[i] = _fields[i].lub(); + return make(_name,flds); + } + private boolean _lub() { + if( _fields!=null ) + for( Field f : _fields ) + if( f.lub() != f ) + return false; + return true; + } + + + // Is forward-reference + @Override public boolean isFRef() { return _fields==null; } + // All fields are final + @Override public boolean isFinal() { + if( _fields==null ) return true; + for( Field fld : _fields ) + if( !fld.isFinal() ) + return false; + return true; + } + + @Override + boolean eq(Type t) { + TypeStruct ts = (TypeStruct)t; // Invariant + if( !_name.equals(ts._name) ) + return false; + if( _fields == ts._fields ) return true; + if( _fields==null || ts._fields==null ) return false; + if( _fields.length!=ts._fields.length ) + return false; + for( int i = 0; i < _fields.length; i++ ) + if( _fields[i] != ts._fields[i] ) + return false; + return true; + } + + @Override + int hash() { + long hash = _name.hashCode(); + if( _fields != null ) + for( Field f : _fields ) + hash = Utils.rot(hash,13) ^ f.hashCode(); + return Utils.fold(hash); + } + + @Override + public StringBuilder print(StringBuilder sb) { + sb.append(_name); + if( _fields == null ) return sb; // Forward reference struct, just print the name + sb.append(" {"); + for( Field f : _fields ) + f._type.print(sb).append(f._final ? " " : " !").append(f._fname).append("; "); + return sb.append("}"); + } + + @Override public String str() { return _name; } + + + public boolean isAry() { return _fields.length==2 && _fields[1]._fname.equals("[]"); } + + public int aryBase() { + assert isAry(); + if( _offs==null ) _offs = offsets(); + return _offs[1]; + } + public int aryScale() { + assert isAry(); + return _fields[1]._type.log_size(); + } + + + // Field offsets as packed byte offsets + private int[] _offs; // Lazily computed and not part of the type semantics + public int offset(int idx) { + if( _offs==null ) _offs = offsets(); + return _offs[idx]; + } + private int[] offsets() { // Field byte offsets + // Compute a layout for a collection of fields + assert _fields != null; // No forward refs + // Compute a layout + int[] cnts = new int[4]; // Count of fields at log field size + for( Field f : _fields ) + cnts[f._type.log_size()]++; // Log size is 0(byte), 1(i16/u16), 2(i32/f32), 3(i64/dbl) + int off = 0, idx = 0; // Base common struct fields go here, e.g. Mark/Klass + // Compute offsets to the start of each power-of-2 aligned fields. + int[] offs = new int[4]; + for( int i=3; i>=0; i-- ) { + offs[i] = off; + off += cnts[i]<