+// Copyright 2006 all rights reserved; see LICENSE file for BSD-style license
+
package edu.berkeley.sbp;
-import edu.berkeley.sbp.*;
import edu.berkeley.sbp.util.*;
import edu.berkeley.sbp.Sequence.Position;
import java.io.*;
import java.util.*;
-/** a parser which translates streams of Tokens of type T into a Forest<R> */
-public abstract class Parser<Tok, Result> {
+// FEATURE: try harder to "fuse" states together along two dimensions:
+// - identical (equivalent) states, or states that subsume each other
+// - unnecessary intermediate states ("short cut" GLR)
+
+/** a parser which translates an Input<Token> into a Forest<NodeType> */
+public abstract class Parser<Token, NodeType> {
- protected final Table<Tok> pt;
+ final Table pt;
/** create a parser to parse the grammar with start symbol <tt>u</tt> */
- protected Parser(Union u, Topology<Tok> top) { this.pt = new Table<Tok>(u, top); }
- protected Parser(Table<Tok> pt) { this.pt = pt; }
+ public Parser(Union u) { this.pt = new Table(u); }
/** implement this method to create the output forest corresponding to a lone shifted input token */
- public abstract Forest<Result> shiftToken(Tok t, Input.Location loc);
-
- /** parse <tt>input</tt>, using the table <tt>pt</tt> to drive the parser */
- public Forest<Result> parse(Input<Tok> input) throws IOException, ParseFailed {
- GSS gss = new GSS();
- Input.Location loc = input.getLocation();
- GSS.Phase current = gss.new Phase<Tok>(null, this, null, input.next(1, 0, 0), loc, null);
- current.newNode(null, Forest.leaf(null, null, null), pt.start, true);
- int count = 1;
- for(;;) {
- loc = input.getLocation();
- current.reduce();
- Forest forest = current.token==null ? null : shiftToken((Tok)current.token, loc);
- GSS.Phase next = gss.new Phase<Tok>(current, this, current, input.next(count, gss.resets, gss.waits), loc, forest);
- count = next.size();
- if (current.isDone()) return (Forest<Result>)gss.finalResult;
- current = next;
+ public abstract Forest<NodeType> shiftToken(Token t, Input.Region region);
+
+ public abstract Topology<Token> emptyTopology();
+
+ public String toString() { return pt.toString(); }
+ Grammar cache() { return pt; }
+
+ /** parse <tt>input</tt>, and return the shared packed parse forest (or throw an exception) */
+ public Forest<NodeType> parse(Input<Token> input) throws IOException, ParseFailed {
+ verbose = System.getProperty("sbp.verbose", null) != null;
+ spinpos = 0;
+ GSS gss = new GSS(input, this);
+ try {
+ for(GSS.Phase current = gss.new Phase<Token>(pt.start); ;) {
+ if (verbose) debug(current.token, gss, input);
+ if (current.isDone()) return (Forest<NodeType>)current.finalResult;
+ Input.Region region = current.getLocation().createRegion(current.getNextLocation());
+ Forest forest = shiftToken((Token)current.token, region);
+ current = gss.new Phase<Token>(current, forest);
+ }
+ } finally {
+ if (verbose) {
+ System.err.print("\r"+ANSI.clreol());
+ debug(null, gss, input);
+ }
}
}
+ // Spinner //////////////////////////////////////////////////////////////////////////////
+
+ private boolean verbose = false;
+ private static final char[] spin = new char[] { '-', '\\', '|', '/' };
+ private int spinpos = 0;
+ private long last = 0;
+ void spin() {
+ if (!verbose) return;
+ long now = System.currentTimeMillis();
+ if (now-last < 70) return;
+ last = now;
+ System.err.print("\r " + spin[spinpos++ % (spin.length)]+"\r");
+ }
+
+ private int _last = -1;
+ private String buf = "";
+ private void debug(Object t, GSS gss, Input input) {
+ //FIXME
+ int c = t==null ? -1 : ((t+"").charAt(0));
+ int last = _last;
+ _last = c;
+ switch(c) {
