@Override
public void seekExact(long ord) throws IOException {
// System.out.println("BTR.seek by ord ord=" + ord);
if (indexEnum == null) {
throw new IllegalStateException("terms index was not loaded");
}
assert ord < numTerms;
// TODO: if ord is in same terms block and
// after current ord, we should avoid this seek just
// like we do in the seek(BytesRef) case
in.seek(indexEnum.seek(ord));
boolean result = nextBlock();
// Block must exist since ord < numTerms:
assert result;
indexIsCurrent = true;
didIndexNext = false;
seekPending = false;
state.ord = indexEnum.ord() - 1;
assert state.ord >= -1 : "ord=" + state.ord;
term.copyBytes(indexEnum.term());
// Now, scan:
int left = (int) (ord - state.ord);
while (left > 0) {
final BytesRef term = _next();
assert term != null;
left--;
assert indexIsCurrent;
}
}
@Override
public void seekExact(BytesRef target, TermState otherState) {
// System.out.println("BTR.seekExact termState target=" + target.utf8ToString() + " " +
// target + " this=" + this);
assert otherState != null && otherState instanceof BlockTermState;
assert !doOrd || ((BlockTermState) otherState).ord < numTerms;
state.copyFrom(otherState);
seekPending = true;
indexIsCurrent = false;
term.copyBytes(target);
}
@Override
public void seekExact(BytesRef target, TermState otherState) {
// if (DEBUG) {
// System.out.println("BTTR.seekExact termState seg=" + segment + " target=" +
// target.utf8ToString() + " " + target + " state=" + otherState);
// }
assert clearEOF();
if (target.compareTo(term.get()) != 0 || !termExists) {
assert otherState != null && otherState instanceof BlockTermState;
currentFrame = staticFrame;
currentFrame.state.copyFrom(otherState);
term.copyBytes(target);
currentFrame.metaDataUpto = currentFrame.getTermBlockOrd();
assert currentFrame.metaDataUpto > 0;
validIndexPrefix = 0;
} else {
// if (DEBUG) {
// System.out.println(" skip seek: already on target state=" + currentFrame.state);
// }
}
}
@Override
public SeekStatus seekCeil(final BytesRef target) throws IOException {
if (fr.index == null) {
throw new IllegalStateException("terms index was not loaded");
}
term.grow(1 + target.length);
assert clearEOF();
// if (DEBUG) {
// System.out.println("\nBTTR.seekCeil seg=" + segment + " target=" + fieldInfo.name + ":" +
// target.utf8ToString() + " " + target + " current=" + brToString(term) + " (exists?=" +
// termExists + ") validIndexPrefix= " + validIndexPrefix);
// printSeekState();
// }
FST.Arc<Pair<BytesRef, Long>> arc;
int targetUpto;
Pair<BytesRef, Long> output;
targetBeforeCurrentLength = currentFrame.ord;
if (currentFrame != staticFrame) {
// We are already seek'd; find the common
// prefix of new seek term vs current term and
// re-use the corresponding seek state. For
// example, if app first seeks to foobar, then
// seeks to foobaz, we can re-use the seek state
// for the first 5 bytes.
// if (DEBUG) {
// System.out.println(" re-use current seek state validIndexPrefix=" + validIndexPrefix);
// }
arc = arcs[0];
assert arc.isFinal();
output = arc.output;
targetUpto = 0;
IDVersionSegmentTermsEnumFrame lastFrame = stack[0];
assert validIndexPrefix <= term.length();
final int targetLimit = Math.min(target.length, validIndexPrefix);
int cmp = 0;
// TODO: we should write our vLong backwards (MSB
// first) to get better sharing from the FST
// First compare up to valid seek frames:
while (targetUpto < targetLimit) {
cmp = (term.byteAt(targetUpto) & 0xFF) - (target.bytes[target.offset + targetUpto] & 0xFF);
// if (DEBUG) {
// System.out.println(" cycle targetUpto=" + targetUpto + " (vs limit=" + targetLimit +
// ") cmp=" + cmp + " (targetLabel=" + (char) (target.bytes[target.offset + targetUpto]) + "
// vs termLabel=" + (char) (term.bytes[targetUpto]) + ")" + " arc.output=" + arc.output +
// " output=" + output);
// }
if (cmp != 0) {
break;
}
arc = arcs[1 + targetUpto];
assert arc.label == (target.bytes[target.offset + targetUpto] & 0xFF)
: "arc.label="
+ (char) arc.label
+ " targetLabel="
+ (char) (target.bytes[target.offset + targetUpto] & 0xFF);
// TODO: we could save the outputs in local
// byte[][] instead of making new objs ever
// seek; but, often the FST doesn't have any
