diff options
Diffstat (limited to 'chapter12/src/main/java/com')
42 files changed, 5232 insertions, 0 deletions
diff --git a/chapter12/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java b/chapter12/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java new file mode 100644 index 00000000..054a489c --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java @@ -0,0 +1,238 @@ + package com.seaofnodes.simple; + +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 ) { + fixLoops(stop); + schedEarly(); + Parser.SCHEDULED = true; + schedLate(Parser.START); + } + + // ------------------------------------------------------------------------ + // Backwards walk on the CFG only, looking for unreachable code - which has + // to be an infinite loop. Insert a bogus never-taken exit to Stop, so the + // loop becomes reachable. Also, set loop nesting depth + private static void fixLoops(StopNode stop) { + // Backwards walk from Stop, looking for unreachable code + BitSet visit = new BitSet(); + HashSet<CFGNode> unreach = new HashSet<>(); + unreach.add(Parser.START); + for( Node ret : stop._inputs ) + ((ReturnNode)ret).walkUnreach(visit,unreach); + if( unreach.isEmpty() ) return; + + // Forwards walk from unreachable, looking for loops with no exit test. + visit.clear(); + for( CFGNode cfg : unreach ) + walkInfinite(cfg,visit,stop); + // Set loop depth on remaining graph + unreach.clear(); + visit.clear(); + for( Node ret : stop._inputs ) + ((ReturnNode)ret).walkUnreach(visit,unreach); + assert unreach.isEmpty(); + } + + + // Forwards walk over previously unreachable, looking for loops with no + // exit test. + static private void walkInfinite( CFGNode n, BitSet visit, StopNode stop ) { + if( visit.get(n._nid) ) return; // Been there, done that + visit.set(n._nid); + if( n instanceof LoopNode loop ) + loop.forceExit(stop); + for( Node use : n._outputs ) + if( use instanceof CFGNode cfg ) + walkInfinite(cfg,visit,stop); + } + + // ------------------------------------------------------------------------ + private static void schedEarly() { + ArrayList<CFGNode> 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); + // Strictly for dead infinite loops, we can have entire code blocks + // not reachable from below - so we reach down, from above, one + // step. Since _schedEarly modifies the output arrays, the normal + // region._outputs ArrayList iterator throws CME. The extra edges + // are always *added* after any Phis, so just walk the Phi prefix. + if( cfg instanceof RegionNode region ) { + int len = region.nOuts(); + for( int i=0; i<len; i++ ) + if( region.out(i) instanceof PhiNode phi ) + _schedEarly(phi,visit); + } + } + } + + // Post-Order of CFG + private static void _rpo_cfg(Node n, BitSet visit, ArrayList<CFGNode> 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 + 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.isPinned() ) + _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 + for( int i=1; i<n.nIns(); i++ ) + if( n.in(i).cfg0().idepth() > early.idepth() ) + early = n.in(i).cfg0(); // Latest/deepest input + n.setDef(0,early); // First place this can go + } + } + + // ------------------------------------------------------------------------ + private static void schedLate(StartNode start) { + CFGNode[] late = new CFGNode[Node.UID()]; + Node[] ns = new Node[Node.UID()]; + _schedLate(start,ns,late); + for( int i=0; i<late.length; i++ ) + if( ns[i] != null ) + ns[i].setDef(0,late[i]); + } + + // Forwards post-order pass. Schedule all outputs first, then draw a + // idom-tree line from the LCA of uses to the early schedule. Schedule is + // legal anywhere on this line; pick the most control-dependent (largest + // idepth) in the shallowest loop nest. + private static void _schedLate(Node n, Node[] ns, CFGNode[] late) { + if( late[n._nid]!=null ) return; // Been there, done that + // 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); + if( n instanceof PhiNode phi ) late[n._nid] = phi.region(); + + // Walk Stores before Loads, so we can get the anti-deps right + for( Node use : n._outputs ) + if( isForwardsEdge(use,n) && + use._type instanceof TypeMem ) + _schedLate(use,ns,late); + // Walk everybody now + for( Node use : n._outputs ) + if( isForwardsEdge(use,n) ) + _schedLate(use,ns,late); + // Already implicitly scheduled + if( n.isPinned() ) return; + // Need to schedule n + + // Walk uses, gathering the LCA (Least Common Ancestor) of uses + CFGNode early = (CFGNode)n.in(0); + assert early != null; + CFGNode lca = null; + for( Node use : n._outputs ) + lca = use_block(n,use, late).idom(lca); + + // 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 + // 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<phi.nIns(); i++ ) + if( phi.in(i)==n ) + if( found==null ) found = phi.region().cfg(i); + else Utils.TODO(); // Can be more than once + assert found!=null; + return found; + } + + + // Least loop depth first, then largest idepth + private static boolean better( CFGNode lca, CFGNode best ) { + return lca._loopDepth < best._loopDepth || + (lca.idepth() > best.idepth() || best instanceof IfNode); + } + + // Skip iteration if a backedge + private static boolean isForwardsEdge(Node use, Node def) { + return use != null && def != null && + !(use.nIns()>2 && use.in(2)==def && (use instanceof LoopNode || (use instanceof PhiNode phi && phi.region() instanceof LoopNode))); + } + + 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: + 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<phi.nIns(); i++ ) + if( phi.in(i)==load.mem() ) + lca = anti_dep(load,phi.region().cfg(i),load.mem().cfg0(),lca,null); + break; + case LoadNode ld: break; // Loads do not cause anti-deps on other loads + case ReturnNode ret: break; // Load must already be ahead of Return + default: throw Utils.TODO(); + } + } + return lca; + } + + // + private static CFGNode anti_dep( LoadNode load, CFGNode stblk, CFGNode defblk, CFGNode lca, Node st ) { + // Walk store blocks "reach" from its scheduled location to its earliest + for( ; stblk != defblk.idom(); stblk = stblk.idom() ) { + // Store and Load overlap, need anti-dependence + if( stblk._anti==load._nid ) { + lca = stblk.idom(lca); // Raise Loads LCA + if( lca == stblk && st != null && Utils.find(st._inputs,load) == -1 ) // And if something moved, + st.addDef(load); // Add anti-dep as well + return lca; // Cap this stores' anti-dep to here + } + } + return lca; + } + +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/GraphVisualizer.java b/chapter12/src/main/java/com/seaofnodes/simple/GraphVisualizer.java new file mode 100644 index 00000000..604770a4 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/GraphVisualizer.java @@ -0,0 +1,280 @@ +package com.seaofnodes.simple; + +import com.seaofnodes.simple.Parser; +import com.seaofnodes.simple.node.*; + +import java.util.*; + +/** + * Simple visualizer that outputs GraphViz dot format. + * The dot output must be saved to a file and run manually via dot to generate the SVG output. + * Currently, this is done manually. + */ +public class GraphVisualizer { + + /** + * If set to true we put the control nodes in a separate cluster from + * data nodes. + */ + boolean _separateControlCluster = false; + + public GraphVisualizer(boolean separateControlCluster) { this._separateControlCluster = separateControlCluster; } + public GraphVisualizer() { this(false); } + + public String generateDotOutput(Parser parse) { return generateDotOutput(parse.STOP,parse._scope,parse._xScopes); } + public String generateDotOutput(StopNode stop, Node scope, Stack<ScopeNode> xScopes) { + + // Since the graph has cycles, we need to create a flat list of all the + // nodes in the graph. + Collection<Node> 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<Node> 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 ScopeNode || n==Parser.XCTRL) + continue; // Do not emit, rolled into MultiNode or Scope cluster already + if (_separateControlCluster && doCtrl && !n.isCFG()) continue; + if (_separateControlCluster && !doCtrl && n.isCFG()) 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<TABLE BORDER=\"0\" CELLBORDER=\"1\" CELLSPACING=\"0\" CELLPADDING=\"4\">\n"); + sb.append("\t\t\t<TR><TD BGCOLOR=\"yellow\">").append(lab).append("</TD></TR>\n"); + sb.append("\t\t\t<TR>"); + boolean doProjTable = false; + for (Node use : n._outputs) { + if (use instanceof ProjNode proj) { + if (!doProjTable) { + doProjTable = true; + sb.append("<TD>").append("\n"); + sb.append("\t\t\t\t<TABLE BORDER=\"0\" CELLBORDER=\"1\" CELLSPACING=\"0\">").append("\n"); + sb.append("\t\t\t\t<TR>"); + } + sb.append("<TD PORT=\"p").append(proj._idx).append("\">").append(proj.glabel()).append("</TD>"); + } + if (use instanceof CProjNode proj) { + if (!doProjTable) { + doProjTable = true; + sb.append("<TD>").append("\n"); + sb.append("\t\t\t\t<TABLE BORDER=\"0\" CELLBORDER=\"1\" CELLSPACING=\"0\">").append("\n"); + sb.append("\t\t\t\t<TR>"); + } + sb.append("<TD PORT=\"p").append(proj._idx).append("\" BGCOLOR=\"yellow\">").append(proj.glabel()).append("</TD>"); + } + } + if (doProjTable) { + sb.append("</TR>").append("\n"); + sb.append("\t\t\t\t</TABLE>").append("\n"); + sb.append("\t\t\t</TD>"); + } + sb.append("</TR>\n"); + sb.append("\t\t\t</TABLE>>\n\t\t"); + + } else { + // control nodes have box shape + // other nodes are ellipses, i.e. default shape + if (n.isCFG()) 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<Node> 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 level=1; + for( int idx = scope._scopes.size()-1; idx>=0; idx-- ) { + var syms = scope._scopes.get(idx); + String scopeName = makeScopeName(scope, level); + sb.append("\tsubgraph cluster_").append(scopeName).append(" {\n"); // Magic "cluster_" in the subgraph name + sb.append("\t\t").append(scopeName).append(" [label=<\n"); + sb.append("\t\t\t<TABLE BORDER=\"0\" CELLBORDER=\"1\" CELLSPACING=\"0\">\n"); + // Add the scope level + int scopeLevel = scope._scopes.size()-level; + sb.append("\t\t\t<TR><TD BGCOLOR=\"cyan\">").append(scopeLevel).append("</TD>"); + for(String name: syms.keySet()) + sb.append("<TD PORT=\"").append(makePortName(scopeName, name)).append("\">").append(name).append("</TD>"); + sb.append("</TR>\n"); + sb.append("\t\t\t</TABLE>>];\n"); + level++; + } + // 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( level-1 ) ); // 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<Node> 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; i<n.nIns(); i++ ) { + Node def = n.in(i); + if( n instanceof PhiNode && def instanceof RegionNode ) { + // Draw a dotted use->def 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.isCFG() ) + 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=1; + for( int i = scope._scopes.size()-1; i>=0; i-- ) { + var syms = scope._scopes.get(i); + String scopeName = makeScopeName(scope, level); + for( String name : syms.keySet() ) { + int idx = syms.get(name); + Node def = scope.in(idx); + while( def instanceof ScopeNode lazy ) + def = lazy.in(idx); + if( def==null ) continue; + sb.append("\t") + .append(scopeName).append(":") + .append('"').append(makePortName(scopeName, name)).append('"') // wrap port name with quotes because $ctrl is not valid unquoted + .append(" -> "); + if( def instanceof CProjNode proj ) { + String mname = proj.ctrl().uniqueName(); + sb.append(mname).append(":p").append(proj._idx); + } else sb.append(def.uniqueName()); + sb.append(";\n"); + } + level++; + } + } + + /** + * Finds all nodes in the graph. + */ + private Collection<Node> findAll(Node stop, Node scope) { + final HashMap<Integer, Node> 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<Integer,Node> 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/chapter12/src/main/java/com/seaofnodes/simple/IRPrinter.java b/chapter12/src/main/java/com/seaofnodes/simple/IRPrinter.java new file mode 100644 index 00000000..43c6c9f7 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/IRPrinter.java @@ -0,0 +1,266 @@ +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<lim; i++ ) + sb.append(" "); + sb.append(" ]] "); + if( n._type!= null ) sb.append(n._type.str()); + sb.append("\n"); + } + + private static StringBuilder nodeId(StringBuilder sb, Node n) { + sb.append("%%%d".formatted(n._nid)); + if (n instanceof ProjNode proj) { + sb.append(".").append(proj._idx); + } + return sb; + } + + // Print a node on 1 line, format is inspired by LLVM + // %id: TYPE = NODE(inputs ....) + // Nodes as referred to as %id + public static void _printLineLlvmFormat( Node n, StringBuilder sb ) { + nodeId(sb, n).append(": "); + if( n._inputs==null ) { + sb.append("DEAD\n"); + return; + } + if( n._type!