From: dan Date: Thu, 3 Dec 2020 21:22:37 +0000 (+0000) Subject: When merging prefix lists in fts5, use 16-way merges instead of 2-way merges. This... X-Git-Tag: version-3.35.0~204 X-Git-Url: http://git.ipfire.org/?a=commitdiff_plain;h=3190b88e756baa82b7e41bc1ca721192784d7bcd;p=thirdparty%2Fsqlite.git When merging prefix lists in fts5, use 16-way merges instead of 2-way merges. This faster. FossilOrigin-Name: 026a93508ec392ca5cd2578ae9eab64974f58beccda088e10d4cc951f237632f --- diff --git a/ext/fts5/fts5_index.c b/ext/fts5/fts5_index.c index 60dc03c3b4..88709f2c35 100644 --- a/ext/fts5/fts5_index.c +++ b/ext/fts5/fts5_index.c @@ -3069,7 +3069,7 @@ static void fts5ChunkIterate( int pgno = pSeg->iLeafPgno; int pgnoSave = 0; - /* This function does notmwork with detail=none databases. */ + /* This function does not work with detail=none databases. */ assert( p->pConfig->eDetail!=FTS5_DETAIL_NONE ); if( (pSeg->flags & FTS5_SEGITER_REVERSE)==0 ){ @@ -3082,6 +3082,9 @@ static void fts5ChunkIterate( fts5DataRelease(pData); if( nRem<=0 ){ break; + }else if( pSeg->pSeg==0 ){ + p->rc = FTS5_CORRUPT; + return; }else{ pgno++; pData = fts5LeafRead(p, FTS5_SEGMENT_ROWID(pSeg->pSeg->iSegid, pgno)); @@ -4803,7 +4806,7 @@ static void fts5AppendPoslist( static void fts5DoclistIterNext(Fts5DoclistIter *pIter){ u8 *p = pIter->aPoslist + pIter->nSize + pIter->nPoslist; - assert( pIter->aPoslist ); + assert( pIter->aPoslist || (p==0 && pIter->aPoslist==0) ); if( p>=pIter->aEof ){ pIter->aPoslist = 0; }else{ @@ -4887,16 +4890,19 @@ static void fts5NextRowid(Fts5Buffer *pBuf, int *piOff, i64 *piRowid){ static void fts5MergeRowidLists( Fts5Index *p, /* FTS5 backend object */ Fts5Buffer *p1, /* First list to merge */ - Fts5Buffer *p2 /* Second list to merge */ + int nBuf, /* Number of entries in apBuf[] */ + Fts5Buffer *aBuf /* Array of other lists to merge into p1 */ ){ int i1 = 0; int i2 = 0; i64 iRowid1 = 0; i64 iRowid2 = 0; i64 iOut = 0; + Fts5Buffer *p2 = &aBuf[0]; Fts5Buffer out; memset(&out, 0, sizeof(out)); + assert( nBuf==1 ); sqlite3Fts5BufferSize(&p->rc, &out, p1->n + p2->n); if( p->rc ) return; @@ -4923,160 +4929,189 @@ static void fts5MergeRowidLists( fts5BufferFree(&out); } +typedef struct PrefixMerger PrefixMerger; +struct PrefixMerger { + Fts5DoclistIter iter; /* Doclist iterator */ + i64 iPos; /* For iterating through a position list */ + int iOff; + u8 *aPos; + PrefixMerger *pNext; /* Next in docid/poslist order */ +}; + +static void fts5PrefixMergerInsertByRowid( + PrefixMerger **ppHead, + PrefixMerger *p +){ + if( p->iter.aPoslist ){ + PrefixMerger **pp = ppHead; + while( *pp && p->iter.iRowid>(*pp)->iter.iRowid ){ + pp = &(*pp)->pNext; + } + p->pNext = *pp; + *pp = p; + } +} + +static void fts5PrefixMergerInsertByPosition( + PrefixMerger **ppHead, + PrefixMerger *p +){ + if( p->iPos>=0 ){ + PrefixMerger **pp = ppHead; + while( *pp && p->iPos>(*pp)->iPos ){ + pp = &(*pp)->pNext; + } + p->pNext = *pp; + *pp = p; + } +} + + /* -** Buffers p1 and p2 contain doclists. This