// shared bytes (but this could change if we
// reverse vLong byte order)
if (arc.output != VersionBlockTreeTermsWriter.NO_OUTPUT) {
output = VersionBlockTreeTermsWriter.FST_OUTPUTS.add(output, arc.output);
}
if (arc.isFinal()) {
lastFrame = stack[1 + lastFrame.ord];
}
targetUpto++;
}
if (cmp == 0) {
final int targetUptoMid = targetUpto;
// Second compare the rest of the term, but
// don't save arc/output/frame:
final int targetLimit2 = Math.min(target.length, term.length());
while (targetUpto < targetLimit2) {
cmp =
(term.byteAt(targetUpto) & 0xFF) - (target.bytes[target.offset + targetUpto] & 0xFF);
// if (DEBUG) {
// System.out.println(" cycle2 targetUpto=" + targetUpto + " (vs limit=" + targetLimit
// + ") cmp=" + cmp + " (targetLabel=" + (char) (target.bytes[target.offset + targetUpto])
// + " vs termLabel=" + (char) (term.bytes[targetUpto]) + ")");
// }
if (cmp != 0) {
break;
}
targetUpto++;
}
if (cmp == 0) {
cmp = term.length() - target.length;
}
targetUpto = targetUptoMid;
}
if (cmp < 0) {
// Common case: target term is after current
// term, ie, app is seeking multiple terms
// in sorted order
// if (DEBUG) {
// System.out.println(" target is after current (shares prefixLen=" + targetUpto + ");
// clear frame.scanned ord=" + lastFrame.ord);
// }
currentFrame = lastFrame;
} else if (cmp > 0) {
// Uncommon case: target term
// is before current term; this means we can
// keep the currentFrame but we must rewind it
// (so we scan from the start)
targetBeforeCurrentLength = 0;
// if (DEBUG) {
// System.out.println(" target is before current (shares prefixLen=" + targetUpto + ");
// rewind frame ord=" + lastFrame.ord);
// }
currentFrame = lastFrame;
currentFrame.rewind();
} else {
// Target is exactly the same as current term
assert term.length() == target.length;
if (termExists) {
// if (DEBUG) {
// System.out.println(" target is same as current; return FOUND");
// }
return SeekStatus.FOUND;
} else {
// if (DEBUG) {
// System.out.println(" target is same as current but term doesn't exist");
// }
}
}
} else {
targetBeforeCurrentLength = -1;
arc = fr.index.getFirstArc(arcs[0]);
// Empty string prefix must have an output (block) in the index!
assert arc.isFinal();
assert arc.output != null;
// if (DEBUG) {
// System.out.println(" no seek state; push root frame");
// }
output = arc.output;
currentFrame = staticFrame;
// term.length = 0;
targetUpto = 0;
currentFrame =
pushFrame(
arc, VersionBlockTreeTermsWriter.FST_OUTPUTS.add(output, arc.nextFinalOutput), 0);
}
// if (DEBUG) {
// System.out.println(" start index loop targetUpto=" + targetUpto + " output=" + output + "
// currentFrame.ord+1=" + currentFrame.ord + " targetBeforeCurrentLength=" +
// targetBeforeCurrentLength);
// }
// We are done sharing the common prefix with the incoming target and where we are currently
// seek'd; now continue walking the index:
while (targetUpto < target.length) {
final int targetLabel = target.bytes[target.offset + targetUpto] & 0xFF;
final FST.Arc<Pair<BytesRef, Long>> nextArc =
fr.index.findTargetArc(targetLabel, arc, getArc(1 + targetUpto), fstReader);
if (nextArc == null) {
// Index is exhausted
// if (DEBUG) {
// System.out.println(" index: index exhausted label=" + ((char) targetLabel) + " " +
// toHex(targetLabel));
// }
validIndexPrefix = currentFrame.prefix;
// validIndexPrefix = targetUpto;
currentFrame.scanToFloorFrame(target);
currentFrame.loadBlock();
final SeekStatus result = currentFrame.scanToTerm(target, false);
if (result == SeekStatus.END) {
term.copyBytes(target);
termExists = false;
if (next() != null) {
// if (DEBUG) {
// System.out.println(" return NOT_FOUND term=" + brToString(term) + " " + term);
// }
return SeekStatus.NOT_FOUND;
} else {
// if (DEBUG) {
// System.out.println(" return END");
// }
return SeekStatus.END;
}
} else {
// if (DEBUG) {
// System.out.println(" return " + result + " term=" + brToString(term) + " " + term);
// }
return result;
}
} else {
// Follow this arc
term.setByteAt(targetUpto, (byte) targetLabel);
arc = nextArc;
// Aggregate output as we go:
assert arc.output != null;
if (arc.output != VersionBlockTreeTermsWriter.NO_OUTPUT) {
output = VersionBlockTreeTermsWriter.FST_OUTPUTS.add(output, arc.output);
