1 package edu.berkeley.sbp;
2 import edu.berkeley.sbp.*;
3 import edu.berkeley.sbp.*;
4 import edu.berkeley.sbp.*;
5 import edu.berkeley.sbp.util.*;
8 import java.lang.reflect.*;
10 //////////////////////////////////////////////////////////////////////////////
13 // - fix public/package/private status
16 //////////////////////////////////////////////////////////////////////////////
17 // Optimizations to add
19 // ** NOTE: not all of these are appropriate for this class -- it is
20 // simply a list of optimizations not implemented. This
21 // class is meant to remain simple and easy to understand;
22 // optimizations which obscure that do not belong here (they
23 // should go into the compiled version instead)
25 // - most of our time is now spent creating and storing Reduct instances
26 // - we should be able to perform Reduct's immediately after creating them...
29 /** implements Tomita's Graph Structured Stack */
34 /** corresponds to a positions <i>between tokens</i> the input stream; same as Tomita's U_i's */
37 /** the token immediately after this phase */
38 public final Token token;
40 /** currently this is necessary only for the code() hack -- it doesn't actually correspond to the input */
41 private final int pos;
44 public Forest.Ref finalResult = null;
46 /** all reductions (pending and completed) */
47 private HashSet<Phase.Reduct> reductions = new HashSet<Phase.Reduct>(); /* ALLOC */
49 /** all nodes, keyed by the value returned by code() */
50 private HashMap<Long,Phase.Node> hash = new HashMap<Long,Phase.Node>(); /* ALLOC */
52 /** the number of pending reductions */
53 private int pendingReductions = 0;
54 private int totalReductions = 0;
55 private HashSet<Reduct> pendingReduct = new HashSet<Reduct>();
57 /** the number of nodes in this phase */
58 private int numNodes = 0;
60 boolean closed = false;
62 private Token.Location location;
63 public Phase(Phase previous, Token token, Token.Location location) {
64 this.pos = previous==null ? 0 : previous.pos+1;
66 this.location = location;
69 public boolean isDone() { return token == null; }
71 private String error = "generic syntax error";
72 public void checkFailure() throws Parser.Failed {
74 throw new Parser.Failed(error, getLocation());
77 public Token.Location getLocation() { return location; }
79 /** add a new node (merging with existing nodes if possible)
80 * @param parent the parent of the new node
81 * @param result the SPPF result corresponding to the new node
82 * @param state the state that the new node is in
83 * @param fromEmptyReduction true iff this node is being created as a result of a reduction of length zero (see GRMLR paper)
84 * @param start the earliest part of the input contributing to this node (used to make merging decisions)
86 public void newNode(Node parent, Forest pending, Parser.Table.State state, boolean fromEmptyReduction, Phase start) {
87 Node p = hash.get(code(state, start));
88 if (p != null) newNode2(p, parent, pending, state, fromEmptyReduction, start);
89 else newNode3(parent, pending, state, fromEmptyReduction, start);
91 private void newNode2(Node p, Node parent, Forest pending, Parser.Table.State state, boolean fromEmptyReduction, Phase start) {
92 p.holder.merge(pending);
93 if (p.parents.contains(parent)) return;
94 p.parents.add(parent, true);
95 if (p!=parent && !fromEmptyReduction) p.queueReductions(parent);
97 private void newNode3(Node parent, Forest pending, Parser.Table.State state, boolean fromEmptyReduction, Phase start) {
99 if (token != null && state.canShift(token)) break;
100 if (state.isAccepting()) break;
101 if (token==null) break;
103 Parser.Table.Reduction r = null;
104 for(Parser.Table.Reduction red : token==null ? state.getEofReductions() : state.getReductions(token)) { r = red; count++; }
105 //if (count==0) return; -- BEWARE! this optimization is suspected to cause really nasty heisenbugs
106 if (count > 1) break;
107 //if (r.numPop == 0) break;
108 //r.reduce(pending, parent, null, Phase.this, null);
112 Node n = new Node(parent, pending, state, start); // ALLOC
113 n.queueEmptyReductions();
114 if (!fromEmptyReduction) n.queueReductions();
117 /** perform all reduction operations */
118 public void reduce() {
119 for(Phase.Node n : hash.values()) {
120 n.queueEmptyReductions();
123 while(pendingReduct.size()>0)
124 pendingReduct.iterator().next().go();
127 /** perform all shift operations, adding promoted nodes to <tt>next</tt> */
128 public void shift(Phase next, Forest result) {
132 for(Phase.Node n : hash.values()) {
134 if (token == null && n.state.isAccepting()) {
136 if (finalResult==null) finalResult = new Forest.Ref();
137 finalResult.merge(n.holder);
139 if (!n.holder.valid()) continue;
140 if (token == null) continue;
141 for(Parser.Table.State st : n.state.getShifts(token)) {
142 if (res == null) res = result;
143 next.newNode(n, res, st, true, this);
148 if (!ok && token != null) {
149 StringBuffer error = new StringBuffer();
150 error.append("error: unable to shift token \"" + token + "\"\n");
151 error.append(" before: " +pendingReductions+ "\n");
152 error.append(" before: " +totalReductions+ "\n");
153 for(Phase.Node n : hash.values()) {
155 n.queueEmptyReductions();
157 error.append(" after: " +pendingReductions+ "\n");