function merges the content -** of the two doclists together and sets buffer p1 to the result before -** returning. -** -** If an error occurs, an error code is left in p->rc. If an error has -** already occurred, this function is a no-op. +** Array aBuf[] contains nBuf doclists. These are all merged in with the +** doclist in buffer p1. */ static void fts5MergePrefixLists( Fts5Index *p, /* FTS5 backend object */ Fts5Buffer *p1, /* First list to merge */ - Fts5Buffer *p2 /* Second list to merge */ + int nBuf, /* Number of buffers in array aBuf[] */ + Fts5Buffer *aBuf /* Other lists to merge in */ ){ - if( p2->n ){ - i64 iLastRowid = 0; - Fts5DoclistIter i1; - Fts5DoclistIter i2; - Fts5Buffer out = {0, 0, 0}; - Fts5Buffer tmp = {0, 0, 0}; - - /* The maximum size of the output is equal to the sum of the two - ** input sizes + 1 varint (9 bytes). The extra varint is because if the - ** first rowid in one input is a large negative number, and the first in - ** the other a non-negative number, the delta for the non-negative - ** number will be larger on disk than the literal integer value - ** was. - ** - ** Or, if the input position-lists are corrupt, then the output might - ** include up to 2 extra 10-byte positions created by interpreting -1 - ** (the value PoslistNext64() uses for EOF) as a position and appending - ** it to the output. This can happen at most once for each input - ** position-list, hence two 10 byte paddings. */ - if( sqlite3Fts5BufferSize(&p->rc, &out, p1->n + p2->n + 9+10+10) ) return; - fts5DoclistIterInit(p1, &i1); - fts5DoclistIterInit(p2, &i2); - - while( 1 ){ - if( i1.iRowidp) + (i2.aPoslist-p2->p)+9+10+10) ); - } - else if( i2.iRowid!=i1.iRowid ){ - /* Copy entry from i2 */ - fts5MergeAppendDocid(&out, iLastRowid, i2.iRowid); - fts5BufferSafeAppendBlob(&out, i2.aPoslist, i2.nPoslist+i2.nSize); - fts5DoclistIterNext(&i2); - if( i2.aPoslist==0 ) break; - assert( out.n<=((i1.aPoslist-p1->p) + (i2.aPoslist-p2->p)+9+10+10) ); +#define fts5PrefixMergerNextPosition(p) \ + sqlite3Fts5PoslistNext64((p)->aPos,(p)->iter.nPoslist,&(p)->iOff,&(p)->iPos); +#define FTS5_MERGE_NLIST 16 + PrefixMerger aMerger[FTS5_MERGE_NLIST]; + PrefixMerger *pHead = 0; + int i; + int nOut = 0; + Fts5Buffer out = {0, 0, 0}; + Fts5Buffer tmp = {0, 0, 0}; + i64 iLastRowid = 0; + + /* Initialize a doclist-iterator for each input buffer. Arrange them in + ** a linked-list starting at pHead in ascending order of rowid. Avoid + ** linking any iterators already at EOF into the linked list at all. */ + assert( nBuf+1<=sizeof(aMerger)/sizeof(aMerger[0]) ); + memset(aMerger, 0, sizeof(PrefixMerger)*(nBuf+1)); + pHead = &aMerger[nBuf]; + fts5DoclistIterInit(p1, &pHead->iter); + for(i=0; in + 9 + 10*nBuf; + + /* The maximum size of the output is equal to the sum of the + ** input sizes + 1 varint (9 bytes). The extra varint is because if the + ** first rowid in one input is a large negative number, and the first in + ** the other a non-negative number, the delta for the non-negative + ** number will be larger on disk than the literal integer value + ** was. + ** + ** Or, if the input position-lists are corrupt, then the output might + ** include up to (nBuf+1) extra 10-byte positions created by interpreting -1 + ** (the value PoslistNext64() uses for EOF) as a position and appending + ** it to the output. This can happen at most once for each input + ** position-list, hence (nBuf+1) 10 byte paddings. */ + if( sqlite3Fts5BufferSize(&p->rc, &out, nOut) ) return; + + while( pHead ){ + fts5MergeAppendDocid(&out, iLastRowid, pHead->iter.iRowid); + + if( pHead->pNext && iLastRowid==pHead->pNext->iter.iRowid ){ + /* Merge data from two or more poslists */ + i64 iPrev = 0; + int nTmp = FTS5_DATA_ZERO_PADDING; + int nMerge = 0; + PrefixMerger *pSave = pHead; + PrefixMerger *pThis = 0; + int nTail = 0; + + pHead = 0; + while( pSave && pSave->iter.iRowid==iLastRowid ){ + PrefixMerger *pNext = pSave->pNext; + pSave->iOff = 0; + pSave->iPos = 0; + pSave->aPos = &pSave->iter.aPoslist[pSave->iter.nSize]; + fts5PrefixMergerNextPosition(pSave); + nTmp += pSave->iter.nPoslist + 10; + nMerge++; + fts5PrefixMergerInsertByPosition(&pHead, pSave); + pSave = pNext; } - else{ - /* Merge the two position lists. */ - i64 iPos1 = 0; - i64 iPos2 = 0; - int iOff1 = 0; - int iOff2 = 0; - u8 *a1 = &i1.aPoslist[i1.nSize]; - u8 *a2 = &i2.aPoslist[i2.nSize]; - int nCopy; - u8 *aCopy; - - i64 iPrev = 0; - Fts5PoslistWriter writer; - memset(&writer, 0, sizeof(writer)); - - /* See the earlier comment in this function for an explanation of why - ** corrupt input position lists might cause the output to consume - ** at most 20 bytes of unexpected space. */ - fts5MergeAppendDocid(&out, iLastRowid, i2.iRowid); - fts5BufferZero(&tmp); - sqlite3Fts5BufferSize(&p->rc, &tmp, - i1.nPoslist + i2.nPoslist + 10 + 10 + FTS5_DATA_ZERO_PADDING - ); - if( p->rc ) break; - sqlite3Fts5PoslistNext64(a1, i1.nPoslist, &iOff1, &iPos1); - sqlite3Fts5PoslistNext64(a2, i2.nPoslist, &iOff2, &iPos2); - assert_nc( iPos1>=0 && iPos2>=0 ); + if( pHead==0 || pHead->pNext==0 ){ + p->rc = FTS5_CORRUPT; + break; + } - if( iPos1=0 && iPos2>=0 ){ - while( 1 ){ - if( iPos1rc, &tmp, nTmp+nMerge*10) ){ + break; + } + fts5BufferZero(&tmp); + + pThis = pHead; + pHead = pThis->pNext; + sqlite3Fts5PoslistSafeAppend(&tmp, &iPrev, pThis->iPos); + fts5PrefixMergerNextPosition(pThis); + fts5PrefixMergerInsertByPosition(&pHead, pThis); + + while( pHead->pNext ){ + pThis = pHead; + if( pThis->iPos!=iPrev ){ + sqlite3Fts5PoslistSafeAppend(&tmp, &iPrev, pThis->iPos); } + fts5PrefixMergerNextPosition(pThis); + pHead = pThis->pNext; + fts5PrefixMergerInsertByPosition(&pHead, pThis); + } - if( iPos1>=0 ){ - if( iPos1!=iPrev ){ - sqlite3Fts5PoslistSafeAppend(&tmp, &iPrev, iPos1); - } - aCopy = &a1[iOff1]; - nCopy = i1.nPoslist - iOff1; - }else{ - assert_nc( iPos2>=0 && iPos2!=iPrev ); - sqlite3Fts5PoslistSafeAppend(&tmp, &iPrev, iPos2); - aCopy = &a2[iOff2]; - nCopy = i2.nPoslist - iOff2; - } - if( nCopy>0 ){ - fts5BufferSafeAppendBlob(&tmp, aCopy, nCopy); - } + if( pHead->iPos!