= null ) n._type.typeName(sb); + sb.append(" = ").append( n.label() ).append( "(" ); + for( int i = 0; i < n._inputs.size(); i++ ) { + Node def = n.in(i); + if (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<Node> 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.isCFG() || 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<Node> 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.isCFG() ) { + for( Node use : n._outputs ) + if( use!=null && use.isCFG() && use.nOuts()>=1 && !(use._outputs.get(0) instanceof LoopNode) ) + postOrd(use, rpos,visit,bfs); + for( Node use : n._outputs ) + if( use!=null && use.isCFG() ) + 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<Node> _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) { + Node n = Utils.del(_bfs, idx); + _bs.clear(n._nid); + } + 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<Node,Integer> ds = new HashMap<>(); + _walk(ds,node,depth); + // Print by block with least idepth + StringBuilder sb = new StringBuilder(); + ArrayList<Node> bns = new ArrayList<>(); + while( !ds.isEmpty() ) { + CFGNode blk = null; + for( Node n : ds.keySet() ) { + CFGNode cfg = n instanceof CFGNode cfg0 && cfg0.blockHead() ? cfg0 : (CFGNode)n.in(0); + if( blk==null || cfg.idepth() < blk.idepth() ) + blk = cfg; + } + ds.remove(blk); + + // Print block header + sb.append("%-13.13s".formatted(label(blk)+":")); + sb.append( " ".repeat(4) ).append(" [[ "); + if( blk instanceof RegionNode || blk instanceof StopNode ) + for( int i=(blk instanceof StopNode ? 0 : 1); i<blk.nIns(); i++ ) + label(sb,blk.cfg(i)); + else if( !(blk instanceof StartNode) ) + label(sb,blk.cfg(0)); + sb.append(" ]] \n"); + + // Collect block contents that are in the depth limit + bns.clear(); + int xd = Integer.MAX_VALUE; + for( Node use : blk._outputs ) { + Integer i = ds.get(use); + if( i!=null && !(use instanceof CFGNode cfg && cfg.blockHead()) ) { + bns.add(use); + xd = Math.min(xd,i); + } + } + // Print Phis up front, if any + for( int i=0; i<bns.size(); i++ ) + if( bns.get(i) instanceof PhiNode phi ) + printLine( phi, sb, llvmFormat,bns,i--,ds); + + // Print block contents in depth order, bumping depth until whole block printed + for( ; !bns.isEmpty(); xd++ ) + for( int i=0; i<bns.size(); i++ ) + if( ds.get(bns.get(i))==xd ) + printLine( bns.get(i), sb, llvmFormat, bns, i--, ds ); + sb.append("\n"); + } + return sb.toString(); + } + + private static void _walk( HashMap<Node,Integer> ds, Node node, int d ) { + Integer nd = ds.get(node); + if( nd!=null && d <= nd ) return; // Been there, done that + ds.put(node,d); + if( d == 0 ) return; // Depth cutoff + for( Node def : node._inputs ) + if( def != null ) + _walk(ds,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<Node> bns, int i, HashMap<Node,Integer> ds ) { + printLine( n, sb, llvmFormat ); + Utils.del(bns,i); + ds.remove(n); + } + +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/IterPeeps.java b/chapter12/src/main/java/com/seaofnodes/simple/IterPeeps.java new file mode 100644 index 00000000..b2b76777 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/IterPeeps.java @@ -0,0 +1,205 @@ +package com.seaofnodes.simple; + +import com.seaofnodes.simple.node.*; + +import java.util.ArrayList; +import java.util.Arrays; +import java.util.BitSet; +import java.util.Random; + +/** + * The IterOptim 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: + * + * <ul> + * <li>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.</li> + * + * <li>Changing a Node also changes the graph "neighborhood". The neigbors 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.</li> + * + * <li>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 `IterOptim.iterate` as `assert + * progressOnList(stop);`</li> + * </ul> + */ +public abstract class IterPeeps { + + private static final WorkList<Node> WORK = new WorkList<>(); + + public static <N extends Node> N add( N n ) { + return (N)WORK.push(n); + } + + public static void addAll( ArrayList<Node> ary ) { + WORK.addAll(ary); + } + + /** + * Iterate peepholes to a fixed point + */ + public static StopNode iterate(StopNode stop, boolean show) { + 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 + } + + if( show ) + System.out.println(new GraphVisualizer().generateDotOutput(stop,null,null)); + GlobalCodeMotion.buildCFG(stop); + 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 -> { + if( WORK.on(n) ) return null; + Node 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<E extends Node> { + + 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( ArrayList<E> 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/chapter12/src/main/java/com/seaofnodes/simple/Parser.java b/chapter12/src/main/java/com/seaofnodes/simple/Parser.java new file mode 100644 index 00000000..7e957d9d --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/Parser.java @@ -0,0 +1,853 @@ +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. + * <p> + * 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. + * <p> + * To make the compiler multithreaded, this field will have to move into a TLS. + */ + public static StartNode START; + + public static ConstantNode ZERO; + public static XCtrlNode XCTRL; + + 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()}. + * <p> + * 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. + * <p> + * 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<String> KEYWORDS = new HashSet<>(){{ + add("break"); + add("continue"); + add("else"); + add("false"); + add("if"); + add("int"); + add("new"); + add("null"); + add("return"); + add("struct"); + add("true"); + 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<ScopeNode> _xScopes = new Stack<>(); + + ScopeNode _continueScope; + ScopeNode _breakScope; + + // Mapping from a Struct name to a Struct. + public static Map<String, TypeStruct> OBJS = new HashMap<>(); + + public Parser(String source, TypeInteger arg) { + Node.reset(); + IterPeeps.reset(); + OBJS.clear(); + SCHEDULED = false; + _lexer = new Lexer(source); + _scope = new ScopeNode(); + _continueScope = _breakScope = null; + START = new StartNode(new Type[]{ Type.CONTROL, arg }); + STOP = new StopNode(source); + ZERO = new ConstantNode(TypeInteger.constant(0)).peephole().keep(); + XCTRL= new XCtrlNode().peephole().keep(); + } + + public Parser(String source) { + this(source, TypeInteger.BOT); + } + + @Override + public String toString() { return _lexer.toString(); } + + String src() { return new String( _lexer._input ); } + + // 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 Node ctrl(Node 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(); + _scope.define(ScopeNode.CTRL, Type.CONTROL , new CProjNode(START, 0, ScopeNode.CTRL).peephole()); + _scope.define(ScopeNode.ARG0, TypeInteger.BOT, new ProjNode(START, 1, ScopeNode.ARG0).peephole()); + parseBlock(); + if( ctrl()._type==Type.CONTROL ) + STOP.addReturn(new ReturnNode(ctrl(), new ConstantNode(TypeInteger.constant(0)).peephole(), _scope).peephole()); + _scope.pop(); + _xScopes.pop(); + if (!_lexer.isEOF()) throw error("Syntax error, unexpected " + _lexer.getAnyNextToken()); + STOP.peephole(); + if( show ) showGraph(); + return STOP; + } + + /** + * Parses a block + * + * <pre> + * '{' statements '}' + * </pre> + * Does not parse the opening or closing '{}' + * @return a {@link Node} or {@code null} + */ + private Node parseBlock() { + // Enter a new scope + _scope.push(); + while (!peek('}') && !_lexer.isEOF()) + parseStatement(); + // Exit scope + _scope.pop(); + return null; + } + + /** + * Parses a statement + * + * <pre> + * returnStatement | declStatement | blockStatement | ifStatement | expressionStatement + * </pre> + * @return a {@link Node} or {@code null} + */ + private Node parseStatement() { + if( false ) return null; + else if (matchx("return") ) return parseReturn(); + else if (matchx("int") ) return parseDecl(TypeInteger.BOT); + else if (match ("{") ) return require(parseBlock(),"}"); + else if (matchx("if") ) return parseIf(); + else if (matchx("while") ) return parseWhile(); + 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(),";"); + // declarations of vars with struct type are handled in parseExpressionStatement due + // to ambiguity + else return parseExpressionStatement(); + } + + /** + * Parse a struct field. + * <pre> + * int IDENTIFIER ; + * </pre> + */ + private Field parseField(String sname) { + if( matchx("int") ) // Currently only parsing "int" type fields + return require(Field.make(requireId().intern(),TypeInteger.BOT),";"); + throw errorSyntax("A field type is expected, only type 'int' is supported at present"); + } + + /** + * 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(); + if ( OBJS.containsKey(typeName)) throw errorSyntax("struct '" + typeName + "' cannot be redefined"); + ArrayList<Field> fields = new ArrayList<>(); + require("{"); + while (!peek('}') && !_lexer.isEOF()) { + Field field = parseField(typeName); + if (fields.contains(field)) throw errorSyntax("Field '" + field + "' already defined in struct '" + typeName + "'"); + fields.add(field); + } + require("}"); + // Build and install the TypeStruct + TypeStruct ts = TypeStruct.make(typeName, fields); + OBJS.put(typeName, ts); // Insert the struct name in the collection of all struct names + START.addMemProj(ts, _scope); // Insert memory edges + return parseStatement(); + } + + /** + * Parses a while statement + * + * <pre> + * while ( expression ) statement + * </pre> + * @return a {@link Node}, never {@code null} + */ + private Node parseWhile() { + + var savedContinueScope = _continueScope; + var savedBreakScope = _breakScope; + + require("("); + + // 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 = require(parseExpression(), ")"); + // 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(); + + // 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()); + + // 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 + parseStatement(); // Parse loop body + + // Merge the loop bottom into other continue statements + if (_continueScope != null) { + _continueScope = jumpTo(_continueScope); + _scope.kill(); + _scope = _continueScope; + } + + // 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. + return _scope = exit; + } + + 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._scopes.size() > _breakScope._scopes.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._scopes.size() <= _breakScope._scopes.size(); + toScope.ctrl(toScope.mergeScopes(cur)); + return toScope; + } + + private void checkLoopActive() { if (_breakScope == null) throw Parser.error("No active loop for a break or continue"); } + + private Node parseBreak () { checkLoopActive(); return ( _breakScope = require(jumpTo( _breakScope ),";")); } + private Node parseContinue() { checkLoopActive(); return (_continueScope = require(jumpTo( _continueScope ),";")); } + + /** + * Parses a statement + * + * <pre> + * if ( expression ) statement [else statement] + * </pre> + * @return a {@link Node}, never {@code null} + */ + private Node parseIf() { + require("("); + // Parse predicate + var pred = require(parseExpression(), ")").