+ case edu.berkeley.sbp.chr.CharAtom.left:
+ buf += "\033[31m{\033[0m";
+ break;
+ case edu.berkeley.sbp.chr.CharAtom.right:
+ buf += "\033[31m}\033[0m";
+ break;
+ case -1: // FIXME
+ case '\n':
+ if (verbose) {
+ if (last==' ') buf += ANSI.blue("\\n");
+ System.err.println("\r"+ANSI.clreol()+"\r"+buf);
+ buf = "";
+ }
+ break;
+ default:
+ buf += ANSI.cyan(""+((char)c));
+ break;
+ }
+ if (t==null) return;
+
+ // FIXME: clean this up
+ String s;
+ s = " " + spin[spinpos++ % (spin.length)]+" parsing ";
+ s += input.getName();
+ s += " "+input.getLocation();
+ while(s.indexOf(':') != -1 && s.indexOf(':') < 8) s = " " + s;
+ String y = "@"+gss.viewPos+" ";
+ while(y.length() < 9) y = " " + y;
+ s += y;
+ s += " nodes="+gss.numOldNodes;
+ while(s.length() < 50) s = s + " ";
+ s += " shifted="+gss.numNewNodes;
+ while(s.length() < 60) s = s + " ";
+ s += " reductions="+gss.numReductions;
+ while(s.length() < 78) s = s + " ";
+ System.err.print("\r"+ANSI.invert(s+ANSI.clreol())+"\r");
+ }
+
// Table //////////////////////////////////////////////////////////////////////////////
/** an SLR(1) parse table which may contain conflicts */
- public static class Table<Tok> extends Walk.Cache {
-
- public final Walk.Cache cache = this;
-
- private void walk(Element e, HashSet<Element> hs) {
- if (e==null) return;
- if (hs.contains(e)) return;
- hs.add(e);
- if (e instanceof Atom) return;
- for(Sequence s : (Union)e) {
- hs.add(s);
- for(Position p = s.firstp(); p != null; p = p.next())
- walk(p.element(), hs);
- }
- }
+ class Table extends Grammar<Token> {
/** the start state */
- public final State<Tok> start;
+ final State<Token> start;
+
+ /** a dummy state from which no reductions can be performed */
+ private final State<Token> dead_state;
/** used to generate unique values for State.idx */
private int master_state_idx = 0;
- HashMap<HashSet<Position>,State<Tok>> all_states = new HashMap<HashSet<Position>,State<Tok>>();
+ /** all the states for this table */
+ HashSet<State<Token>> all_states = new HashSet<State<Token>>();
+
+ /** all the doomed states in this table */
+ HashMap<HashSet<Position>,State<Token>> doomed_states = new HashMap<HashSet<Position>,State<Token>>();
+
+ /** all the non-doomed states in this table */
+ HashMap<HashSet<Position>,State<Token>> normal_states = new HashMap<HashSet<Position>,State<Token>>();
+
+ Topology<Token> emptyTopology() { return Parser.this.emptyTopology(); }
+
/** construct a parse table for the given grammar */
- public Table(Topology top) { this("s", top); }
- public Table(String startSymbol, Topology top) { this(new Union(startSymbol), top); }
- public Table(Union ux, Topology top) {
- Union start0 = new Union("0");
- start0.add(new Sequence.Singleton(ux));
-
- for(Sequence s : start0) cache.eof.put(s, true);
- cache.eof.put(start0, true);
-
- // construct the set of states
- HashSet<Element> all_elements = new HashSet<Element>();
- walk(start0, all_elements);
- for(Element e : all_elements)
- cache.ys.addAll(e, new Walk.YieldSet(e, cache).walk());
- HashSet<Position> hp = new HashSet<Position>();
- reachable(start0, hp);
- this.start = new State<Tok>(hp, all_states, all_elements);
+ Table(Union ux) {
+ super(new Union("0", Sequence.create(ux), true));
+ // create the "dead state"
+ this.dead_state = new State<Token>(new HashSet<Position>(), true);
+
+ // construct the start state; this will recursively create *all* the states
+ this.start = new State<Token>(reachable(rootUnion), false);
+
+ buildReductions();
+ sortReductions();
+ }
+
+ /** fill in the reductions table */
+ private void buildReductions() {
// for each state, fill in the corresponding "row" of the parse table
