diff options
Diffstat (limited to 'lib/libsqlite3/src/analyze.c')
-rw-r--r-- | lib/libsqlite3/src/analyze.c | 258 |
1 files changed, 169 insertions, 89 deletions
diff --git a/lib/libsqlite3/src/analyze.c b/lib/libsqlite3/src/analyze.c index 2a03e292c88..f9c03dc848a 100644 --- a/lib/libsqlite3/src/analyze.c +++ b/lib/libsqlite3/src/analyze.c @@ -246,6 +246,7 @@ static void openStatTable( assert( i<ArraySize(aTable) ); sqlite3VdbeAddOp4Int(v, OP_OpenWrite, iStatCur+i, aRoot[i], iDb, 3); sqlite3VdbeChangeP5(v, aCreateTbl[i]); + VdbeComment((v, aTable[i].zName)); } } @@ -281,7 +282,8 @@ struct Stat4Sample { struct Stat4Accum { tRowcnt nRow; /* Number of rows in the entire table */ tRowcnt nPSample; /* How often to do a periodic sample */ - int nCol; /* Number of columns in index + rowid */ + int nCol; /* Number of columns in index + pk/rowid */ + int nKeyCol; /* Number of index columns w/o the pk/rowid */ int mxSample; /* Maximum number of samples to accumulate */ Stat4Sample current; /* Current row as a Stat4Sample */ u32 iPrn; /* Pseudo-random number used for sampling */ @@ -367,9 +369,22 @@ static void stat4Destructor(void *pOld){ } /* -** Implementation of the stat_init(N,C) SQL function. The two parameters -** are the number of rows in the table or index (C) and the number of columns -** in the index (N). The second argument (C) is only used for STAT3 and STAT4. +** Implementation of the stat_init(N,K,C) SQL function. The three parameters +** are: +** N: The number of columns in the index including the rowid/pk (note 1) +** K: The number of columns in the index excluding the rowid/pk. +** C: The number of rows in the index (note 2) +** +** Note 1: In the special case of the covering index that implements a +** WITHOUT ROWID table, N is the number of PRIMARY KEY columns, not the +** total number of columns in the table. +** +** Note 2: C is only used for STAT3 and STAT4. +** +** For indexes on ordinary rowid tables, N==K+1. But for indexes on +** WITHOUT ROWID tables, N=K+P where P is the number of columns in the +** PRIMARY KEY of the table. The covering index that implements the +** original WITHOUT ROWID table as N==K as a special case. ** ** This routine allocates the Stat4Accum object in heap memory. The return ** value is a pointer to the the Stat4Accum object encoded as a blob (i.e. @@ -382,6 +397,7 @@ static void statInit( ){ Stat4Accum *p; int nCol; /* Number of columns in index being sampled */ + int nKeyCol; /* Number of key columns */ int nColUp; /* nCol rounded up for alignment */ int n; /* Bytes of space to allocate */ sqlite3 *db; /* Database connection */ @@ -392,8 +408,11 @@ static void statInit( /* Decode the three function arguments */ UNUSED_PARAMETER(argc); nCol = sqlite3_value_int(argv[0]); - assert( nCol>1 ); /* >1 because it includes the rowid column */ + assert( nCol>0 ); nColUp = sizeof(tRowcnt)<8 ? (nCol+1)&~1 : nCol; + nKeyCol = sqlite3_value_int(argv[1]); + assert( nKeyCol<=nCol ); + assert( nKeyCol>0 ); /* Allocate the space required for the Stat4Accum object */ n = sizeof(*p) @@ -415,6 +434,7 @@ static void