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.upcast(ifT,pred,false); // Up-cast predicate + parseStatement(); // Parse true-side + 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 + if (matchx("else")) { + _scope.upcast(ifF,pred,true); // Up-cast predicate + parseStatement(); + fScope = _scope; + } + pred.unkeep(); + + if( tScope.nIns() != ndefs || fScope.nIns() != ndefs ) + throw error("Cannot define a new name on one arm of an if"); + + // Merge results + _scope = tScope; + _xScopes.pop(); // Discard pushed from graph display + + return ctrl(tScope.mergeScopes(fScope)); + } + + + /** + * Parses a return statement; "return" already parsed. + * The $ctrl edge is killed. + * + * <pre> + * 'return' expr ; + * </pre> + * @return an expression {@link Node}, never {@code null} + */ + private Node parseReturn() { + var expr = require(parseExpression(), ";"); + 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; + } + + /** + * Parses an expression statement or a declaration statement where type is a struct + * + * <pre> + * name; // Error + * type name; // Define name with default initial value + * type name = expr; // Define name with given initial value + * name = expr; // Reassign existing + * expr // Something else + * </pre> + * @return an expression {@link Node}, never {@code null} + */ + private Node parseExpressionStatement() { + int old = _lexer._position; + Type t = type(); + String name = requireId(); + Node expr; + if( match(";") ) { // Assign a default value + // No type and no expr is an error + if( t==null ) throw errorSyntax("expression"); + expr = new ConstantNode(t.makeInit()).peephole(); + } else if( match("=") ) { // Assign "= expr;" + expr = require(parseExpression(), ";"); + } else { // Neither, so just a normal expression parse + _lexer._position = old; + return require(parseExpression(),";"); + } + + // Defining a new variable vs updating an old one + if( t != null ) { + if( _scope.define(name,t,expr) == null ) + throw error("Redefining name '" + name + "'"); + } else { + Node n = _scope.lookup(name); + t = _scope.lookupDeclaredType(name); + if( n==null ) + throw error("Undefined name '" + name + "'"); + _scope.update(name,expr); + } + if( !expr._type.isa(t) ) + throw error("Type " + expr._type.str() + " is not of declared type " + t.str()); + return expr; + } + + // Parse a type-or-null + private Type type() { + int old = _lexer._position; + String tname = _lexer.matchId(); + if( tname==null ) return null; + if( tname.equals("int") ) return TypeInteger.BOT; + TypeStruct obj = OBJS.get(tname); + if( obj != null ) + return TypeMemPtr.make(obj,match("?")); + // Not a type; unwind the parse + _lexer._position = old; + return null; + } + + /** + * Parses a declStatement + * + * <pre> + * type name = expression ';' + * </pre> + * @return an expression {@link Node} + */ + private Node parseDecl(Type t) { + var name = requireId(); + var expr = match(";") + // Assign a null value + ? new ConstantNode(t.makeInit()).peephole() + // Assign "= expr;" + : require(require("=").parseExpression(), ";"); + if( !expr._type.isa(t) ) + throw error("Type " + expr._type.str() + " is not of declared type " + t.str()); + if( _scope.define(name,t,expr) == null ) + throw error("Redefining name '" + name + "'"); + return expr; + } + + + /** + * Parse an expression of the form: + * + * <pre> + * expr : compareExpr + * </pre> + * @return an expression {@link Node}, never {@code null} + */ + private Node parseExpression() { return parseComparison(); } + + /** + * Parse an expression of the form: + * + * <pre> + * expr : additiveExpr op additiveExpr + * </pre> + * @return an comparator expression {@link Node}, never {@code null} + */ + private Node parseComparison() { + var lhs = parseAddition(); + 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,parseAddition()); + lhs = lhs.peephole(); + if( negate ) // Extra negate for != + lhs = new NotNode(lhs).peephole(); + } + return lhs; + } + + /** + * Parse an additive expression + * + * <pre> + * additiveExpr : multiplicativeExpr (('+' | '-') multiplicativeExpr)* + * </pre> + * @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 = lhs.peephole(); + } + return lhs; + } + + /** + * Parse an multiplicativeExpr expression + * + * <pre> + * multiplicativeExpr : unaryExpr (('*' | '/') unaryExpr)* + * </pre> + * @return a multiply expression {@link Node}, never {@code null} + */ + private Node parseMultiplication() { + var lhs = parseUnary(); + 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 = lhs.peephole(); + } + return lhs; + } + + /** + * Parse a unary minus expression. + * + * <pre> + * unaryExpr : ('-') | '!') unaryExpr | postfixExpr | primaryExpr + * </pre> + * @return a unary expression {@link Node}, never {@code null} + */ + private Node parseUnary() { + if (match("-")) return new MinusNode(parseUnary()).peephole(); + if (match("!")) return new NotNode(parseUnary()).peephole(); + return parsePostfix(parsePrimary()); + } + + /** + * Parse a primary expression: + * + * <pre> + * primaryExpr : integerLiteral | Identifier | true | false | null | new Identifier | '(' expression ')' + * </pre> + * @return a primary {@link Node}, never {@code null} + */ + private Node parsePrimary() { + if( _lexer.isNumber() ) return parseIntegerLiteral(); + if( match("(") ) return require(parseExpression(), ")"); + if( matchx("true" ) ) return new ConstantNode(TypeInteger.constant(1)).peephole(); + if( matchx("false") ) return ZERO; + if( matchx("null" ) ) return new ConstantNode(TypeMemPtr.NULLPTR).peephole(); + if( matchx("new") ) { + String structName = requireId(); + TypeStruct obj = OBJS.get(structName); + if( obj == null) throw errorSyntax("Unknown struct type '" + structName + "'"); + return newStruct(obj); + } + String name = _lexer.matchId(); + if( name == null) throw errorSyntax("an identifier or expression"); + Node n = _scope.lookup(name); + if( n!=null ) return n; + throw error("Undefined name '" + name + "'"); + } + + /** + * Return a NewNode but also generate instructions to initialize it. + */ + private Node newStruct(TypeStruct obj) { + Node n = new NewNode(TypeMemPtr.make(obj), ctrl()).peephole().keep(); + int alias = START._aliasStarts.get(obj._name); + for( Field field : obj._fields ) { + memAlias(alias, new StoreNode(field._fname, alias, memAlias(alias), n, ZERO).peephole()); + alias++; + } + return n.unkeep(); + } + + // 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.lookup(memName(alias) ); } + private Node memAlias(int alias, Node st) { return _scope.update(memName(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. + * + * <pre> + * expr ('.' IDENTIFIER)* [ = expr ] + * </pre> + */ + private Node parsePostfix(Node expr) { + if( !match(".") ) return expr; + + if( !(expr._type instanceof TypeMemPtr ptr) ) + throw error("Expected struct reference but got " + expr._type.str()); + + String name = requireId().intern(); + int idx = ptr._obj==null ? -1 : ptr._obj.find(name); + if( idx == -1 ) throw error("Accessing unknown field '" + name + "' from '" + ptr.str() + "'"); + int alias = START._aliasStarts.get(ptr._obj._name)+idx; + + if( match("=") ) { + // Disambiguate "obj.fld==x" boolean test from "obj.fld=x" field assignment + if( peek('=') ) _lexer._position--; + else { + Node val = parseExpression(); + memAlias(alias, new StoreNode(name, alias, memAlias(alias), expr, val).peephole()); + return expr; // "obj.a = expr" returns the expression while updating memory + } + } + + Type declaredType = ptr._obj._fields[idx]._type; + return parsePostfix(new LoadNode(name, alias, declaredType, memAlias(alias), expr).peephole()); + } + + /** + * Parse integer literal + * + * <pre> + * integerLiteral: [1-9][0-9]* | [0] + * </pre> + */ + private ConstantNode parseIntegerLiteral() { + return (ConstantNode) new ConstantNode(_lexer.parseNumber()).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); } + // 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(); } + + // Require and return an identifier + private String requireId() { + String id = _lexer.matchId(); + if (id != null && !KEYWORDS.contains(id) ) return id; + throw error("Expected an identifier, found '"+id+"'"); + } + + + + // Require an exact match + private Parser require(String syntax) { require(null, syntax); return this; } + private <N> 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]; + } + + 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; + } + } + + + // 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; + } + + 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() {return isNumber(peek());} + boolean isNumber(char ch) {return Character.isDigit(ch);} + + private Type parseNumber() { + String snum = parseNumberString(); + if (snum.length() > 1 && snum.charAt(0) == '0') + throw error("Syntax error: integer values cannot start with '0'"); + return TypeInteger.constant(Long.parseLong(snum)); + } + private String parseNumberString() { + int start = _position; + while (isNumber(nextChar())) ; + return new String(_input, start, --_position - start); + } + + // 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); + } + } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/Utils.java b/chapter12/src/main/java/com/seaofnodes/simple/Utils.java new file mode 100644 index 00000000..f3ac25b1 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/Utils.java @@ -0,0 +1,58 @@ +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 <E> E del(ArrayList<E> array, int i) { + if ( i >= 0 && i < array.size() ) { + E tmp = array.get(i); + E last = array.removeLast(); + if (i < array.size()) array.set(i, last); + return tmp; + } + return null; + } + + /** + * 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 <E> int find( ArrayList<E> ary, E x ) { + for( int i=0; i<ary.size(); i++ ) + if( ary.get(i)==x ) + return i; + return -1; + } + /** + * 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 <E> int find( E[] ary, E x ) { + for( int i=0; i<ary.length; i++ ) + if( ary[i]==x ) + return i; + return -1; + } + + // Rotate a long, nice for hashes + public static long rot( long x, int n ) { return (x<<n) | (x>>>n); } + + // Fold a 64bit hash into 32 bits + public static int fold( long x ) { return (int)((x>>32) ^ x); } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/AddNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/AddNode.java new file mode 100644 index 00000000..2cc0b465 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/AddNode.java @@ -0,0 +1,174 @@ +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() { + if( in(1)._type instanceof TypeInteger i0 && + in(2)._type instanceof TypeInteger i1 ) { + if (i0.isConstant() && i1.isConstant()) + return TypeInteger.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 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.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. + // 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); + // 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<ns.length; i++ ) + ns[i] = op.copy(lphi.in(i), rhs instanceof PhiNode ? rhs.in(i) : rhs).peephole(); + String label = lphi._label + (rhs instanceof PhiNode rphi ? rphi._label : ""); + Node phi = new PhiNode(label,lphi._declaredType,ns).peephole(); + // Rotate needs another op, otherwise just the phi + return lhs==lphi ? phi : op.copy(lhs.in(1),phi); + } + + /** + * Tests if the op is a phi and has all constant inputs. + * If not, returns null. + * If op is a phi, but its inputs are not all constants, then dep is added as + * a dependency to the phi's non-const input, because if later the input turn into a constant + * dep can make progress. + */ + static PhiNode pcon(Node op, Node dep) { + return op instanceof PhiNode phi && phi.allCons(dep) ? phi : null; + } + + // Compare two off-spine nodes and decide what order they should be in. + // Do we rotate ((x + hi) + lo) into ((x + lo) + hi) ? + // Generally constants always go right, then Phi-of-constants, then muls, then others. + // Ties with in a category sort by node ID. + // TRUE if swapping hi and lo. + static boolean spine_cmp( Node hi, Node lo, Node dep ) { + if( lo._type.isConstant() ) return false; + if( hi._type.isConstant() ) return true ; + + if( lo instanceof PhiNode lphi && lphi.region()._type==Type.XCONTROL ) return false; + if( hi instanceof PhiNode hphi && hphi.region()._type==Type.XCONTROL ) return false; + + if( lo instanceof PhiNode && lo.allCons(dep) ) return false; + if( hi instanceof PhiNode && hi.allCons(dep) ) return true ; + + if( lo instanceof PhiNode && !(hi instanceof PhiNode) ) return true; + if( hi instanceof PhiNode && !(lo instanceof PhiNode) ) return false; + + // Same category of "others" + return lo._nid > hi._nid; + } + + @Override Node copy(Node lhs, Node rhs) { return new AddNode(lhs,rhs); } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/BoolNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/BoolNode.java new file mode 100644 index 00000000..9c15f01b --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/BoolNode.java @@ -0,0 +1,75 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeInteger; +import com.seaofnodes.simple.type.TypeMemPtr; + +import java.util.BitSet; + +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 Type compute() { + if( in(1)._type instanceof TypeInteger i0 && + in(2)._type instanceof TypeInteger i1 ) { + if (i0.isConstant() && i1.isConstant()) + return TypeInteger.constant(doOp(i0.value(), i1.value()) ? 1 : 0); + } + return in(1)._type.meet(in(2)._type); + } + + abstract boolean doOp(long lhs, long rhs); + + @Override + public Node idealize() { + // Compare of same + if( in(1)==in(2) ) + return new ConstantNode(TypeInteger.constant(doOp(3,3)?1:0)); + + // 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 ) + return new EQ(in(2),in(1)); + // Equals sorts by NID otherwise: non.high == non.low becomes non.low == non.high + else if( in(1)._nid > in(2)._nid ) + return new EQ(in(2),in(1)); + } + // Equals X==0 becomes a !X + if( (in(2)._type == TypeInteger.