Change prefix search from O(N*M) to O(NlogM).

Backports (4599) from fts3. (CVS 5455)

FossilOrigin-Name: 3f614453d2d7c753a5963b027fe8618b50b4f6b9
This commit is contained in:
shess
2008-07-22 23:08:40 +00:00
parent b2822a2b5e
commit deca811cb5
3 changed files with 128 additions and 25 deletions
+121 -18
View File
@@ -455,6 +455,7 @@ static int getVarint32(const char *p, int *pi){
** dataBufferInit - create a buffer with given initial capacity.
** dataBufferReset - forget buffer's data, retaining capacity.
** dataBufferDestroy - free buffer's data.
** dataBufferSwap - swap contents of two buffers.
** dataBufferExpand - expand capacity without adding data.
** dataBufferAppend - append data.
** dataBufferAppend2 - append two pieces of data at once.
@@ -479,6 +480,11 @@ static void dataBufferDestroy(DataBuffer *pBuffer){
if( pBuffer->pData!=NULL ) sqlite3_free(pBuffer->pData);
SCRAMBLE(pBuffer);
}
static void dataBufferSwap(DataBuffer *pBuffer1, DataBuffer *pBuffer2){
DataBuffer tmp = *pBuffer1;
*pBuffer1 = *pBuffer2;
*pBuffer2 = tmp;
}
static void dataBufferExpand(DataBuffer *pBuffer, int nAddCapacity){
assert( nAddCapacity>0 );
/* TODO(shess) Consider expanding more aggressively. Note that the
@@ -5219,6 +5225,26 @@ static int segmentMerge(fulltext_vtab *v, int iLevel){
return rc;
}
/* Accumulate the union of *acc and *pData into *acc. */
static void docListAccumulateUnion(DataBuffer *acc,
const char *pData, int nData) {
DataBuffer tmp = *acc;
dataBufferInit(acc, tmp.nData+nData);
docListUnion(tmp.pData, tmp.nData, pData, nData, acc);
dataBufferDestroy(&tmp);
}
/* TODO(shess) It might be interesting to explore different merge
** strategies, here. For instance, since this is a sorted merge, we
** could easily merge many doclists in parallel. With some
** comprehension of the storage format, we could merge all of the
** doclists within a leaf node directly from the leaf node's storage.
** It may be worthwhile to merge smaller doclists before larger
** doclists, since they can be traversed more quickly - but the
** results may have less overlap, making them more expensive in a
** different way.
*/
/* Scan pReader for pTerm/nTerm, and merge the term's doclist over
** *out (any doclists with duplicate docids overwrite those in *out).
** Internal function for loadSegmentLeaf().
@@ -5226,39 +5252,116 @@ static int segmentMerge(fulltext_vtab *v, int iLevel){
static int loadSegmentLeavesInt(fulltext_vtab *v, LeavesReader *pReader,
const char *pTerm, int nTerm, int isPrefix,
DataBuffer *out){
/* doclist data is accumulated into pBuffers similar to how one does
** increment in binary arithmetic. If index 0 is empty, the data is
** stored there. If there is data there, it is merged and the
** results carried into position 1, with further merge-and-carry
** until an empty position is found.
*/
DataBuffer *pBuffers = NULL;
int nBuffers = 0, nMaxBuffers = 0, rc;
assert( nTerm>0 );
/* Process while the prefix matches. */
while( !leavesReaderAtEnd(pReader) ){
for(rc=SQLITE_OK; rc==SQLITE_OK && !leavesReaderAtEnd(pReader);
rc=leavesReaderStep(v, pReader)){
/* TODO(shess) Really want leavesReaderTermCmp(), but that name is
** already taken to compare the terms of two LeavesReaders. Think
** on a better name. [Meanwhile, break encapsulation rather than
** use a confusing name.]
*/
int rc;
int c = leafReaderTermCmp(&pReader->leafReader, pTerm, nTerm, isPrefix);
if( c>0 ) break; /* Past any possible matches. */
if( c==0 ){
const char *pData = leavesReaderData(pReader);
int nData = leavesReaderDataBytes(pReader);
if( out->nData==0 ){
dataBufferReplace(out, pData, nData);
int iBuffer, nData = leavesReaderDataBytes(pReader);
/* Find the first empty buffer. */
for(iBuffer=0; iBuffer<nBuffers; ++iBuffer){
if( 0==pBuffers[iBuffer].nData ) break;
}
/* Out of buffers, add an empty one. */
if( iBuffer==nBuffers ){
if( nBuffers==nMaxBuffers ){
DataBuffer *p;
nMaxBuffers += 20;
/* Manual realloc so we can handle NULL appropriately. */
p = sqlite3_malloc(nMaxBuffers*sizeof(*pBuffers));
if( p==NULL ){
rc = SQLITE_NOMEM;
break;
}
if( nBuffers>0 ){
assert(pBuffers!=NULL);
memcpy(p, pBuffers, nBuffers*sizeof(*pBuffers));
sqlite3_free(pBuffers);
}
pBuffers = p;
}
dataBufferInit(&(pBuffers[nBuffers]), 0);
nBuffers++;
}
/* At this point, must have an empty at iBuffer. */
assert(iBuffer<nBuffers && pBuffers[iBuffer].nData==0);
/* If empty was first buffer, no need for merge logic. */
if( iBuffer==0 ){
dataBufferReplace(&(pBuffers[0]), pData, nData);
}else{
DataBuffer result;
dataBufferInit(&result, out->nData+nData);
docListUnion(out->pData, out->nData, pData, nData, &result);
dataBufferDestroy(out);
*out = result;