statInit( p->db = db; p->nRow = 0; p->nCol = nCol; + p->nKeyCol = nKeyCol; p->current.anDLt = (tRowcnt*)&p[1]; p->current.anEq = &p->current.anDLt[nColUp]; @@ -425,9 +445,9 @@ static void statInit( p->iGet = -1; p->mxSample = mxSample; - p->nPSample = (tRowcnt)(sqlite3_value_int64(argv[1])/(mxSample/3+1) + 1); + p->nPSample = (tRowcnt)(sqlite3_value_int64(argv[2])/(mxSample/3+1) + 1); p->current.anLt = &p->current.anEq[nColUp]; - p->iPrn = nCol*0x689e962d ^ sqlite3_value_int(argv[1])*0xd0944565; + p->iPrn = nCol*0x689e962d ^ sqlite3_value_int(argv[2])*0xd0944565; /* Set up the Stat4Accum.a[] and aBest[] arrays */ p->a = (struct Stat4Sample*)&p->current.anLt[nColUp]; @@ -450,7 +470,7 @@ static void statInit( sqlite3_result_blob(context, p, sizeof(p), stat4Destructor); } static const FuncDef statInitFuncdef = { - 1+IsStat34, /* nArg */ + 2+IsStat34, /* nArg */ SQLITE_UTF8, /* funcFlags */ 0, /* pUserData */ 0, /* pNext */ @@ -674,7 +694,10 @@ static void samplePushPrevious(Stat4Accum *p, int iChng){ ** R Rowid for the current row. Might be a key record for ** WITHOUT ROWID tables. ** -** The SQL function always returns NULL. +** This SQL function always returns NULL. It's purpose it to accumulate +** statistical data and/or samples in the Stat4Accum object about the +** index being analyzed. The stat_get() SQL function will later be used to +** extract relevant information for constructing the sqlite_statN tables. ** ** The R parameter is only used for STAT3 and STAT4 */ @@ -691,7 +714,7 @@ static void statPush( UNUSED_PARAMETER( argc ); UNUSED_PARAMETER( context ); - assert( p->nCol>1 ); /* Includes rowid field */ + assert( p->nCol>0 ); assert( iChng<p->nCol ); if( p->nRow==0 ){ @@ -768,7 +791,10 @@ static const FuncDef statPushFuncdef = { /* ** Implementation of the stat_get(P,J) SQL function. This routine is -** used to query the results. Content is returned for parameter J +** used to query statistical information that has been gathered into +** the Stat4Accum object by prior calls to stat_push(). The P parameter +** is a BLOB which is decoded into a pointer to the Stat4Accum objects. +** The content to returned is determined by the parameter J ** which is one of the STAT_GET_xxxx values defined above. ** ** If neither STAT3 nor STAT4 are enabled, then J is always @@ -819,7 +845,7 @@ static void statGet( char *z; int i; - char *zRet = sqlite3MallocZero(p->nCol * 25); + char *zRet = sqlite3MallocZero( (p->nKeyCol+1)*25 ); if( zRet==0 ){ sqlite3_result_error_nomem(context); return; @@ -827,7 +853,7 @@ static void statGet( sqlite3_snprintf(24, zRet, "%llu", (u64)p->nRow); z = zRet + sqlite3Strlen30(zRet); - for(i=0; i<(p->nCol-1); i++){ + for(i=0; i<p->nKeyCol; i++){ u64 nDistinct = p->current.anDLt[i] + 1; u64 iVal = (p->nRow + nDistinct - 1) / nDistinct; sqlite3_snprintf(24, z, " %llu", iVal); @@ -987,27 +1013,27 @@ static void analyzeOneTable( sqlite3VdbeAddOp4(v, OP_String8, 0, regTabname, 0, pTab->zName, 0); for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ - int nCol; /* Number of columns indexed by pIdx */ - int *aGotoChng; /* Array of jump instruction addresses */ + int nCol; /* Number of columns in pIdx. "N" */ int addrRewind; /* Address of "OP_Rewind iIdxCur" */ - int addrGotoChng0; /* Address of "Goto addr_chng_0" */ int addrNextRow; /* Address of "next_row:" */ const char *zIdxName; /* Name of the index */ + int nColTest; /* Number of columns to test for changes */ if( pOnlyIdx && pOnlyIdx!