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 "=="; } boolean doOp(long lhs, long rhs) { return lhs == rhs; } Node copy(Node lhs, Node rhs) { return new EQ(lhs,rhs); } } + public static class LT extends BoolNode { public LT(Node lhs, Node rhs) { super(lhs,rhs); } String op() { return "<" ; } boolean doOp(long lhs, long rhs) { return lhs < rhs; } Node copy(Node lhs, Node rhs) { return new LT(lhs,rhs); } } + public static class LE extends BoolNode { public LE(Node lhs, Node rhs) { super(lhs,rhs); } String op() { return "<="; } boolean doOp(long lhs, long rhs) { return lhs <= rhs; } Node copy(Node lhs, Node rhs) { return new LE(lhs,rhs); } } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/CFGNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/CFGNode.java new file mode 100644 index 00000000..b605071b --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/CFGNode.java @@ -0,0 +1,92 @@ +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 + * <p> + * CFG nodes have a immediate dominator depth (idepth) and a loop nesting + * depth(loop_depth). + * <p> + * idepth is computed lazily upon first request, and is valid even in the + * Parser, and is used by peepholes during parsing and afterward. + * <p> + * loop_depth is computed after optimization as part of scheduling. + * + */ +public abstract class CFGNode extends Node { + + public CFGNode(Node... nodes) { super(nodes); } + + @Override public boolean isCFG() { return true; } + @Override public boolean isPinned() { return true; } + + 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; } + + // Should be exactly 1 tail from a block head + public CFGNode blockTail() { + assert blockHead(); + for( Node n : _outputs ) + if( n instanceof CFGNode cfg ) + return cfg; + return null; + } + + // ------------------------------------------------------------------------ + /** + * 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. + * <p> + * See {@link <a href="https://en.wikipedia.org/wiki/Dominator_(graph_theory)">...</a>} + */ + 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() { return cfg(0); } + + // Return the LCA of two idoms + public CFGNode idom(CFGNode rhs) { + if( rhs==null ) return this; + CFGNode lhs = this; + while( lhs != rhs ) { + var comp = lhs.idepth() - rhs.idepth(); + if( comp >= 0 ) lhs = lhs.idom(); + if( comp <= 0 ) rhs = rhs.idom(); + } + return lhs; + } + + // Loop nesting depth + public int _loopDepth; + public int loopDepth() { return _loopDepth==0 ? (_loopDepth = cfg(0).loopDepth()) : _loopDepth; } + + // Anti-dependence field support + public int _anti; // Per-CFG field to help find anti-deps + + // ------------------------------------------------------------------------ + // Support routines for Global Code Motion + + // Tik-tok recursion pattern. This method is final, and every caller does + // this work. + public final void walkUnreach( BitSet visit, HashSet<CFGNode> unreach ) { + if( visit.get(_nid) ) return; + visit.set(_nid); + _walkUnreach(visit,unreach); + unreach.remove(this); // Since we reached here... Node was not unreachable + } + + // Tik-tok recursion pattern; not-final; callers override this. + void _walkUnreach( BitSet visit, HashSet<CFGNode> unreach ) { + cfg(0).walkUnreach(visit,unreach); + } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/CProjNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/CProjNode.java new file mode 100644 index 00000000..3d6fd1fa --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/CProjNode.java @@ -0,0 +1,64 @@ +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 + public Node getBlockStart() { + return ctrl() instanceof IfNode ? this : ctrl().getBlockStart(); + } + @Override + boolean eq( Node n ) { return _idx == ((CProjNode)n)._idx; } + + @Override + int hash() { return _idx; } + +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/CastNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/CastNode.java new file mode 100644 index 00000000..fec540f7 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/CastNode.java @@ -0,0 +1,38 @@ +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 + 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/chapter12/src/main/java/com/seaofnodes/simple/node/ConstantNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/ConstantNode.java new file mode 100644 index 00000000..cf1a76b0 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/ConstantNode.java @@ -0,0 +1,56 @@ +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. + * <p> + * 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 <em>solely</em> to allow visitation. + * <p> + * 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; } + // The special forever-zero is pinned, all others can sink + @Override public boolean isPinned() { return this==Parser.ZERO; } + + @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/chapter12/src/main/java/com/seaofnodes/simple/node/DivNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/DivNode.java new file mode 100644 index 00000000..ba06d037 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/DivNode.java @@ -0,0 +1,38 @@ +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() { + if (in(1)._type instanceof TypeInteger i0 && + in(2)._type instanceof TypeInteger i1) { + if (i0.isConstant() && i1.isConstant()) + return i1.value() == 0 + ? TypeInteger.ZERO + : TypeInteger.constant(i0.value()/i1.value()); + } + return in(1)._type.meet(in(2)._type); + } + + @Override + public Node idealize() { return null; } + + @Override Node copy(Node lhs, Node rhs) { return new DivNode(lhs,rhs); } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/IfNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/IfNode.java new file mode 100644 index 00000000..5457ac6e --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/IfNode.java @@ -0,0 +1,71 @@ +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 instanceof TypeInteger ti && ti.isConstant()) + return ti==TypeInteger.ZERO ? TypeTuple.IF_FALSE : 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; + } + + @Override void _walkUnreach( BitSet visit, HashSet<CFGNode> unreach ) { + for( Node proj : _outputs ) + if( ((CProjNode)proj)._loopDepth == 0 ) + unreach.add((CProjNode)proj); + super._walkUnreach(visit,unreach); + } + + @Override public Node getBlockStart() { return ctrl().getBlockStart(); } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/LoadNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/LoadNode.java new file mode 100644 index 00000000..4ed06da4 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/LoadNode.java @@ -0,0 +1,78 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.Utils; +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.Field; + +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 { + + Type _declaredType; + /** + * Load a value from a ptr.field. + * + * @param name The field we are loading + * @param memSlice The memory alias node - this is updated after a Store + * @param memPtr The ptr to the struct from where we load a field + */ + public LoadNode(String name, int alias, Type glb, Node memSlice, Node memPtr) { + super(name, alias, memSlice, memPtr); + _declaredType = glb; + } + + @Override + public String label() { return "Load"; } + + @Override + public String glabel() { return "."+_name; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { return sb.append(".").append(_name); } + + @Override + public Type compute() { + return _declaredType; + } + + @Override + public Node idealize() { + + // Simple Load-after-Store on same address. + if( mem() instanceof StoreNode st && + ptr() == st.ptr() ) { // Must check same object + assert _name.equals(st._name); // Equiv class aliasing is perfect + return st.val(); + } + + // 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 ) { + // 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 = new LoadNode(_name,_alias,_declaredType,memphi.in(1),ptr()).peephole(); + Node ld2 = new LoadNode(_name,_alias,_declaredType,memphi.in(2),ptr()).peephole(); + return new PhiNode(_name,_type,memphi.region(),ld1,ld2); + } + } + + return null; + } + + // Profitable if we find a matching Store on this Phi arm. + private boolean profit(PhiNode phi, int idx) { + Node px = phi.in(idx); + return px!=null && px.addDep(this) instanceof StoreNode st1 && ptr()==st1.ptr(); + } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/LoopNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/LoopNode.java new file mode 100644 index 00000000..c662dee6 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/LoopNode.java @@ -0,0 +1,77 @@ +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() { return entry(); } + + @Override public int loopDepth() { + if( _loopDepth!=0 ) return _loopDepth; // Was already set + _loopDepth = entry()._loopDepth+1; // Entry depth plus one + // One-time tag loop exits + for( CFGNode idom = back(); idom!=this; idom = idom.idom() ) { + // Walk idom in loop, setting depth + idom._loopDepth = _loopDepth; + // Loop exit hits the CProj before the If, instead of jumping from + // Region directly to If. + if( idom instanceof CProjNode proj ) { + assert proj.in(0) instanceof IfNode; // Loop exit test + // Find the loop exit CProj, and set loop_depth + for( Node use : proj.in(0)._outputs ) + if( use instanceof CProjNode proj2 && proj2 != idom ) + proj2._loopDepth = _loopDepth-1; + } + } + return _loopDepth; + } + + // 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 void 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 ) + return; // 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)); + } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/MemOpNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/MemOpNode.java new file mode 100644 index 00000000..7f8d060a --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/MemOpNode.java @@ -0,0 +1,44 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Field; +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; + + public MemOpNode(String name, int alias, Node memSlice, Node memPtr) { + super(null, memSlice, memPtr); + _name = name; + _alias = alias; + } + public MemOpNode(String name, int alias, Node memSlice, Node memPtr, Node value) { + this(name, alias, memSlice, memPtr); + addDef(value); + } + + public Node mem() { return in(1); } + public Node ptr() { return in(2); } + + @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; + return (ptr==Type.BOTTOM || (ptr instanceof TypeMemPtr tmp && tmp._nil) ) + ? "Might be null accessing '" + _name + "'" + : null; + } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/MinusNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/MinusNode.java new file mode 100644 index 00000000..9b687613 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/MinusNode.java @@ -0,0 +1,36 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeInteger; + +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) + return i0.isConstant() ? TypeInteger.constant(-i0.value()) : i0; + 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; + } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/MulNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/MulNode.java new file mode 100644 index 00000000..7e9f88a2 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/MulNode.java @@ -0,0 +1,57 @@ +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() { + if (in(1)._type instanceof TypeInteger i0 && + in(2)._type instanceof TypeInteger i1) { + if (i0.isConstant() && i1.isConstant()) + return TypeInteger.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 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); } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/MultiNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/MultiNode.java new file mode 100644 index 00000000..379e140d --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/MultiNode.java @@ -0,0 +1,12 @@ +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; + } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/NeverNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/NeverNode.java new file mode 100644 index 00000000..f50d3503 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/NeverNode.java @@ -0,0 +1,19 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.Parser; +import com.seaofnodes.simple.type.TypeInteger; +import com.seaofnodes.simple.type.Type; +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 TypeInteger.BOT; } + + @Override public Node idealize() { return null; } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/NewNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/NewNode.java new file mode 100644 index 00000000..0724ec56 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/NewNode.java @@ -0,0 +1,42 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeMemPtr; + +import java.util.BitSet; + +public class NewNode extends Node { + + TypeMemPtr _ptr; + + public NewNode(TypeMemPtr ptr, Node ctrl) { + super(ctrl); + this._ptr = ptr; + } + + @Override + public String label() { return "new"; } + + @Override + public String glabel() { return "new " + _ptr._obj._name; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + sb.append("new "); + return sb.append(_ptr._obj.str()); + } + + @Override public boolean isPinned() { return true; } + + @Override + public Type compute() { return _ptr; } + + @Override + public Node idealize() { return null; } + + @Override + boolean eq(Node n) { return this == n; } + + @Override + int hash() { return _ptr.hashCode(); } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/Node.java b/chapter12/src/main/java/com/seaofnodes/simple/node/Node.java new file mode 100644 index 00000000..fc49ad6a --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/Node.java @@ -0,0 +1,624 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.IRPrinter; +import com.seaofnodes.simple.Parser; +import com.seaofnodes.simple.IterPeeps; +import com.seaofnodes.simple.Utils; +import com.seaofnodes.simple.type.Type; + +import java.util.