/* TODO(shess) Rather than destroy out, we could retain it for
** later reuse.
/* pAcc is the empty buffer the merged data will end up in. */
DataBuffer *pAcc = &(pBuffers[iBuffer]);
DataBuffer *p = &(pBuffers[0]);
/* Handle position 0 specially to avoid need to prime pAcc
** with pData/nData.
*/
dataBufferSwap(p, pAcc);
docListAccumulateUnion(pAcc, pData, nData);
/* Accumulate remaining doclists into pAcc. */
for(++p; p<pAcc; ++p){
docListAccumulateUnion(pAcc, p->pData, p->nData);
/* dataBufferReset() could allow a large doclist to blow up
** our memory requirements.
*/
if( p->nCapacity<1024 ){
dataBufferReset(p);
}else{
dataBufferDestroy(p);
dataBufferInit(p, 0);
}
}
}
}
if( c>0 ) break; /* Past any possible matches. */
rc = leavesReaderStep(v, pReader);
if( rc!=SQLITE_OK ) return rc;
}
return SQLITE_OK;
/* Union all the doclists together into *out. */
/* TODO(shess) What if *out is big? Sigh. */
if( rc==SQLITE_OK && nBuffers>0 ){
int iBuffer;
for(iBuffer=0; iBuffer<nBuffers; ++iBuffer){
if( pBuffers[iBuffer].nData>0 ){
if( out->nData==0 ){
dataBufferSwap(out, &(pBuffers[iBuffer]));
}else{
docListAccumulateUnion(out, pBuffers[iBuffer].pData,
pBuffers[iBuffer].nData);
}
}
}
}
while( nBuffers-- ){
dataBufferDestroy(&(pBuffers[nBuffers]));
}
if( pBuffers!=NULL ) sqlite3_free(pBuffers);
return rc;
}
/* Call loadSegmentLeavesInt() with pData/nData as input. */
+6 -6
View File
@@ -1,5 +1,5 @@
C Changes\sfts2\sto\suse\sonly\ssqlite3_malloc()\sand\snot\ssystem\smalloc.\r\nBackports\s(4554)\sand\s(4555)\sfrom\sfts3.\s(CVS\s5454)
D 2008-07-22T22:57:54
C Change\sprefix\ssearch\sfrom\sO(N*M)\sto\sO(NlogM).\r\nBackports\s(4599)\sfrom\sfts3.\s(CVS\s5455)
D 2008-07-22T23:08:40
F Makefile.arm-wince-mingw32ce-gcc fcd5e9cd67fe88836360bb4f9ef4cb7f8e2fb5a0
F Makefile.in 77ff156061bb870aa0a8b3d545c670d08070f7e6
F Makefile.linux-gcc d53183f4aa6a9192d249731c90dbdffbd2c68654
@@ -39,7 +39,7 @@ F ext/fts1/simple_tokenizer.c 1844d72f7194c3fd3d7e4173053911bf0661b70d
F ext/fts1/tokenizer.h 0c53421b832366d20d720d21ea3e1f6e66a36ef9
F ext/fts2/README.tokenizers 21e3684ea5a095b55d70f6878b4ce6af5932dfb7
F ext/fts2/README.txt 8c18f41574404623b76917b9da66fcb0ab38328d
F ext/fts2/fts2.c f50c7faca742f40b14a2b279d652594c532e6523
F ext/fts2/fts2.c 7a2e88d110d059c986234c3d7734133d59a709ef
F ext/fts2/fts2.h da5f76c65163301d1068a971fd32f4119e3c95fa
F ext/fts2/fts2_hash.c 2689e42e1107ea67207f725cf69cf8972d00cf93
F ext/fts2/fts2_hash.h 9a5b1be94664139f93217a0770d7144425cffb3a
@@ -609,7 +609,7 @@ F tool/speedtest16.c c8a9c793df96db7e4933f0852abb7a03d48f2e81
F tool/speedtest2.tcl ee2149167303ba8e95af97873c575c3e0fab58ff
F tool/speedtest8.c 1dbced29de5f59ba2ebf877edcadf171540374d1
F tool/speedtest8inst1.c 293327bc76823f473684d589a8160bde1f52c14e
P d562515e1cdd05212674516033c64b5f5668b799
R 2682d383bee51312a260e3583fb091fa
P ecf2dec66cb979cb7d8db3b7ce5c64cab57fe2bb
R 89694177ea5ff7c878e6daa15283840d
U shess
Z 7b59d5a322f77dacbe20c5139c2b6d35
Z 48033bdd7abd5465e3195f281a547d67
+1 -1
View File
@@ -1 +1 @@
ecf2dec66cb979cb7d8db3b7ce5c64cab57fe2bb
3f614453d2d7c753a5963b027fe8618b50b4f6b9