}
// if (DEBUG) {
// System.out.println(" index: follow label=" + toHex(target.bytes[target.offset +
// targetUpto]&0xff) + " arc.output=" + arc.output + " arc.nfo=" + arc.nextFinalOutput);
// }
targetUpto++;
if (arc.isFinal()) {
// if (DEBUG) System.out.println(" arc is final!");
currentFrame =
pushFrame(
arc,
VersionBlockTreeTermsWriter.FST_OUTPUTS.add(output, arc.nextFinalOutput),
targetUpto);
// if (DEBUG) System.out.println(" curFrame.ord=" + currentFrame.ord + " hasTerms=" +
// currentFrame.hasTerms);
}
}
}
// validIndexPrefix = targetUpto;
validIndexPrefix = currentFrame.prefix;
currentFrame.scanToFloorFrame(target);
currentFrame.loadBlock();
final SeekStatus result = currentFrame.scanToTerm(target, false);
if (result == SeekStatus.END) {
term.copyBytes(target);
termExists = false;
if (next() != null) {
// if (DEBUG) {
// System.out.println(" return NOT_FOUND term=" + term.utf8ToString() + " " + term);
// }
return SeekStatus.NOT_FOUND;
} else {
// if (DEBUG) {
// System.out.println(" return END");
// }
return SeekStatus.END;
}
} else {
return result;
}
}
// TODO: we may want an alternate mode here which is
// "if you are about to return NOT_FOUND I won't use
// the terms data from that"; eg FuzzyTermsEnum will
// (usually) just immediately call seek again if we
// return NOT_FOUND so it's a waste for us to fill in
// the term that was actually NOT_FOUND
@Override
public SeekStatus seekCeil(final BytesRef target) throws IOException {
if (indexEnum == null) {
throw new IllegalStateException("terms index was not loaded");
}
// System.out.println("BTR.seek seg=" + segment + " target=" + fieldInfo.name + ":" +
// target.utf8ToString() + " " + target + " current=" + term().utf8ToString() + " " + term()
// + " indexIsCurrent=" + indexIsCurrent + " didIndexNext=" + didIndexNext + " seekPending="
// + seekPending + " divisor=" + indexReader.getDivisor() + " this=" + this);
if (didIndexNext) {
if (nextIndexTerm == null) {
// System.out.println(" nextIndexTerm=null");
} else {
// System.out.println(" nextIndexTerm=" + nextIndexTerm.utf8ToString());
}
}
boolean doSeek = true;
// See if we can avoid seeking, because target term
// is after current term but before next index term:
if (indexIsCurrent) {
final int cmp = BytesRef.getUTF8SortedAsUnicodeComparator().compare(term.get(), target);
if (cmp == 0) {
// Already at the requested term
return SeekStatus.FOUND;
} else if (cmp < 0) {
// Target term is after current term
if (!didIndexNext) {
if (indexEnum.next() == -1) {
nextIndexTerm = null;
} else {
nextIndexTerm = indexEnum.term();
}
// System.out.println(" now do index next() nextIndexTerm=" + (nextIndexTerm == null
// ? "null" : nextIndexTerm.utf8ToString()));
didIndexNext = true;
}
if (nextIndexTerm == null
|| BytesRef.getUTF8SortedAsUnicodeComparator().compare(target, nextIndexTerm) < 0) {
// Optimization: requested term is within the
// same term block we are now in; skip seeking
// (but do scanning):
doSeek = false;
// System.out.println(" skip seek: nextIndexTerm=" + (nextIndexTerm == null ? "null"
// : nextIndexTerm.utf8ToString()));
}
}
}
if (doSeek) {
// System.out.println(" seek");
// Ask terms index to find biggest indexed term (=
// first term in a block) that's <= our text:
in.seek(indexEnum.seek(target));
boolean result = nextBlock();
// Block must exist since, at least, the indexed term
// is in the block:
assert result;
indexIsCurrent = true;
didIndexNext = false;
if (doOrd) {
state.ord = indexEnum.ord() - 1;
}
term.copyBytes(indexEnum.term());
// System.out.println(" seek: term=" + term.utf8ToString());
} else {
// System.out.println(" skip seek");
if (state.termBlockOrd == blockTermCount && !nextBlock()) {
indexIsCurrent = false;
return SeekStatus.END;
}
}
seekPending = false;
int common = 0;
// Scan within block. We could do this by calling
// _next() and testing the resulting term, but this
// is wasteful. Instead, we first confirm the
// target matches the common prefix of this block,
// and then we scan the term bytes directly from the
// termSuffixesreader's byte[], saving a copy into
// the BytesRef term per term. Only when we return
// do we then copy the bytes into the term.