*; +import java.util.function.Function; + +/** + * All Nodes in the Sea of Nodes IR inherit from the Node class. + * The Node class provides common functionality used by all subtypes. + * Subtypes of Node specialize by overriding methods. + */ +public abstract class Node implements OutNode { + + /** + * Each node has a unique dense Node ID within a compilation context + * The ID is useful for debugging, for using as an offset in a bitvector, + * as well as for computing equality of nodes (to be implemented later). + */ + public final int _nid; + + /** + * Inputs to the node. These are use-def references to Nodes. + * <p> + * 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 #isCFG} node. + */ + public final ArrayList<Node> _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. + * <p> + * 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 ArrayList<Node> _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 ArrayList<>(); + Collections.addAll(_inputs,inputs); + _outputs = new ArrayList<>(); + 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. This version will fail on cycles, which we will + // correct later when we can parse programs with loops. 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 ArrayList<Node> 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 <em>def</em> into a Node. Keeps the edges correct, by removing + * the corresponding <em>use->def</em> edge. This may make the original + * <em>def</em> go dead. This function is co-recursive with {@link #kill}. + * <p> + + * This method is the normal path for altering a Node, because it does the + * proper default edge maintenance. It also <em>immediately</em> 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 Node setDef(int idx, Node 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 self 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); + 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 + old_def.moveDepsToWorklist(); + Utils.del(_inputs, idx); + return this; + } + + /** + * Add a new def to an existing Node. Keep the edges correct by + * adding the corresponding <em>def->use</em> 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 extends Node> 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 ) { + Utils.del(_outputs, Utils.find(_outputs, use)); + return _outputs.isEmpty(); + } + + // Shortcut for "popping" n nodes. A "pop" is basically a + // setDef(last,null) followed by lowering the nIns() count. + void popN(int n) { + unlock(); + for( int i=0; i<n; i++ ) { + Node old_def = _inputs.removeLast(); + 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 <em>uses</em>, by setting all of its <em>defs</em> + * to null. This may recursively kill more Nodes, and is basically dead + * code elimination. + */ + public void kill( ) { + unlock(); + 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 extends Node> N keep() { return addUse(null); } + // Remove bogus null. + public <N extends Node> N unkeep() { + delUse(null); + return (N)this; + } + // Test "keep" status + public boolean iskeep() { return Utils.find(_outputs,null) != -1; } + + + // 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 = Utils.find(n._inputs, this); + n._inputs.set(idx,nnn); + nnn.addUse(n); + } + 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: + * <ul> + * <li>if the Type {@link Type#isConstant}, we replace with a {@link ConstantNode}</li> + * <li>in a future chapter we will look for a + * <a href="https://en.wikipedia.org/wiki/Common_subexpression_elimination">Common Subexpression</a> + * to eliminate.</li> + * <li>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. </li> + * + * Unlike peephole above, this explicitly returns null for no-change, or not-null + * for a better replacement (which can be this). + * </ul> + */ + 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 + * <a href="https://en.wikipedia.org/wiki/Monotonic_function">Monotonic</a> + * as it is part of a Monotone Analysis Framework. + * <a href="https://www.cse.psu.edu/~gxt29/teaching/cse597s21/slides/08monotoneFramework.pdf">See for example this set of slides</a>. + * <p> + * 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. + * <p> + * {@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: + * <pre> ary[idx+1]</pre> + * might turn into Sea-of-Nodes IR: + * <pre> (ary+12)+((idx+1) * 4)</pre> + * This expression can then be idealized into: + * <pre> ary + ((idx*4) + (12 + (1*4)))</pre> + * And more folding: + * <pre> ary + ((idx<<2) + 16)</pre> + * And during code-gen: + * <pre> MOV4 Rary,Ridx,16 // or some such hardware-specific notation </pre> + * <p> + * {@link #idealize} has a very specific calling convention: + * <ul> + * <li>If NO change is made, return {@code null} + * <li>If ANY change is made, return not-null; this can be {@code this} + * <li>The returned Node does NOT call {@link #peephole} on itself; the {@link #peephole} call will recursively peephole it. + * <li>Any NEW nodes that are not directly returned DO call {@link #peephole}. + * </ul> + * <p> + * Examples: + * <table border="3"> + * <tr><th> before </th><th> after </th><th>return </th><th>comment </th></tr> + * <tr><td>{@code (x+5) }</td><td>{@code (x+5) }</td><td>{@code null }</td><td>No change</td></tr> + * <tr><td>{@code (5+x) }</td><td>{@code (x+5) }</td><td>{@code this }</td><td>Swapped arguments</td></tr> + * <tr><td>{@code ((x+1)+2)}</td><td>{@code (x+(1+2))}</td><td>{@code (x+_) }</td><td>Returns 2 new Nodes</td></tr> + * </table> + * + * 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. + * <p> + * Since idealize calls peephole and peephole calls idealize, you must be + * careful that all idealizations are <em>monotonic</em>: 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. + ArrayList<Node> _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 ArrayList<>(); + if( Utils.find(_deps ,dep) != -1 ) return this; // Already on list + if( Utils.find(_inputs,dep) != -1 ) return this; // No need for deps on immediate neighbors + if( Utils.find(_outputs,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<Node,Node> 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<len; i++ ) + if( in(i) != n.in(i) ) + return false; + return eq(n); + } + // Subclasses add extra checks (such as ConstantNodes have same constant), + // and can assume "this!=n" and has the same Java class. + boolean eq( Node n ) { return true; } + + + // Cached hash. If zero, then not computed AND this Node is NOT in the GVN + // table - and can have its edges hacked (which will change his hash + // anyway). If Non-Zero then this Node is IN the GVN table, or is being + // probed to see if it can be inserted. His edges are "locked", because + // hacking his edges will change his hash. + int _hash; + + // If the _hash is set, then the Node is in the GVN table; remove it. + void unlock() { + if( _hash==0 ) return; + Node old = GVN.remove(this); // Pull from table + assert old==this; + _hash=0; // Out of table now + } + + + // Hash of opcode and inputs + @Override public final int hashCode() { + if( _hash != 0 ) return _hash; + int hash = hash(); + for( Node n : _inputs ) + if( n != null ) + hash = hash ^ (hash<<17) ^ (hash>>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 control-flow-graph related */ + public boolean isCFG() { return false; } + + /** Is this Node Memory related */ + public boolean isMem() { return false; } + + /** Return block start from a isCFG() */ + public Node getBlockStart() { return null; } + + /** Pinned in the schedule */ + public boolean isPinned() { return false; } + + // ------------------------------------------------------------------------ + // Peephole utilities + + // Swap inputs without letting either input go dead during the swap. + Node swap12() { + unlock(); // Hash is order dependent + Node tmp = in(1); + _inputs.set(1,in(2)); + _inputs.set(2,tmp); + return this; + } + + /** + * Does this node contain all constants? Ignores in(0), as is usually + * control. In an input is not a constant, we add dep as a dependency to + * it because dep can make progress if the input becomes a constant later. + * It is sufficient for one of the non-const inputs to have the dependency, + * so we don't bother checking the rest. + */ + boolean allCons(Node dep) { + for( int i=1; i<nIns(); i++ ) + if( !(in(i)._type.isConstant()) ) { + in(i).addDep(dep); // If in(i) becomes a constant later, will trigger some peephole + return false; + } + return true; + } + + + // Make a shallow copy (same class) of this Node, with given inputs and + // empty outputs and a new Node ID. The original inputs are ignored. + // Does not need to be implemented in isCFG() nodes. + Node copy(Node lhs, Node rhs) { throw Utils.TODO("Binary ops need to implement copy"); } + + // Report any post-optimize errors + String err() { return null; } + + /** + * Used to allow repeating tests in the same JVM. This just resets the + * Node unique id generator, and is done as part of making a new Parser. + */ + public static void reset() { + UNIQUE_ID = 1; + _disablePeephole=false; + GVN.clear(); + ITER_CNT = ITER_NOP_CNT = 0; + } + + + // Utility to walk the entire graph applying a function; return the first + // not-null result. + private static final BitSet WVISIT = new BitSet(); + final public <E> E walk( Function<Node,E> pred ) { + assert WVISIT.isEmpty(); + E rez = _walk(pred); + WVISIT.clear(); + return rez; + } + + private <E> E _walk( Function<Node,E> 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/chapter12/src/main/java/com/seaofnodes/simple/node/NotNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/NotNode.java new file mode 100644 index 00000000..05e44eb0 --- /dev/null +++ b/chapter12/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; + switch( t0 ) { + case TypeInteger 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.TOP ) return TypeInteger.TOP; + if( p0 == TypeMemPtr.NULLPTR ) return TypeInteger.constant(1); + if( !p0._nil ) return TypeInteger.constant(0); + return TypeInteger.BOT; + case Type t: + if( t0.getClass() != Type.class ) + // Only doing NOT on ints and ptrs + throw Utils.TODO(); + return t0==Type.TOP ? Type.TOP : Type.BOTTOM; + } + } + + @Override + public Node idealize() { return null; } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/OutNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/OutNode.java new file mode 100644 index 00000000..7e531145 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/OutNode.java @@ -0,0 +1,7 @@ +package com.seaofnodes.simple.node; + +import java.util.ArrayList; + +public interface OutNode { + ArrayList<Node> outs(); +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/PhiNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/PhiNode.java new file mode 100644 index 00000000..a07dc140 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/PhiNode.java @@ -0,0 +1,159 @@ +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_"+_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.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.glb().dual(); + 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 && !op.isCFG() && same_op() ) { + Node[] lhss = new Node[nIns()]; + Node[] rhss = new Node[nIns()]; + lhss[0] = rhss[0] = in(0); // Set Region + for( int i=1; i<nIns(); i++ ) { + lhss[i] = in(i).in(1); + rhss[i] = in(i).in(2); + } + Node phi_lhs = new PhiNode(_label, _declaredType,lhss).peephole(); + Node phi_rhs = new PhiNode(_label, _declaredType,rhss).peephole(); + return op.copy(phi_lhs,phi_rhs); + } + + // If merging Phi(N, cast(N)) - we are losing the cast JOIN effects, so just remove. + if( nIns()==3 ) { + if( in(1) instanceof CastNode cast && cast.in(1).addDep(this)==in(2) ) return in(2); + if( in(2) instanceof CastNode cast && cast.in(1).addDep(this)==in(1) ) return in(1); + } + // If merging a null-checked null and the checked value, just use the value. + // if( val ) ..; phi(Region,False=0/null,True=val); + // then replace with plain val. + if( nIns()==3 ) { + int nullx = -1; + if( in(1)._type == in(1)._type.makeInit() ) nullx = 1; + if( in(2)._type == in(2)._type.makeInit() ) nullx = 2; + if( nullx != -1 ) { + Node val = in(3-nullx); + if( region().idom() instanceof IfNode iff && iff.pred().addDep(this)==val ) { + // Must walk the idom on the null side to make sure we hit False. + CFGNode idom = (CFGNode)region().in(nullx); + while( 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; i<nIns(); i++ ) + if( in(1).getClass() != in(i).getClass() ) + return false; + return true; + } + + /** + * If only single unique input, return it + */ + private Node singleUniqueInput() { + if( region() instanceof LoopNode loop && loop.entry()._type == Type.XCONTROL ) + return null; // Dead entry loops just ignore and let the loop collapse + Node live = null; + 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( region().in(i).addDep(this)._type != Type.XCONTROL && in(i) != this ) + if( live == null || live == in(i) ) live = in(i); + else return null; + } + return live; + } + + @Override + boolean allCons(Node dep) { + if( !