158 error.append(" candidate states:\n");
159 for(Phase.Node n : hash.values()) {
160 //for(Sequence.Position p : n.state) error.append(" " + p + "\n");
161 //error.append(" --\n");
162 for(Parser.Table.Reduction r : n.state.getReductions(token)) error.append(" " + r + "\n");
163 //error.append(" ==\n");
165 next.error = error.toString();
168 // this massively improves GC performance
174 // GSS Nodes //////////////////////////////////////////////////////////////////////////////
176 private HashMap<Parser.Table.Reduction,Forest> pcache = new HashMap<Parser.Table.Reduction,Forest>();
177 /** a node in the GSS */
180 private Forest.Ref holder = null;
181 private HashMap<Parser.Table.Reduction,Forest> cache = null;
183 public HashMap<Parser.Table.Reduction,Forest> cache() { return cache==null ? (cache = new HashMap<Parser.Table.Reduction,Forest>()) : cache; }
184 public Forest.Ref holder() { return holder==null ? (holder = new Forest.Ref()) : holder; }
185 public Forest pending() { return Phase.this.closed ? holder().resolve() : holder; }
186 public FastSet<Node> parents() { return parents; }
188 /** which Phase this Node belongs to (node that Node is also a non-static inner class of Phase) */
189 public final Phase phase = Phase.this;
191 /** what state this node is in */
192 public final Parser.Table.State state;
194 /** the set of nodes to which there is an edge starting at this node */
195 public final FastSet<Node> parents = new FastSet<Node>(); /* ALLOC */
198 public void queueReductions() {
199 for(Node n2 : parents)
204 public void queueReductions(Node n2) {
205 new Reduct(this, n2, null);
206 for(Parser.Table.Reduction r : token==null ? state.getEofReductions() : state.getReductions(token)) {
208 // currently we have this weird problem where we
209 // have to do an individual reduct for each child
210 // when the reduction length is one (ie the
211 // children wind up being children of the newly
212 // created node rather than part of the popped
215 if (r.numPop == 1) new Reduct(this, n2, r);
221 public void queueEmptyReductions() {
222 for(Parser.Table.Reduction r : token==null ? state.getEofReductions() : state.getReductions(token)) {
224 new Reduct(this, null, r); /* ALLOC */
228 private Node(Node parent, Forest pending, Parser.Table.State state, Phase start) {
230 if (pending != null) this.holder().merge(pending);
231 if (parent != null) parents.add(parent, true);
232 if (Phase.this.hash.get(code(state, start)) != null) throw new Error("severe problem!");
233 Phase.this.hash.put(code(state, start), this);
234 Phase.this.numNodes++;
235 if (parent==null) holder().valid = true; // hack to make sure that the "base" node is always considered valid
240 // Forest / Completed Reductions //////////////////////////////////////////////////////////////////////////////
242 /** a pending or completed reduction */
245 /** the node from which the reduction should begin */
246 public Node n = null;
248 /** the node on the other end of the edge to be reduced along (either: null, the second node of the reduction,
249 * or the parent of the result of a length-one reduction)
251 public Node n2 = null;
253 /** true iff the reduction has already been performed */
254 private boolean done = false;
256 /** the reduction to be applied */
257 public Parser.Table.Reduction r;
259 public Tree<String> result = null;
261 public Reduct(Node n, Node n2, Parser.Table.Reduction r) {
265 if (reductions.contains(this)) { done = true; return; }
266 reductions.add(this);
267 pendingReduct.add(this);
271 /** perform the reduction */
275 pendingReduct.remove(this);
278 // FIXME: explain this
280 for(Parser.Table.Reduction r : token==null ? n.state.getEofReductions() : n.state.getReductions(token)) {
281 if (r.numPop <= 1) continue;
282 r.reduce(n, n2, Phase.this, null);
284 } else if (r.numPop<=1) {
286 // The problem here is that a "reduction of length 0/1"
287 // performed twice with different values of n2 needs
288 // to only create a *single* new result, but must add
289 // multiple parents to the node holding that result.
290 // The current reducer doesn't differentiate between
291 // the next node of an n-pop reduction and the
292 // ultimate parent of the last pop, so we need to
293 // cache instances here as a way of avoiding
296 Forest ret = (r.numPop==0 ? pcache : n.cache()).get(r);
297 if (ret != null) r.reduce(n, n2, n.phase, ret);
298 else (r.numPop==0 ? pcache : n.cache()).put(r, r.reduce(n, n2, n.phase, null));
301 r.reduce(n, n2, Phase.this, null);
305 // FIXME: this is a PITA
306 public int hashCode() { return n.hashCode() ^ (r==null ? 0 : r.hashCode()) ^ (n2==null ? 0 : n2.hashCode()); }
307 public boolean equals(Object o) {
308 if (o==null) return false;
309 if (o==this) return true;
310 if (!(o instanceof Reduct)) return false;
311 Reduct other = (Reduct)o;
312 return equal(r, other.r) && equal(n, other.n) && equal(n2, other.n2);
319 private static boolean equal(Object a, Object b) {
320 if (a==null && b==null) return true;
321 if (a==null || b==null) return false;
325 /** this is something of a hack right now */
326 private static long code(Parser.Table.State state, Phase start) {
327 return (((long)state.idx) << 32) | (start==null ? 0 : start.pos);