while (true) {
// First, see if target term matches common prefix
// in this block:
if (common < termBlockPrefix) {
final int cmp =
(term.byteAt(common) & 0xFF) - (target.bytes[target.offset + common] & 0xFF);
if (cmp < 0) {
// TODO: maybe we should store common prefix
// in block header? (instead of relying on
// last term of previous block)
// Target's prefix is after the common block
// prefix, so term cannot be in this block
// but it could be in next block. We
// must scan to end-of-block to set common
// prefix for next block:
if (state.termBlockOrd < blockTermCount) {
while (state.termBlockOrd < blockTermCount - 1) {
state.termBlockOrd++;
state.ord++;
termSuffixesReader.skipBytes(termSuffixesReader.readVInt());
}
final int suffix = termSuffixesReader.readVInt();
term.setLength(termBlockPrefix + suffix);
term.grow(term.length());
termSuffixesReader.readBytes(term.bytes(), termBlockPrefix, suffix);
}
state.ord++;
if (!nextBlock()) {
indexIsCurrent = false;
return SeekStatus.END;
}
common = 0;
} else if (cmp > 0) {
// Target's prefix is before the common prefix
// of this block, so we position to start of
// block and return NOT_FOUND:
assert state.termBlockOrd == 0;
final int suffix = termSuffixesReader.readVInt();
term.setLength(termBlockPrefix + suffix);
term.grow(term.length());
termSuffixesReader.readBytes(term.bytes(), termBlockPrefix, suffix);
return SeekStatus.NOT_FOUND;
} else {
common++;
}
continue;
}
// Test every term in this block
while (true) {
state.termBlockOrd++;
state.ord++;
final int suffix = termSuffixesReader.readVInt();
// We know the prefix matches, so just compare the new suffix:
final int termLen = termBlockPrefix + suffix;
int bytePos = termSuffixesReader.getPosition();
boolean next = false;
final int limit = target.offset + (termLen < target.length ? termLen : target.length);
int targetPos = target.offset + termBlockPrefix;
while (targetPos < limit) {
final int cmp = (termSuffixes[bytePos++] & 0xFF) - (target.bytes[targetPos++] & 0xFF);
if (cmp < 0) {
// Current term is still before the target;
// keep scanning
next = true;
break;
} else if (cmp > 0) {
// Done! Current term is after target. Stop
// here, fill in real term, return NOT_FOUND.
term.setLength(termBlockPrefix + suffix);
term.grow(term.length());
termSuffixesReader.readBytes(term.bytes(), termBlockPrefix, suffix);
// System.out.println(" NOT_FOUND");
return SeekStatus.NOT_FOUND;
}
}
if (!next && target.length <= termLen) {
term.setLength(termBlockPrefix + suffix);
term.grow(term.length());
termSuffixesReader.readBytes(term.bytes(), termBlockPrefix, suffix);
if (target.length == termLen) {
// Done! Exact match. Stop here, fill in
// real term, return FOUND.
// System.out.println(" FOUND");
return SeekStatus.FOUND;
} else {
// System.out.println(" NOT_FOUND");
return SeekStatus.NOT_FOUND;
}
}
if (state.termBlockOrd == blockTermCount) {
// Must pre-fill term for next block's common prefix
term.setLength(termBlockPrefix + suffix);
term.grow(term.length());
termSuffixesReader.readBytes(term.bytes(), termBlockPrefix, suffix);
break;
} else {
termSuffixesReader.skipBytes(suffix);
}
}
// The purpose of the terms dict index is to seek
// the enum to the closest index term before the
// term we are looking for. So, we should never
// cross another index term (besides the first
// one) while we are scanning:
assert indexIsCurrent;
if (!nextBlock()) {
// System.out.println(" END");
indexIsCurrent = false;
return SeekStatus.END;
}
common = 0;
}
}