(region() instanceof RegionNode r) ) return false; + // When the region completes (is no longer in progress) the Phi can + // become a "all constants" Phi, and the "dep" might make progress. + addDep(dep); + if( r.inProgress() ) return false; + return super.allCons(dep); + } + + // True if last input is null + public final boolean inProgress() { + return in(nIns()-1) == null; + } + + // Never equal if inProgress + @Override boolean eq( Node n ) { + return !inProgress(); + } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/ProjNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/ProjNode.java new file mode 100644 index 00000000..f94e98a7 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/ProjNode.java @@ -0,0 +1,46 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeMem; +import com.seaofnodes.simple.type.TypeTuple; + +import java.util.BitSet; + +public class ProjNode extends Node { + + // Which slice of the incoming multipart value + public final int _idx; + + // Debugging label + public final String _label; + + public ProjNode(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 isMem() { return _type instanceof TypeMem; } + @Override public boolean isPinned() { return true; } + + @Override + public Type compute() { + Type t = in(0)._type; + return t instanceof TypeTuple tt ? tt._types[_idx] : Type.BOTTOM; + } + + @Override public Node idealize() { return null; } + + @Override + boolean eq( Node n ) { return _idx == ((ProjNode)n)._idx; } + + @Override + int hash() { return _idx; } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/RegionNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/RegionNode.java new file mode 100644 index 00000000..dcc7ccef --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/RegionNode.java @@ -0,0 +1,129 @@ +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.Type; +import java.util.ArrayList; +import java.util.Arrays; +import java.util.BitSet; +import java.util.HashSet; + +public class RegionNode extends CFGNode { + + public RegionNode(Node... nodes) { super(nodes); } + + @Override + public String label() { return "Region"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + return sb.append(label()).append(_nid); + } + + @Override public boolean isMultiHead() { return true; } + @Override public boolean blockHead() { return true; } + + @Override + public Type compute() { + if( inProgress() ) return Type.CONTROL; + Type t = Type.XCONTROL; + for (int i = 1; i < nIns(); i++) + t = t.meet(in(i)._type); + return t; + } + + @Override + public Node idealize() { + if( inProgress() ) return null; + // Delete dead paths into a Region + int path = findDeadInput(); + if( path != 0 && + // Do not delete the entry path of a loop (ok to remove the back + // edge and make the loop a single-entry Region which folds away + // the Loop). Folding the entry path confused the loop structure, + // moving the backedge to the entry point. + !(this instanceof LoopNode loop && loop.entry()==in(path)) ) { + // Cannot use the obvious output iterator here, because a Phi + // deleting an input might recursively delete *itself*. This + // shuffles the output array, and we might miss iterating an + // unrelated Phi. So on rare occasions we repeat the loop to get + // all the Phis. + int nouts = 0; + while( nouts != nOuts() ) { + nouts = nOuts(); + for( int i=0; i<nOuts(); i++ ) + if( out(i) instanceof PhiNode phi && phi.nIns()==nIns() ) { + phi.delDef(path); + IterPeeps.addAll(phi._outputs); + } + } + return isDead() ? Parser.XCTRL : delDef(path); + } + // If down to a single input, become that input + if( nIns()==2 && !hasPhi() ) + return in(1); // Collapse if no Phis; 1-input Phis will collapse on their own + + // If a CFG diamond with no merging, delete: "if( pred ) {} else {};" + if( !hasPhi() && // No Phi users, just a control user + in(1) instanceof ProjNode p1 && + in(2) instanceof ProjNode p2 && + p1.in(0)==p2.in(0) && + p1.in(0) instanceof IfNode iff ) + return iff.ctrl(); + + return null; + } + + private int findDeadInput() { + for( int i=1; i<nIns(); i++ ) + if( in(i)._type==Type.XCONTROL ) + return i; + return 0; // All inputs alive + } + + private boolean hasPhi() { + for( Node phi : _outputs ) + if( phi instanceof PhiNode ) + return true; + return false; + } + + // Immediate dominator of Region is a little more complicated. + @Override public int idepth() { + if( _idepth!=0 ) return _idepth; + int d=0; + for( Node n : _inputs ) + if( n!=null ) + d = Math.max(d,((CFGNode)n).idepth()+1); + return _idepth=d; + } + + @Override public CFGNode idom() { + CFGNode lca = null; + // Walk the LHS & RHS idom trees in parallel until they match, or either fails. + // Because this does not cache, it can be linear in the size of the program. + for( int i=1; i<nIns(); i++ ) + lca = cfg(i).idom(lca); + return lca; + } + + @Override void _walkUnreach( BitSet visit, HashSet<CFGNode> unreach ) { + for( int i=1; i<nIns(); i++ ) + cfg(i).walkUnreach(visit,unreach); + } + + @Override public int loopDepth() { return _loopDepth==0 ? (_loopDepth = cfg(1).loopDepth()) : _loopDepth; } + + // True if last input is null + public final boolean inProgress() { + return nIns()>1 && in(nIns()-1) == null; + } + + // Never equal if inProgress + @Override boolean eq( Node n ) { + return !inProgress(); + } + + @Override public Node getBlockStart() { return this; } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/ReturnNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/ReturnNode.java new file mode 100644 index 00000000..444b76a8 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/ReturnNode.java @@ -0,0 +1,57 @@ +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. + * <p> + * In this presentation, Return functions as a Stop node, since multiple <code>return</code> statements are not possible. + * The Stop node will be introduced in Chapter 6 when we implement <code>if</code> statements. + * <p> + * The Return's output is the value from the data node. + */ +public class ReturnNode extends CFGNode { + + public ReturnNode(Node ctrl, Node data, ScopeNode scope) { + super(ctrl, data); + // We lookup memory slices by the naming convention that they start with $ + // We could also use implicit knowledge that all memory projects are at offset >= 2 + if( scope != null ) { + String[] names = scope.reverseNames(); + for (String name: names) { + if (!name.equals("$ctrl") && name.startsWith("$")) + addDef(scope.lookup(name)); + } + } + } + + public Node ctrl() { return in(0); } + public Node expr() { return in(1); } + + @Override + public String label() { return "Return"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + sb.append("return "); + expr()._print0(sb, visited); + return sb.append(";"); + } + + @Override + public Type compute() { + return TypeTuple.make(ctrl()._type,expr()._type); + } + + @Override + public Node idealize() { + if( ctrl()._type==Type.XCONTROL ) + return ctrl(); + return null; + } + + @Override public Node getBlockStart() { return ctrl().getBlockStart(); } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/ScopeNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/ScopeNode.java new file mode 100644 index 00000000..843ce733 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/ScopeNode.java @@ -0,0 +1,297 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.IterPeeps; +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.Utils; +import com.seaofnodes.simple.type.TypeMemPtr; + +import java.util.*; + +/** + * The Scope node is purely a parser helper - it tracks names to nodes with a + * stack of scopes. + */ +public class ScopeNode extends Node { + + /** + * The control is a name that binds to the currently active control + * node in the graph + */ + public static final String CTRL = "$ctrl"; + public static final String ARG0 = "arg"; + + /** + * Names for every input edge + */ + public final Stack<HashMap<String, Integer>> _scopes; + + /** + * Tracks declared types for every name + */ + public Stack<HashMap<String, Type>> _types; + + + // A new ScopeNode + public ScopeNode() { + _scopes = new Stack<>(); + _types = new Stack<>(); + _type = Type.BOTTOM; + } + + + @Override public String label() { return "Scope"; } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + sb.append("Scope[ "); + String[] names = reverseNames(); + for( int j=0; j<nIns(); j++ ) { + sb.append(names[j]).append(":"); + Node n = in(j); + while( n instanceof ScopeNode loop ) { + sb.append("Lazy_"); + n = loop.in(j); + } + n._print0(sb, visited).append(" "); + } + sb.setLength(sb.length()-1); + return sb.append("]"); + } + + /** + * Recover the names for all variable bindings. + * The result is an array of names that is aligned with the + * inputs to the Node. + * + * This is an expensive operation. + */ + public String[] reverseNames() { + String[] names = new String[nIns()]; + for( HashMap<String,Integer> syms : _scopes ) + for( String name : syms.keySet() ) + names[syms.get(name)] = name; + return names; + } + + @Override public Type compute() { return Type.BOTTOM; } + + @Override public Node idealize() { return null; } + + public void push() { _scopes.push(new HashMap<>()); _types.push(new HashMap<>()); } + public void pop() { popN(_scopes.pop().size()); _types.pop(); } + + /** + * Create a new name in the current scope + */ + public Node define( String name, Type declaredType, Node n ) { + HashMap<String,Integer> syms = _scopes.lastElement(); + _types.lastElement().put(name,declaredType); + if( syms.put(name,nIns()) != null ) + return null; // Double define + return addDef(n); + } + /** + * Lookup a name in all scopes starting from most deeply nested. + * + * @param name Name to be looked up + * @see #update(String, Node, int) + */ + public Node lookup( String name ) { return update(name,null,_scopes.size()-1); } + /** + * 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 Node update( String name, Node n ) { return update(name,n,_scopes.size()-1); } + /** + * Both recursive lookup and update. + * <p> + * A shared implementation allows us to create lazy phis both during + * lookups and updates; the lazy phi creation is part of chapter 8. + * + * @param name Name whose binding is being updated or looked up + * @param n If null, do a lookup, else update binding + * @param nestingLevel The starting nesting level + * @return node being looked up, or the one that was updated + */ + private Node update( String name, Node n, int nestingLevel ) { + if( nestingLevel<0 ) return null; // Missed in all scopes, not found + var syms = _scopes.get(nestingLevel); // Get the symbol table for nesting level + var idx = syms.get(name); + if( idx == null ) return update(name,n,nestingLevel-1); // Not found in this scope, recursively look in parent scope + Node old = in(idx); + if( old instanceof ScopeNode loop ) { + // Lazy Phi! + old = loop.in(idx) instanceof PhiNode phi && loop.ctrl()==phi.region() + // Loop already has a real Phi, use it + ? loop.in(idx) + // 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(idx,new PhiNode(name, lookupDeclaredType(name),loop.ctrl(),loop.update(name,null,nestingLevel),null).peephole()); + setDef(idx,old); + } + return n==null ? old : setDef(idx,n); // Not lazy, so this is the answer + } + + // Return declared type + public Type lookupDeclaredType( String name ) { + for( int i=_types.size(); i>0; i-- ) { + Type t = _types.get(i-1).get(name); + if( t != null ) return t; + } + return null; + } + + public Node ctrl() { return in(0); } + + /** + * The ctrl of a ScopeNode is always bound to the currently active + * control node in the graph, via a special name '$ctrl' that is not + * a valid identifier in the language grammar and hence cannot be + * referenced in Simple code. + * + * @param n The node to be bound to '$ctrl' + * + * @return Node that was bound + */ + public Node ctrl(Node n) { return setDef(0,n); } + + /** + * 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). + * <p> + * If the {@code loop} flag is set, the edges are filled in as the original + * Scope, as a 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. + * <p> + * 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 + for( HashMap<String,Integer> syms : _scopes ) + dup._scopes.push(new HashMap<>(syms)); + for( HashMap<String,Type> ts : _types ) + dup._types.push(ts); // Types don't change, just keep stacks aligned + + dup.addDef(ctrl()); // Control input is just copied + for( int i=1; i<nIns(); i++ ) + // For lazy phis on loops we use a sentinel + // that will trigger phi creation on update + dup.addDef(loop ? this : in(i)); + return dup; + } + + /** + * Merges the names whose node bindings differ, by creating Phi node for such names + * The names could occur at all stack levels, but a given name can only differ in the + * innermost stack level where the name is bound. + * + * @param that The ScopeNode to be merged into this + * @return A new node representing the merge point + */ + public Node mergeScopes(ScopeNode that) { + RegionNode r = (RegionNode) ctrl(new RegionNode(null,ctrl(), that.ctrl()).keep()); + String[] ns = reverseNames(); + // Note that we skip i==0, which is bound to '$ctrl' + for (int i = 1; i < nIns(); i++) + if( in(i) != that.in(i) ) // No need for redundant Phis + // If we are in lazy phi mode we need to a lookup + // by name as it will trigger a phi creation + setDef(i, new PhiNode(ns[i], this.lookupDeclaredType(ns[i]), r, this.lookup(ns[i]), that.lookup(ns[i])).peephole()); + that.kill(); // Kill merged scope + IterPeeps.add(r); + return r.unkeep().peephole(); + } + + // Merge the backedge scope into this loop head scope + // We set the second input to the phi from the back edge (i.e. loop body) + public void endLoop(ScopeNode back, ScopeNode exit ) { + Node ctrl = ctrl(); + assert ctrl instanceof LoopNode loop && loop.inProgress(); + ctrl.setDef(2,back.ctrl()); + for( int i=1; i<nIns(); i++ ) { + if( back.in(i) != this ) { + PhiNode phi = (PhiNode)in(i); + assert phi.region()==ctrl && phi.in(2)==null; + phi.setDef(2,back.in(i)); + } + if( exit.in(i) == this ) // Replace a lazy-phi on the exit path also + exit.setDef(i,in(i)); + } + back.kill(); // Loop backedge is dead + // Now one-time do a useless-phi removal + for( int i=1; i<nIns(); i++ ) { + if( in(i) instanceof PhiNode phi ) { + // Do an eager useless-phi removal + Node in = phi.peephole(); + IterPeeps.addAll(phi._outputs); + phi.moveDepsToWorklist(); + if( in != phi ) { + if( !phi.iskeep() ) // Keeping phi around for parser elsewhere + phi.subsume(in); + setDef(i,in); // Set the update back into Scope + } + } + } + + } + + + // Up-casting: using the results of an If to improve a value. + // E.g. "if( ptr ) ptr.field;" is legal because ptr is known not-null. + + // This Scope looks for direct variable uses, or certain simple + // combinations, and replaces the variable with the upcast variant. + public Node upcast( Node ctrl, Node pred, boolean invert ) { + if( ctrl._type==Type.XCONTROL ) return null; + // Invert the If conditional + if( invert ) + pred = pred instanceof NotNode not ? not.in(1) : IterPeeps.add(new NotNode(pred).peephole()); + + // Direct use of a value as predicate. This is a zero/null test. + if( Utils.find(_inputs, pred) != -1 ) { + if( !(pred._type instanceof TypeMemPtr tmp) ) + // Must be an `int`, since int and ptr are the only two value types + // being tested. No representation for a generic not-null int, so no upcast. + return null; + if( tmp.isa(TypeMemPtr.VOIDPTR) ) + return null; // Already not-null, no reason to upcast + // Upcast the ptr to not-null ptr, and replace in scope + return replace(pred,new CastNode(TypeMemPtr.VOIDPTR,ctrl,pred).peephole()); + } + + if( pred instanceof NotNode not ) { + // Direct use of a !value as predicate. This is a zero/null test. + if( Utils.find(_inputs, not.in(1)) != -1 ) { + Type tinit = not.in(1)._type.makeInit(); + if( not.in(1)._type.isa(tinit) ) return null; // Already zero/null, no reason to upcast + return replace(not.in(1), new ConstantNode(tinit).peephole()); + } + } + // Apr/9/2024: Attempted to replace X with Y if guarded by a test of + // X==Y. This didn't seem to help very much, or at least in the test + // cases seen so far was a very minor help. + + // No upcast + return null; + } + + private Node replace( Node old, Node cast ) { + assert old!=null && old!=cast; + for( int i=0; i<nIns(); i++ ) + if( in(i)==old ) + setDef(i,cast); + return cast; + } + +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/StartNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/StartNode.java new file mode 100644 index 00000000..62fc8a6b --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/StartNode.java @@ -0,0 +1,85 @@ +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.HashMap; +import java.util.HashSet; + +/** + * 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 CFGNode implements MultiNode { + + // This field is NOT final, because the tuple expands as we add memory aliases + TypeTuple _args; + + // The alias number of the first field in the struct. + // + public final HashMap<String,Integer> _aliasStarts; + + + public StartNode(Type[] args) { + super(); + _args = TypeTuple.make(args); + _type = _args; + _aliasStarts = new HashMap<>(); + } + + /** + * Creates a projection for each of the struct's fields, using the field alias + * as the key. + */ + public void addMemProj( TypeStruct ts, ScopeNode scope) { + int len = _args._types.length; + _aliasStarts.put(ts._name,len); + + // resize the tuple's type array to include all fields of the struct + int max = len + ts._fields.length; + Type[] args = Arrays.copyOf(_args._types, max); + + // The new members of the tuple get a mem type with an alias + for( int alias = len; alias < max; alias++ ) + args[alias] = TypeMem.make(alias); + _type = _args = TypeTuple.make(args); + // For each of the fields we now add a mem projection. Note that the + // alias matches the slot of the field in the tuple + for( int alias = len; alias < max; alias++ ) { + String name = Parser.memName(alias); + Node n = new ProjNode(this, alias, name).peephole(); + scope.define(name, args[alias], n); + } + } + + @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 TypeTuple compute() { return _args; } + + @Override + public Node idealize() { return null; } + + // No immediate dominator, and idepth==0 + @Override public int idepth() { return 0; } + @Override public CFGNode idom() { return null; } + + @Override void _walkUnreach( BitSet visit, HashSet<CFGNode> unreach ) { } + + @Override public int loopDepth() { return (_loopDepth=1); } + + @Override public Node getBlockStart() { return this; } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/StopNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/StopNode.java new file mode 100644 index 00000000..94b2ad6d --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/StopNode.java @@ -0,0 +1,79 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.IterPeeps; + +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<nIns(); i++ ) + if( in(i)._type==Type.XCONTROL ) + delDef(i--); + if( len != nIns() ) return this; + return null; + } + + @Override public int idepth() { + if( _idepth!=0 ) return _idepth; + int d=0; + for( Node n : _inputs ) + if( n!=null ) + d = Math.max(d,((CFGNode)n).idepth()+1); + return _idepth=d; + } + @Override public CFGNode idom() { return null; } + + @Override public int loopDepth() { return (_loopDepth=1); } + + public Node addReturn(Node node) { + return addDef(node); + } + + public StopNode iterate( ) { return IterPeeps.iterate(this,false).typeCheck(); } + public StopNode iterate(boolean show) { return IterPeeps.iterate(this,show ).typeCheck(); } + StopNode typeCheck() { + String err = walk( Node::err ); + if( err != null ) throw new RuntimeException(err); + return this; + } + + @Override public Node getBlockStart() { return this; } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/StoreNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/StoreNode.java new file mode 100644 index 00000000..7c41e767 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/StoreNode.java @@ -0,0 +1,70 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.Utils; +import com.seaofnodes.simple.type.*; + +import java.util.BitSet; + +/** + * Store represents setting a value to a memory based object, in chapter 10 + * this means a field inside a struct. + */ +public class StoreNode extends MemOpNode { + + /** + * @param name The struct field we are assigning to + * @param memSlice The memory alias node - this is updated after a Store + * @param memPtr The ptr to the struct where we will store a value + * @param value Value to be stored + */ + public StoreNode(String name, int alias, Node memSlice, Node memPtr, Node value) { + super(name, alias, memSlice, memPtr, value); + } + + @Override + public String label() { return "Store"; } + @Override + public String glabel() { return "."+_name+" ="; } + @Override + public boolean isMem() { return true; } + public Node val() { return in(3); } + + @Override + StringBuilder _print1(StringBuilder sb, BitSet visited) { + return sb.append(".").append(_name).append("=").append( val()).append(";"); + } + + @Override + public Type compute() { return TypeMem.make(_alias); } + + @Override + public Node idealize() { + + // Simple store-after-store on same address. Should pick up the + // required init-store being stomped by a first user store. + if( mem() instanceof StoreNode st && + ptr()==st.ptr() && // Must check same object + ptr()._type instanceof TypeMemPtr && // No bother if weird dead pointers + // Must have exactly one use of "this" or you get weird + // non-serializable memory effects in the worse case. + st.checkNoUseBeyond(this) ) { + assert _name.equals(st._name); // Equiv class aliasing is perfect + setDef(1,st.mem()); + return this; + } + + return null; + } + + // Check that `this` has no uses beyond `that` + private boolean checkNoUseBeyond(Node that) { + if( nOuts()==1 ) return true; + // Add deps on the other uses (can be e.g. ScopeNode mid-parse) so that + // when the other uses go away we can retry. + for( Node use : _outputs ) + if( use != that ) + use.addDep(that); + return false; + } + +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/SubNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/SubNode.java new file mode 100644 index 00000000..7137f38e --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/SubNode.java @@ -0,0 +1,51 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.Parser; +import com.seaofnodes.simple.type.Type; +import com.seaofnodes.simple.type.TypeInteger; + +import java.util.BitSet; + +public class SubNode extends Node { + public SubNode(Node lhs, Node rhs) { super(null, lhs, rhs); } + + @Override public String label() { return "Sub"; } + + @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 TypeInteger i0 && + in(2)._type instanceof TypeInteger i1) { + if (i0.isConstant() && i1.isConstant()) + return TypeInteger.constant(i0.value()-i1.value()); + } + return in(1)._type.meet(in(2)._type); + } + + @Override + public Node idealize() { + // Sub of same is 0 + if( in(1)==in(2) ) + return Parser.ZERO; + + // x - (-y) is x+y + if( in(2) instanceof MinusNode minus ) + return new AddNode(in(1),minus.in(1)); + + // (-x) - y is -(x+y) + if( in(1) instanceof MinusNode minus ) + return new MinusNode(new AddNode(minus.in(1),in(2)).peephole()); + + return null; + } + + @Override Node copy(Node lhs, Node rhs) { return new SubNode(lhs,rhs); } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/node/XCtrlNode.java b/chapter12/src/main/java/com/seaofnodes/simple/node/XCtrlNode.java new file mode 100644 index 00000000..83779dd2 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/node/XCtrlNode.java @@ -0,0 +1,13 @@ +package com.seaofnodes.simple.node; + +import com.seaofnodes.simple.Parser; +import com.seaofnodes.simple.type.Type; +import java.util.BitSet; + +public class XCtrlNode extends CFGNode { + public XCtrlNode( ) { super(Parser.START); } + @Override public String label() { return "Xctrl"; } + @Override StringBuilder _print1(StringBuilder sb, BitSet visited) { return sb.append("Xctrl"); } + @Override public Type compute() { return Type.XCONTROL; } + @Override public Node idealize() { return null; } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/type/Field.java b/chapter12/src/main/java/com/seaofnodes/simple/type/Field.java new file mode 100644 index 00000000..e1c337d4 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/type/Field.java @@ -0,0 +1,60 @@ +package com.seaofnodes.simple.type; + +import com.seaofnodes.simple.Utils; +import java.util.ArrayList; + +/** + * Represents a field in a struct. This is not a Type in the type system. + */ +public class Field extends Type { + + // The pair {fieldName,type} uniquely identifies a field. + + // Field name + public final String _fname; + // Type of the field + public final Type _type; + + private Field(String fname, Type type ) { + super(TFLD); + _fname = fname; + _type = type; + } + // Make with existing alias + public static Field make( String fname, Type type ) { + return new Field(fname,type).intern(); + } + + public static final Field TEST = make("test",TypeInteger.ZERO); + public static void gather(ArrayList<Type> ts) { ts.add(TEST); } + + @Override Field xmeet( Type that ) { + Field fld = (Field)that; // Invariant + assert _fname.equals(fld._fname); + return make(_fname,_type.meet(fld._type)); + } + + @Override + public Field dual() { return make(_fname,_type.dual()); } + + @Override + public Field glb() { + return make(_fname,_type.glb()); + } + + // Override in subclasses + int hash() { return _fname.hashCode() ^ _type.hashCode(); } + + boolean eq(Type t) { + Field f = (Field)t; + return _fname.equals(f._fname) && _type==f._type; + } + + + @Override + public StringBuilder _print( StringBuilder sb ) { + return _type._print(sb.append(_fname).append(":")); + } + + @Override public String str() { return _fname; } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/type/Type.java b/chapter12/src/main/java/com/seaofnodes/simple/type/Type.java new file mode 100644 index 00000000..aa0ba1ea --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/type/Type.java @@ -0,0 +1,187 @@ +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 <a href="https://www.cse.psu.edu/~gxt29/teaching/cse597s21/slides/08monotoneFramework.pdf">see for example this set of slides</a>. + * <p> + * The types form a lattice; @see <a href="https://en.wikipedia.org/wiki/Lattice_(order)">a symmetric complete bounded (ranked) lattice.