- for(State<Tok> state : all_states.values())
+ for(State<Token> state : all_states)
for(Position p : state.hs) {
- // the Grammar's designated "last position" is the only accepting state
- if (start0.contains(p.owner()) && p.next()==null)
- state.accept = true;
-
- if (isRightNullable(p)) {
- Walk.Follow wf = new Walk.Follow(top.empty(), p.owner(), all_elements, cache);
- Topology follow = wf.walk(p.owner());
- for(Position p2 = p; p2 != null && p2.element() != null; p2 = p2.next())
- follow = follow.intersect(new Walk.Follow(top.empty(), p2.element(), all_elements, cache).walk(p2.element()));
- state.reductions.put(follow, p);
- if (wf.includesEof()) state.eofReductions.add(p);
- }
-
// if the element following this position is an atom, copy the corresponding
// set of rows out of the "master" goto table and into this state's shift table
if (p.element() != null && p.element() instanceof Atom)
- state.shifts.addAll(state.gotoSetTerminals.subset(((Atom)p.element())));
+ state.shifts.addAll(state.gotoSetTerminals.subset(((Atom)p.element()).getTokenTopology()));
+
+ // RNGLR: we can potentially reduce from any "right-nullable" position -- that is,
+ // any position for which all Elements after it in the Sequence are capable of
+ // matching the empty string.
+ if (!isRightNullable(p)) continue;
+ Topology<Token> follow = follow(p.owner());
+ for(Position p2 = p; p2 != null && p2.element() != null; p2 = p2.next()) {
+ if (!(p2.element() instanceof Union))
+ throw new Error("impossible -- only Unions can be nullable");
+
+ // interesting RNGLR-followRestriction interaction: we must intersect
+ // not just the follow-set of the last non-nullable element, but the
+ // follow-sets of the nulled elements as well.
+ for(Sequence s : ((Union)p2.element()))
+ follow = follow.intersect(follow(s));
+ Topology<Token> set = epsilonFollowSet((Union)p2.element());
+ if (set != null) follow = follow.intersect(set);
+ }
+
+ // indicate that when the next token is in the set "follow", nodes in this
+ // state should reduce according to Position "p"
+ state.reductions.put(follow, p);
+ if (followEof.contains(p.owner())) state.eofReductions.add(p);
}
- if (top instanceof IntegerTopology)
- for(State<Tok> state : all_states.values()) {
- state.oreductions = state.reductions.optimize(((IntegerTopology)top).functor());
- state.oshifts = state.shifts.optimize(((IntegerTopology)top).functor());
+
+ // optimize the reductions table
+ if (emptyTopology() instanceof IntegerTopology)
+ for(State<Token> state : all_states) {
+ // FIXME: this is pretty ugly
+ state.oreductions = state.reductions.optimize(((IntegerTopology)emptyTopology()).functor());
+ state.oshifts = state.shifts.optimize(((IntegerTopology)emptyTopology()).functor());
}
}
- private boolean isRightNullable(Position p) {
- if (p.isLast()) return true;
- if (!possiblyEpsilon(p.element())) return false;
- return isRightNullable(p.next());
- }
+ // FIXME: this method needs to be cleaned up and documented
+ private void sortReductions() {
+ // crude algorithm to assing an ordinal ordering to every position
+ // al will be sorted in DECREASING order (al[0] >= al[1])
+ ArrayList<Sequence.Position> al = new ArrayList<Sequence.Position>();
+ for(State s : all_states) {
+ for(Object po : s) {
+ Sequence.Position p = (Sequence.Position)po;
+ if (al.contains(p)) continue;
+ int i=0;
+ for(; i<al.size(); i++) {
+ if (comparePositions(p, al.get(i)) < 0)
+ break;
+ }
+ al.add(i, p);
+ }
+ }
+ // FIXME: this actually pollutes the "pure" objects (the ones that should not be modified by the Parser)
+ // sort in increasing order...