=pIdx ) continue; if( pIdx->pPartIdxWhere==0 ) needTableCnt = 0; - VdbeNoopComment((v, "Begin analysis of %s", pIdx->zName)); - nCol = pIdx->nKeyCol; - aGotoChng = sqlite3DbMallocRaw(db, sizeof(int)*(nCol+1)); - if( aGotoChng==0 ) continue; - - /* Populate the register containing the index name. */ - if( pIdx->autoIndex==2 && !HasRowid(pTab) ){ + if( !HasRowid(pTab) && IsPrimaryKeyIndex(pIdx) ){ + nCol = pIdx->nKeyCol; zIdxName = pTab->zName; + nColTest = nCol - 1; }else{ + nCol = pIdx->nColumn; zIdxName = pIdx->zName; + nColTest = pIdx->uniqNotNull ? pIdx->nKeyCol-1 : nCol-1; } + + /* Populate the register containing the index name. */ sqlite3VdbeAddOp4(v, OP_String8, 0, regIdxname, 0, zIdxName, 0); + VdbeComment((v, "Analysis for %s.%s", pTab->zName, zIdxName)); /* ** Pseudo-code for loop that calls stat_push(): @@ -1032,7 +1058,7 @@ static void analyzeOneTable( ** regPrev(1) = idx(1) ** ... ** - ** chng_addr_N: + ** endDistinctTest: ** regRowid = idx(rowid) ** stat_push(P, regChng, regRowid) ** Next csr @@ -1045,7 +1071,7 @@ static void analyzeOneTable( ** the regPrev array and a trailing rowid (the rowid slot is required ** when building a record to insert into the sample column of ** the sqlite_stat4 table. */ - pParse->nMem = MAX(pParse->nMem, regPrev+nCol); + pParse->nMem = MAX(pParse->nMem, regPrev+nColTest); /* Open a read-only cursor on the index being analyzed. */ assert( iDb==sqlite3SchemaToIndex(db, pIdx->pSchema) ); @@ -1055,18 +1081,22 @@ static void analyzeOneTable( /* Invoke the stat_init() function. The arguments are: ** - ** (1) the number of columns in the index including the rowid, - ** (2) the number of rows in the index, + ** (1) the number of columns in the index including the rowid + ** (or for a WITHOUT ROWID table, the number of PK columns), + ** (2) the number of columns in the key without the rowid/pk + ** (3) the number of rows in the index, ** - ** The second argument is only used for STAT3 and STAT4 + ** + ** The third argument is only used for STAT3 and STAT4 */ #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 - sqlite3VdbeAddOp2(v, OP_Count, iIdxCur, regStat4+2); + sqlite3VdbeAddOp2(v, OP_Count, iIdxCur, regStat4+3); #endif - sqlite3VdbeAddOp2(v, OP_Integer, nCol+1, regStat4+1); + sqlite3VdbeAddOp2(v, OP_Integer, nCol, regStat4+1); + sqlite3VdbeAddOp2(v, OP_Integer, pIdx->nKeyCol, regStat4+2); sqlite3VdbeAddOp3(v, OP_Function, 0, regStat4+1, regStat4); sqlite3VdbeChangeP4(v, -1, (char*)&statInitFuncdef, P4_FUNCDEF); - sqlite3VdbeChangeP5(v, 1+IsStat34); + sqlite3VdbeChangeP5(v, 2+IsStat34); /* Implementation of the following: ** @@ -1079,44 +1109,62 @@ static void analyzeOneTable( addrRewind = sqlite3VdbeAddOp1(v, OP_Rewind, iIdxCur); VdbeCoverage(v); sqlite3VdbeAddOp2(v, OP_Integer, 