=iPrev ){ + sqlite3Fts5PoslistSafeAppend(&tmp, &iPrev, pHead->iPos); + } + nTail = pHead->iter.nPoslist - pHead->iOff; - /* WRITEPOSLISTSIZE */ - assert_nc( tmp.n<=i1.nPoslist+i2.nPoslist ); - assert( tmp.n<=i1.nPoslist+i2.nPoslist+10+10 ); - if( tmp.n>i1.nPoslist+i2.nPoslist ){ - if( p->rc==SQLITE_OK ) p->rc = FTS5_CORRUPT; - break; + /* WRITEPOSLISTSIZE */ + assert( tmp.n+nTail<=nTmp ); + if( tmp.n+nTail>nTmp-FTS5_DATA_ZERO_PADDING ){ + if( p->rc==SQLITE_OK ) p->rc = FTS5_CORRUPT; + break; + } + fts5BufferSafeAppendVarint(&out, (tmp.n+nTail) * 2); + fts5BufferSafeAppendBlob(&out, tmp.p, tmp.n); + if( nTail>0 ){ + fts5BufferSafeAppendBlob(&out, &pHead->aPos[pHead->iOff], nTail); + } + + pHead = pSave; + for(i=0; iiter.aPoslist && pThis->iter.iRowid==iLastRowid ){ + fts5DoclistIterNext(&pThis->iter); + fts5PrefixMergerInsertByRowid(&pHead, pThis); } - fts5BufferSafeAppendVarint(&out, tmp.n * 2); - fts5BufferSafeAppendBlob(&out, tmp.p, tmp.n); - fts5DoclistIterNext(&i1); - fts5DoclistIterNext(&i2); - assert_nc( out.n<=(p1->n+p2->n+9) ); - if( i1.aPoslist==0 || i2.aPoslist==0 ) break; - assert( out.n<=((i1.aPoslist-p1->p) + (i2.aPoslist-p2->p)+9+10+10) ); } - } - if( i1.aPoslist ){ - fts5MergeAppendDocid(&out, iLastRowid, i1.iRowid); - fts5BufferSafeAppendBlob(&out, i1.aPoslist, i1.aEof - i1.aPoslist); - } - else if( i2.aPoslist ){ - fts5MergeAppendDocid(&out, iLastRowid, i2.iRowid); - fts5BufferSafeAppendBlob(&out, i2.aPoslist, i2.aEof - i2.aPoslist); + }else{ + /* Copy poslist from pHead to output */ + PrefixMerger *pThis = pHead; + Fts5DoclistIter *pI = &pThis->iter; + fts5BufferSafeAppendBlob(&out, pI->aPoslist, pI->nPoslist+pI->nSize); + fts5DoclistIterNext(pI); + pHead = pThis->pNext; + fts5PrefixMergerInsertByRowid(&pHead, pThis); } - assert_nc( out.n<=(p1->n+p2->n+9) ); - - fts5BufferFree(p1); - fts5BufferFree(&tmp); - memset(&out.p[out.n], 0, FTS5_DATA_ZERO_PADDING); - *p1 = out; } + + fts5BufferFree(p1); + fts5BufferFree(&tmp); + memset(&out.p[out.n], 0, FTS5_DATA_ZERO_PADDING); + *p1 = out; } static void fts5SetupPrefixIter( @@ -5090,14 +5125,17 @@ static void fts5SetupPrefixIter( ){ Fts5Structure *pStruct; Fts5Buffer *aBuf; - const int nBuf = 32; + int nBuf = 32; + int nMerge = 1; - void (*xMerge)(Fts5Index*, Fts5Buffer*, Fts5Buffer*); + void (*xMerge)(Fts5Index*, Fts5Buffer*, int, Fts5Buffer*); void (*xAppend)(Fts5Index*, i64, Fts5Iter*, Fts5Buffer*); if( p->pConfig->eDetail==FTS5_DETAIL_NONE ){ xMerge = fts5MergeRowidLists; xAppend = fts5AppendRowid; }else{ + nMerge = FTS5_MERGE_NLIST-1; + nBuf = nMerge*8; /* Sufficient to merge (16^8)==(2^32) lists */ xMerge = fts5MergePrefixLists; xAppend = fts5AppendPoslist; } @@ -5158,13 +5196,21 @@ static void fts5SetupPrefixIter( if( p1->base.iRowid<=iLastRowid && doclist.n>0 ){ for(i=0; p->rc==SQLITE_OK && doclist.n; i++){ - assert( ibase.iRowid; } - for(i=0; irc==SQLITE_OK ){ - xMerge(p, &doclist, &aBuf[i]); + xMerge(p, &doclist, nMerge, &aBuf[i]); + } + for(iFree=i; iFree