+ OUTER: while(true) {
+ for(int i=0; i<al.size(); i++)
+ for(int j=i+1; j<al.size(); j++)
+ if (comparePositions(al.get(i), al.get(j)) > 0) {
+ Sequence.Position p = al.remove(j);
+ al.add(i, p);
+ continue OUTER;
+ }
+ break;
+ }
- /** a single state in the LR table and the transitions possible from it */
+ int j = 1;
+ int pk = 0;
+ for(int i=0; i<al.size(); i++) {
+ boolean inc = false;
+ for(int k=pk; k<i; k++) {
+ if (comparePositions(al.get(k), al.get(i)) > 0)
+ { inc = true; break; }
+ }
+ inc = true;
+ if (inc) {
+ j++;
+ pk = i;
+ }
+ al.get(i).ord = j;
+ }
+ }
- public class State<Tok> implements Comparable<State<Tok>>, IntegerMappable, Iterable<Position> {
+ /**
+ * A single state in the LR table and the transitions
+ * possible from it
+ *
+ * A state corresponds to a set of Sequence.Position's. Each
+ * Node in the GSS has a State; the Node represents a set of
+ * possible parses, one for each Position in the State.
+ *
+ * Every state is either "doomed" or "normal". If a Position
+ * is part of a Sequence which is a conjunct (that is, it was
+ * passed to Sequence.{and(),andnot()}), then that Position
+ * will appear only in doomed States. Furthermore, any set
+ * of Positions reachable from a doomed State also forms a
+ * doomed State. Note that in this latter case, a doomed
+ * state might have exactly the same set of Positions as a
+ * non-doomed state.
+ *
+ * Nodes with non-doomed states represent nodes which
+ * contribute to actual valid parses. Nodes with doomed
+ * States exist for no other purpose than to enable/disable
+ * some future reduction from a non-doomed Node. Because of
+ * this, we "garbage-collect" Nodes with doomed states if
+ * there are no more non-doomed Nodes which they could
+ * affect (see Result, Reduction, and Node for details).
+ *
+ * Without this optimization, many seemingly-innocuous uses
+ * of positive and negative conjuncts can trigger O(n^2)
+ * space+time complexity in otherwise simple grammars. There
+ * is an example of this in the regression suite.
+ */
+ class State<Token> implements IntegerMappable, Iterable<Position> {
public final int idx = master_state_idx++;
private final HashSet<Position> hs;
+ public HashSet<State<Token>> conjunctStates = new HashSet<State<Token>>();
- public transient HashMap<Element,State<Tok>> gotoSetNonTerminals = new HashMap<Element,State<Tok>>();
- private transient TopologicalBag<Tok,State<Tok>> gotoSetTerminals = new TopologicalBag<Tok,State<Tok>>();
+ HashMap<Sequence,State<Token>> gotoSetNonTerminals = new HashMap<Sequence,State<Token>>();
+ private transient TopologicalBag<Token,State<Token>> gotoSetTerminals = new TopologicalBag<Token,State<Token>>();
- private TopologicalBag<Tok,Position> reductions = new TopologicalBag<Tok,Position>();
- private HashSet<Position> eofReductions = new HashSet<Position>();
- private TopologicalBag<Tok,State<Tok>> shifts = new TopologicalBag<Tok,State<Tok>>();
- private boolean accept = false;
+ private TopologicalBag<Token,Position> reductions = new TopologicalBag<Token,Position>();
+ private HashSet<Position> eofReductions = new HashSet<Position>();
+ private TopologicalBag<Token,State<Token>> shifts = new TopologicalBag<Token,State<Token>>();
+ private boolean accept = false;
- private VisitableMap<Tok,State<Tok>> oshifts = null;
- private VisitableMap<Tok,Position> oreductions = null;
+ private VisitableMap<Token,State<Token>> oshifts = null;
+ private VisitableMap<Token,Position> oreductions = null;
+ public final boolean doomed;
// Interface Methods //////////////////////////////////////////////////////////////////////////////
- boolean isAccepting() { return accept; }
- public Iterator<Position> iterator() { return hs.iterator(); }
-
- boolean canShift(Tok t) { return oshifts.contains(t); }
- <B,C> void invokeShifts(Tok t, Invokable<State<Tok>,B,C> irbc, B b, C c) {
- oshifts.invoke(t, irbc, b, c);
+ public boolean doomed() { return doomed; }
+ boolean isAccepting() { return accept; }
+ public Iterator<Position> iterator() { return hs.iterator(); }