0, regChng); - addrGotoChng0 = sqlite3VdbeAddOp0(v, OP_Goto); - - /* - ** next_row: - ** regChng = 0 - ** if( idx(0) != regPrev(0) ) goto chng_addr_0 - ** regChng = 1 - ** if( idx(1) != regPrev(1) ) goto chng_addr_1 - ** ... - ** regChng = N - ** goto chng_addr_N - */ addrNextRow = sqlite3VdbeCurrentAddr(v); - for(i=0; i<nCol; i++){ - char *pColl = (char*)sqlite3LocateCollSeq(pParse, pIdx->azColl[i]); - sqlite3VdbeAddOp2(v, OP_Integer, i, regChng); - sqlite3VdbeAddOp3(v, OP_Column, iIdxCur, i, regTemp); - aGotoChng[i] = - sqlite3VdbeAddOp4(v, OP_Ne, regTemp, 0, regPrev+i, pColl, P4_COLLSEQ); - sqlite3VdbeChangeP5(v, SQLITE_NULLEQ); - VdbeCoverage(v); - } - sqlite3VdbeAddOp2(v, OP_Integer, nCol, regChng); - aGotoChng[nCol] = sqlite3VdbeAddOp0(v, OP_Goto); - /* - ** chng_addr_0: - ** regPrev(0) = idx(0) - ** chng_addr_1: - ** regPrev(1) = idx(1) - ** ... - */ - sqlite3VdbeJumpHere(v, addrGotoChng0); - for(i=0; i<nCol; i++){ - sqlite3VdbeJumpHere(v, aGotoChng[i]); - sqlite3VdbeAddOp3(v, OP_Column, iIdxCur, i, regPrev+i); + if( nColTest>0 ){ + int endDistinctTest = sqlite3VdbeMakeLabel(v); + int *aGotoChng; /* Array of jump instruction addresses */ + aGotoChng = sqlite3DbMallocRaw(db, sizeof(int)*nColTest); + if( aGotoChng==0 ) continue; + + /* + ** next_row: + ** regChng = 0 + ** if( idx(0) != regPrev(0) ) goto chng_addr_0 + ** regChng = 1 + ** if( idx(1) != regPrev(1) ) goto chng_addr_1 + ** ... + ** regChng = N + ** goto endDistinctTest + */ + sqlite3VdbeAddOp0(v, OP_Goto); + addrNextRow = sqlite3VdbeCurrentAddr(v); + if( nColTest==1 && pIdx->nKeyCol==1 && IsUniqueIndex(pIdx) ){ + /* For a single-column UNIQUE index, once we have found a non-NULL + ** row, we know that all the rest will be distinct, so skip + ** subsequent distinctness tests. */ + sqlite3VdbeAddOp2(v, OP_NotNull, regPrev, endDistinctTest); + VdbeCoverage(v); + } + for(i=0; i<nColTest; i++){ + char *pColl = (char*)sqlite3LocateCollSeq(pParse, pIdx->azColl[i]); + sqlite3VdbeAddOp2(v, OP_Integer, i, regChng); + sqlite3VdbeAddOp3(v, OP_Column, iIdxCur, i, regTemp); + aGotoChng[i] = + sqlite3VdbeAddOp4(v, OP_Ne, regTemp, 0, regPrev+i, pColl, P4_COLLSEQ); + sqlite3VdbeChangeP5(v, SQLITE_NULLEQ); + VdbeCoverage(v); + } + sqlite3VdbeAddOp2(v, OP_Integer, nColTest, regChng); + sqlite3VdbeAddOp2(v, OP_Goto, 0, endDistinctTest); + + + /* + ** chng_addr_0: + ** regPrev(0) = idx(0) + ** chng_addr_1: + ** regPrev(1) = idx(1) + ** ... + */ + sqlite3VdbeJumpHere(v, addrNextRow-1); + for(i=0; i<nColTest; i++){ + sqlite3VdbeJumpHere(v, aGotoChng[i]); + sqlite3VdbeAddOp3(v, OP_Column, iIdxCur, i, regPrev+i); + } + sqlite3VdbeResolveLabel(v, endDistinctTest); + sqlite3DbFree(db, aGotoChng); } - + /* ** chng_addr_N: ** regRowid = idx(rowid) // STAT34 only @@ -1124,7 +1172,6 @@ static void analyzeOneTable( ** Next csr ** if !eof(csr) goto next_row; */ - sqlite3VdbeJumpHere(v, aGotoChng[nCol]); #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 assert( regRowid==(regStat4+2) ); if( HasRowid(pTab) ){ @@ -1168,7 +1215,7 @@ static void analyzeOneTable( int addrIsNull; u8 