</a> + * <p> + * 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. + * <p>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<Type,Type> 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 TTUPLE = 6; // Tuples; finite collections of unrelated Types, kept in parallel + static final byte TMEM = 7; // All memory (alias 0) or A slice of memory - with specific alias + static final byte TMEMPTR = 8; // Memory pointer type + static final byte TSTRUCT = 9; // Structs; tuples with named fields + static final byte TFLD =10; // Fields into struct + + 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<Type> ts = new ArrayList<>(); + ts.add(BOTTOM); + ts.add(CONTROL); + Field.gather(ts); + TypeInteger.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 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); } + + public Type makeInit() { return null; } + + // ---------------------------------------------------------- + + // 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 performance. + + // Factory method which interns "this" + public <T extends Type> 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 ? XCONTROL : BOTTOM; } + + // ---------------------------------------------------------- + // 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/chapter12/src/main/java/com/seaofnodes/simple/type/TypeInteger.java b/chapter12/src/main/java/com/seaofnodes/simple/type/TypeInteger.java new file mode 100644 index 00000000..f2a93a3a --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/type/TypeInteger.java @@ -0,0 +1,101 @@ +package com.seaofnodes.simple.type; + +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(true , 0); + + public final boolean _is_con; + + /** + * The constant value or + * if not constant then 1=bottom, 0=top. + */ + public final long _con; + + private TypeInteger(boolean is_con, long con) { + super(TINT); + _is_con = is_con; + _con = con; + } + public static TypeInteger make(boolean is_con, long con) { + return new TypeInteger(is_con,con).intern(); + } + + public static TypeInteger constant(long con) { return make(true, con); } + + public static void gather(ArrayList<Type> ts) { ts.add(ZERO); ts.add(BOT); } + + // 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("IntTop"); + if( this==BOT ) return sb.append("IntBot"); + return sb.append(_con); + } + + @Override public String str() { + if( this==TOP ) return "~int"; + if( this==BOT ) return "int"; + return ""+_con; + } + + /** + * 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 isHighOrConst() { return _is_con || _con==0; } + + @Override + public boolean isConstant() { return _is_con; } + + public long value() { return _con; } + + @Override + public Type xmeet(Type other) { + // Invariant from caller: 'this' != 'other' and same class (TypeInteger) + TypeInteger i = (TypeInteger)other; // Contract + // BOT wins + if ( this==BOT ) return this; + if ( i ==BOT ) return i ; + // TOP loses + if ( i ==TOP ) return this; + if ( this==TOP ) return i ; + // Since both are constants, and are never equals (contract) unequals + // constants fall to bottom + return BOT; + } + + @Override + public Type dual() { + if( isConstant() ) return this; // Constants are a self-dual + return _con==0 ? BOT : TOP; + } + + @Override + public Type glb() { return BOT; } + + @Override public TypeInteger makeInit() { return ZERO; } + + @Override + int hash() { return (int)(_con ^ (_is_con ? 0 : 0x4000)); } + @Override + public boolean eq( Type t ) { + TypeInteger i = (TypeInteger)t; // Contract + return _con==i._con && _is_con==i._is_con; + } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/type/TypeMem.java b/chapter12/src/main/java/com/seaofnodes/simple/type/TypeMem.java new file mode 100644 index 00000000..198d49ee --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/type/TypeMem.java @@ -0,0 +1,64 @@ +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. + private final int _alias; + + private TypeMem(int alias) { super(TMEM); _alias = alias; } + + public static TypeMem make(int alias) { return new TypeMem(alias).intern(); } + public static final TypeMem TOP = make( 0); + public static final TypeMem BOT = make(-1); + + + public static void gather(ArrayList<Type> ts) { ts.add(make(1)); ts.add(BOT); } + + @Override + TypeMem xmeet(Type t) { + TypeMem that = (TypeMem) t; // Invariant: TypeMem and unequal + return _alias==0 ? that : (that._alias==0 ? this : BOT); + } + + @Override + public Type dual() { + if( _alias== 0 ) return BOT; + if( _alias==-1 ) return TOP; + return this; + } + + @Override + public Type glb() { return TypeMem.BOT; } + + @Override + int hash() { return 9876543 + _alias; } + + @Override + boolean eq(Type t) { + TypeMem that = (TypeMem) t; // Invariant + return _alias == that._alias; + } + + @Override + public StringBuilder _print(StringBuilder sb) { + return sb.append("MEM#").append( switch(_alias) { + case 0 -> "TOP"; + case -1 -> "BOT"; + default -> ""+_alias; + }); + } + + @Override public String str() { return _print(new StringBuilder()).toString(); } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/type/TypeMemPtr.java b/chapter12/src/main/java/com/seaofnodes/simple/type/TypeMemPtr.java new file mode 100644 index 00000000..65352402 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/type/TypeMemPtr.java @@ -0,0 +1,78 @@ +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 structs. + * Pointers can be to specific struct 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 one of + // (null,TypeStruct). + // where nil is one of + // (true,false) meaning an explicit null is allowed or not + + public final TypeStruct _obj; // null, a TypeStruct, or sentinel TypeStruct.MANY + public final boolean _nil; + + private TypeMemPtr(TypeStruct obj, boolean nil) { + super(TMEMPTR); + _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 TypeMemPtr.make(obj, false); } + + public static TypeMemPtr BOT = make(TypeStruct.BOT,true); + public static TypeMemPtr TOP = BOT.dual(); + 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<Type> 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==null ? null : _obj.dual(), !_nil); } + + @Override + public Type glb() { + if( _obj==null ) return BOT; + return make(_obj.glb(),true); + } + @Override public TypeMemPtr makeInit() { return NULLPTR; } + + @Override + int hash() { return (_obj==null ? 0xDEADBEEF : _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/chapter12/src/main/java/com/seaofnodes/simple/type/TypeStruct.java b/chapter12/src/main/java/com/seaofnodes/simple/type/TypeStruct.java new file mode 100644 index 00000000..29102675 --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/type/TypeStruct.java @@ -0,0 +1,121 @@ +package com.seaofnodes.simple.type; + +import com.seaofnodes.simple.Utils; +import java.util.ArrayList; + +/** + * Represents a struct type. + */ +public class TypeStruct extends Type { + + // For now in terms of the lattice + // a struct type just stays as is + + public final String _name; + public final Field[] _fields; + + private TypeStruct(String name, Field[] fields) { + super(Type.TSTRUCT); + _name = name; + _fields = fields; + } + + public static TypeStruct make(String name, Field[] fields) { return new TypeStruct(name, fields).intern(); } + public static TypeStruct make(String name, ArrayList<Field> fields) { return make(name,fields.toArray(new Field[fields.size()])); } + + 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}); + public static void gather(ArrayList<Type> ts) { ts.add(TEST); ts.add(BOT); } + + public int find(String fname) { + for( int i=0; i<_fields.length; i++ ) + if( _fields[i]._fname.equals(fname) ) + 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 + // Now all fields should be the same, so just do field meets + assert _fields.length == that._fields.length; + Field[] flds = new Field[_fields.length]; + for( int i=0; i<_fields.length; i++ ) + flds[i] = (Field)_fields[i].meet(that._fields[i]); + return make(_name,flds); + } + + @Override + public TypeStruct dual() { + if( this==TOP ) return BOT; + if( this==BOT ) return TOP; + 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() { + for( Field f : _fields ) + if( f.glb() != f ) + return false; + return true; + } + + @Override + boolean eq(Type t) { + TypeStruct ts = (TypeStruct)t; // Invariant + if( !_name.equals(ts._name) ) + 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(); + for( Field f : _fields ) hash = Utils.rot(hash,13) ^ f.hashCode(); + return Utils.fold(hash); + } + + @Override + public StringBuilder _print(StringBuilder sb) { + sb.append(_name).append(" {\n"); + for( Field f : _fields ) { + sb.append(" ").append(f._fname).append(":"); + f._type._print(sb); + sb.append(";\n"); + } + return sb.append("}"); + } + + @Override public String str() { return _name; } +} diff --git a/chapter12/src/main/java/com/seaofnodes/simple/type/TypeTuple.java b/chapter12/src/main/java/com/seaofnodes/simple/type/TypeTuple.java new file mode 100644 index 00000000..01bcbe9f --- /dev/null +++ b/chapter12/src/main/java/com/seaofnodes/simple/type/TypeTuple.java @@ -0,0 +1,87 @@ +package com.seaofnodes.simple.type; + +import java.util.ArrayList; + +public class TypeTuple extends Type { + + public final Type[] _types; + + private TypeTuple(Type[] types) { super(TTUPLE); _types = types; } + public static TypeTuple make(Type... types) { return new TypeTuple(types).intern(); } + + private static TypeTuple TEST = make(TypeInteger.BOT,TypeMemPtr.TEST); + public static void gather(ArrayList<Type> ts) { ts.add(TEST); } + + @Override + Type xmeet(Type other) { + TypeTuple tt = (TypeTuple)other; // contract from xmeet + assert _types.length == tt._types.length; + Type[] ts = new Type[_types.length]; + for( int i=0; i<_types.length; i++ ) + ts[i] = _types[i].meet(tt._types[i]); + return make(ts); + } + + @Override public Type dual() { + Type[] ts = new Type[_types.length]; + for( int i=0; i<_types.length; i++ ) + ts[i] = _types[i].dual(); + return make(ts); + } + + @Override + public Type glb() { + Type[] ts = new Type[_types.length]; + for( int i=0; i<_types.length; i++ ) + ts[i] = _types[i].glb(); + return make(ts); + } + + @Override + public StringBuilder _print(StringBuilder sb) { + sb.append("["); + for( Type t : _types ) + t._print(sb).append(","); + sb.setLength(sb.length()-1); + sb.append("]"); + return sb; + } + + @Override public String str() { return _print(new StringBuilder()).toString(); } + + /** + * Display Type name in a format that's good for IR printer + */ + @Override + public StringBuilder typeName(StringBuilder sb) { + sb.append("["); + for( Type t : _types ) + t.typeName(sb).append(","); + sb.setLength(sb.length()-1); + return sb.append("]"); + } + + public static final TypeTuple IF_BOTH = make(new Type[]{Type. CONTROL,Type. CONTROL}); + public static final TypeTuple IF_NEITHER = make(new Type[]{Type.XCONTROL,Type.XCONTROL}); + public static final TypeTuple IF_TRUE = make(new Type[]{Type. CONTROL,Type.XCONTROL}); + public static final TypeTuple IF_FALSE = make(new Type[]{Type.XCONTROL,Type. CONTROL}); + + @Override + int hash() { + int sum = 0; + for( Type type : _types ) sum ^= type.hashCode(); + return sum; + } + + @Override + boolean eq( Type t ) { + TypeTuple tt = (TypeTuple)t; // Contract + if( _types.length != tt._types.length ) return false; + for( int i=0; i<_types.length; i++ ) + if( _types[i]!=tt._types[i] ) + return false; + return true; + } + + +} |