+ boolean canShift(Token t) { return oshifts!=null && oshifts.contains(t); }
+ void invokeShifts(Token t, GSS.Phase phase, Result r) { oshifts.invoke(t, phase, r); }
+ boolean canReduce(Token t) {
+ return oreductions != null && (t==null ? eofReductions.size()>0 : oreductions.contains(t)); }
+ void invokeEpsilonReductions(Token t, Node node) {
+ if (t==null) for(Position r : eofReductions) node.invoke(r, null);
+ else oreductions.invoke(t, node, null);
}
-
- boolean canReduce(Tok t) { return t==null ? eofReductions.size()>0 : oreductions.contains(t); }
- <B,C> void invokeReductions(Tok t, Invokable<Position,B,C> irbc, B b, C c) {
- if (t==null) for(Position r : eofReductions) irbc.invoke(r, b, c);
- else oreductions.invoke(t, irbc, b, c);
+ void invokeReductions(Token t, Node node, Result b) {
+ if (t==null) for(Position r : eofReductions) node.invoke(r, b);
+ else oreductions.invoke(t, node, b);
}
// Constructor //////////////////////////////////////////////////////////////////////////////
/**
* create a new state consisting of all the <tt>Position</tt>s in <tt>hs</tt>
* @param hs the set of <tt>Position</tt>s comprising this <tt>State</tt>
- * @param all_states the set of states already constructed (to avoid recreating states)
- * @param all_elements the set of all elements (Atom instances need not be included)
+ * @param all the set of all elements (Atom instances need not be included)
*
* In principle these two steps could be merged, but they
* are written separately to highlight these two facts:
* for non-Atom Elements.
* </ul>
*/
- public State(HashSet<Position> hs,
- HashMap<HashSet<Position>,State<Tok>> all_states,
- HashSet<Element> all_elements) {
+ public State(HashSet<Position> hs, boolean doomed) {
this.hs = hs;
+ this.doomed = doomed;
- // register ourselves in the all_states hash so that no
- // two states are ever created with an identical position set
- all_states.put(hs, this);
+ // register ourselves so that no two states are ever
+ // created with an identical position set (termination depends on this)
+ ((HashMap)(doomed ? doomed_states : normal_states)).put(hs, this);
+ ((HashSet)all_states).add(this);
+
+ for(Position p : hs) {
+ // Step 1a: take note if we are an accepting state
+ // (last position of the root Union's sequence)
+ if (p.next()==null && !doomed && rootUnion.contains(p.owner()))
+ accept = true;
+
+ // Step 1b: If any Position in the set is the first position of its sequence, then this
+ // state is responsible for spawning the "doomed" states for each of the
+ // Sequence's conjuncts. This obligation is recorded by adding the to-be-spawned
+ // states to conjunctStates.
+ if (!p.isFirst()) continue;
+ for(Sequence s : p.owner().needs())
+ if (!hs.contains(s.firstp()))
+ conjunctStates.add(mkstate(reachable(s.firstp()), true));
+ for(Sequence s : p.owner().hates())
+ if (!hs.contains(s.firstp()))
+ conjunctStates.add(mkstate(reachable(s.firstp()), true));
+ }
- // Step 1a: examine all Position's in this state and compute the mappings from
+ // Step 2a: examine all Position's in this state and compute the mappings from
// sets of follow tokens (tokens which could follow this position) to sets
// of _new_ positions (positions after shifting). These mappings are
// collectively known as the _closure_
- TopologicalBag<Tok,Position> bag0 = new TopologicalBag<Tok,Position>();
+ TopologicalBag<Token,Position> bag0 = new TopologicalBag<Token,Position>();
for(Position position : hs) {
if (position.isLast() || !(position.element() instanceof Atom)) continue;
Atom a = (Atom)position.element();
HashSet<Position> hp = new HashSet<Position>();
reachable(position.next(), hp);
- bag0.addAll(a, hp);
+ bag0.addAll(a.getTokenTopology(), hp);
}
- // Step 1b: for each _minimal, contiguous_ set of characters having an identical next-position
+ // Step 2b: for each _minimal, contiguous_ set of characters having an identical next-position
// set, add that character set to the goto table (with the State corresponding to the
// computed next-position set).