seekOp = HasRowid(pTab) ? OP_NotExists : OP_NotFound; - pParse->nMem = MAX(pParse->nMem, regCol+nCol+1); + pParse->nMem = MAX(pParse->nMem, regCol+nCol); addrNext = sqlite3VdbeCurrentAddr(v); callStatGet(v, regStat4, STAT_GET_ROWID, regSampleRowid); @@ -1190,7 +1237,7 @@ static void analyzeOneTable( i16 iCol = pIdx->aiColumn[i]; sqlite3ExprCodeGetColumnOfTable(v, pTab, iTabCur, iCol, regCol+i); } - sqlite3VdbeAddOp3(v, OP_MakeRecord, regCol, nCol+1, regSample); + sqlite3VdbeAddOp3(v, OP_MakeRecord, regCol, nCol, regSample); #endif sqlite3VdbeAddOp3(v, OP_MakeRecord, regTabname, 6, regTemp); sqlite3VdbeAddOp2(v, OP_NewRowid, iStatCur+1, regNewRowid); @@ -1202,7 +1249,6 @@ static void analyzeOneTable( /* End of analysis */ sqlite3VdbeJumpHere(v, addrRewind); - sqlite3DbFree(db, aGotoChng); } @@ -1303,6 +1349,7 @@ void sqlite3Analyze(Parse *pParse, Token *pName1, Token *pName2){ Table *pTab; Index *pIdx; Token *pTableName; + Vdbe *v; /* Read the database schema. If an error occurs, leave an error message ** and code in pParse and return NULL. */ @@ -1350,6 +1397,8 @@ void sqlite3Analyze(Parse *pParse, Token *pName1, Token *pName2){ } } } + v = sqlite3GetVdbe(pParse); + if( v ) sqlite3VdbeAddOp0(v, OP_Expire); } /* @@ -1371,6 +1420,7 @@ static void decodeIntArray( char *zIntArray, /* String containing int array to decode */ int nOut, /* Number of slots in aOut[] */ tRowcnt *aOut, /* Store integers here */ + LogEst *aLog, /* Or, if aOut==0, here */ Index *pIndex /* Handle extra flags for this index, if not NULL */ ){ char *z = zIntArray; @@ -1389,7 +1439,17 @@ static void decodeIntArray( v = v*10 + c - '0'; z++; } - aOut[i] = v; +#ifdef SQLITE_ENABLE_STAT3_OR_STAT4 + if( aOut ){ + aOut[i] = v; + }else +#else + assert( aOut==0 ); + UNUSED_PARAMETER(aOut); +#endif + { + aLog[i] = sqlite3LogEst(v); + } if( *z==' ' ) z++; } #ifndef SQLITE_ENABLE_STAT3_OR_STAT4 @@ -1397,14 +1457,19 @@ static void decodeIntArray( #else if( pIndex ) #endif - { - if( strcmp(z, "unordered")==0 ){ + while( z[0] ){ + if( sqlite3_strglob("unordered*", z)==0 ){ pIndex->bUnordered = 1; }else if( sqlite3_strglob("sz=[0-9]*", z)==0 ){ - int v32 = 0; - sqlite3GetInt32(z+3, &v32); - pIndex->szIdxRow = sqlite3LogEst(v32); + pIndex->szIdxRow = sqlite3LogEst(sqlite3Atoi(z+3)); } +#ifdef SQLITE_ENABLE_COSTMULT + else if( sqlite3_strglob("costmult=[0-9]*",z)==0 ){ + pIndex->pTable->costMult = sqlite3LogEst(sqlite3Atoi(z+9)); + } +#endif + while( z[0]!=0 && z[0]!=' ' ) z++; + while( z[0]==' ' ) z++; } } @@ -1445,12 +1510,16 @@ static int analysisLoader(void *pData, int argc, char **argv, char **NotUsed){ z = argv[2]; if( pIndex ){ - decodeIntArray((char*)z, pIndex->nKeyCol+1, pIndex->aiRowEst, pIndex); - if( pIndex->pPartIdxWhere==0 ) pTable->nRowEst = pIndex->aiRowEst[0]; + pIndex->bUnordered = 0; + decodeIntArray((char*)z, pIndex->nKeyCol+1, 0, pIndex->aiRowLogEst, pIndex); + if( pIndex->pPartIdxWhere==0 ) pTable->nRowLogEst = pIndex->aiRowLogEst[0]; }else{ Index fakeIdx; fakeIdx.szIdxRow = pTable->szTabRow; - decodeIntArray((char*)z, 1, &pTable->nRowEst, &fakeIdx); +#ifdef SQLITE_ENABLE_COSTMULT + fakeIdx.pTable = pTable; +#endif + decodeIntArray((char*)z, 1, 0, &pTable->nRowLogEst, &fakeIdx); pTable->szTabRow = fakeIdx.szIdxRow; } @@ -1491,7 +1560,16 @@ static void initAvgEq(Index *pIdx){ IndexSample *aSample = pIdx->aSample; IndexSample *pFinal = &aSample[pIdx->nSample-1]; int iCol; - for(iCol=0; iCol<pIdx->nKeyCol; iCol++){ + int nCol = 1; + if( pIdx->nSampleCol>1 ){ + /* If this is stat4 data, then calculate aAvgEq[] values for all + ** sample columns except the last. The last is always set to 1, as + ** once the trailing PK fields are considered all index keys are + ** unique. */ + nCol = pIdx->nSampleCol-1; + pIdx->aAvgEq[nCol] = 1; + } + for(iCol=0; iCol<nCol; iCol++){ int i; /* Used to iterate through samples */ tRowcnt sumEq = 0; /* Sum of the nEq values */ tRowcnt nSum = 0; /* Number of terms contributing to sumEq */ @@ -1514,7 +1592,6 @@ static void initAvgEq(Index *pIdx){ } if( avgEq==0 ) avgEq = 1; pIdx->aAvgEq[iCol] = avgEq; - if( pIdx->nSampleCol==1 ) break; } } } @@ -1573,7 +1650,6 @@ static int loadStatTbl( while( sqlite3_step(pStmt)==SQLITE_ROW ){ int nIdxCol = 1; /* Number of columns in stat4 records */ - int nAvgCol = 1; /* Number of entries in Index.aAvgEq */ char *zIndex; /* Index name */ Index *pIdx; /* Pointer to the index object */ @@ -1591,13 +1667,17 @@ static int loadStatTbl( ** loaded from the stat4 table. In this case ignore stat3 data. */ if( pIdx==0 || pIdx->nSample ) continue; if( bStat3==0 ){ - nIdxCol = pIdx->nKeyCol+1; - nAvgCol = pIdx->nKeyCol; + assert( !HasRowid(pIdx->pTable) || pIdx->nColumn==pIdx->nKeyCol+1 ); + if( !HasRowid(pIdx->pTable) && IsPrimaryKeyIndex(pIdx) ){ + nIdxCol = pIdx->nKeyCol; + }else{ + nIdxCol = pIdx->nColumn; + } } pIdx->nSampleCol = nIdxCol; nByte = sizeof(IndexSample) * nSample; nByte += sizeof(tRowcnt) * nIdxCol * 3 * nSample; - nByte += nAvgCol * sizeof(tRowcnt); /* Space for Index.aAvgEq[] */ + nByte += nIdxCol * sizeof(tRowcnt); /* Space for Index.aAvgEq[] */ pIdx->aSample = sqlite3DbMallocZero(db, nByte); if( pIdx->aSample==0 ){ @@ -1605,7 +1685,7 @@ static int loadStatTbl( return SQLITE_NOMEM; } pSpace = (tRowcnt*)&pIdx->aSample[nSample]; - pIdx->aAvgEq = pSpace; pSpace += nAvgCol; + pIdx->aAvgEq = pSpace; pSpace += nIdxCol; for(i=0; i<nSample; i++){ pIdx->aSample[i].anEq = pSpace; pSpace += nIdxCol; pIdx->aSample[i].anLt = pSpace; pSpace += nIdxCol; @@ -1642,9 +1722,9 @@ static int loadStatTbl( pPrevIdx = pIdx; } pSample = &pIdx->aSample[pIdx->nSample]; - decodeIntArray((char*)sqlite3_column_text(pStmt,1), nCol, pSample->anEq, 0); - decodeIntArray((char*)sqlite3_column_text(pStmt,2), nCol, pSample->anLt, 0); - decodeIntArray((char*)sqlite3_column_text(pStmt,3), nCol, pSample->anDLt,0); + decodeIntArray((char*)sqlite3_column_text(pStmt,1),nCol,pSample->anEq,0,0); + decodeIntArray((char*)sqlite3_column_text(pStmt,2),nCol,pSample->anLt,0,0); + decodeIntArray((char*)sqlite3_column_text(pStmt,3),nCol,pSample->anDLt,0,0); /* Take a copy of the sample. Add two 0x00 bytes the end of the buffer. ** This is in case the sample record is corrupted. In that case, the |