- for(Topology<Tok> r : bag0) {
+ for(Topology<Token> r : bag0) {
HashSet<Position> h = new HashSet<Position>();
for(Position p : bag0.getAll(r)) h.add(p);
- gotoSetTerminals.put(r, all_states.get(h) == null ? new State<Tok>(h, all_states, all_elements) : all_states.get(h));
+ ((TopologicalBag)gotoSetTerminals).put(r, mkstate(h, doomed));
}
- // Step 2: for every non-Atom element (ie every Element which has a corresponding reduction),
- // compute the closure over every position in this set which is followed by a symbol
- // which could yield the Element in question.
+ // Step 3: for every Sequence, compute the closure over every position in this set which
+ // is followed by a symbol which could yield the Sequence.
//
// "yields" [in one or more step] is used instead of "produces" [in exactly one step]
// to avoid having to iteratively construct our set of States as shown in most
// expositions of the algorithm (ie "keep doing XYZ until things stop changing").
- HashMapBag<Element,Position> move = new HashMapBag<Element,Position>();
- for(Position p : hs) {
- Element e = p.element();
- if (e==null) continue;
- for(Element y : cache.ys.getAll(e)) {
- HashSet<Position> hp = new HashSet<Position>();
- reachable(p.next(), hp);
- move.addAll(y, hp);
- }
- }
- for(Element y : move) {
+
+ HashMapBag<Sequence,Position> move = new HashMapBag<Sequence,Position>();
+ for(Position p : hs)
+ if (!p.isLast() && p.element() instanceof Union)
+ for(Sequence s : ((Union)p.element())) {
+ HashSet<Position> hp = new HashSet<Position>();
+ reachable(p.next(), hp);
+ move.addAll(s, hp);
+ }
+ OUTER: for(Sequence y : move) {
+ // if a reduction is "lame", it should wind up in the dead_state after reducing
HashSet<Position> h = move.getAll(y);
- State<Tok> s = all_states.get(h) == null ? new State<Tok>(h, all_states, all_elements) : all_states.get(h);
+ State<Token> s = mkstate(h, doomed);
+ for(Position p : hs)
+ if (p.element() != null && (p.element() instanceof Union))
+ for(Sequence seq : ((Union)p.element()))
+ if (seq.needs.contains(y) || seq.hates.contains(y)) {
+ // FIXME: assumption that no sequence is ever both usefully (non-lamely) matched
+ // and also directly lamely matched
+ ((HashMap)gotoSetNonTerminals).put(y, dead_state);
+ continue OUTER;
+ }
gotoSetNonTerminals.put(y, s);
}
}
+ private State<Token> mkstate(HashSet<Position> h, boolean b) {
+ State ret = (b?doomed_states:normal_states).get(h);
+ if (ret==null) ret = new State<Token>(h,b);
+ return ret;
+ }
+
+ public int toInt() { return idx; }
public String toString() {
StringBuffer ret = new StringBuffer();
- ret.append("state["+idx+"]: ");
- for(Position p : this) ret.append("{"+p+"} ");
+ for(Position p : hs)
+ ret.append(p+"\n");
return ret.toString();
}
-
- public int compareTo(State<Tok> s) { return idx==s.idx ? 0 : idx < s.idx ? -1 : 1; }
- public int toInt() { return idx; }
}
+
}
// Helpers //////////////////////////////////////////////////////////////////////////////
+ private static HashSet<Position> reachable(Element e) {
+ HashSet<Position> h = new HashSet<Position>();
+ reachable(e, h);
+ return h;
+ }
private static void reachable(Element e, HashSet<Position> h) {
if (e instanceof Atom) return;
for(Sequence s : ((Union)e))
h.add(p);
if (p.element() != null) reachable(p.element(), h);
}
+ private static HashSet<Position> reachable(Position p) {
+ HashSet<Position> ret = new HashSet<Position>();
+ reachable(p, ret);
+ return ret;
+ }
}
projects / sbp.git / blobdiff